Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * costsize.c
4 : * Routines to compute (and set) relation sizes and path costs
5 : *
6 : * Path costs are measured in arbitrary units established by these basic
7 : * parameters:
8 : *
9 : * seq_page_cost Cost of a sequential page fetch
10 : * random_page_cost Cost of a non-sequential page fetch
11 : * cpu_tuple_cost Cost of typical CPU time to process a tuple
12 : * cpu_index_tuple_cost Cost of typical CPU time to process an index tuple
13 : * cpu_operator_cost Cost of CPU time to execute an operator or function
14 : * parallel_tuple_cost Cost of CPU time to pass a tuple from worker to leader backend
15 : * parallel_setup_cost Cost of setting up shared memory for parallelism
16 : *
17 : * We expect that the kernel will typically do some amount of read-ahead
18 : * optimization; this in conjunction with seek costs means that seq_page_cost
19 : * is normally considerably less than random_page_cost. (However, if the
20 : * database is fully cached in RAM, it is reasonable to set them equal.)
21 : *
22 : * We also use a rough estimate "effective_cache_size" of the number of
23 : * disk pages in Postgres + OS-level disk cache. (We can't simply use
24 : * NBuffers for this purpose because that would ignore the effects of
25 : * the kernel's disk cache.)
26 : *
27 : * Obviously, taking constants for these values is an oversimplification,
28 : * but it's tough enough to get any useful estimates even at this level of
29 : * detail. Note that all of these parameters are user-settable, in case
30 : * the default values are drastically off for a particular platform.
31 : *
32 : * seq_page_cost and random_page_cost can also be overridden for an individual
33 : * tablespace, in case some data is on a fast disk and other data is on a slow
34 : * disk. Per-tablespace overrides never apply to temporary work files such as
35 : * an external sort or a materialize node that overflows work_mem.
36 : *
37 : * We compute two separate costs for each path:
38 : * total_cost: total estimated cost to fetch all tuples
39 : * startup_cost: cost that is expended before first tuple is fetched
40 : * In some scenarios, such as when there is a LIMIT or we are implementing
41 : * an EXISTS(...) sub-select, it is not necessary to fetch all tuples of the
42 : * path's result. A caller can estimate the cost of fetching a partial
43 : * result by interpolating between startup_cost and total_cost. In detail:
44 : * actual_cost = startup_cost +
45 : * (total_cost - startup_cost) * tuples_to_fetch / path->rows;
46 : * Note that a base relation's rows count (and, by extension, plan_rows for
47 : * plan nodes below the LIMIT node) are set without regard to any LIMIT, so
48 : * that this equation works properly. (Note: while path->rows is never zero
49 : * for ordinary relations, it is zero for paths for provably-empty relations,
50 : * so beware of division-by-zero.) The LIMIT is applied as a top-level
51 : * plan node.
52 : *
53 : * Each path stores the total number of disabled nodes that exist at or
54 : * below that point in the plan tree. This is regarded as a component of
55 : * the cost, and paths with fewer disabled nodes should be regarded as
56 : * cheaper than those with more. Disabled nodes occur when the user sets
57 : * a GUC like enable_seqscan=false. We can't necessarily respect such a
58 : * setting in every part of the plan tree, but we want to respect in as many
59 : * parts of the plan tree as possible. Simpler schemes like storing a Boolean
60 : * here rather than a count fail to do that. We used to disable nodes by
61 : * adding a large constant to the startup cost, but that distorted planning
62 : * in other ways.
63 : *
64 : * For largely historical reasons, most of the routines in this module use
65 : * the passed result Path only to store their results (rows, startup_cost and
66 : * total_cost) into. All the input data they need is passed as separate
67 : * parameters, even though much of it could be extracted from the Path.
68 : * An exception is made for the cost_XXXjoin() routines, which expect all
69 : * the other fields of the passed XXXPath to be filled in, and similarly
70 : * cost_index() assumes the passed IndexPath is valid except for its output
71 : * values.
72 : *
73 : *
74 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
75 : * Portions Copyright (c) 1994, Regents of the University of California
76 : *
77 : * IDENTIFICATION
78 : * src/backend/optimizer/path/costsize.c
79 : *
80 : *-------------------------------------------------------------------------
81 : */
82 :
83 : #include "postgres.h"
84 :
85 : #include <limits.h>
86 : #include <math.h>
87 :
88 : #include "access/amapi.h"
89 : #include "access/htup_details.h"
90 : #include "access/tsmapi.h"
91 : #include "executor/executor.h"
92 : #include "executor/nodeAgg.h"
93 : #include "executor/nodeHash.h"
94 : #include "executor/nodeMemoize.h"
95 : #include "miscadmin.h"
96 : #include "nodes/makefuncs.h"
97 : #include "nodes/nodeFuncs.h"
98 : #include "optimizer/clauses.h"
99 : #include "optimizer/cost.h"
100 : #include "optimizer/optimizer.h"
101 : #include "optimizer/pathnode.h"
102 : #include "optimizer/paths.h"
103 : #include "optimizer/placeholder.h"
104 : #include "optimizer/plancat.h"
105 : #include "optimizer/restrictinfo.h"
106 : #include "parser/parsetree.h"
107 : #include "utils/lsyscache.h"
108 : #include "utils/selfuncs.h"
109 : #include "utils/spccache.h"
110 : #include "utils/tuplesort.h"
111 :
112 :
113 : #define LOG2(x) (log(x) / 0.693147180559945)
114 :
115 : /*
116 : * Append and MergeAppend nodes are less expensive than some other operations
117 : * which use cpu_tuple_cost; instead of adding a separate GUC, estimate the
118 : * per-tuple cost as cpu_tuple_cost multiplied by this value.
119 : */
120 : #define APPEND_CPU_COST_MULTIPLIER 0.5
121 :
122 : /*
123 : * Maximum value for row estimates. We cap row estimates to this to help
124 : * ensure that costs based on these estimates remain within the range of what
125 : * double can represent. add_path() wouldn't act sanely given infinite or NaN
126 : * cost values.
127 : */
128 : #define MAXIMUM_ROWCOUNT 1e100
129 :
130 : double seq_page_cost = DEFAULT_SEQ_PAGE_COST;
131 : double random_page_cost = DEFAULT_RANDOM_PAGE_COST;
132 : double cpu_tuple_cost = DEFAULT_CPU_TUPLE_COST;
133 : double cpu_index_tuple_cost = DEFAULT_CPU_INDEX_TUPLE_COST;
134 : double cpu_operator_cost = DEFAULT_CPU_OPERATOR_COST;
135 : double parallel_tuple_cost = DEFAULT_PARALLEL_TUPLE_COST;
136 : double parallel_setup_cost = DEFAULT_PARALLEL_SETUP_COST;
137 : double recursive_worktable_factor = DEFAULT_RECURSIVE_WORKTABLE_FACTOR;
138 :
139 : int effective_cache_size = DEFAULT_EFFECTIVE_CACHE_SIZE;
140 :
141 : Cost disable_cost = 1.0e10;
142 :
143 : int max_parallel_workers_per_gather = 2;
144 :
145 : bool enable_seqscan = true;
146 : bool enable_indexscan = true;
147 : bool enable_indexonlyscan = true;
148 : bool enable_bitmapscan = true;
149 : bool enable_tidscan = true;
150 : bool enable_sort = true;
151 : bool enable_incremental_sort = true;
152 : bool enable_hashagg = true;
153 : bool enable_nestloop = true;
154 : bool enable_material = true;
155 : bool enable_memoize = true;
156 : bool enable_mergejoin = true;
157 : bool enable_hashjoin = true;
158 : bool enable_gathermerge = true;
159 : bool enable_partitionwise_join = false;
160 : bool enable_partitionwise_aggregate = false;
161 : bool enable_parallel_append = true;
162 : bool enable_parallel_hash = true;
163 : bool enable_partition_pruning = true;
164 : bool enable_presorted_aggregate = true;
165 : bool enable_async_append = true;
166 :
167 : typedef struct
168 : {
169 : PlannerInfo *root;
170 : QualCost total;
171 : } cost_qual_eval_context;
172 :
173 : static List *extract_nonindex_conditions(List *qual_clauses, List *indexclauses);
174 : static MergeScanSelCache *cached_scansel(PlannerInfo *root,
175 : RestrictInfo *rinfo,
176 : PathKey *pathkey);
177 : static void cost_rescan(PlannerInfo *root, Path *path,
178 : Cost *rescan_startup_cost, Cost *rescan_total_cost);
179 : static bool cost_qual_eval_walker(Node *node, cost_qual_eval_context *context);
180 : static void get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel,
181 : ParamPathInfo *param_info,
182 : QualCost *qpqual_cost);
183 : static bool has_indexed_join_quals(NestPath *path);
184 : static double approx_tuple_count(PlannerInfo *root, JoinPath *path,
185 : List *quals);
186 : static double calc_joinrel_size_estimate(PlannerInfo *root,
187 : RelOptInfo *joinrel,
188 : RelOptInfo *outer_rel,
189 : RelOptInfo *inner_rel,
190 : double outer_rows,
191 : double inner_rows,
192 : SpecialJoinInfo *sjinfo,
193 : List *restrictlist);
194 : static Selectivity get_foreign_key_join_selectivity(PlannerInfo *root,
195 : Relids outer_relids,
196 : Relids inner_relids,
197 : SpecialJoinInfo *sjinfo,
198 : List **restrictlist);
199 : static Cost append_nonpartial_cost(List *subpaths, int numpaths,
200 : int parallel_workers);
201 : static void set_rel_width(PlannerInfo *root, RelOptInfo *rel);
202 : static int32 get_expr_width(PlannerInfo *root, const Node *expr);
203 : static double relation_byte_size(double tuples, int width);
204 : static double page_size(double tuples, int width);
205 : static double get_parallel_divisor(Path *path);
206 :
207 :
208 : /*
209 : * clamp_row_est
210 : * Force a row-count estimate to a sane value.
211 : */
212 : double
213 8540282 : clamp_row_est(double nrows)
214 : {
215 : /*
216 : * Avoid infinite and NaN row estimates. Costs derived from such values
217 : * are going to be useless. Also force the estimate to be at least one
218 : * row, to make explain output look better and to avoid possible
219 : * divide-by-zero when interpolating costs. Make it an integer, too.
220 : */
221 8540282 : if (nrows > MAXIMUM_ROWCOUNT || isnan(nrows))
222 0 : nrows = MAXIMUM_ROWCOUNT;
223 8540282 : else if (nrows <= 1.0)
224 3052440 : nrows = 1.0;
225 : else
226 5487842 : nrows = rint(nrows);
227 :
228 8540282 : return nrows;
229 : }
230 :
231 : /*
232 : * clamp_width_est
233 : * Force a tuple-width estimate to a sane value.
234 : *
235 : * The planner represents datatype width and tuple width estimates as int32.
236 : * When summing column width estimates to create a tuple width estimate,
237 : * it's possible to reach integer overflow in edge cases. To ensure sane
238 : * behavior, we form such sums in int64 arithmetic and then apply this routine
239 : * to clamp to int32 range.
240 : */
241 : int32
242 1842524 : clamp_width_est(int64 tuple_width)
243 : {
244 : /*
245 : * Anything more than MaxAllocSize is clearly bogus, since we could not
246 : * create a tuple that large.
247 : */
248 1842524 : if (tuple_width > MaxAllocSize)
249 0 : return (int32) MaxAllocSize;
250 :
251 : /*
252 : * Unlike clamp_row_est, we just Assert that the value isn't negative,
253 : * rather than masking such errors.
254 : */
255 : Assert(tuple_width >= 0);
256 :
257 1842524 : return (int32) tuple_width;
258 : }
259 :
260 : /*
261 : * clamp_cardinality_to_long
262 : * Cast a Cardinality value to a sane long value.
263 : */
264 : long
265 45294 : clamp_cardinality_to_long(Cardinality x)
266 : {
267 : /*
268 : * Just for paranoia's sake, ensure we do something sane with negative or
269 : * NaN values.
270 : */
271 45294 : if (isnan(x))
272 0 : return LONG_MAX;
273 45294 : if (x <= 0)
274 584 : return 0;
275 :
276 : /*
277 : * If "long" is 64 bits, then LONG_MAX cannot be represented exactly as a
278 : * double. Casting it to double and back may well result in overflow due
279 : * to rounding, so avoid doing that. We trust that any double value that
280 : * compares strictly less than "(double) LONG_MAX" will cast to a
281 : * representable "long" value.
282 : */
283 44710 : return (x < (double) LONG_MAX) ? (long) x : LONG_MAX;
284 : }
285 :
286 :
287 : /*
288 : * cost_seqscan
289 : * Determines and returns the cost of scanning a relation sequentially.
290 : *
291 : * 'baserel' is the relation to be scanned
292 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
293 : */
294 : void
295 415624 : cost_seqscan(Path *path, PlannerInfo *root,
296 : RelOptInfo *baserel, ParamPathInfo *param_info)
297 : {
298 415624 : Cost startup_cost = 0;
299 : Cost cpu_run_cost;
300 : Cost disk_run_cost;
301 : double spc_seq_page_cost;
302 : QualCost qpqual_cost;
303 : Cost cpu_per_tuple;
304 :
305 : /* Should only be applied to base relations */
306 : Assert(baserel->relid > 0);
307 : Assert(baserel->rtekind == RTE_RELATION);
308 :
309 : /* Mark the path with the correct row estimate */
310 415624 : if (param_info)
311 840 : path->rows = param_info->ppi_rows;
312 : else
313 414784 : path->rows = baserel->rows;
314 :
315 : /* fetch estimated page cost for tablespace containing table */
316 415624 : get_tablespace_page_costs(baserel->reltablespace,
317 : NULL,
318 : &spc_seq_page_cost);
319 :
320 : /*
321 : * disk costs
322 : */
323 415624 : disk_run_cost = spc_seq_page_cost * baserel->pages;
324 :
325 : /* CPU costs */
326 415624 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
327 :
328 415624 : startup_cost += qpqual_cost.startup;
329 415624 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
330 415624 : cpu_run_cost = cpu_per_tuple * baserel->tuples;
331 : /* tlist eval costs are paid per output row, not per tuple scanned */
332 415624 : startup_cost += path->pathtarget->cost.startup;
333 415624 : cpu_run_cost += path->pathtarget->cost.per_tuple * path->rows;
334 :
335 : /* Adjust costing for parallelism, if used. */
336 415624 : if (path->parallel_workers > 0)
337 : {
338 26246 : double parallel_divisor = get_parallel_divisor(path);
339 :
340 : /* The CPU cost is divided among all the workers. */
341 26246 : cpu_run_cost /= parallel_divisor;
342 :
343 : /*
344 : * It may be possible to amortize some of the I/O cost, but probably
345 : * not very much, because most operating systems already do aggressive
346 : * prefetching. For now, we assume that the disk run cost can't be
347 : * amortized at all.
348 : */
349 :
350 : /*
351 : * In the case of a parallel plan, the row count needs to represent
352 : * the number of tuples processed per worker.
353 : */
354 26246 : path->rows = clamp_row_est(path->rows / parallel_divisor);
355 : }
356 :
357 415624 : path->disabled_nodes = enable_seqscan ? 0 : 1;
358 415624 : path->startup_cost = startup_cost;
359 415624 : path->total_cost = startup_cost + cpu_run_cost + disk_run_cost;
360 415624 : }
361 :
362 : /*
363 : * cost_samplescan
364 : * Determines and returns the cost of scanning a relation using sampling.
365 : *
366 : * 'baserel' is the relation to be scanned
367 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
368 : */
369 : void
370 306 : cost_samplescan(Path *path, PlannerInfo *root,
371 : RelOptInfo *baserel, ParamPathInfo *param_info)
372 : {
373 306 : Cost startup_cost = 0;
374 306 : Cost run_cost = 0;
375 : RangeTblEntry *rte;
376 : TableSampleClause *tsc;
377 : TsmRoutine *tsm;
378 : double spc_seq_page_cost,
379 : spc_random_page_cost,
380 : spc_page_cost;
381 : QualCost qpqual_cost;
382 : Cost cpu_per_tuple;
383 :
384 : /* Should only be applied to base relations with tablesample clauses */
385 : Assert(baserel->relid > 0);
386 306 : rte = planner_rt_fetch(baserel->relid, root);
387 : Assert(rte->rtekind == RTE_RELATION);
388 306 : tsc = rte->tablesample;
389 : Assert(tsc != NULL);
390 306 : tsm = GetTsmRoutine(tsc->tsmhandler);
391 :
392 : /* Mark the path with the correct row estimate */
393 306 : if (param_info)
394 72 : path->rows = param_info->ppi_rows;
395 : else
396 234 : path->rows = baserel->rows;
397 :
398 : /* fetch estimated page cost for tablespace containing table */
399 306 : get_tablespace_page_costs(baserel->reltablespace,
400 : &spc_random_page_cost,
401 : &spc_seq_page_cost);
402 :
403 : /* if NextSampleBlock is used, assume random access, else sequential */
404 612 : spc_page_cost = (tsm->NextSampleBlock != NULL) ?
405 306 : spc_random_page_cost : spc_seq_page_cost;
406 :
407 : /*
408 : * disk costs (recall that baserel->pages has already been set to the
409 : * number of pages the sampling method will visit)
410 : */
411 306 : run_cost += spc_page_cost * baserel->pages;
412 :
413 : /*
414 : * CPU costs (recall that baserel->tuples has already been set to the
415 : * number of tuples the sampling method will select). Note that we ignore
416 : * execution cost of the TABLESAMPLE parameter expressions; they will be
417 : * evaluated only once per scan, and in most usages they'll likely be
418 : * simple constants anyway. We also don't charge anything for the
419 : * calculations the sampling method might do internally.
420 : */
421 306 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
422 :
423 306 : startup_cost += qpqual_cost.startup;
424 306 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
425 306 : run_cost += cpu_per_tuple * baserel->tuples;
426 : /* tlist eval costs are paid per output row, not per tuple scanned */
427 306 : startup_cost += path->pathtarget->cost.startup;
428 306 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
429 :
430 306 : path->disabled_nodes = 0;
431 306 : path->startup_cost = startup_cost;
432 306 : path->total_cost = startup_cost + run_cost;
433 306 : }
434 :
435 : /*
436 : * cost_gather
437 : * Determines and returns the cost of gather path.
438 : *
439 : * 'rel' is the relation to be operated upon
440 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
441 : * 'rows' may be used to point to a row estimate; if non-NULL, it overrides
442 : * both 'rel' and 'param_info'. This is useful when the path doesn't exactly
443 : * correspond to any particular RelOptInfo.
444 : */
445 : void
446 22834 : cost_gather(GatherPath *path, PlannerInfo *root,
447 : RelOptInfo *rel, ParamPathInfo *param_info,
448 : double *rows)
449 : {
450 22834 : Cost startup_cost = 0;
451 22834 : Cost run_cost = 0;
452 :
453 : /* Mark the path with the correct row estimate */
454 22834 : if (rows)
455 5400 : path->path.rows = *rows;
456 17434 : else if (param_info)
457 0 : path->path.rows = param_info->ppi_rows;
458 : else
459 17434 : path->path.rows = rel->rows;
460 :
461 22834 : startup_cost = path->subpath->startup_cost;
462 :
463 22834 : run_cost = path->subpath->total_cost - path->subpath->startup_cost;
464 :
465 : /* Parallel setup and communication cost. */
466 22834 : startup_cost += parallel_setup_cost;
467 22834 : run_cost += parallel_tuple_cost * path->path.rows;
468 :
469 22834 : path->path.disabled_nodes = path->subpath->disabled_nodes;
470 22834 : path->path.startup_cost = startup_cost;
471 22834 : path->path.total_cost = (startup_cost + run_cost);
472 22834 : }
473 :
474 : /*
475 : * cost_gather_merge
476 : * Determines and returns the cost of gather merge path.
477 : *
478 : * GatherMerge merges several pre-sorted input streams, using a heap that at
479 : * any given instant holds the next tuple from each stream. If there are N
480 : * streams, we need about N*log2(N) tuple comparisons to construct the heap at
481 : * startup, and then for each output tuple, about log2(N) comparisons to
482 : * replace the top heap entry with the next tuple from the same stream.
483 : */
484 : void
485 14210 : cost_gather_merge(GatherMergePath *path, PlannerInfo *root,
486 : RelOptInfo *rel, ParamPathInfo *param_info,
487 : int input_disabled_nodes,
488 : Cost input_startup_cost, Cost input_total_cost,
489 : double *rows)
490 : {
491 14210 : Cost startup_cost = 0;
492 14210 : Cost run_cost = 0;
493 : Cost comparison_cost;
494 : double N;
495 : double logN;
496 :
497 : /* Mark the path with the correct row estimate */
498 14210 : if (rows)
499 8576 : path->path.rows = *rows;
500 5634 : else if (param_info)
501 0 : path->path.rows = param_info->ppi_rows;
502 : else
503 5634 : path->path.rows = rel->rows;
504 :
505 : /*
506 : * Add one to the number of workers to account for the leader. This might
507 : * be overgenerous since the leader will do less work than other workers
508 : * in typical cases, but we'll go with it for now.
509 : */
510 : Assert(path->num_workers > 0);
511 14210 : N = (double) path->num_workers + 1;
512 14210 : logN = LOG2(N);
513 :
514 : /* Assumed cost per tuple comparison */
515 14210 : comparison_cost = 2.0 * cpu_operator_cost;
516 :
517 : /* Heap creation cost */
518 14210 : startup_cost += comparison_cost * N * logN;
519 :
520 : /* Per-tuple heap maintenance cost */
521 14210 : run_cost += path->path.rows * comparison_cost * logN;
522 :
523 : /* small cost for heap management, like cost_merge_append */
524 14210 : run_cost += cpu_operator_cost * path->path.rows;
525 :
526 : /*
527 : * Parallel setup and communication cost. Since Gather Merge, unlike
528 : * Gather, requires us to block until a tuple is available from every
529 : * worker, we bump the IPC cost up a little bit as compared with Gather.
530 : * For lack of a better idea, charge an extra 5%.
531 : */
532 14210 : startup_cost += parallel_setup_cost;
533 14210 : run_cost += parallel_tuple_cost * path->path.rows * 1.05;
534 :
535 14210 : path->path.disabled_nodes = input_disabled_nodes
536 14210 : + (enable_gathermerge ? 0 : 1);
537 14210 : path->path.startup_cost = startup_cost + input_startup_cost;
538 14210 : path->path.total_cost = (startup_cost + run_cost + input_total_cost);
539 14210 : }
540 :
541 : /*
542 : * cost_index
543 : * Determines and returns the cost of scanning a relation using an index.
544 : *
545 : * 'path' describes the indexscan under consideration, and is complete
546 : * except for the fields to be set by this routine
547 : * 'loop_count' is the number of repetitions of the indexscan to factor into
548 : * estimates of caching behavior
549 : *
550 : * In addition to rows, startup_cost and total_cost, cost_index() sets the
551 : * path's indextotalcost and indexselectivity fields. These values will be
552 : * needed if the IndexPath is used in a BitmapIndexScan.
553 : *
554 : * NOTE: path->indexquals must contain only clauses usable as index
555 : * restrictions. Any additional quals evaluated as qpquals may reduce the
556 : * number of returned tuples, but they won't reduce the number of tuples
557 : * we have to fetch from the table, so they don't reduce the scan cost.
558 : */
559 : void
560 756226 : cost_index(IndexPath *path, PlannerInfo *root, double loop_count,
561 : bool partial_path)
562 : {
563 756226 : IndexOptInfo *index = path->indexinfo;
564 756226 : RelOptInfo *baserel = index->rel;
565 756226 : bool indexonly = (path->path.pathtype == T_IndexOnlyScan);
566 : amcostestimate_function amcostestimate;
567 : List *qpquals;
568 756226 : Cost startup_cost = 0;
569 756226 : Cost run_cost = 0;
570 756226 : Cost cpu_run_cost = 0;
571 : Cost indexStartupCost;
572 : Cost indexTotalCost;
573 : Selectivity indexSelectivity;
574 : double indexCorrelation,
575 : csquared;
576 : double spc_seq_page_cost,
577 : spc_random_page_cost;
578 : Cost min_IO_cost,
579 : max_IO_cost;
580 : QualCost qpqual_cost;
581 : Cost cpu_per_tuple;
582 : double tuples_fetched;
583 : double pages_fetched;
584 : double rand_heap_pages;
585 : double index_pages;
586 :
587 : /* Should only be applied to base relations */
588 : Assert(IsA(baserel, RelOptInfo) &&
589 : IsA(index, IndexOptInfo));
590 : Assert(baserel->relid > 0);
591 : Assert(baserel->rtekind == RTE_RELATION);
592 :
593 : /*
594 : * Mark the path with the correct row estimate, and identify which quals
595 : * will need to be enforced as qpquals. We need not check any quals that
596 : * are implied by the index's predicate, so we can use indrestrictinfo not
597 : * baserestrictinfo as the list of relevant restriction clauses for the
598 : * rel.
599 : */
600 756226 : if (path->path.param_info)
601 : {
602 137896 : path->path.rows = path->path.param_info->ppi_rows;
603 : /* qpquals come from the rel's restriction clauses and ppi_clauses */
604 137896 : qpquals = list_concat(extract_nonindex_conditions(path->indexinfo->indrestrictinfo,
605 : path->indexclauses),
606 137896 : extract_nonindex_conditions(path->path.param_info->ppi_clauses,
607 : path->indexclauses));
608 : }
609 : else
610 : {
611 618330 : path->path.rows = baserel->rows;
612 : /* qpquals come from just the rel's restriction clauses */
613 618330 : qpquals = extract_nonindex_conditions(path->indexinfo->indrestrictinfo,
614 : path->indexclauses);
615 : }
616 :
617 : /* we don't need to check enable_indexonlyscan; indxpath.c does that */
618 756226 : path->path.disabled_nodes = enable_indexscan ? 0 : 1;
619 :
620 : /*
621 : * Call index-access-method-specific code to estimate the processing cost
622 : * for scanning the index, as well as the selectivity of the index (ie,
623 : * the fraction of main-table tuples we will have to retrieve) and its
624 : * correlation to the main-table tuple order. We need a cast here because
625 : * pathnodes.h uses a weak function type to avoid including amapi.h.
626 : */
627 756226 : amcostestimate = (amcostestimate_function) index->amcostestimate;
628 756226 : amcostestimate(root, path, loop_count,
629 : &indexStartupCost, &indexTotalCost,
630 : &indexSelectivity, &indexCorrelation,
631 : &index_pages);
632 :
633 : /*
634 : * Save amcostestimate's results for possible use in bitmap scan planning.
635 : * We don't bother to save indexStartupCost or indexCorrelation, because a
636 : * bitmap scan doesn't care about either.
637 : */
638 756226 : path->indextotalcost = indexTotalCost;
639 756226 : path->indexselectivity = indexSelectivity;
640 :
641 : /* all costs for touching index itself included here */
642 756226 : startup_cost += indexStartupCost;
643 756226 : run_cost += indexTotalCost - indexStartupCost;
644 :
645 : /* estimate number of main-table tuples fetched */
646 756226 : tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
647 :
648 : /* fetch estimated page costs for tablespace containing table */
649 756226 : get_tablespace_page_costs(baserel->reltablespace,
650 : &spc_random_page_cost,
651 : &spc_seq_page_cost);
652 :
653 : /*----------
654 : * Estimate number of main-table pages fetched, and compute I/O cost.
655 : *
656 : * When the index ordering is uncorrelated with the table ordering,
657 : * we use an approximation proposed by Mackert and Lohman (see
658 : * index_pages_fetched() for details) to compute the number of pages
659 : * fetched, and then charge spc_random_page_cost per page fetched.
660 : *
661 : * When the index ordering is exactly correlated with the table ordering
662 : * (just after a CLUSTER, for example), the number of pages fetched should
663 : * be exactly selectivity * table_size. What's more, all but the first
664 : * will be sequential fetches, not the random fetches that occur in the
665 : * uncorrelated case. So if the number of pages is more than 1, we
666 : * ought to charge
667 : * spc_random_page_cost + (pages_fetched - 1) * spc_seq_page_cost
668 : * For partially-correlated indexes, we ought to charge somewhere between
669 : * these two estimates. We currently interpolate linearly between the
670 : * estimates based on the correlation squared (XXX is that appropriate?).
671 : *
672 : * If it's an index-only scan, then we will not need to fetch any heap
673 : * pages for which the visibility map shows all tuples are visible.
674 : * Hence, reduce the estimated number of heap fetches accordingly.
675 : * We use the measured fraction of the entire heap that is all-visible,
676 : * which might not be particularly relevant to the subset of the heap
677 : * that this query will fetch; but it's not clear how to do better.
678 : *----------
679 : */
680 756226 : if (loop_count > 1)
681 : {
682 : /*
683 : * For repeated indexscans, the appropriate estimate for the
684 : * uncorrelated case is to scale up the number of tuples fetched in
685 : * the Mackert and Lohman formula by the number of scans, so that we
686 : * estimate the number of pages fetched by all the scans; then
687 : * pro-rate the costs for one scan. In this case we assume all the
688 : * fetches are random accesses.
689 : */
690 79294 : pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
691 : baserel->pages,
692 79294 : (double) index->pages,
693 : root);
694 :
695 79294 : if (indexonly)
696 9324 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
697 :
698 79294 : rand_heap_pages = pages_fetched;
699 :
700 79294 : max_IO_cost = (pages_fetched * spc_random_page_cost) / loop_count;
701 :
702 : /*
703 : * In the perfectly correlated case, the number of pages touched by
704 : * each scan is selectivity * table_size, and we can use the Mackert
705 : * and Lohman formula at the page level to estimate how much work is
706 : * saved by caching across scans. We still assume all the fetches are
707 : * random, though, which is an overestimate that's hard to correct for
708 : * without double-counting the cache effects. (But in most cases
709 : * where such a plan is actually interesting, only one page would get
710 : * fetched per scan anyway, so it shouldn't matter much.)
711 : */
712 79294 : pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
713 :
714 79294 : pages_fetched = index_pages_fetched(pages_fetched * loop_count,
715 : baserel->pages,
716 79294 : (double) index->pages,
717 : root);
718 :
719 79294 : if (indexonly)
720 9324 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
721 :
722 79294 : min_IO_cost = (pages_fetched * spc_random_page_cost) / loop_count;
723 : }
724 : else
725 : {
726 : /*
727 : * Normal case: apply the Mackert and Lohman formula, and then
728 : * interpolate between that and the correlation-derived result.
729 : */
730 676932 : pages_fetched = index_pages_fetched(tuples_fetched,
731 : baserel->pages,
732 676932 : (double) index->pages,
733 : root);
734 :
735 676932 : if (indexonly)
736 64136 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
737 :
738 676932 : rand_heap_pages = pages_fetched;
739 :
740 : /* max_IO_cost is for the perfectly uncorrelated case (csquared=0) */
741 676932 : max_IO_cost = pages_fetched * spc_random_page_cost;
742 :
743 : /* min_IO_cost is for the perfectly correlated case (csquared=1) */
744 676932 : pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
745 :
746 676932 : if (indexonly)
747 64136 : pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
748 :
749 676932 : if (pages_fetched > 0)
750 : {
751 615974 : min_IO_cost = spc_random_page_cost;
752 615974 : if (pages_fetched > 1)
753 179320 : min_IO_cost += (pages_fetched - 1) * spc_seq_page_cost;
754 : }
755 : else
756 60958 : min_IO_cost = 0;
757 : }
758 :
759 756226 : if (partial_path)
760 : {
761 : /*
762 : * For index only scans compute workers based on number of index pages
763 : * fetched; the number of heap pages we fetch might be so small as to
764 : * effectively rule out parallelism, which we don't want to do.
765 : */
766 262442 : if (indexonly)
767 23224 : rand_heap_pages = -1;
768 :
769 : /*
770 : * Estimate the number of parallel workers required to scan index. Use
771 : * the number of heap pages computed considering heap fetches won't be
772 : * sequential as for parallel scans the pages are accessed in random
773 : * order.
774 : */
775 262442 : path->path.parallel_workers = compute_parallel_worker(baserel,
776 : rand_heap_pages,
777 : index_pages,
778 : max_parallel_workers_per_gather);
779 :
780 : /*
781 : * Fall out if workers can't be assigned for parallel scan, because in
782 : * such a case this path will be rejected. So there is no benefit in
783 : * doing extra computation.
784 : */
785 262442 : if (path->path.parallel_workers <= 0)
786 252328 : return;
787 :
788 10114 : path->path.parallel_aware = true;
789 : }
790 :
791 : /*
792 : * Now interpolate based on estimated index order correlation to get total
793 : * disk I/O cost for main table accesses.
794 : */
795 503898 : csquared = indexCorrelation * indexCorrelation;
796 :
797 503898 : run_cost += max_IO_cost + csquared * (min_IO_cost - max_IO_cost);
798 :
799 : /*
800 : * Estimate CPU costs per tuple.
801 : *
802 : * What we want here is cpu_tuple_cost plus the evaluation costs of any
803 : * qual clauses that we have to evaluate as qpquals.
804 : */
805 503898 : cost_qual_eval(&qpqual_cost, qpquals, root);
806 :
807 503898 : startup_cost += qpqual_cost.startup;
808 503898 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
809 :
810 503898 : cpu_run_cost += cpu_per_tuple * tuples_fetched;
811 :
812 : /* tlist eval costs are paid per output row, not per tuple scanned */
813 503898 : startup_cost += path->path.pathtarget->cost.startup;
814 503898 : cpu_run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
815 :
816 : /* Adjust costing for parallelism, if used. */
817 503898 : if (path->path.parallel_workers > 0)
818 : {
819 10114 : double parallel_divisor = get_parallel_divisor(&path->path);
820 :
821 10114 : path->path.rows = clamp_row_est(path->path.rows / parallel_divisor);
822 :
823 : /* The CPU cost is divided among all the workers. */
824 10114 : cpu_run_cost /= parallel_divisor;
825 : }
826 :
827 503898 : run_cost += cpu_run_cost;
828 :
829 503898 : path->path.startup_cost = startup_cost;
830 503898 : path->path.total_cost = startup_cost + run_cost;
831 : }
832 :
833 : /*
834 : * extract_nonindex_conditions
835 : *
836 : * Given a list of quals to be enforced in an indexscan, extract the ones that
837 : * will have to be applied as qpquals (ie, the index machinery won't handle
838 : * them). Here we detect only whether a qual clause is directly redundant
839 : * with some indexclause. If the index path is chosen for use, createplan.c
840 : * will try a bit harder to get rid of redundant qual conditions; specifically
841 : * it will see if quals can be proven to be implied by the indexquals. But
842 : * it does not seem worth the cycles to try to factor that in at this stage,
843 : * since we're only trying to estimate qual eval costs. Otherwise this must
844 : * match the logic in create_indexscan_plan().
845 : *
846 : * qual_clauses, and the result, are lists of RestrictInfos.
847 : * indexclauses is a list of IndexClauses.
848 : */
849 : static List *
850 894122 : extract_nonindex_conditions(List *qual_clauses, List *indexclauses)
851 : {
852 894122 : List *result = NIL;
853 : ListCell *lc;
854 :
855 1877810 : foreach(lc, qual_clauses)
856 : {
857 983688 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
858 :
859 983688 : if (rinfo->pseudoconstant)
860 9696 : continue; /* we may drop pseudoconstants here */
861 973992 : if (is_redundant_with_indexclauses(rinfo, indexclauses))
862 568800 : continue; /* dup or derived from same EquivalenceClass */
863 : /* ... skip the predicate proof attempt createplan.c will try ... */
864 405192 : result = lappend(result, rinfo);
865 : }
866 894122 : return result;
867 : }
868 :
869 : /*
870 : * index_pages_fetched
871 : * Estimate the number of pages actually fetched after accounting for
872 : * cache effects.
873 : *
874 : * We use an approximation proposed by Mackert and Lohman, "Index Scans
875 : * Using a Finite LRU Buffer: A Validated I/O Model", ACM Transactions
876 : * on Database Systems, Vol. 14, No. 3, September 1989, Pages 401-424.
877 : * The Mackert and Lohman approximation is that the number of pages
878 : * fetched is
879 : * PF =
880 : * min(2TNs/(2T+Ns), T) when T <= b
881 : * 2TNs/(2T+Ns) when T > b and Ns <= 2Tb/(2T-b)
882 : * b + (Ns - 2Tb/(2T-b))*(T-b)/T when T > b and Ns > 2Tb/(2T-b)
883 : * where
884 : * T = # pages in table
885 : * N = # tuples in table
886 : * s = selectivity = fraction of table to be scanned
887 : * b = # buffer pages available (we include kernel space here)
888 : *
889 : * We assume that effective_cache_size is the total number of buffer pages
890 : * available for the whole query, and pro-rate that space across all the
891 : * tables in the query and the index currently under consideration. (This
892 : * ignores space needed for other indexes used by the query, but since we
893 : * don't know which indexes will get used, we can't estimate that very well;
894 : * and in any case counting all the tables may well be an overestimate, since
895 : * depending on the join plan not all the tables may be scanned concurrently.)
896 : *
897 : * The product Ns is the number of tuples fetched; we pass in that
898 : * product rather than calculating it here. "pages" is the number of pages
899 : * in the object under consideration (either an index or a table).
900 : * "index_pages" is the amount to add to the total table space, which was
901 : * computed for us by make_one_rel.
902 : *
903 : * Caller is expected to have ensured that tuples_fetched is greater than zero
904 : * and rounded to integer (see clamp_row_est). The result will likewise be
905 : * greater than zero and integral.
906 : */
907 : double
908 1053104 : index_pages_fetched(double tuples_fetched, BlockNumber pages,
909 : double index_pages, PlannerInfo *root)
910 : {
911 : double pages_fetched;
912 : double total_pages;
913 : double T,
914 : b;
915 :
916 : /* T is # pages in table, but don't allow it to be zero */
917 1053104 : T = (pages > 1) ? (double) pages : 1.0;
918 :
919 : /* Compute number of pages assumed to be competing for cache space */
920 1053104 : total_pages = root->total_table_pages + index_pages;
921 1053104 : total_pages = Max(total_pages, 1.0);
922 : Assert(T <= total_pages);
923 :
924 : /* b is pro-rated share of effective_cache_size */
925 1053104 : b = (double) effective_cache_size * T / total_pages;
926 :
927 : /* force it positive and integral */
928 1053104 : if (b <= 1.0)
929 0 : b = 1.0;
930 : else
931 1053104 : b = ceil(b);
932 :
933 : /* This part is the Mackert and Lohman formula */
934 1053104 : if (T <= b)
935 : {
936 1053104 : pages_fetched =
937 1053104 : (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
938 1053104 : if (pages_fetched >= T)
939 609514 : pages_fetched = T;
940 : else
941 443590 : pages_fetched = ceil(pages_fetched);
942 : }
943 : else
944 : {
945 : double lim;
946 :
947 0 : lim = (2.0 * T * b) / (2.0 * T - b);
948 0 : if (tuples_fetched <= lim)
949 : {
950 0 : pages_fetched =
951 0 : (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
952 : }
953 : else
954 : {
955 0 : pages_fetched =
956 0 : b + (tuples_fetched - lim) * (T - b) / T;
957 : }
958 0 : pages_fetched = ceil(pages_fetched);
959 : }
960 1053104 : return pages_fetched;
961 : }
962 :
963 : /*
964 : * get_indexpath_pages
965 : * Determine the total size of the indexes used in a bitmap index path.
966 : *
967 : * Note: if the same index is used more than once in a bitmap tree, we will
968 : * count it multiple times, which perhaps is the wrong thing ... but it's
969 : * not completely clear, and detecting duplicates is difficult, so ignore it
970 : * for now.
971 : */
972 : static double
973 178756 : get_indexpath_pages(Path *bitmapqual)
974 : {
975 178756 : double result = 0;
976 : ListCell *l;
977 :
978 178756 : if (IsA(bitmapqual, BitmapAndPath))
979 : {
980 22580 : BitmapAndPath *apath = (BitmapAndPath *) bitmapqual;
981 :
982 67740 : foreach(l, apath->bitmapquals)
983 : {
984 45160 : result += get_indexpath_pages((Path *) lfirst(l));
985 : }
986 : }
987 156176 : else if (IsA(bitmapqual, BitmapOrPath))
988 : {
989 70 : BitmapOrPath *opath = (BitmapOrPath *) bitmapqual;
990 :
991 222 : foreach(l, opath->bitmapquals)
992 : {
993 152 : result += get_indexpath_pages((Path *) lfirst(l));
994 : }
995 : }
996 156106 : else if (IsA(bitmapqual, IndexPath))
997 : {
998 156106 : IndexPath *ipath = (IndexPath *) bitmapqual;
999 :
1000 156106 : result = (double) ipath->indexinfo->pages;
1001 : }
1002 : else
1003 0 : elog(ERROR, "unrecognized node type: %d", nodeTag(bitmapqual));
1004 :
1005 178756 : return result;
1006 : }
1007 :
1008 : /*
1009 : * cost_bitmap_heap_scan
1010 : * Determines and returns the cost of scanning a relation using a bitmap
1011 : * index-then-heap plan.
1012 : *
1013 : * 'baserel' is the relation to be scanned
1014 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1015 : * 'bitmapqual' is a tree of IndexPaths, BitmapAndPaths, and BitmapOrPaths
1016 : * 'loop_count' is the number of repetitions of the indexscan to factor into
1017 : * estimates of caching behavior
1018 : *
1019 : * Note: the component IndexPaths in bitmapqual should have been costed
1020 : * using the same loop_count.
1021 : */
1022 : void
1023 516524 : cost_bitmap_heap_scan(Path *path, PlannerInfo *root, RelOptInfo *baserel,
1024 : ParamPathInfo *param_info,
1025 : Path *bitmapqual, double loop_count)
1026 : {
1027 516524 : Cost startup_cost = 0;
1028 516524 : Cost run_cost = 0;
1029 : Cost indexTotalCost;
1030 : QualCost qpqual_cost;
1031 : Cost cpu_per_tuple;
1032 : Cost cost_per_page;
1033 : Cost cpu_run_cost;
1034 : double tuples_fetched;
1035 : double pages_fetched;
1036 : double spc_seq_page_cost,
1037 : spc_random_page_cost;
1038 : double T;
1039 :
1040 : /* Should only be applied to base relations */
1041 : Assert(IsA(baserel, RelOptInfo));
1042 : Assert(baserel->relid > 0);
1043 : Assert(baserel->rtekind == RTE_RELATION);
1044 :
1045 : /* Mark the path with the correct row estimate */
1046 516524 : if (param_info)
1047 213574 : path->rows = param_info->ppi_rows;
1048 : else
1049 302950 : path->rows = baserel->rows;
1050 :
1051 516524 : pages_fetched = compute_bitmap_pages(root, baserel, bitmapqual,
1052 : loop_count, &indexTotalCost,
1053 : &tuples_fetched);
1054 :
1055 516524 : startup_cost += indexTotalCost;
1056 516524 : T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
1057 :
1058 : /* Fetch estimated page costs for tablespace containing table. */
1059 516524 : get_tablespace_page_costs(baserel->reltablespace,
1060 : &spc_random_page_cost,
1061 : &spc_seq_page_cost);
1062 :
1063 : /*
1064 : * For small numbers of pages we should charge spc_random_page_cost
1065 : * apiece, while if nearly all the table's pages are being read, it's more
1066 : * appropriate to charge spc_seq_page_cost apiece. The effect is
1067 : * nonlinear, too. For lack of a better idea, interpolate like this to
1068 : * determine the cost per page.
1069 : */
1070 516524 : if (pages_fetched >= 2.0)
1071 108224 : cost_per_page = spc_random_page_cost -
1072 108224 : (spc_random_page_cost - spc_seq_page_cost)
1073 108224 : * sqrt(pages_fetched / T);
1074 : else
1075 408300 : cost_per_page = spc_random_page_cost;
1076 :
1077 516524 : run_cost += pages_fetched * cost_per_page;
1078 :
1079 : /*
1080 : * Estimate CPU costs per tuple.
1081 : *
1082 : * Often the indexquals don't need to be rechecked at each tuple ... but
1083 : * not always, especially not if there are enough tuples involved that the
1084 : * bitmaps become lossy. For the moment, just assume they will be
1085 : * rechecked always. This means we charge the full freight for all the
1086 : * scan clauses.
1087 : */
1088 516524 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1089 :
1090 516524 : startup_cost += qpqual_cost.startup;
1091 516524 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1092 516524 : cpu_run_cost = cpu_per_tuple * tuples_fetched;
1093 :
1094 : /* Adjust costing for parallelism, if used. */
1095 516524 : if (path->parallel_workers > 0)
1096 : {
1097 4190 : double parallel_divisor = get_parallel_divisor(path);
1098 :
1099 : /* The CPU cost is divided among all the workers. */
1100 4190 : cpu_run_cost /= parallel_divisor;
1101 :
1102 4190 : path->rows = clamp_row_est(path->rows / parallel_divisor);
1103 : }
1104 :
1105 :
1106 516524 : run_cost += cpu_run_cost;
1107 :
1108 : /* tlist eval costs are paid per output row, not per tuple scanned */
1109 516524 : startup_cost += path->pathtarget->cost.startup;
1110 516524 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1111 :
1112 516524 : path->disabled_nodes = enable_bitmapscan ? 0 : 1;
1113 516524 : path->startup_cost = startup_cost;
1114 516524 : path->total_cost = startup_cost + run_cost;
1115 516524 : }
1116 :
1117 : /*
1118 : * cost_bitmap_tree_node
1119 : * Extract cost and selectivity from a bitmap tree node (index/and/or)
1120 : */
1121 : void
1122 967374 : cost_bitmap_tree_node(Path *path, Cost *cost, Selectivity *selec)
1123 : {
1124 967374 : if (IsA(path, IndexPath))
1125 : {
1126 913902 : *cost = ((IndexPath *) path)->indextotalcost;
1127 913902 : *selec = ((IndexPath *) path)->indexselectivity;
1128 :
1129 : /*
1130 : * Charge a small amount per retrieved tuple to reflect the costs of
1131 : * manipulating the bitmap. This is mostly to make sure that a bitmap
1132 : * scan doesn't look to be the same cost as an indexscan to retrieve a
1133 : * single tuple.
1134 : */
1135 913902 : *cost += 0.1 * cpu_operator_cost * path->rows;
1136 : }
1137 53472 : else if (IsA(path, BitmapAndPath))
1138 : {
1139 49936 : *cost = path->total_cost;
1140 49936 : *selec = ((BitmapAndPath *) path)->bitmapselectivity;
1141 : }
1142 3536 : else if (IsA(path, BitmapOrPath))
1143 : {
1144 3536 : *cost = path->total_cost;
1145 3536 : *selec = ((BitmapOrPath *) path)->bitmapselectivity;
1146 : }
1147 : else
1148 : {
1149 0 : elog(ERROR, "unrecognized node type: %d", nodeTag(path));
1150 : *cost = *selec = 0; /* keep compiler quiet */
1151 : }
1152 967374 : }
1153 :
1154 : /*
1155 : * cost_bitmap_and_node
1156 : * Estimate the cost of a BitmapAnd node
1157 : *
1158 : * Note that this considers only the costs of index scanning and bitmap
1159 : * creation, not the eventual heap access. In that sense the object isn't
1160 : * truly a Path, but it has enough path-like properties (costs in particular)
1161 : * to warrant treating it as one. We don't bother to set the path rows field,
1162 : * however.
1163 : */
1164 : void
1165 49730 : cost_bitmap_and_node(BitmapAndPath *path, PlannerInfo *root)
1166 : {
1167 : Cost totalCost;
1168 : Selectivity selec;
1169 : ListCell *l;
1170 :
1171 : /*
1172 : * We estimate AND selectivity on the assumption that the inputs are
1173 : * independent. This is probably often wrong, but we don't have the info
1174 : * to do better.
1175 : *
1176 : * The runtime cost of the BitmapAnd itself is estimated at 100x
1177 : * cpu_operator_cost for each tbm_intersect needed. Probably too small,
1178 : * definitely too simplistic?
1179 : */
1180 49730 : totalCost = 0.0;
1181 49730 : selec = 1.0;
1182 149190 : foreach(l, path->bitmapquals)
1183 : {
1184 99460 : Path *subpath = (Path *) lfirst(l);
1185 : Cost subCost;
1186 : Selectivity subselec;
1187 :
1188 99460 : cost_bitmap_tree_node(subpath, &subCost, &subselec);
1189 :
1190 99460 : selec *= subselec;
1191 :
1192 99460 : totalCost += subCost;
1193 99460 : if (l != list_head(path->bitmapquals))
1194 49730 : totalCost += 100.0 * cpu_operator_cost;
1195 : }
1196 49730 : path->bitmapselectivity = selec;
1197 49730 : path->path.rows = 0; /* per above, not used */
1198 49730 : path->path.disabled_nodes = 0;
1199 49730 : path->path.startup_cost = totalCost;
1200 49730 : path->path.total_cost = totalCost;
1201 49730 : }
1202 :
1203 : /*
1204 : * cost_bitmap_or_node
1205 : * Estimate the cost of a BitmapOr node
1206 : *
1207 : * See comments for cost_bitmap_and_node.
1208 : */
1209 : void
1210 1016 : cost_bitmap_or_node(BitmapOrPath *path, PlannerInfo *root)
1211 : {
1212 : Cost totalCost;
1213 : Selectivity selec;
1214 : ListCell *l;
1215 :
1216 : /*
1217 : * We estimate OR selectivity on the assumption that the inputs are
1218 : * non-overlapping, since that's often the case in "x IN (list)" type
1219 : * situations. Of course, we clamp to 1.0 at the end.
1220 : *
1221 : * The runtime cost of the BitmapOr itself is estimated at 100x
1222 : * cpu_operator_cost for each tbm_union needed. Probably too small,
1223 : * definitely too simplistic? We are aware that the tbm_unions are
1224 : * optimized out when the inputs are BitmapIndexScans.
1225 : */
1226 1016 : totalCost = 0.0;
1227 1016 : selec = 0.0;
1228 2850 : foreach(l, path->bitmapquals)
1229 : {
1230 1834 : Path *subpath = (Path *) lfirst(l);
1231 : Cost subCost;
1232 : Selectivity subselec;
1233 :
1234 1834 : cost_bitmap_tree_node(subpath, &subCost, &subselec);
1235 :
1236 1834 : selec += subselec;
1237 :
1238 1834 : totalCost += subCost;
1239 1834 : if (l != list_head(path->bitmapquals) &&
1240 818 : !IsA(subpath, IndexPath))
1241 0 : totalCost += 100.0 * cpu_operator_cost;
1242 : }
1243 1016 : path->bitmapselectivity = Min(selec, 1.0);
1244 1016 : path->path.rows = 0; /* per above, not used */
1245 1016 : path->path.startup_cost = totalCost;
1246 1016 : path->path.total_cost = totalCost;
1247 1016 : }
1248 :
1249 : /*
1250 : * cost_tidscan
1251 : * Determines and returns the cost of scanning a relation using TIDs.
1252 : *
1253 : * 'baserel' is the relation to be scanned
1254 : * 'tidquals' is the list of TID-checkable quals
1255 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1256 : */
1257 : void
1258 860 : cost_tidscan(Path *path, PlannerInfo *root,
1259 : RelOptInfo *baserel, List *tidquals, ParamPathInfo *param_info)
1260 : {
1261 860 : Cost startup_cost = 0;
1262 860 : Cost run_cost = 0;
1263 : QualCost qpqual_cost;
1264 : Cost cpu_per_tuple;
1265 : QualCost tid_qual_cost;
1266 : double ntuples;
1267 : ListCell *l;
1268 : double spc_random_page_cost;
1269 :
1270 : /* Should only be applied to base relations */
1271 : Assert(baserel->relid > 0);
1272 : Assert(baserel->rtekind == RTE_RELATION);
1273 : Assert(tidquals != NIL);
1274 :
1275 : /* Mark the path with the correct row estimate */
1276 860 : if (param_info)
1277 144 : path->rows = param_info->ppi_rows;
1278 : else
1279 716 : path->rows = baserel->rows;
1280 :
1281 : /* Count how many tuples we expect to retrieve */
1282 860 : ntuples = 0;
1283 1744 : foreach(l, tidquals)
1284 : {
1285 884 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
1286 884 : Expr *qual = rinfo->clause;
1287 :
1288 : /*
1289 : * We must use a TID scan for CurrentOfExpr; in any other case, we
1290 : * should be generating a TID scan only if enable_tidscan=true. Also,
1291 : * if CurrentOfExpr is the qual, there should be only one.
1292 : */
1293 : Assert(enable_tidscan || IsA(qual, CurrentOfExpr));
1294 : Assert(list_length(tidquals) == 1 || !IsA(qual, CurrentOfExpr));
1295 :
1296 884 : if (IsA(qual, ScalarArrayOpExpr))
1297 : {
1298 : /* Each element of the array yields 1 tuple */
1299 50 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) qual;
1300 50 : Node *arraynode = (Node *) lsecond(saop->args);
1301 :
1302 50 : ntuples += estimate_array_length(root, arraynode);
1303 : }
1304 834 : else if (IsA(qual, CurrentOfExpr))
1305 : {
1306 : /* CURRENT OF yields 1 tuple */
1307 404 : ntuples++;
1308 : }
1309 : else
1310 : {
1311 : /* It's just CTID = something, count 1 tuple */
1312 430 : ntuples++;
1313 : }
1314 : }
1315 :
1316 : /*
1317 : * The TID qual expressions will be computed once, any other baserestrict
1318 : * quals once per retrieved tuple.
1319 : */
1320 860 : cost_qual_eval(&tid_qual_cost, tidquals, root);
1321 :
1322 : /* fetch estimated page cost for tablespace containing table */
1323 860 : get_tablespace_page_costs(baserel->reltablespace,
1324 : &spc_random_page_cost,
1325 : NULL);
1326 :
1327 : /* disk costs --- assume each tuple on a different page */
1328 860 : run_cost += spc_random_page_cost * ntuples;
1329 :
1330 : /* Add scanning CPU costs */
1331 860 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1332 :
1333 : /* XXX currently we assume TID quals are a subset of qpquals */
1334 860 : startup_cost += qpqual_cost.startup + tid_qual_cost.per_tuple;
1335 860 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple -
1336 860 : tid_qual_cost.per_tuple;
1337 860 : run_cost += cpu_per_tuple * ntuples;
1338 :
1339 : /* tlist eval costs are paid per output row, not per tuple scanned */
1340 860 : startup_cost += path->pathtarget->cost.startup;
1341 860 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1342 :
1343 : /*
1344 : * There are assertions above verifying that we only reach this function
1345 : * either when enable_tidscan=true or when the TID scan is the only legal
1346 : * path, so it's safe to set disabled_nodes to zero here.
1347 : */
1348 860 : path->disabled_nodes = 0;
1349 860 : path->startup_cost = startup_cost;
1350 860 : path->total_cost = startup_cost + run_cost;
1351 860 : }
1352 :
1353 : /*
1354 : * cost_tidrangescan
1355 : * Determines and sets the costs of scanning a relation using a range of
1356 : * TIDs for 'path'
1357 : *
1358 : * 'baserel' is the relation to be scanned
1359 : * 'tidrangequals' is the list of TID-checkable range quals
1360 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1361 : */
1362 : void
1363 1940 : cost_tidrangescan(Path *path, PlannerInfo *root,
1364 : RelOptInfo *baserel, List *tidrangequals,
1365 : ParamPathInfo *param_info)
1366 : {
1367 : Selectivity selectivity;
1368 : double pages;
1369 1940 : Cost startup_cost = 0;
1370 1940 : Cost run_cost = 0;
1371 : QualCost qpqual_cost;
1372 : Cost cpu_per_tuple;
1373 : QualCost tid_qual_cost;
1374 : double ntuples;
1375 : double nseqpages;
1376 : double spc_random_page_cost;
1377 : double spc_seq_page_cost;
1378 :
1379 : /* Should only be applied to base relations */
1380 : Assert(baserel->relid > 0);
1381 : Assert(baserel->rtekind == RTE_RELATION);
1382 :
1383 : /* Mark the path with the correct row estimate */
1384 1940 : if (param_info)
1385 0 : path->rows = param_info->ppi_rows;
1386 : else
1387 1940 : path->rows = baserel->rows;
1388 :
1389 : /* Count how many tuples and pages we expect to scan */
1390 1940 : selectivity = clauselist_selectivity(root, tidrangequals, baserel->relid,
1391 : JOIN_INNER, NULL);
1392 1940 : pages = ceil(selectivity * baserel->pages);
1393 :
1394 1940 : if (pages <= 0.0)
1395 42 : pages = 1.0;
1396 :
1397 : /*
1398 : * The first page in a range requires a random seek, but each subsequent
1399 : * page is just a normal sequential page read. NOTE: it's desirable for
1400 : * TID Range Scans to cost more than the equivalent Sequential Scans,
1401 : * because Seq Scans have some performance advantages such as scan
1402 : * synchronization and parallelizability, and we'd prefer one of them to
1403 : * be picked unless a TID Range Scan really is better.
1404 : */
1405 1940 : ntuples = selectivity * baserel->tuples;
1406 1940 : nseqpages = pages - 1.0;
1407 :
1408 : /*
1409 : * The TID qual expressions will be computed once, any other baserestrict
1410 : * quals once per retrieved tuple.
1411 : */
1412 1940 : cost_qual_eval(&tid_qual_cost, tidrangequals, root);
1413 :
1414 : /* fetch estimated page cost for tablespace containing table */
1415 1940 : get_tablespace_page_costs(baserel->reltablespace,
1416 : &spc_random_page_cost,
1417 : &spc_seq_page_cost);
1418 :
1419 : /* disk costs; 1 random page and the remainder as seq pages */
1420 1940 : run_cost += spc_random_page_cost + spc_seq_page_cost * nseqpages;
1421 :
1422 : /* Add scanning CPU costs */
1423 1940 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1424 :
1425 : /*
1426 : * XXX currently we assume TID quals are a subset of qpquals at this
1427 : * point; they will be removed (if possible) when we create the plan, so
1428 : * we subtract their cost from the total qpqual cost. (If the TID quals
1429 : * can't be removed, this is a mistake and we're going to underestimate
1430 : * the CPU cost a bit.)
1431 : */
1432 1940 : startup_cost += qpqual_cost.startup + tid_qual_cost.per_tuple;
1433 1940 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple -
1434 1940 : tid_qual_cost.per_tuple;
1435 1940 : run_cost += cpu_per_tuple * ntuples;
1436 :
1437 : /* tlist eval costs are paid per output row, not per tuple scanned */
1438 1940 : startup_cost += path->pathtarget->cost.startup;
1439 1940 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1440 :
1441 : /* we should not generate this path type when enable_tidscan=false */
1442 : Assert(enable_tidscan);
1443 1940 : path->disabled_nodes = 0;
1444 1940 : path->startup_cost = startup_cost;
1445 1940 : path->total_cost = startup_cost + run_cost;
1446 1940 : }
1447 :
1448 : /*
1449 : * cost_subqueryscan
1450 : * Determines and returns the cost of scanning a subquery RTE.
1451 : *
1452 : * 'baserel' is the relation to be scanned
1453 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1454 : * 'trivial_pathtarget' is true if the pathtarget is believed to be trivial.
1455 : */
1456 : void
1457 48404 : cost_subqueryscan(SubqueryScanPath *path, PlannerInfo *root,
1458 : RelOptInfo *baserel, ParamPathInfo *param_info,
1459 : bool trivial_pathtarget)
1460 : {
1461 : Cost startup_cost;
1462 : Cost run_cost;
1463 : List *qpquals;
1464 : QualCost qpqual_cost;
1465 : Cost cpu_per_tuple;
1466 :
1467 : /* Should only be applied to base relations that are subqueries */
1468 : Assert(baserel->relid > 0);
1469 : Assert(baserel->rtekind == RTE_SUBQUERY);
1470 :
1471 : /*
1472 : * We compute the rowcount estimate as the subplan's estimate times the
1473 : * selectivity of relevant restriction clauses. In simple cases this will
1474 : * come out the same as baserel->rows; but when dealing with parallelized
1475 : * paths we must do it like this to get the right answer.
1476 : */
1477 48404 : if (param_info)
1478 606 : qpquals = list_concat_copy(param_info->ppi_clauses,
1479 606 : baserel->baserestrictinfo);
1480 : else
1481 47798 : qpquals = baserel->baserestrictinfo;
1482 :
1483 48404 : path->path.rows = clamp_row_est(path->subpath->rows *
1484 48404 : clauselist_selectivity(root,
1485 : qpquals,
1486 : 0,
1487 : JOIN_INNER,
1488 : NULL));
1489 :
1490 : /*
1491 : * Cost of path is cost of evaluating the subplan, plus cost of evaluating
1492 : * any restriction clauses and tlist that will be attached to the
1493 : * SubqueryScan node, plus cpu_tuple_cost to account for selection and
1494 : * projection overhead.
1495 : */
1496 48404 : path->path.disabled_nodes = path->subpath->disabled_nodes;
1497 48404 : path->path.startup_cost = path->subpath->startup_cost;
1498 48404 : path->path.total_cost = path->subpath->total_cost;
1499 :
1500 : /*
1501 : * However, if there are no relevant restriction clauses and the
1502 : * pathtarget is trivial, then we expect that setrefs.c will optimize away
1503 : * the SubqueryScan plan node altogether, so we should just make its cost
1504 : * and rowcount equal to the input path's.
1505 : *
1506 : * Note: there are some edge cases where createplan.c will apply a
1507 : * different targetlist to the SubqueryScan node, thus falsifying our
1508 : * current estimate of whether the target is trivial, and making the cost
1509 : * estimate (though not the rowcount) wrong. It does not seem worth the
1510 : * extra complication to try to account for that exactly, especially since
1511 : * that behavior falsifies other cost estimates as well.
1512 : */
1513 48404 : if (qpquals == NIL && trivial_pathtarget)
1514 24620 : return;
1515 :
1516 23784 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1517 :
1518 23784 : startup_cost = qpqual_cost.startup;
1519 23784 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1520 23784 : run_cost = cpu_per_tuple * path->subpath->rows;
1521 :
1522 : /* tlist eval costs are paid per output row, not per tuple scanned */
1523 23784 : startup_cost += path->path.pathtarget->cost.startup;
1524 23784 : run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
1525 :
1526 23784 : path->path.startup_cost += startup_cost;
1527 23784 : path->path.total_cost += startup_cost + run_cost;
1528 : }
1529 :
1530 : /*
1531 : * cost_functionscan
1532 : * Determines and returns the cost of scanning a function RTE.
1533 : *
1534 : * 'baserel' is the relation to be scanned
1535 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1536 : */
1537 : void
1538 51118 : cost_functionscan(Path *path, PlannerInfo *root,
1539 : RelOptInfo *baserel, ParamPathInfo *param_info)
1540 : {
1541 51118 : Cost startup_cost = 0;
1542 51118 : Cost run_cost = 0;
1543 : QualCost qpqual_cost;
1544 : Cost cpu_per_tuple;
1545 : RangeTblEntry *rte;
1546 : QualCost exprcost;
1547 :
1548 : /* Should only be applied to base relations that are functions */
1549 : Assert(baserel->relid > 0);
1550 51118 : rte = planner_rt_fetch(baserel->relid, root);
1551 : Assert(rte->rtekind == RTE_FUNCTION);
1552 :
1553 : /* Mark the path with the correct row estimate */
1554 51118 : if (param_info)
1555 8200 : path->rows = param_info->ppi_rows;
1556 : else
1557 42918 : path->rows = baserel->rows;
1558 :
1559 : /*
1560 : * Estimate costs of executing the function expression(s).
1561 : *
1562 : * Currently, nodeFunctionscan.c always executes the functions to
1563 : * completion before returning any rows, and caches the results in a
1564 : * tuplestore. So the function eval cost is all startup cost, and per-row
1565 : * costs are minimal.
1566 : *
1567 : * XXX in principle we ought to charge tuplestore spill costs if the
1568 : * number of rows is large. However, given how phony our rowcount
1569 : * estimates for functions tend to be, there's not a lot of point in that
1570 : * refinement right now.
1571 : */
1572 51118 : cost_qual_eval_node(&exprcost, (Node *) rte->functions, root);
1573 :
1574 51118 : startup_cost += exprcost.startup + exprcost.per_tuple;
1575 :
1576 : /* Add scanning CPU costs */
1577 51118 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1578 :
1579 51118 : startup_cost += qpqual_cost.startup;
1580 51118 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1581 51118 : run_cost += cpu_per_tuple * baserel->tuples;
1582 :
1583 : /* tlist eval costs are paid per output row, not per tuple scanned */
1584 51118 : startup_cost += path->pathtarget->cost.startup;
1585 51118 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1586 :
1587 51118 : path->disabled_nodes = 0;
1588 51118 : path->startup_cost = startup_cost;
1589 51118 : path->total_cost = startup_cost + run_cost;
1590 51118 : }
1591 :
1592 : /*
1593 : * cost_tablefuncscan
1594 : * Determines and returns the cost of scanning a table function.
1595 : *
1596 : * 'baserel' is the relation to be scanned
1597 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1598 : */
1599 : void
1600 626 : cost_tablefuncscan(Path *path, PlannerInfo *root,
1601 : RelOptInfo *baserel, ParamPathInfo *param_info)
1602 : {
1603 626 : Cost startup_cost = 0;
1604 626 : Cost run_cost = 0;
1605 : QualCost qpqual_cost;
1606 : Cost cpu_per_tuple;
1607 : RangeTblEntry *rte;
1608 : QualCost exprcost;
1609 :
1610 : /* Should only be applied to base relations that are functions */
1611 : Assert(baserel->relid > 0);
1612 626 : rte = planner_rt_fetch(baserel->relid, root);
1613 : Assert(rte->rtekind == RTE_TABLEFUNC);
1614 :
1615 : /* Mark the path with the correct row estimate */
1616 626 : if (param_info)
1617 234 : path->rows = param_info->ppi_rows;
1618 : else
1619 392 : path->rows = baserel->rows;
1620 :
1621 : /*
1622 : * Estimate costs of executing the table func expression(s).
1623 : *
1624 : * XXX in principle we ought to charge tuplestore spill costs if the
1625 : * number of rows is large. However, given how phony our rowcount
1626 : * estimates for tablefuncs tend to be, there's not a lot of point in that
1627 : * refinement right now.
1628 : */
1629 626 : cost_qual_eval_node(&exprcost, (Node *) rte->tablefunc, root);
1630 :
1631 626 : startup_cost += exprcost.startup + exprcost.per_tuple;
1632 :
1633 : /* Add scanning CPU costs */
1634 626 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1635 :
1636 626 : startup_cost += qpqual_cost.startup;
1637 626 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1638 626 : run_cost += cpu_per_tuple * baserel->tuples;
1639 :
1640 : /* tlist eval costs are paid per output row, not per tuple scanned */
1641 626 : startup_cost += path->pathtarget->cost.startup;
1642 626 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1643 :
1644 626 : path->disabled_nodes = 0;
1645 626 : path->startup_cost = startup_cost;
1646 626 : path->total_cost = startup_cost + run_cost;
1647 626 : }
1648 :
1649 : /*
1650 : * cost_valuesscan
1651 : * Determines and returns the cost of scanning a VALUES RTE.
1652 : *
1653 : * 'baserel' is the relation to be scanned
1654 : * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1655 : */
1656 : void
1657 8218 : cost_valuesscan(Path *path, PlannerInfo *root,
1658 : RelOptInfo *baserel, ParamPathInfo *param_info)
1659 : {
1660 8218 : Cost startup_cost = 0;
1661 8218 : Cost run_cost = 0;
1662 : QualCost qpqual_cost;
1663 : Cost cpu_per_tuple;
1664 :
1665 : /* Should only be applied to base relations that are values lists */
1666 : Assert(baserel->relid > 0);
1667 : Assert(baserel->rtekind == RTE_VALUES);
1668 :
1669 : /* Mark the path with the correct row estimate */
1670 8218 : if (param_info)
1671 66 : path->rows = param_info->ppi_rows;
1672 : else
1673 8152 : path->rows = baserel->rows;
1674 :
1675 : /*
1676 : * For now, estimate list evaluation cost at one operator eval per list
1677 : * (probably pretty bogus, but is it worth being smarter?)
1678 : */
1679 8218 : cpu_per_tuple = cpu_operator_cost;
1680 :
1681 : /* Add scanning CPU costs */
1682 8218 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1683 :
1684 8218 : startup_cost += qpqual_cost.startup;
1685 8218 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1686 8218 : run_cost += cpu_per_tuple * baserel->tuples;
1687 :
1688 : /* tlist eval costs are paid per output row, not per tuple scanned */
1689 8218 : startup_cost += path->pathtarget->cost.startup;
1690 8218 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1691 :
1692 8218 : path->disabled_nodes = 0;
1693 8218 : path->startup_cost = startup_cost;
1694 8218 : path->total_cost = startup_cost + run_cost;
1695 8218 : }
1696 :
1697 : /*
1698 : * cost_ctescan
1699 : * Determines and returns the cost of scanning a CTE RTE.
1700 : *
1701 : * Note: this is used for both self-reference and regular CTEs; the
1702 : * possible cost differences are below the threshold of what we could
1703 : * estimate accurately anyway. Note that the costs of evaluating the
1704 : * referenced CTE query are added into the final plan as initplan costs,
1705 : * and should NOT be counted here.
1706 : */
1707 : void
1708 5176 : cost_ctescan(Path *path, PlannerInfo *root,
1709 : RelOptInfo *baserel, ParamPathInfo *param_info)
1710 : {
1711 5176 : Cost startup_cost = 0;
1712 5176 : Cost run_cost = 0;
1713 : QualCost qpqual_cost;
1714 : Cost cpu_per_tuple;
1715 :
1716 : /* Should only be applied to base relations that are CTEs */
1717 : Assert(baserel->relid > 0);
1718 : Assert(baserel->rtekind == RTE_CTE);
1719 :
1720 : /* Mark the path with the correct row estimate */
1721 5176 : if (param_info)
1722 0 : path->rows = param_info->ppi_rows;
1723 : else
1724 5176 : path->rows = baserel->rows;
1725 :
1726 : /* Charge one CPU tuple cost per row for tuplestore manipulation */
1727 5176 : cpu_per_tuple = cpu_tuple_cost;
1728 :
1729 : /* Add scanning CPU costs */
1730 5176 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1731 :
1732 5176 : startup_cost += qpqual_cost.startup;
1733 5176 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1734 5176 : run_cost += cpu_per_tuple * baserel->tuples;
1735 :
1736 : /* tlist eval costs are paid per output row, not per tuple scanned */
1737 5176 : startup_cost += path->pathtarget->cost.startup;
1738 5176 : run_cost += path->pathtarget->cost.per_tuple * path->rows;
1739 :
1740 5176 : path->disabled_nodes = 0;
1741 5176 : path->startup_cost = startup_cost;
1742 5176 : path->total_cost = startup_cost + run_cost;
1743 5176 : }
1744 :
1745 : /*
1746 : * cost_namedtuplestorescan
1747 : * Determines and returns the cost of scanning a named tuplestore.
1748 : */
1749 : void
1750 478 : cost_namedtuplestorescan(Path *path, PlannerInfo *root,
1751 : RelOptInfo *baserel, ParamPathInfo *param_info)
1752 : {
1753 478 : Cost startup_cost = 0;
1754 478 : Cost run_cost = 0;
1755 : QualCost qpqual_cost;
1756 : Cost cpu_per_tuple;
1757 :
1758 : /* Should only be applied to base relations that are Tuplestores */
1759 : Assert(baserel->relid > 0);
1760 : Assert(baserel->rtekind == RTE_NAMEDTUPLESTORE);
1761 :
1762 : /* Mark the path with the correct row estimate */
1763 478 : if (param_info)
1764 0 : path->rows = param_info->ppi_rows;
1765 : else
1766 478 : path->rows = baserel->rows;
1767 :
1768 : /* Charge one CPU tuple cost per row for tuplestore manipulation */
1769 478 : cpu_per_tuple = cpu_tuple_cost;
1770 :
1771 : /* Add scanning CPU costs */
1772 478 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1773 :
1774 478 : startup_cost += qpqual_cost.startup;
1775 478 : cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1776 478 : run_cost += cpu_per_tuple * baserel->tuples;
1777 :
1778 478 : path->disabled_nodes = 0;
1779 478 : path->startup_cost = startup_cost;
1780 478 : path->total_cost = startup_cost + run_cost;
1781 478 : }
1782 :
1783 : /*
1784 : * cost_resultscan
1785 : * Determines and returns the cost of scanning an RTE_RESULT relation.
1786 : */
1787 : void
1788 4268 : cost_resultscan(Path *path, PlannerInfo *root,
1789 : RelOptInfo *baserel, ParamPathInfo *param_info)
1790 : {
1791 4268 : Cost startup_cost = 0;
1792 4268 : Cost run_cost = 0;
1793 : QualCost qpqual_cost;
1794 : Cost cpu_per_tuple;
1795 :
1796 : /* Should only be applied to RTE_RESULT base relations */
1797 : Assert(baserel->relid > 0);
1798 : Assert(baserel->rtekind == RTE_RESULT);
1799 :
1800 : /* Mark the path with the correct row estimate */
1801 4268 : if (param_info)
1802 156 : path->rows = param_info->ppi_rows;
1803 : else
1804 4112 : path->rows = baserel->rows;
1805 :
1806 : /* We charge qual cost plus cpu_tuple_cost */
1807 4268 : get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1808 :
1809 4268 : startup_cost += qpqual_cost.startup;
1810 4268 : cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1811 4268 : run_cost += cpu_per_tuple * baserel->tuples;
1812 :
1813 4268 : path->disabled_nodes = 0;
1814 4268 : path->startup_cost = startup_cost;
1815 4268 : path->total_cost = startup_cost + run_cost;
1816 4268 : }
1817 :
1818 : /*
1819 : * cost_recursive_union
1820 : * Determines and returns the cost of performing a recursive union,
1821 : * and also the estimated output size.
1822 : *
1823 : * We are given Paths for the nonrecursive and recursive terms.
1824 : */
1825 : void
1826 920 : cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
1827 : {
1828 : Cost startup_cost;
1829 : Cost total_cost;
1830 : double total_rows;
1831 :
1832 : /* We probably have decent estimates for the non-recursive term */
1833 920 : startup_cost = nrterm->startup_cost;
1834 920 : total_cost = nrterm->total_cost;
1835 920 : total_rows = nrterm->rows;
1836 :
1837 : /*
1838 : * We arbitrarily assume that about 10 recursive iterations will be
1839 : * needed, and that we've managed to get a good fix on the cost and output
1840 : * size of each one of them. These are mighty shaky assumptions but it's
1841 : * hard to see how to do better.
1842 : */
1843 920 : total_cost += 10 * rterm->total_cost;
1844 920 : total_rows += 10 * rterm->rows;
1845 :
1846 : /*
1847 : * Also charge cpu_tuple_cost per row to account for the costs of
1848 : * manipulating the tuplestores. (We don't worry about possible
1849 : * spill-to-disk costs.)
1850 : */
1851 920 : total_cost += cpu_tuple_cost * total_rows;
1852 :
1853 920 : runion->disabled_nodes = nrterm->disabled_nodes + rterm->disabled_nodes;
1854 920 : runion->startup_cost = startup_cost;
1855 920 : runion->total_cost = total_cost;
1856 920 : runion->rows = total_rows;
1857 920 : runion->pathtarget->width = Max(nrterm->pathtarget->width,
1858 : rterm->pathtarget->width);
1859 920 : }
1860 :
1861 : /*
1862 : * cost_tuplesort
1863 : * Determines and returns the cost of sorting a relation using tuplesort,
1864 : * not including the cost of reading the input data.
1865 : *
1866 : * If the total volume of data to sort is less than sort_mem, we will do
1867 : * an in-memory sort, which requires no I/O and about t*log2(t) tuple
1868 : * comparisons for t tuples.
1869 : *
1870 : * If the total volume exceeds sort_mem, we switch to a tape-style merge
1871 : * algorithm. There will still be about t*log2(t) tuple comparisons in
1872 : * total, but we will also need to write and read each tuple once per
1873 : * merge pass. We expect about ceil(logM(r)) merge passes where r is the
1874 : * number of initial runs formed and M is the merge order used by tuplesort.c.
1875 : * Since the average initial run should be about sort_mem, we have
1876 : * disk traffic = 2 * relsize * ceil(logM(p / sort_mem))
1877 : * cpu = comparison_cost * t * log2(t)
1878 : *
1879 : * If the sort is bounded (i.e., only the first k result tuples are needed)
1880 : * and k tuples can fit into sort_mem, we use a heap method that keeps only
1881 : * k tuples in the heap; this will require about t*log2(k) tuple comparisons.
1882 : *
1883 : * The disk traffic is assumed to be 3/4ths sequential and 1/4th random
1884 : * accesses (XXX can't we refine that guess?)
1885 : *
1886 : * By default, we charge two operator evals per tuple comparison, which should
1887 : * be in the right ballpark in most cases. The caller can tweak this by
1888 : * specifying nonzero comparison_cost; typically that's used for any extra
1889 : * work that has to be done to prepare the inputs to the comparison operators.
1890 : *
1891 : * 'tuples' is the number of tuples in the relation
1892 : * 'width' is the average tuple width in bytes
1893 : * 'comparison_cost' is the extra cost per comparison, if any
1894 : * 'sort_mem' is the number of kilobytes of work memory allowed for the sort
1895 : * 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
1896 : */
1897 : static void
1898 1627052 : cost_tuplesort(Cost *startup_cost, Cost *run_cost,
1899 : double tuples, int width,
1900 : Cost comparison_cost, int sort_mem,
1901 : double limit_tuples)
1902 : {
1903 1627052 : double input_bytes = relation_byte_size(tuples, width);
1904 : double output_bytes;
1905 : double output_tuples;
1906 1627052 : int64 sort_mem_bytes = sort_mem * (int64) 1024;
1907 :
1908 : /*
1909 : * We want to be sure the cost of a sort is never estimated as zero, even
1910 : * if passed-in tuple count is zero. Besides, mustn't do log(0)...
1911 : */
1912 1627052 : if (tuples < 2.0)
1913 486288 : tuples = 2.0;
1914 :
1915 : /* Include the default cost-per-comparison */
1916 1627052 : comparison_cost += 2.0 * cpu_operator_cost;
1917 :
1918 : /* Do we have a useful LIMIT? */
1919 1627052 : if (limit_tuples > 0 && limit_tuples < tuples)
1920 : {
1921 1848 : output_tuples = limit_tuples;
1922 1848 : output_bytes = relation_byte_size(output_tuples, width);
1923 : }
1924 : else
1925 : {
1926 1625204 : output_tuples = tuples;
1927 1625204 : output_bytes = input_bytes;
1928 : }
1929 :
1930 1627052 : if (output_bytes > sort_mem_bytes)
1931 : {
1932 : /*
1933 : * We'll have to use a disk-based sort of all the tuples
1934 : */
1935 17234 : double npages = ceil(input_bytes / BLCKSZ);
1936 17234 : double nruns = input_bytes / sort_mem_bytes;
1937 17234 : double mergeorder = tuplesort_merge_order(sort_mem_bytes);
1938 : double log_runs;
1939 : double npageaccesses;
1940 :
1941 : /*
1942 : * CPU costs
1943 : *
1944 : * Assume about N log2 N comparisons
1945 : */
1946 17234 : *startup_cost = comparison_cost * tuples * LOG2(tuples);
1947 :
1948 : /* Disk costs */
1949 :
1950 : /* Compute logM(r) as log(r) / log(M) */
1951 17234 : if (nruns > mergeorder)
1952 4710 : log_runs = ceil(log(nruns) / log(mergeorder));
1953 : else
1954 12524 : log_runs = 1.0;
1955 17234 : npageaccesses = 2.0 * npages * log_runs;
1956 : /* Assume 3/4ths of accesses are sequential, 1/4th are not */
1957 17234 : *startup_cost += npageaccesses *
1958 17234 : (seq_page_cost * 0.75 + random_page_cost * 0.25);
1959 : }
1960 1609818 : else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
1961 : {
1962 : /*
1963 : * We'll use a bounded heap-sort keeping just K tuples in memory, for
1964 : * a total number of tuple comparisons of N log2 K; but the constant
1965 : * factor is a bit higher than for quicksort. Tweak it so that the
1966 : * cost curve is continuous at the crossover point.
1967 : */
1968 1370 : *startup_cost = comparison_cost * tuples * LOG2(2.0 * output_tuples);
1969 : }
1970 : else
1971 : {
1972 : /* We'll use plain quicksort on all the input tuples */
1973 1608448 : *startup_cost = comparison_cost * tuples * LOG2(tuples);
1974 : }
1975 :
1976 : /*
1977 : * Also charge a small amount (arbitrarily set equal to operator cost) per
1978 : * extracted tuple. We don't charge cpu_tuple_cost because a Sort node
1979 : * doesn't do qual-checking or projection, so it has less overhead than
1980 : * most plan nodes. Note it's correct to use tuples not output_tuples
1981 : * here --- the upper LIMIT will pro-rate the run cost so we'd be double
1982 : * counting the LIMIT otherwise.
1983 : */
1984 1627052 : *run_cost = cpu_operator_cost * tuples;
1985 1627052 : }
1986 :
1987 : /*
1988 : * cost_incremental_sort
1989 : * Determines and returns the cost of sorting a relation incrementally, when
1990 : * the input path is presorted by a prefix of the pathkeys.
1991 : *
1992 : * 'presorted_keys' is the number of leading pathkeys by which the input path
1993 : * is sorted.
1994 : *
1995 : * We estimate the number of groups into which the relation is divided by the
1996 : * leading pathkeys, and then calculate the cost of sorting a single group
1997 : * with tuplesort using cost_tuplesort().
1998 : */
1999 : void
2000 11060 : cost_incremental_sort(Path *path,
2001 : PlannerInfo *root, List *pathkeys, int presorted_keys,
2002 : int input_disabled_nodes,
2003 : Cost input_startup_cost, Cost input_total_cost,
2004 : double input_tuples, int width, Cost comparison_cost, int sort_mem,
2005 : double limit_tuples)
2006 : {
2007 : Cost startup_cost,
2008 : run_cost,
2009 11060 : input_run_cost = input_total_cost - input_startup_cost;
2010 : double group_tuples,
2011 : input_groups;
2012 : Cost group_startup_cost,
2013 : group_run_cost,
2014 : group_input_run_cost;
2015 11060 : List *presortedExprs = NIL;
2016 : ListCell *l;
2017 11060 : bool unknown_varno = false;
2018 :
2019 : Assert(presorted_keys > 0 && presorted_keys < list_length(pathkeys));
2020 :
2021 : /*
2022 : * We want to be sure the cost of a sort is never estimated as zero, even
2023 : * if passed-in tuple count is zero. Besides, mustn't do log(0)...
2024 : */
2025 11060 : if (input_tuples < 2.0)
2026 6558 : input_tuples = 2.0;
2027 :
2028 : /* Default estimate of number of groups, capped to one group per row. */
2029 11060 : input_groups = Min(input_tuples, DEFAULT_NUM_DISTINCT);
2030 :
2031 : /*
2032 : * Extract presorted keys as list of expressions.
2033 : *
2034 : * We need to be careful about Vars containing "varno 0" which might have
2035 : * been introduced by generate_append_tlist, which would confuse
2036 : * estimate_num_groups (in fact it'd fail for such expressions). See
2037 : * recurse_set_operations which has to deal with the same issue.
2038 : *
2039 : * Unlike recurse_set_operations we can't access the original target list
2040 : * here, and even if we could it's not very clear how useful would that be
2041 : * for a set operation combining multiple tables. So we simply detect if
2042 : * there are any expressions with "varno 0" and use the default
2043 : * DEFAULT_NUM_DISTINCT in that case.
2044 : *
2045 : * We might also use either 1.0 (a single group) or input_tuples (each row
2046 : * being a separate group), pretty much the worst and best case for
2047 : * incremental sort. But those are extreme cases and using something in
2048 : * between seems reasonable. Furthermore, generate_append_tlist is used
2049 : * for set operations, which are likely to produce mostly unique output
2050 : * anyway - from that standpoint the DEFAULT_NUM_DISTINCT is defensive
2051 : * while maintaining lower startup cost.
2052 : */
2053 11156 : foreach(l, pathkeys)
2054 : {
2055 11156 : PathKey *key = (PathKey *) lfirst(l);
2056 11156 : EquivalenceMember *member = (EquivalenceMember *)
2057 11156 : linitial(key->pk_eclass->ec_members);
2058 :
2059 : /*
2060 : * Check if the expression contains Var with "varno 0" so that we
2061 : * don't call estimate_num_groups in that case.
2062 : */
2063 11156 : if (bms_is_member(0, pull_varnos(root, (Node *) member->em_expr)))
2064 : {
2065 10 : unknown_varno = true;
2066 10 : break;
2067 : }
2068 :
2069 : /* expression not containing any Vars with "varno 0" */
2070 11146 : presortedExprs = lappend(presortedExprs, member->em_expr);
2071 :
2072 11146 : if (foreach_current_index(l) + 1 >= presorted_keys)
2073 11050 : break;
2074 : }
2075 :
2076 : /* Estimate the number of groups with equal presorted keys. */
2077 11060 : if (!unknown_varno)
2078 11050 : input_groups = estimate_num_groups(root, presortedExprs, input_tuples,
2079 : NULL, NULL);
2080 :
2081 11060 : group_tuples = input_tuples / input_groups;
2082 11060 : group_input_run_cost = input_run_cost / input_groups;
2083 :
2084 : /*
2085 : * Estimate the average cost of sorting of one group where presorted keys
2086 : * are equal.
2087 : */
2088 11060 : cost_tuplesort(&group_startup_cost, &group_run_cost,
2089 : group_tuples, width, comparison_cost, sort_mem,
2090 : limit_tuples);
2091 :
2092 : /*
2093 : * Startup cost of incremental sort is the startup cost of its first group
2094 : * plus the cost of its input.
2095 : */
2096 11060 : startup_cost = group_startup_cost + input_startup_cost +
2097 : group_input_run_cost;
2098 :
2099 : /*
2100 : * After we started producing tuples from the first group, the cost of
2101 : * producing all the tuples is given by the cost to finish processing this
2102 : * group, plus the total cost to process the remaining groups, plus the
2103 : * remaining cost of input.
2104 : */
2105 11060 : run_cost = group_run_cost + (group_run_cost + group_startup_cost) *
2106 11060 : (input_groups - 1) + group_input_run_cost * (input_groups - 1);
2107 :
2108 : /*
2109 : * Incremental sort adds some overhead by itself. Firstly, it has to
2110 : * detect the sort groups. This is roughly equal to one extra copy and
2111 : * comparison per tuple.
2112 : */
2113 11060 : run_cost += (cpu_tuple_cost + comparison_cost) * input_tuples;
2114 :
2115 : /*
2116 : * Additionally, we charge double cpu_tuple_cost for each input group to
2117 : * account for the tuplesort_reset that's performed after each group.
2118 : */
2119 11060 : run_cost += 2.0 * cpu_tuple_cost * input_groups;
2120 :
2121 11060 : path->rows = input_tuples;
2122 :
2123 : /* should not generate these paths when enable_incremental_sort=false */
2124 : Assert(enable_incremental_sort);
2125 11060 : path->disabled_nodes = input_disabled_nodes;
2126 :
2127 11060 : path->startup_cost = startup_cost;
2128 11060 : path->total_cost = startup_cost + run_cost;
2129 11060 : }
2130 :
2131 : /*
2132 : * cost_sort
2133 : * Determines and returns the cost of sorting a relation, including
2134 : * the cost of reading the input data.
2135 : *
2136 : * NOTE: some callers currently pass NIL for pathkeys because they
2137 : * can't conveniently supply the sort keys. Since this routine doesn't
2138 : * currently do anything with pathkeys anyway, that doesn't matter...
2139 : * but if it ever does, it should react gracefully to lack of key data.
2140 : * (Actually, the thing we'd most likely be interested in is just the number
2141 : * of sort keys, which all callers *could* supply.)
2142 : */
2143 : void
2144 1615992 : cost_sort(Path *path, PlannerInfo *root,
2145 : List *pathkeys, int input_disabled_nodes,
2146 : Cost input_cost, double tuples, int width,
2147 : Cost comparison_cost, int sort_mem,
2148 : double limit_tuples)
2149 :
2150 : {
2151 : Cost startup_cost;
2152 : Cost run_cost;
2153 :
2154 1615992 : cost_tuplesort(&startup_cost, &run_cost,
2155 : tuples, width,
2156 : comparison_cost, sort_mem,
2157 : limit_tuples);
2158 :
2159 1615992 : startup_cost += input_cost;
2160 :
2161 1615992 : path->rows = tuples;
2162 1615992 : path->disabled_nodes = input_disabled_nodes + (enable_sort ? 0 : 1);
2163 1615992 : path->startup_cost = startup_cost;
2164 1615992 : path->total_cost = startup_cost + run_cost;
2165 1615992 : }
2166 :
2167 : /*
2168 : * append_nonpartial_cost
2169 : * Estimate the cost of the non-partial paths in a Parallel Append.
2170 : * The non-partial paths are assumed to be the first "numpaths" paths
2171 : * from the subpaths list, and to be in order of decreasing cost.
2172 : */
2173 : static Cost
2174 18394 : append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
2175 : {
2176 : Cost *costarr;
2177 : int arrlen;
2178 : ListCell *l;
2179 : ListCell *cell;
2180 : int path_index;
2181 : int min_index;
2182 : int max_index;
2183 :
2184 18394 : if (numpaths == 0)
2185 14254 : return 0;
2186 :
2187 : /*
2188 : * Array length is number of workers or number of relevant paths,
2189 : * whichever is less.
2190 : */
2191 4140 : arrlen = Min(parallel_workers, numpaths);
2192 4140 : costarr = (Cost *) palloc(sizeof(Cost) * arrlen);
2193 :
2194 : /* The first few paths will each be claimed by a different worker. */
2195 4140 : path_index = 0;
2196 11946 : foreach(cell, subpaths)
2197 : {
2198 8650 : Path *subpath = (Path *) lfirst(cell);
2199 :
2200 8650 : if (path_index == arrlen)
2201 844 : break;
2202 7806 : costarr[path_index++] = subpath->total_cost;
2203 : }
2204 :
2205 : /*
2206 : * Since subpaths are sorted by decreasing cost, the last one will have
2207 : * the minimum cost.
2208 : */
2209 4140 : min_index = arrlen - 1;
2210 :
2211 : /*
2212 : * For each of the remaining subpaths, add its cost to the array element
2213 : * with minimum cost.
2214 : */
2215 4622 : for_each_cell(l, subpaths, cell)
2216 : {
2217 1028 : Path *subpath = (Path *) lfirst(l);
2218 :
2219 : /* Consider only the non-partial paths */
2220 1028 : if (path_index++ == numpaths)
2221 546 : break;
2222 :
2223 482 : costarr[min_index] += subpath->total_cost;
2224 :
2225 : /* Update the new min cost array index */
2226 482 : min_index = 0;
2227 1482 : for (int i = 0; i < arrlen; i++)
2228 : {
2229 1000 : if (costarr[i] < costarr[min_index])
2230 184 : min_index = i;
2231 : }
2232 : }
2233 :
2234 : /* Return the highest cost from the array */
2235 4140 : max_index = 0;
2236 11946 : for (int i = 0; i < arrlen; i++)
2237 : {
2238 7806 : if (costarr[i] > costarr[max_index])
2239 200 : max_index = i;
2240 : }
2241 :
2242 4140 : return costarr[max_index];
2243 : }
2244 :
2245 : /*
2246 : * cost_append
2247 : * Determines and returns the cost of an Append node.
2248 : */
2249 : void
2250 54764 : cost_append(AppendPath *apath, PlannerInfo *root)
2251 : {
2252 : ListCell *l;
2253 :
2254 54764 : apath->path.disabled_nodes = 0;
2255 54764 : apath->path.startup_cost = 0;
2256 54764 : apath->path.total_cost = 0;
2257 54764 : apath->path.rows = 0;
2258 :
2259 54764 : if (apath->subpaths == NIL)
2260 1762 : return;
2261 :
2262 53002 : if (!apath->path.parallel_aware)
2263 : {
2264 34608 : List *pathkeys = apath->path.pathkeys;
2265 :
2266 34608 : if (pathkeys == NIL)
2267 : {
2268 32416 : Path *firstsubpath = (Path *) linitial(apath->subpaths);
2269 :
2270 : /*
2271 : * For an unordered, non-parallel-aware Append we take the startup
2272 : * cost as the startup cost of the first subpath.
2273 : */
2274 32416 : apath->path.startup_cost = firstsubpath->startup_cost;
2275 :
2276 : /*
2277 : * Compute rows, number of disabled nodes, and total cost as sums
2278 : * of underlying subplan values.
2279 : */
2280 125274 : foreach(l, apath->subpaths)
2281 : {
2282 92858 : Path *subpath = (Path *) lfirst(l);
2283 :
2284 92858 : apath->path.rows += subpath->rows;
2285 92858 : apath->path.disabled_nodes += subpath->disabled_nodes;
2286 92858 : apath->path.total_cost += subpath->total_cost;
2287 : }
2288 : }
2289 : else
2290 : {
2291 : /*
2292 : * For an ordered, non-parallel-aware Append we take the startup
2293 : * cost as the sum of the subpath startup costs. This ensures
2294 : * that we don't underestimate the startup cost when a query's
2295 : * LIMIT is such that several of the children have to be run to
2296 : * satisfy it. This might be overkill --- another plausible hack
2297 : * would be to take the Append's startup cost as the maximum of
2298 : * the child startup costs. But we don't want to risk believing
2299 : * that an ORDER BY LIMIT query can be satisfied at small cost
2300 : * when the first child has small startup cost but later ones
2301 : * don't. (If we had the ability to deal with nonlinear cost
2302 : * interpolation for partial retrievals, we would not need to be
2303 : * so conservative about this.)
2304 : *
2305 : * This case is also different from the above in that we have to
2306 : * account for possibly injecting sorts into subpaths that aren't
2307 : * natively ordered.
2308 : */
2309 8508 : foreach(l, apath->subpaths)
2310 : {
2311 6316 : Path *subpath = (Path *) lfirst(l);
2312 : int presorted_keys;
2313 : Path sort_path; /* dummy for result of
2314 : * cost_sort/cost_incremental_sort */
2315 :
2316 6316 : if (!pathkeys_count_contained_in(pathkeys, subpath->pathkeys,
2317 : &presorted_keys))
2318 : {
2319 : /*
2320 : * We'll need to insert a Sort node, so include costs for
2321 : * that. We choose to use incremental sort if it is
2322 : * enabled and there are presorted keys; otherwise we use
2323 : * full sort.
2324 : *
2325 : * We can use the parent's LIMIT if any, since we
2326 : * certainly won't pull more than that many tuples from
2327 : * any child.
2328 : */
2329 56 : if (enable_incremental_sort && presorted_keys > 0)
2330 : {
2331 12 : cost_incremental_sort(&sort_path,
2332 : root,
2333 : pathkeys,
2334 : presorted_keys,
2335 : subpath->disabled_nodes,
2336 : subpath->startup_cost,
2337 : subpath->total_cost,
2338 : subpath->rows,
2339 12 : subpath->pathtarget->width,
2340 : 0.0,
2341 : work_mem,
2342 : apath->limit_tuples);
2343 : }
2344 : else
2345 : {
2346 44 : cost_sort(&sort_path,
2347 : root,
2348 : pathkeys,
2349 : subpath->disabled_nodes,
2350 : subpath->total_cost,
2351 : subpath->rows,
2352 44 : subpath->pathtarget->width,
2353 : 0.0,
2354 : work_mem,
2355 : apath->limit_tuples);
2356 : }
2357 :
2358 56 : subpath = &sort_path;
2359 : }
2360 :
2361 6316 : apath->path.rows += subpath->rows;
2362 6316 : apath->path.disabled_nodes += subpath->disabled_nodes;
2363 6316 : apath->path.startup_cost += subpath->startup_cost;
2364 6316 : apath->path.total_cost += subpath->total_cost;
2365 : }
2366 : }
2367 : }
2368 : else /* parallel-aware */
2369 : {
2370 18394 : int i = 0;
2371 18394 : double parallel_divisor = get_parallel_divisor(&apath->path);
2372 :
2373 : /* Parallel-aware Append never produces ordered output. */
2374 : Assert(apath->path.pathkeys == NIL);
2375 :
2376 : /* Calculate startup cost. */
2377 71806 : foreach(l, apath->subpaths)
2378 : {
2379 53412 : Path *subpath = (Path *) lfirst(l);
2380 :
2381 : /*
2382 : * Append will start returning tuples when the child node having
2383 : * lowest startup cost is done setting up. We consider only the
2384 : * first few subplans that immediately get a worker assigned.
2385 : */
2386 53412 : if (i == 0)
2387 18394 : apath->path.startup_cost = subpath->startup_cost;
2388 35018 : else if (i < apath->path.parallel_workers)
2389 17836 : apath->path.startup_cost = Min(apath->path.startup_cost,
2390 : subpath->startup_cost);
2391 :
2392 : /*
2393 : * Apply parallel divisor to subpaths. Scale the number of rows
2394 : * for each partial subpath based on the ratio of the parallel
2395 : * divisor originally used for the subpath to the one we adopted.
2396 : * Also add the cost of partial paths to the total cost, but
2397 : * ignore non-partial paths for now.
2398 : */
2399 53412 : if (i < apath->first_partial_path)
2400 8288 : apath->path.rows += subpath->rows / parallel_divisor;
2401 : else
2402 : {
2403 : double subpath_parallel_divisor;
2404 :
2405 45124 : subpath_parallel_divisor = get_parallel_divisor(subpath);
2406 45124 : apath->path.rows += subpath->rows * (subpath_parallel_divisor /
2407 : parallel_divisor);
2408 45124 : apath->path.total_cost += subpath->total_cost;
2409 : }
2410 :
2411 53412 : apath->path.disabled_nodes += subpath->disabled_nodes;
2412 53412 : apath->path.rows = clamp_row_est(apath->path.rows);
2413 :
2414 53412 : i++;
2415 : }
2416 :
2417 : /* Add cost for non-partial subpaths. */
2418 18394 : apath->path.total_cost +=
2419 18394 : append_nonpartial_cost(apath->subpaths,
2420 : apath->first_partial_path,
2421 : apath->path.parallel_workers);
2422 : }
2423 :
2424 : /*
2425 : * Although Append does not do any selection or projection, it's not free;
2426 : * add a small per-tuple overhead.
2427 : */
2428 53002 : apath->path.total_cost +=
2429 53002 : cpu_tuple_cost * APPEND_CPU_COST_MULTIPLIER * apath->path.rows;
2430 : }
2431 :
2432 : /*
2433 : * cost_merge_append
2434 : * Determines and returns the cost of a MergeAppend node.
2435 : *
2436 : * MergeAppend merges several pre-sorted input streams, using a heap that
2437 : * at any given instant holds the next tuple from each stream. If there
2438 : * are N streams, we need about N*log2(N) tuple comparisons to construct
2439 : * the heap at startup, and then for each output tuple, about log2(N)
2440 : * comparisons to replace the top entry.
2441 : *
2442 : * (The effective value of N will drop once some of the input streams are
2443 : * exhausted, but it seems unlikely to be worth trying to account for that.)
2444 : *
2445 : * The heap is never spilled to disk, since we assume N is not very large.
2446 : * So this is much simpler than cost_sort.
2447 : *
2448 : * As in cost_sort, we charge two operator evals per tuple comparison.
2449 : *
2450 : * 'pathkeys' is a list of sort keys
2451 : * 'n_streams' is the number of input streams
2452 : * 'input_disabled_nodes' is the sum of the input streams' disabled node counts
2453 : * 'input_startup_cost' is the sum of the input streams' startup costs
2454 : * 'input_total_cost' is the sum of the input streams' total costs
2455 : * 'tuples' is the number of tuples in all the streams
2456 : */
2457 : void
2458 4248 : cost_merge_append(Path *path, PlannerInfo *root,
2459 : List *pathkeys, int n_streams,
2460 : int input_disabled_nodes,
2461 : Cost input_startup_cost, Cost input_total_cost,
2462 : double tuples)
2463 : {
2464 4248 : Cost startup_cost = 0;
2465 4248 : Cost run_cost = 0;
2466 : Cost comparison_cost;
2467 : double N;
2468 : double logN;
2469 :
2470 : /*
2471 : * Avoid log(0)...
2472 : */
2473 4248 : N = (n_streams < 2) ? 2.0 : (double) n_streams;
2474 4248 : logN = LOG2(N);
2475 :
2476 : /* Assumed cost per tuple comparison */
2477 4248 : comparison_cost = 2.0 * cpu_operator_cost;
2478 :
2479 : /* Heap creation cost */
2480 4248 : startup_cost += comparison_cost * N * logN;
2481 :
2482 : /* Per-tuple heap maintenance cost */
2483 4248 : run_cost += tuples * comparison_cost * logN;
2484 :
2485 : /*
2486 : * Although MergeAppend does not do any selection or projection, it's not
2487 : * free; add a small per-tuple overhead.
2488 : */
2489 4248 : run_cost += cpu_tuple_cost * APPEND_CPU_COST_MULTIPLIER * tuples;
2490 :
2491 4248 : path->disabled_nodes = input_disabled_nodes;
2492 4248 : path->startup_cost = startup_cost + input_startup_cost;
2493 4248 : path->total_cost = startup_cost + run_cost + input_total_cost;
2494 4248 : }
2495 :
2496 : /*
2497 : * cost_material
2498 : * Determines and returns the cost of materializing a relation, including
2499 : * the cost of reading the input data.
2500 : *
2501 : * If the total volume of data to materialize exceeds work_mem, we will need
2502 : * to write it to disk, so the cost is much higher in that case.
2503 : *
2504 : * Note that here we are estimating the costs for the first scan of the
2505 : * relation, so the materialization is all overhead --- any savings will
2506 : * occur only on rescan, which is estimated in cost_rescan.
2507 : */
2508 : void
2509 505536 : cost_material(Path *path,
2510 : int input_disabled_nodes,
2511 : Cost input_startup_cost, Cost input_total_cost,
2512 : double tuples, int width)
2513 : {
2514 505536 : Cost startup_cost = input_startup_cost;
2515 505536 : Cost run_cost = input_total_cost - input_startup_cost;
2516 505536 : double nbytes = relation_byte_size(tuples, width);
2517 505536 : double work_mem_bytes = work_mem * (Size) 1024;
2518 :
2519 505536 : path->rows = tuples;
2520 :
2521 : /*
2522 : * Whether spilling or not, charge 2x cpu_operator_cost per tuple to
2523 : * reflect bookkeeping overhead. (This rate must be more than what
2524 : * cost_rescan charges for materialize, ie, cpu_operator_cost per tuple;
2525 : * if it is exactly the same then there will be a cost tie between
2526 : * nestloop with A outer, materialized B inner and nestloop with B outer,
2527 : * materialized A inner. The extra cost ensures we'll prefer
2528 : * materializing the smaller rel.) Note that this is normally a good deal
2529 : * less than cpu_tuple_cost; which is OK because a Material plan node
2530 : * doesn't do qual-checking or projection, so it's got less overhead than
2531 : * most plan nodes.
2532 : */
2533 505536 : run_cost += 2 * cpu_operator_cost * tuples;
2534 :
2535 : /*
2536 : * If we will spill to disk, charge at the rate of seq_page_cost per page.
2537 : * This cost is assumed to be evenly spread through the plan run phase,
2538 : * which isn't exactly accurate but our cost model doesn't allow for
2539 : * nonuniform costs within the run phase.
2540 : */
2541 505536 : if (nbytes > work_mem_bytes)
2542 : {
2543 4928 : double npages = ceil(nbytes / BLCKSZ);
2544 :
2545 4928 : run_cost += seq_page_cost * npages;
2546 : }
2547 :
2548 505536 : path->disabled_nodes = input_disabled_nodes + (enable_material ? 0 : 1);
2549 505536 : path->startup_cost = startup_cost;
2550 505536 : path->total_cost = startup_cost + run_cost;
2551 505536 : }
2552 :
2553 : /*
2554 : * cost_memoize_rescan
2555 : * Determines the estimated cost of rescanning a Memoize node.
2556 : *
2557 : * In order to estimate this, we must gain knowledge of how often we expect to
2558 : * be called and how many distinct sets of parameters we are likely to be
2559 : * called with. If we expect a good cache hit ratio, then we can set our
2560 : * costs to account for that hit ratio, plus a little bit of cost for the
2561 : * caching itself. Caching will not work out well if we expect to be called
2562 : * with too many distinct parameter values. The worst-case here is that we
2563 : * never see any parameter value twice, in which case we'd never get a cache
2564 : * hit and caching would be a complete waste of effort.
2565 : */
2566 : static void
2567 274376 : cost_memoize_rescan(PlannerInfo *root, MemoizePath *mpath,
2568 : Cost *rescan_startup_cost, Cost *rescan_total_cost)
2569 : {
2570 : EstimationInfo estinfo;
2571 : ListCell *lc;
2572 274376 : Cost input_startup_cost = mpath->subpath->startup_cost;
2573 274376 : Cost input_total_cost = mpath->subpath->total_cost;
2574 274376 : double tuples = mpath->subpath->rows;
2575 274376 : Cardinality est_calls = mpath->est_calls;
2576 274376 : int width = mpath->subpath->pathtarget->width;
2577 :
2578 : double hash_mem_bytes;
2579 : double est_entry_bytes;
2580 : Cardinality est_cache_entries;
2581 : Cardinality ndistinct;
2582 : double evict_ratio;
2583 : double hit_ratio;
2584 : Cost startup_cost;
2585 : Cost total_cost;
2586 :
2587 : /* available cache space */
2588 274376 : hash_mem_bytes = get_hash_memory_limit();
2589 :
2590 : /*
2591 : * Set the number of bytes each cache entry should consume in the cache.
2592 : * To provide us with better estimations on how many cache entries we can
2593 : * store at once, we make a call to the executor here to ask it what
2594 : * memory overheads there are for a single cache entry.
2595 : */
2596 274376 : est_entry_bytes = relation_byte_size(tuples, width) +
2597 274376 : ExecEstimateCacheEntryOverheadBytes(tuples);
2598 :
2599 : /* include the estimated width for the cache keys */
2600 586128 : foreach(lc, mpath->param_exprs)
2601 311752 : est_entry_bytes += get_expr_width(root, (Node *) lfirst(lc));
2602 :
2603 : /* estimate on the upper limit of cache entries we can hold at once */
2604 274376 : est_cache_entries = floor(hash_mem_bytes / est_entry_bytes);
2605 :
2606 : /* estimate on the distinct number of parameter values */
2607 274376 : ndistinct = estimate_num_groups(root, mpath->param_exprs, est_calls, NULL,
2608 : &estinfo);
2609 :
2610 : /*
2611 : * When the estimation fell back on using a default value, it's a bit too
2612 : * risky to assume that it's ok to use a Memoize node. The use of a
2613 : * default could cause us to use a Memoize node when it's really
2614 : * inappropriate to do so. If we see that this has been done, then we'll
2615 : * assume that every call will have unique parameters, which will almost
2616 : * certainly mean a MemoizePath will never survive add_path().
2617 : */
2618 274376 : if ((estinfo.flags & SELFLAG_USED_DEFAULT) != 0)
2619 16316 : ndistinct = est_calls;
2620 :
2621 : /* Remember the ndistinct estimate for EXPLAIN */
2622 274376 : mpath->est_unique_keys = ndistinct;
2623 :
2624 : /*
2625 : * Since we've already estimated the maximum number of entries we can
2626 : * store at once and know the estimated number of distinct values we'll be
2627 : * called with, we'll take this opportunity to set the path's est_entries.
2628 : * This will ultimately determine the hash table size that the executor
2629 : * will use. If we leave this at zero, the executor will just choose the
2630 : * size itself. Really this is not the right place to do this, but it's
2631 : * convenient since everything is already calculated.
2632 : */
2633 274376 : mpath->est_entries = Min(Min(ndistinct, est_cache_entries),
2634 : PG_UINT32_MAX);
2635 :
2636 : /*
2637 : * When the number of distinct parameter values is above the amount we can
2638 : * store in the cache, then we'll have to evict some entries from the
2639 : * cache. This is not free. Here we estimate how often we'll incur the
2640 : * cost of that eviction.
2641 : */
2642 274376 : evict_ratio = 1.0 - Min(est_cache_entries, ndistinct) / ndistinct;
2643 :
2644 : /*
2645 : * In order to estimate how costly a single scan will be, we need to
2646 : * attempt to estimate what the cache hit ratio will be. To do that we
2647 : * must look at how many scans are estimated in total for this node and
2648 : * how many of those scans we expect to get a cache hit.
2649 : */
2650 548752 : hit_ratio = ((est_calls - ndistinct) / est_calls) *
2651 274376 : (est_cache_entries / Max(ndistinct, est_cache_entries));
2652 :
2653 : /* Remember the hit ratio estimate for EXPLAIN */
2654 274376 : mpath->est_hit_ratio = hit_ratio;
2655 :
2656 : Assert(hit_ratio >= 0 && hit_ratio <= 1.0);
2657 :
2658 : /*
2659 : * Set the total_cost accounting for the expected cache hit ratio. We
2660 : * also add on a cpu_operator_cost to account for a cache lookup. This
2661 : * will happen regardless of whether it's a cache hit or not.
2662 : */
2663 274376 : total_cost = input_total_cost * (1.0 - hit_ratio) + cpu_operator_cost;
2664 :
2665 : /* Now adjust the total cost to account for cache evictions */
2666 :
2667 : /* Charge a cpu_tuple_cost for evicting the actual cache entry */
2668 274376 : total_cost += cpu_tuple_cost * evict_ratio;
2669 :
2670 : /*
2671 : * Charge a 10th of cpu_operator_cost to evict every tuple in that entry.
2672 : * The per-tuple eviction is really just a pfree, so charging a whole
2673 : * cpu_operator_cost seems a little excessive.
2674 : */
2675 274376 : total_cost += cpu_operator_cost / 10.0 * evict_ratio * tuples;
2676 :
2677 : /*
2678 : * Now adjust for storing things in the cache, since that's not free
2679 : * either. Everything must go in the cache. We don't proportion this
2680 : * over any ratio, just apply it once for the scan. We charge a
2681 : * cpu_tuple_cost for the creation of the cache entry and also a
2682 : * cpu_operator_cost for each tuple we expect to cache.
2683 : */
2684 274376 : total_cost += cpu_tuple_cost + cpu_operator_cost * tuples;
2685 :
2686 : /*
2687 : * Getting the first row must be also be proportioned according to the
2688 : * expected cache hit ratio.
2689 : */
2690 274376 : startup_cost = input_startup_cost * (1.0 - hit_ratio);
2691 :
2692 : /*
2693 : * Additionally we charge a cpu_tuple_cost to account for cache lookups,
2694 : * which we'll do regardless of whether it was a cache hit or not.
2695 : */
2696 274376 : startup_cost += cpu_tuple_cost;
2697 :
2698 274376 : *rescan_startup_cost = startup_cost;
2699 274376 : *rescan_total_cost = total_cost;
2700 274376 : }
2701 :
2702 : /*
2703 : * cost_agg
2704 : * Determines and returns the cost of performing an Agg plan node,
2705 : * including the cost of its input.
2706 : *
2707 : * aggcosts can be NULL when there are no actual aggregate functions (i.e.,
2708 : * we are using a hashed Agg node just to do grouping).
2709 : *
2710 : * Note: when aggstrategy == AGG_SORTED, caller must ensure that input costs
2711 : * are for appropriately-sorted input.
2712 : */
2713 : void
2714 74884 : cost_agg(Path *path, PlannerInfo *root,
2715 : AggStrategy aggstrategy, const AggClauseCosts *aggcosts,
2716 : int numGroupCols, double numGroups,
2717 : List *quals,
2718 : int disabled_nodes,
2719 : Cost input_startup_cost, Cost input_total_cost,
2720 : double input_tuples, double input_width)
2721 : {
2722 : double output_tuples;
2723 : Cost startup_cost;
2724 : Cost total_cost;
2725 74884 : const AggClauseCosts dummy_aggcosts = {0};
2726 :
2727 : /* Use all-zero per-aggregate costs if NULL is passed */
2728 74884 : if (aggcosts == NULL)
2729 : {
2730 : Assert(aggstrategy == AGG_HASHED);
2731 18354 : aggcosts = &dummy_aggcosts;
2732 : }
2733 :
2734 : /*
2735 : * The transCost.per_tuple component of aggcosts should be charged once
2736 : * per input tuple, corresponding to the costs of evaluating the aggregate
2737 : * transfns and their input expressions. The finalCost.per_tuple component
2738 : * is charged once per output tuple, corresponding to the costs of
2739 : * evaluating the finalfns. Startup costs are of course charged but once.
2740 : *
2741 : * If we are grouping, we charge an additional cpu_operator_cost per
2742 : * grouping column per input tuple for grouping comparisons.
2743 : *
2744 : * We will produce a single output tuple if not grouping, and a tuple per
2745 : * group otherwise. We charge cpu_tuple_cost for each output tuple.
2746 : *
2747 : * Note: in this cost model, AGG_SORTED and AGG_HASHED have exactly the
2748 : * same total CPU cost, but AGG_SORTED has lower startup cost. If the
2749 : * input path is already sorted appropriately, AGG_SORTED should be
2750 : * preferred (since it has no risk of memory overflow). This will happen
2751 : * as long as the computed total costs are indeed exactly equal --- but if
2752 : * there's roundoff error we might do the wrong thing. So be sure that
2753 : * the computations below form the same intermediate values in the same
2754 : * order.
2755 : */
2756 74884 : if (aggstrategy == AGG_PLAIN)
2757 : {
2758 36884 : startup_cost = input_total_cost;
2759 36884 : startup_cost += aggcosts->transCost.startup;
2760 36884 : startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2761 36884 : startup_cost += aggcosts->finalCost.startup;
2762 36884 : startup_cost += aggcosts->finalCost.per_tuple;
2763 : /* we aren't grouping */
2764 36884 : total_cost = startup_cost + cpu_tuple_cost;
2765 36884 : output_tuples = 1;
2766 : }
2767 38000 : else if (aggstrategy == AGG_SORTED || aggstrategy == AGG_MIXED)
2768 : {
2769 : /* Here we are able to deliver output on-the-fly */
2770 11260 : startup_cost = input_startup_cost;
2771 11260 : total_cost = input_total_cost;
2772 11260 : if (aggstrategy == AGG_MIXED && !enable_hashagg)
2773 456 : ++disabled_nodes;
2774 : /* calcs phrased this way to match HASHED case, see note above */
2775 11260 : total_cost += aggcosts->transCost.startup;
2776 11260 : total_cost += aggcosts->transCost.per_tuple * input_tuples;
2777 11260 : total_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2778 11260 : total_cost += aggcosts->finalCost.startup;
2779 11260 : total_cost += aggcosts->finalCost.per_tuple * numGroups;
2780 11260 : total_cost += cpu_tuple_cost * numGroups;
2781 11260 : output_tuples = numGroups;
2782 : }
2783 : else
2784 : {
2785 : /* must be AGG_HASHED */
2786 26740 : startup_cost = input_total_cost;
2787 26740 : if (!enable_hashagg)
2788 1578 : ++disabled_nodes;
2789 26740 : startup_cost += aggcosts->transCost.startup;
2790 26740 : startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2791 : /* cost of computing hash value */
2792 26740 : startup_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2793 26740 : startup_cost += aggcosts->finalCost.startup;
2794 :
2795 26740 : total_cost = startup_cost;
2796 26740 : total_cost += aggcosts->finalCost.per_tuple * numGroups;
2797 : /* cost of retrieving from hash table */
2798 26740 : total_cost += cpu_tuple_cost * numGroups;
2799 26740 : output_tuples = numGroups;
2800 : }
2801 :
2802 : /*
2803 : * Add the disk costs of hash aggregation that spills to disk.
2804 : *
2805 : * Groups that go into the hash table stay in memory until finalized, so
2806 : * spilling and reprocessing tuples doesn't incur additional invocations
2807 : * of transCost or finalCost. Furthermore, the computed hash value is
2808 : * stored with the spilled tuples, so we don't incur extra invocations of
2809 : * the hash function.
2810 : *
2811 : * Hash Agg begins returning tuples after the first batch is complete.
2812 : * Accrue writes (spilled tuples) to startup_cost and to total_cost;
2813 : * accrue reads only to total_cost.
2814 : */
2815 74884 : if (aggstrategy == AGG_HASHED || aggstrategy == AGG_MIXED)
2816 : {
2817 : double pages;
2818 27656 : double pages_written = 0.0;
2819 27656 : double pages_read = 0.0;
2820 : double spill_cost;
2821 : double hashentrysize;
2822 : double nbatches;
2823 : Size mem_limit;
2824 : uint64 ngroups_limit;
2825 : int num_partitions;
2826 : int depth;
2827 :
2828 : /*
2829 : * Estimate number of batches based on the computed limits. If less
2830 : * than or equal to one, all groups are expected to fit in memory;
2831 : * otherwise we expect to spill.
2832 : */
2833 27656 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
2834 : input_width,
2835 27656 : aggcosts->transitionSpace);
2836 27656 : hash_agg_set_limits(hashentrysize, numGroups, 0, &mem_limit,
2837 : &ngroups_limit, &num_partitions);
2838 :
2839 27656 : nbatches = Max((numGroups * hashentrysize) / mem_limit,
2840 : numGroups / ngroups_limit);
2841 :
2842 27656 : nbatches = Max(ceil(nbatches), 1.0);
2843 27656 : num_partitions = Max(num_partitions, 2);
2844 :
2845 : /*
2846 : * The number of partitions can change at different levels of
2847 : * recursion; but for the purposes of this calculation assume it stays
2848 : * constant.
2849 : */
2850 27656 : depth = ceil(log(nbatches) / log(num_partitions));
2851 :
2852 : /*
2853 : * Estimate number of pages read and written. For each level of
2854 : * recursion, a tuple must be written and then later read.
2855 : */
2856 27656 : pages = relation_byte_size(input_tuples, input_width) / BLCKSZ;
2857 27656 : pages_written = pages_read = pages * depth;
2858 :
2859 : /*
2860 : * HashAgg has somewhat worse IO behavior than Sort on typical
2861 : * hardware/OS combinations. Account for this with a generic penalty.
2862 : */
2863 27656 : pages_read *= 2.0;
2864 27656 : pages_written *= 2.0;
2865 :
2866 27656 : startup_cost += pages_written * random_page_cost;
2867 27656 : total_cost += pages_written * random_page_cost;
2868 27656 : total_cost += pages_read * seq_page_cost;
2869 :
2870 : /* account for CPU cost of spilling a tuple and reading it back */
2871 27656 : spill_cost = depth * input_tuples * 2.0 * cpu_tuple_cost;
2872 27656 : startup_cost += spill_cost;
2873 27656 : total_cost += spill_cost;
2874 : }
2875 :
2876 : /*
2877 : * If there are quals (HAVING quals), account for their cost and
2878 : * selectivity.
2879 : */
2880 74884 : if (quals)
2881 : {
2882 : QualCost qual_cost;
2883 :
2884 4398 : cost_qual_eval(&qual_cost, quals, root);
2885 4398 : startup_cost += qual_cost.startup;
2886 4398 : total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
2887 :
2888 4398 : output_tuples = clamp_row_est(output_tuples *
2889 4398 : clauselist_selectivity(root,
2890 : quals,
2891 : 0,
2892 : JOIN_INNER,
2893 : NULL));
2894 : }
2895 :
2896 74884 : path->rows = output_tuples;
2897 74884 : path->disabled_nodes = disabled_nodes;
2898 74884 : path->startup_cost = startup_cost;
2899 74884 : path->total_cost = total_cost;
2900 74884 : }
2901 :
2902 : /*
2903 : * get_windowclause_startup_tuples
2904 : * Estimate how many tuples we'll need to fetch from a WindowAgg's
2905 : * subnode before we can output the first WindowAgg tuple.
2906 : *
2907 : * How many tuples need to be read depends on the WindowClause. For example,
2908 : * a WindowClause with no PARTITION BY and no ORDER BY requires that all
2909 : * subnode tuples are read and aggregated before the WindowAgg can output
2910 : * anything. If there's a PARTITION BY, then we only need to look at tuples
2911 : * in the first partition. Here we attempt to estimate just how many
2912 : * 'input_tuples' the WindowAgg will need to read for the given WindowClause
2913 : * before the first tuple can be output.
2914 : */
2915 : static double
2916 2754 : get_windowclause_startup_tuples(PlannerInfo *root, WindowClause *wc,
2917 : double input_tuples)
2918 : {
2919 2754 : int frameOptions = wc->frameOptions;
2920 : double partition_tuples;
2921 : double return_tuples;
2922 : double peer_tuples;
2923 :
2924 : /*
2925 : * First, figure out how many partitions there are likely to be and set
2926 : * partition_tuples according to that estimate.
2927 : */
2928 2754 : if (wc->partitionClause != NIL)
2929 : {
2930 : double num_partitions;
2931 716 : List *partexprs = get_sortgrouplist_exprs(wc->partitionClause,
2932 716 : root->parse->targetList);
2933 :
2934 716 : num_partitions = estimate_num_groups(root, partexprs, input_tuples,
2935 : NULL, NULL);
2936 716 : list_free(partexprs);
2937 :
2938 716 : partition_tuples = input_tuples / num_partitions;
2939 : }
2940 : else
2941 : {
2942 : /* all tuples belong to the same partition */
2943 2038 : partition_tuples = input_tuples;
2944 : }
2945 :
2946 : /* estimate the number of tuples in each peer group */
2947 2754 : if (wc->orderClause != NIL)
2948 : {
2949 : double num_groups;
2950 : List *orderexprs;
2951 :
2952 2274 : orderexprs = get_sortgrouplist_exprs(wc->orderClause,
2953 2274 : root->parse->targetList);
2954 :
2955 : /* estimate out how many peer groups there are in the partition */
2956 2274 : num_groups = estimate_num_groups(root, orderexprs,
2957 : partition_tuples, NULL,
2958 : NULL);
2959 2274 : list_free(orderexprs);
2960 2274 : peer_tuples = partition_tuples / num_groups;
2961 : }
2962 : else
2963 : {
2964 : /* no ORDER BY so only 1 tuple belongs in each peer group */
2965 480 : peer_tuples = 1.0;
2966 : }
2967 :
2968 2754 : if (frameOptions & FRAMEOPTION_END_UNBOUNDED_FOLLOWING)
2969 : {
2970 : /* include all partition rows */
2971 346 : return_tuples = partition_tuples;
2972 : }
2973 2408 : else if (frameOptions & FRAMEOPTION_END_CURRENT_ROW)
2974 : {
2975 1418 : if (frameOptions & FRAMEOPTION_ROWS)
2976 : {
2977 : /* just count the current row */
2978 608 : return_tuples = 1.0;
2979 : }
2980 810 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
2981 : {
2982 : /*
2983 : * When in RANGE/GROUPS mode, it's more complex. If there's no
2984 : * ORDER BY, then all rows in the partition are peers, otherwise
2985 : * we'll need to read the first group of peers.
2986 : */
2987 810 : if (wc->orderClause == NIL)
2988 308 : return_tuples = partition_tuples;
2989 : else
2990 502 : return_tuples = peer_tuples;
2991 : }
2992 : else
2993 : {
2994 : /*
2995 : * Something new we don't support yet? This needs attention.
2996 : * We'll just return 1.0 in the meantime.
2997 : */
2998 : Assert(false);
2999 0 : return_tuples = 1.0;
3000 : }
3001 : }
3002 990 : else if (frameOptions & FRAMEOPTION_END_OFFSET_PRECEDING)
3003 : {
3004 : /*
3005 : * BETWEEN ... AND N PRECEDING will only need to read the WindowAgg's
3006 : * subnode after N ROWS/RANGES/GROUPS. N can be 0, but not negative,
3007 : * so we'll just assume only the current row needs to be read to fetch
3008 : * the first WindowAgg row.
3009 : */
3010 108 : return_tuples = 1.0;
3011 : }
3012 882 : else if (frameOptions & FRAMEOPTION_END_OFFSET_FOLLOWING)
3013 : {
3014 882 : Const *endOffset = (Const *) wc->endOffset;
3015 : double end_offset_value;
3016 :
3017 : /* try and figure out the value specified in the endOffset. */
3018 882 : if (IsA(endOffset, Const))
3019 : {
3020 882 : if (endOffset->constisnull)
3021 : {
3022 : /*
3023 : * NULLs are not allowed, but currently, there's no code to
3024 : * error out if there's a NULL Const. We'll only discover
3025 : * this during execution. For now, just pretend everything is
3026 : * fine and assume that just the first row/range/group will be
3027 : * needed.
3028 : */
3029 0 : end_offset_value = 1.0;
3030 : }
3031 : else
3032 : {
3033 882 : switch (endOffset->consttype)
3034 : {
3035 24 : case INT2OID:
3036 24 : end_offset_value =
3037 24 : (double) DatumGetInt16(endOffset->constvalue);
3038 24 : break;
3039 132 : case INT4OID:
3040 132 : end_offset_value =
3041 132 : (double) DatumGetInt32(endOffset->constvalue);
3042 132 : break;
3043 384 : case INT8OID:
3044 384 : end_offset_value =
3045 384 : (double) DatumGetInt64(endOffset->constvalue);
3046 384 : break;
3047 342 : default:
3048 342 : end_offset_value =
3049 342 : partition_tuples / peer_tuples *
3050 : DEFAULT_INEQ_SEL;
3051 342 : break;
3052 : }
3053 : }
3054 : }
3055 : else
3056 : {
3057 : /*
3058 : * When the end bound is not a Const, we'll just need to guess. We
3059 : * just make use of DEFAULT_INEQ_SEL.
3060 : */
3061 0 : end_offset_value =
3062 0 : partition_tuples / peer_tuples * DEFAULT_INEQ_SEL;
3063 : }
3064 :
3065 882 : if (frameOptions & FRAMEOPTION_ROWS)
3066 : {
3067 : /* include the N FOLLOWING and the current row */
3068 222 : return_tuples = end_offset_value + 1.0;
3069 : }
3070 660 : else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
3071 : {
3072 : /* include N FOLLOWING ranges/group and the initial range/group */
3073 660 : return_tuples = peer_tuples * (end_offset_value + 1.0);
3074 : }
3075 : else
3076 : {
3077 : /*
3078 : * Something new we don't support yet? This needs attention.
3079 : * We'll just return 1.0 in the meantime.
3080 : */
3081 : Assert(false);
3082 0 : return_tuples = 1.0;
3083 : }
3084 : }
3085 : else
3086 : {
3087 : /*
3088 : * Something new we don't support yet? This needs attention. We'll
3089 : * just return 1.0 in the meantime.
3090 : */
3091 : Assert(false);
3092 0 : return_tuples = 1.0;
3093 : }
3094 :
3095 2754 : if (wc->partitionClause != NIL || wc->orderClause != NIL)
3096 : {
3097 : /*
3098 : * Cap the return value to the estimated partition tuples and account
3099 : * for the extra tuple WindowAgg will need to read to confirm the next
3100 : * tuple does not belong to the same partition or peer group.
3101 : */
3102 2474 : return_tuples = Min(return_tuples + 1.0, partition_tuples);
3103 : }
3104 : else
3105 : {
3106 : /*
3107 : * Cap the return value so it's never higher than the expected tuples
3108 : * in the partition.
3109 : */
3110 280 : return_tuples = Min(return_tuples, partition_tuples);
3111 : }
3112 :
3113 : /*
3114 : * We needn't worry about any EXCLUDE options as those only exclude rows
3115 : * from being aggregated, not from being read from the WindowAgg's
3116 : * subnode.
3117 : */
3118 :
3119 2754 : return clamp_row_est(return_tuples);
3120 : }
3121 :
3122 : /*
3123 : * cost_windowagg
3124 : * Determines and returns the cost of performing a WindowAgg plan node,
3125 : * including the cost of its input.
3126 : *
3127 : * Input is assumed already properly sorted.
3128 : */
3129 : void
3130 2754 : cost_windowagg(Path *path, PlannerInfo *root,
3131 : List *windowFuncs, WindowClause *winclause,
3132 : int input_disabled_nodes,
3133 : Cost input_startup_cost, Cost input_total_cost,
3134 : double input_tuples)
3135 : {
3136 : Cost startup_cost;
3137 : Cost total_cost;
3138 : double startup_tuples;
3139 : int numPartCols;
3140 : int numOrderCols;
3141 : ListCell *lc;
3142 :
3143 2754 : numPartCols = list_length(winclause->partitionClause);
3144 2754 : numOrderCols = list_length(winclause->orderClause);
3145 :
3146 2754 : startup_cost = input_startup_cost;
3147 2754 : total_cost = input_total_cost;
3148 :
3149 : /*
3150 : * Window functions are assumed to cost their stated execution cost, plus
3151 : * the cost of evaluating their input expressions, per tuple. Since they
3152 : * may in fact evaluate their inputs at multiple rows during each cycle,
3153 : * this could be a drastic underestimate; but without a way to know how
3154 : * many rows the window function will fetch, it's hard to do better. In
3155 : * any case, it's a good estimate for all the built-in window functions,
3156 : * so we'll just do this for now.
3157 : */
3158 6246 : foreach(lc, windowFuncs)
3159 : {
3160 3492 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc);
3161 : Cost wfunccost;
3162 : QualCost argcosts;
3163 :
3164 3492 : argcosts.startup = argcosts.per_tuple = 0;
3165 3492 : add_function_cost(root, wfunc->winfnoid, (Node *) wfunc,
3166 : &argcosts);
3167 3492 : startup_cost += argcosts.startup;
3168 3492 : wfunccost = argcosts.per_tuple;
3169 :
3170 : /* also add the input expressions' cost to per-input-row costs */
3171 3492 : cost_qual_eval_node(&argcosts, (Node *) wfunc->args, root);
3172 3492 : startup_cost += argcosts.startup;
3173 3492 : wfunccost += argcosts.per_tuple;
3174 :
3175 : /*
3176 : * Add the filter's cost to per-input-row costs. XXX We should reduce
3177 : * input expression costs according to filter selectivity.
3178 : */
3179 3492 : cost_qual_eval_node(&argcosts, (Node *) wfunc->aggfilter, root);
3180 3492 : startup_cost += argcosts.startup;
3181 3492 : wfunccost += argcosts.per_tuple;
3182 :
3183 3492 : total_cost += wfunccost * input_tuples;
3184 : }
3185 :
3186 : /*
3187 : * We also charge cpu_operator_cost per grouping column per tuple for
3188 : * grouping comparisons, plus cpu_tuple_cost per tuple for general
3189 : * overhead.
3190 : *
3191 : * XXX this neglects costs of spooling the data to disk when it overflows
3192 : * work_mem. Sooner or later that should get accounted for.
3193 : */
3194 2754 : total_cost += cpu_operator_cost * (numPartCols + numOrderCols) * input_tuples;
3195 2754 : total_cost += cpu_tuple_cost * input_tuples;
3196 :
3197 2754 : path->rows = input_tuples;
3198 2754 : path->disabled_nodes = input_disabled_nodes;
3199 2754 : path->startup_cost = startup_cost;
3200 2754 : path->total_cost = total_cost;
3201 :
3202 : /*
3203 : * Also, take into account how many tuples we need to read from the
3204 : * subnode in order to produce the first tuple from the WindowAgg. To do
3205 : * this we proportion the run cost (total cost not including startup cost)
3206 : * over the estimated startup tuples. We already included the startup
3207 : * cost of the subnode, so we only need to do this when the estimated
3208 : * startup tuples is above 1.0.
3209 : */
3210 2754 : startup_tuples = get_windowclause_startup_tuples(root, winclause,
3211 : input_tuples);
3212 :
3213 2754 : if (startup_tuples > 1.0)
3214 2466 : path->startup_cost += (total_cost - startup_cost) / input_tuples *
3215 2466 : (startup_tuples - 1.0);
3216 2754 : }
3217 :
3218 : /*
3219 : * cost_group
3220 : * Determines and returns the cost of performing a Group plan node,
3221 : * including the cost of its input.
3222 : *
3223 : * Note: caller must ensure that input costs are for appropriately-sorted
3224 : * input.
3225 : */
3226 : void
3227 1214 : cost_group(Path *path, PlannerInfo *root,
3228 : int numGroupCols, double numGroups,
3229 : List *quals,
3230 : int input_disabled_nodes,
3231 : Cost input_startup_cost, Cost input_total_cost,
3232 : double input_tuples)
3233 : {
3234 : double output_tuples;
3235 : Cost startup_cost;
3236 : Cost total_cost;
3237 :
3238 1214 : output_tuples = numGroups;
3239 1214 : startup_cost = input_startup_cost;
3240 1214 : total_cost = input_total_cost;
3241 :
3242 : /*
3243 : * Charge one cpu_operator_cost per comparison per input tuple. We assume
3244 : * all columns get compared at most of the tuples.
3245 : */
3246 1214 : total_cost += cpu_operator_cost * input_tuples * numGroupCols;
3247 :
3248 : /*
3249 : * If there are quals (HAVING quals), account for their cost and
3250 : * selectivity.
3251 : */
3252 1214 : if (quals)
3253 : {
3254 : QualCost qual_cost;
3255 :
3256 0 : cost_qual_eval(&qual_cost, quals, root);
3257 0 : startup_cost += qual_cost.startup;
3258 0 : total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
3259 :
3260 0 : output_tuples = clamp_row_est(output_tuples *
3261 0 : clauselist_selectivity(root,
3262 : quals,
3263 : 0,
3264 : JOIN_INNER,
3265 : NULL));
3266 : }
3267 :
3268 1214 : path->rows = output_tuples;
3269 1214 : path->disabled_nodes = input_disabled_nodes;
3270 1214 : path->startup_cost = startup_cost;
3271 1214 : path->total_cost = total_cost;
3272 1214 : }
3273 :
3274 : /*
3275 : * initial_cost_nestloop
3276 : * Preliminary estimate of the cost of a nestloop join path.
3277 : *
3278 : * This must quickly produce lower-bound estimates of the path's startup and
3279 : * total costs. If we are unable to eliminate the proposed path from
3280 : * consideration using the lower bounds, final_cost_nestloop will be called
3281 : * to obtain the final estimates.
3282 : *
3283 : * The exact division of labor between this function and final_cost_nestloop
3284 : * is private to them, and represents a tradeoff between speed of the initial
3285 : * estimate and getting a tight lower bound. We choose to not examine the
3286 : * join quals here, since that's by far the most expensive part of the
3287 : * calculations. The end result is that CPU-cost considerations must be
3288 : * left for the second phase; and for SEMI/ANTI joins, we must also postpone
3289 : * incorporation of the inner path's run cost.
3290 : *
3291 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3292 : * other data to be used by final_cost_nestloop
3293 : * 'jointype' is the type of join to be performed
3294 : * 'outer_path' is the outer input to the join
3295 : * 'inner_path' is the inner input to the join
3296 : * 'extra' contains miscellaneous information about the join
3297 : */
3298 : void
3299 2717354 : initial_cost_nestloop(PlannerInfo *root, JoinCostWorkspace *workspace,
3300 : JoinType jointype,
3301 : Path *outer_path, Path *inner_path,
3302 : JoinPathExtraData *extra)
3303 : {
3304 : int disabled_nodes;
3305 2717354 : Cost startup_cost = 0;
3306 2717354 : Cost run_cost = 0;
3307 2717354 : double outer_path_rows = outer_path->rows;
3308 : Cost inner_rescan_start_cost;
3309 : Cost inner_rescan_total_cost;
3310 : Cost inner_run_cost;
3311 : Cost inner_rescan_run_cost;
3312 :
3313 : /* Count up disabled nodes. */
3314 2717354 : disabled_nodes = enable_nestloop ? 0 : 1;
3315 2717354 : disabled_nodes += inner_path->disabled_nodes;
3316 2717354 : disabled_nodes += outer_path->disabled_nodes;
3317 :
3318 : /* estimate costs to rescan the inner relation */
3319 2717354 : cost_rescan(root, inner_path,
3320 : &inner_rescan_start_cost,
3321 : &inner_rescan_total_cost);
3322 :
3323 : /* cost of source data */
3324 :
3325 : /*
3326 : * NOTE: clearly, we must pay both outer and inner paths' startup_cost
3327 : * before we can start returning tuples, so the join's startup cost is
3328 : * their sum. We'll also pay the inner path's rescan startup cost
3329 : * multiple times.
3330 : */
3331 2717354 : startup_cost += outer_path->startup_cost + inner_path->startup_cost;
3332 2717354 : run_cost += outer_path->total_cost - outer_path->startup_cost;
3333 2717354 : if (outer_path_rows > 1)
3334 1914260 : run_cost += (outer_path_rows - 1) * inner_rescan_start_cost;
3335 :
3336 2717354 : inner_run_cost = inner_path->total_cost - inner_path->startup_cost;
3337 2717354 : inner_rescan_run_cost = inner_rescan_total_cost - inner_rescan_start_cost;
3338 :
3339 2717354 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI ||
3340 2656240 : extra->inner_unique)
3341 : {
3342 : /*
3343 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3344 : * executor will stop after the first match.
3345 : *
3346 : * Getting decent estimates requires inspection of the join quals,
3347 : * which we choose to postpone to final_cost_nestloop.
3348 : */
3349 :
3350 : /* Save private data for final_cost_nestloop */
3351 1239392 : workspace->inner_run_cost = inner_run_cost;
3352 1239392 : workspace->inner_rescan_run_cost = inner_rescan_run_cost;
3353 : }
3354 : else
3355 : {
3356 : /* Normal case; we'll scan whole input rel for each outer row */
3357 1477962 : run_cost += inner_run_cost;
3358 1477962 : if (outer_path_rows > 1)
3359 1066164 : run_cost += (outer_path_rows - 1) * inner_rescan_run_cost;
3360 : }
3361 :
3362 : /* CPU costs left for later */
3363 :
3364 : /* Public result fields */
3365 2717354 : workspace->disabled_nodes = disabled_nodes;
3366 2717354 : workspace->startup_cost = startup_cost;
3367 2717354 : workspace->total_cost = startup_cost + run_cost;
3368 : /* Save private data for final_cost_nestloop */
3369 2717354 : workspace->run_cost = run_cost;
3370 2717354 : }
3371 :
3372 : /*
3373 : * final_cost_nestloop
3374 : * Final estimate of the cost and result size of a nestloop join path.
3375 : *
3376 : * 'path' is already filled in except for the rows and cost fields
3377 : * 'workspace' is the result from initial_cost_nestloop
3378 : * 'extra' contains miscellaneous information about the join
3379 : */
3380 : void
3381 1324882 : final_cost_nestloop(PlannerInfo *root, NestPath *path,
3382 : JoinCostWorkspace *workspace,
3383 : JoinPathExtraData *extra)
3384 : {
3385 1324882 : Path *outer_path = path->jpath.outerjoinpath;
3386 1324882 : Path *inner_path = path->jpath.innerjoinpath;
3387 1324882 : double outer_path_rows = outer_path->rows;
3388 1324882 : double inner_path_rows = inner_path->rows;
3389 1324882 : Cost startup_cost = workspace->startup_cost;
3390 1324882 : Cost run_cost = workspace->run_cost;
3391 : Cost cpu_per_tuple;
3392 : QualCost restrict_qual_cost;
3393 : double ntuples;
3394 :
3395 : /* Set the number of disabled nodes. */
3396 1324882 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
3397 :
3398 : /* Protect some assumptions below that rowcounts aren't zero */
3399 1324882 : if (outer_path_rows <= 0)
3400 0 : outer_path_rows = 1;
3401 1324882 : if (inner_path_rows <= 0)
3402 702 : inner_path_rows = 1;
3403 : /* Mark the path with the correct row estimate */
3404 1324882 : if (path->jpath.path.param_info)
3405 30078 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3406 : else
3407 1294804 : path->jpath.path.rows = path->jpath.path.parent->rows;
3408 :
3409 : /* For partial paths, scale row estimate. */
3410 1324882 : if (path->jpath.path.parallel_workers > 0)
3411 : {
3412 13450 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3413 :
3414 13450 : path->jpath.path.rows =
3415 13450 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
3416 : }
3417 :
3418 : /* cost of inner-relation source data (we already dealt with outer rel) */
3419 :
3420 1324882 : if (path->jpath.jointype == JOIN_SEMI || path->jpath.jointype == JOIN_ANTI ||
3421 1282318 : extra->inner_unique)
3422 853894 : {
3423 : /*
3424 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3425 : * executor will stop after the first match.
3426 : */
3427 853894 : Cost inner_run_cost = workspace->inner_run_cost;
3428 853894 : Cost inner_rescan_run_cost = workspace->inner_rescan_run_cost;
3429 : double outer_matched_rows;
3430 : double outer_unmatched_rows;
3431 : Selectivity inner_scan_frac;
3432 :
3433 : /*
3434 : * For an outer-rel row that has at least one match, we can expect the
3435 : * inner scan to stop after a fraction 1/(match_count+1) of the inner
3436 : * rows, if the matches are evenly distributed. Since they probably
3437 : * aren't quite evenly distributed, we apply a fuzz factor of 2.0 to
3438 : * that fraction. (If we used a larger fuzz factor, we'd have to
3439 : * clamp inner_scan_frac to at most 1.0; but since match_count is at
3440 : * least 1, no such clamp is needed now.)
3441 : */
3442 853894 : outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
3443 853894 : outer_unmatched_rows = outer_path_rows - outer_matched_rows;
3444 853894 : inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
3445 :
3446 : /*
3447 : * Compute number of tuples processed (not number emitted!). First,
3448 : * account for successfully-matched outer rows.
3449 : */
3450 853894 : ntuples = outer_matched_rows * inner_path_rows * inner_scan_frac;
3451 :
3452 : /*
3453 : * Now we need to estimate the actual costs of scanning the inner
3454 : * relation, which may be quite a bit less than N times inner_run_cost
3455 : * due to early scan stops. We consider two cases. If the inner path
3456 : * is an indexscan using all the joinquals as indexquals, then an
3457 : * unmatched outer row results in an indexscan returning no rows,
3458 : * which is probably quite cheap. Otherwise, the executor will have
3459 : * to scan the whole inner rel for an unmatched row; not so cheap.
3460 : */
3461 853894 : if (has_indexed_join_quals(path))
3462 : {
3463 : /*
3464 : * Successfully-matched outer rows will only require scanning
3465 : * inner_scan_frac of the inner relation. In this case, we don't
3466 : * need to charge the full inner_run_cost even when that's more
3467 : * than inner_rescan_run_cost, because we can assume that none of
3468 : * the inner scans ever scan the whole inner relation. So it's
3469 : * okay to assume that all the inner scan executions can be
3470 : * fractions of the full cost, even if materialization is reducing
3471 : * the rescan cost. At this writing, it's impossible to get here
3472 : * for a materialized inner scan, so inner_run_cost and
3473 : * inner_rescan_run_cost will be the same anyway; but just in
3474 : * case, use inner_run_cost for the first matched tuple and
3475 : * inner_rescan_run_cost for additional ones.
3476 : */
3477 142284 : run_cost += inner_run_cost * inner_scan_frac;
3478 142284 : if (outer_matched_rows > 1)
3479 19616 : run_cost += (outer_matched_rows - 1) * inner_rescan_run_cost * inner_scan_frac;
3480 :
3481 : /*
3482 : * Add the cost of inner-scan executions for unmatched outer rows.
3483 : * We estimate this as the same cost as returning the first tuple
3484 : * of a nonempty scan. We consider that these are all rescans,
3485 : * since we used inner_run_cost once already.
3486 : */
3487 142284 : run_cost += outer_unmatched_rows *
3488 142284 : inner_rescan_run_cost / inner_path_rows;
3489 :
3490 : /*
3491 : * We won't be evaluating any quals at all for unmatched rows, so
3492 : * don't add them to ntuples.
3493 : */
3494 : }
3495 : else
3496 : {
3497 : /*
3498 : * Here, a complicating factor is that rescans may be cheaper than
3499 : * first scans. If we never scan all the way to the end of the
3500 : * inner rel, it might be (depending on the plan type) that we'd
3501 : * never pay the whole inner first-scan run cost. However it is
3502 : * difficult to estimate whether that will happen (and it could
3503 : * not happen if there are any unmatched outer rows!), so be
3504 : * conservative and always charge the whole first-scan cost once.
3505 : * We consider this charge to correspond to the first unmatched
3506 : * outer row, unless there isn't one in our estimate, in which
3507 : * case blame it on the first matched row.
3508 : */
3509 :
3510 : /* First, count all unmatched join tuples as being processed */
3511 711610 : ntuples += outer_unmatched_rows * inner_path_rows;
3512 :
3513 : /* Now add the forced full scan, and decrement appropriate count */
3514 711610 : run_cost += inner_run_cost;
3515 711610 : if (outer_unmatched_rows >= 1)
3516 675190 : outer_unmatched_rows -= 1;
3517 : else
3518 36420 : outer_matched_rows -= 1;
3519 :
3520 : /* Add inner run cost for additional outer tuples having matches */
3521 711610 : if (outer_matched_rows > 0)
3522 254846 : run_cost += outer_matched_rows * inner_rescan_run_cost * inner_scan_frac;
3523 :
3524 : /* Add inner run cost for additional unmatched outer tuples */
3525 711610 : if (outer_unmatched_rows > 0)
3526 460062 : run_cost += outer_unmatched_rows * inner_rescan_run_cost;
3527 : }
3528 : }
3529 : else
3530 : {
3531 : /* Normal-case source costs were included in preliminary estimate */
3532 :
3533 : /* Compute number of tuples processed (not number emitted!) */
3534 470988 : ntuples = outer_path_rows * inner_path_rows;
3535 : }
3536 :
3537 : /* CPU costs */
3538 1324882 : cost_qual_eval(&restrict_qual_cost, path->jpath.joinrestrictinfo, root);
3539 1324882 : startup_cost += restrict_qual_cost.startup;
3540 1324882 : cpu_per_tuple = cpu_tuple_cost + restrict_qual_cost.per_tuple;
3541 1324882 : run_cost += cpu_per_tuple * ntuples;
3542 :
3543 : /* tlist eval costs are paid per output row, not per tuple scanned */
3544 1324882 : startup_cost += path->jpath.path.pathtarget->cost.startup;
3545 1324882 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
3546 :
3547 1324882 : path->jpath.path.startup_cost = startup_cost;
3548 1324882 : path->jpath.path.total_cost = startup_cost + run_cost;
3549 1324882 : }
3550 :
3551 : /*
3552 : * initial_cost_mergejoin
3553 : * Preliminary estimate of the cost of a mergejoin path.
3554 : *
3555 : * This must quickly produce lower-bound estimates of the path's startup and
3556 : * total costs. If we are unable to eliminate the proposed path from
3557 : * consideration using the lower bounds, final_cost_mergejoin will be called
3558 : * to obtain the final estimates.
3559 : *
3560 : * The exact division of labor between this function and final_cost_mergejoin
3561 : * is private to them, and represents a tradeoff between speed of the initial
3562 : * estimate and getting a tight lower bound. We choose to not examine the
3563 : * join quals here, except for obtaining the scan selectivity estimate which
3564 : * is really essential (but fortunately, use of caching keeps the cost of
3565 : * getting that down to something reasonable).
3566 : * We also assume that cost_sort/cost_incremental_sort is cheap enough to use
3567 : * here.
3568 : *
3569 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3570 : * other data to be used by final_cost_mergejoin
3571 : * 'jointype' is the type of join to be performed
3572 : * 'mergeclauses' is the list of joinclauses to be used as merge clauses
3573 : * 'outer_path' is the outer input to the join
3574 : * 'inner_path' is the inner input to the join
3575 : * 'outersortkeys' is the list of sort keys for the outer path
3576 : * 'innersortkeys' is the list of sort keys for the inner path
3577 : * 'outer_presorted_keys' is the number of presorted keys of the outer path
3578 : * 'extra' contains miscellaneous information about the join
3579 : *
3580 : * Note: outersortkeys and innersortkeys should be NIL if no explicit
3581 : * sort is needed because the respective source path is already ordered.
3582 : */
3583 : void
3584 1184338 : initial_cost_mergejoin(PlannerInfo *root, JoinCostWorkspace *workspace,
3585 : JoinType jointype,
3586 : List *mergeclauses,
3587 : Path *outer_path, Path *inner_path,
3588 : List *outersortkeys, List *innersortkeys,
3589 : int outer_presorted_keys,
3590 : JoinPathExtraData *extra)
3591 : {
3592 : int disabled_nodes;
3593 1184338 : Cost startup_cost = 0;
3594 1184338 : Cost run_cost = 0;
3595 1184338 : double outer_path_rows = outer_path->rows;
3596 1184338 : double inner_path_rows = inner_path->rows;
3597 : Cost inner_run_cost;
3598 : double outer_rows,
3599 : inner_rows,
3600 : outer_skip_rows,
3601 : inner_skip_rows;
3602 : Selectivity outerstartsel,
3603 : outerendsel,
3604 : innerstartsel,
3605 : innerendsel;
3606 : Path sort_path; /* dummy for result of
3607 : * cost_sort/cost_incremental_sort */
3608 :
3609 : /* Protect some assumptions below that rowcounts aren't zero */
3610 1184338 : if (outer_path_rows <= 0)
3611 96 : outer_path_rows = 1;
3612 1184338 : if (inner_path_rows <= 0)
3613 126 : inner_path_rows = 1;
3614 :
3615 : /*
3616 : * A merge join will stop as soon as it exhausts either input stream
3617 : * (unless it's an outer join, in which case the outer side has to be
3618 : * scanned all the way anyway). Estimate fraction of the left and right
3619 : * inputs that will actually need to be scanned. Likewise, we can
3620 : * estimate the number of rows that will be skipped before the first join
3621 : * pair is found, which should be factored into startup cost. We use only
3622 : * the first (most significant) merge clause for this purpose. Since
3623 : * mergejoinscansel() is a fairly expensive computation, we cache the
3624 : * results in the merge clause RestrictInfo.
3625 : */
3626 1184338 : if (mergeclauses && jointype != JOIN_FULL)
3627 1178198 : {
3628 1178198 : RestrictInfo *firstclause = (RestrictInfo *) linitial(mergeclauses);
3629 : List *opathkeys;
3630 : List *ipathkeys;
3631 : PathKey *opathkey;
3632 : PathKey *ipathkey;
3633 : MergeScanSelCache *cache;
3634 :
3635 : /* Get the input pathkeys to determine the sort-order details */
3636 1178198 : opathkeys = outersortkeys ? outersortkeys : outer_path->pathkeys;
3637 1178198 : ipathkeys = innersortkeys ? innersortkeys : inner_path->pathkeys;
3638 : Assert(opathkeys);
3639 : Assert(ipathkeys);
3640 1178198 : opathkey = (PathKey *) linitial(opathkeys);
3641 1178198 : ipathkey = (PathKey *) linitial(ipathkeys);
3642 : /* debugging check */
3643 1178198 : if (opathkey->pk_opfamily != ipathkey->pk_opfamily ||
3644 1178198 : opathkey->pk_eclass->ec_collation != ipathkey->pk_eclass->ec_collation ||
3645 1178198 : opathkey->pk_cmptype != ipathkey->pk_cmptype ||
3646 1178198 : opathkey->pk_nulls_first != ipathkey->pk_nulls_first)
3647 0 : elog(ERROR, "left and right pathkeys do not match in mergejoin");
3648 :
3649 : /* Get the selectivity with caching */
3650 1178198 : cache = cached_scansel(root, firstclause, opathkey);
3651 :
3652 1178198 : if (bms_is_subset(firstclause->left_relids,
3653 1178198 : outer_path->parent->relids))
3654 : {
3655 : /* left side of clause is outer */
3656 628428 : outerstartsel = cache->leftstartsel;
3657 628428 : outerendsel = cache->leftendsel;
3658 628428 : innerstartsel = cache->rightstartsel;
3659 628428 : innerendsel = cache->rightendsel;
3660 : }
3661 : else
3662 : {
3663 : /* left side of clause is inner */
3664 549770 : outerstartsel = cache->rightstartsel;
3665 549770 : outerendsel = cache->rightendsel;
3666 549770 : innerstartsel = cache->leftstartsel;
3667 549770 : innerendsel = cache->leftendsel;
3668 : }
3669 1178198 : if (jointype == JOIN_LEFT ||
3670 : jointype == JOIN_ANTI)
3671 : {
3672 194670 : outerstartsel = 0.0;
3673 194670 : outerendsel = 1.0;
3674 : }
3675 983528 : else if (jointype == JOIN_RIGHT ||
3676 : jointype == JOIN_RIGHT_ANTI)
3677 : {
3678 195136 : innerstartsel = 0.0;
3679 195136 : innerendsel = 1.0;
3680 : }
3681 : }
3682 : else
3683 : {
3684 : /* cope with clauseless or full mergejoin */
3685 6140 : outerstartsel = innerstartsel = 0.0;
3686 6140 : outerendsel = innerendsel = 1.0;
3687 : }
3688 :
3689 : /*
3690 : * Convert selectivities to row counts. We force outer_rows and
3691 : * inner_rows to be at least 1, but the skip_rows estimates can be zero.
3692 : */
3693 1184338 : outer_skip_rows = rint(outer_path_rows * outerstartsel);
3694 1184338 : inner_skip_rows = rint(inner_path_rows * innerstartsel);
3695 1184338 : outer_rows = clamp_row_est(outer_path_rows * outerendsel);
3696 1184338 : inner_rows = clamp_row_est(inner_path_rows * innerendsel);
3697 :
3698 : Assert(outer_skip_rows <= outer_rows);
3699 : Assert(inner_skip_rows <= inner_rows);
3700 :
3701 : /*
3702 : * Readjust scan selectivities to account for above rounding. This is
3703 : * normally an insignificant effect, but when there are only a few rows in
3704 : * the inputs, failing to do this makes for a large percentage error.
3705 : */
3706 1184338 : outerstartsel = outer_skip_rows / outer_path_rows;
3707 1184338 : innerstartsel = inner_skip_rows / inner_path_rows;
3708 1184338 : outerendsel = outer_rows / outer_path_rows;
3709 1184338 : innerendsel = inner_rows / inner_path_rows;
3710 :
3711 : Assert(outerstartsel <= outerendsel);
3712 : Assert(innerstartsel <= innerendsel);
3713 :
3714 1184338 : disabled_nodes = enable_mergejoin ? 0 : 1;
3715 :
3716 : /* cost of source data */
3717 :
3718 1184338 : if (outersortkeys) /* do we need to sort outer? */
3719 : {
3720 : /*
3721 : * We can assert that the outer path is not already ordered
3722 : * appropriately for the mergejoin; otherwise, outersortkeys would
3723 : * have been set to NIL.
3724 : */
3725 : Assert(!pathkeys_contained_in(outersortkeys, outer_path->pathkeys));
3726 :
3727 : /*
3728 : * We choose to use incremental sort if it is enabled and there are
3729 : * presorted keys; otherwise we use full sort.
3730 : */
3731 567280 : if (enable_incremental_sort && outer_presorted_keys > 0)
3732 : {
3733 1660 : cost_incremental_sort(&sort_path,
3734 : root,
3735 : outersortkeys,
3736 : outer_presorted_keys,
3737 : outer_path->disabled_nodes,
3738 : outer_path->startup_cost,
3739 : outer_path->total_cost,
3740 : outer_path_rows,
3741 1660 : outer_path->pathtarget->width,
3742 : 0.0,
3743 : work_mem,
3744 : -1.0);
3745 : }
3746 : else
3747 : {
3748 565620 : cost_sort(&sort_path,
3749 : root,
3750 : outersortkeys,
3751 : outer_path->disabled_nodes,
3752 : outer_path->total_cost,
3753 : outer_path_rows,
3754 565620 : outer_path->pathtarget->width,
3755 : 0.0,
3756 : work_mem,
3757 : -1.0);
3758 : }
3759 :
3760 567280 : disabled_nodes += sort_path.disabled_nodes;
3761 567280 : startup_cost += sort_path.startup_cost;
3762 567280 : startup_cost += (sort_path.total_cost - sort_path.startup_cost)
3763 567280 : * outerstartsel;
3764 567280 : run_cost += (sort_path.total_cost - sort_path.startup_cost)
3765 567280 : * (outerendsel - outerstartsel);
3766 : }
3767 : else
3768 : {
3769 617058 : disabled_nodes += outer_path->disabled_nodes;
3770 617058 : startup_cost += outer_path->startup_cost;
3771 617058 : startup_cost += (outer_path->total_cost - outer_path->startup_cost)
3772 617058 : * outerstartsel;
3773 617058 : run_cost += (outer_path->total_cost - outer_path->startup_cost)
3774 617058 : * (outerendsel - outerstartsel);
3775 : }
3776 :
3777 1184338 : if (innersortkeys) /* do we need to sort inner? */
3778 : {
3779 : /*
3780 : * We can assert that the inner path is not already ordered
3781 : * appropriately for the mergejoin; otherwise, innersortkeys would
3782 : * have been set to NIL.
3783 : */
3784 : Assert(!pathkeys_contained_in(innersortkeys, inner_path->pathkeys));
3785 :
3786 : /*
3787 : * We do not consider incremental sort for inner path, because
3788 : * incremental sort does not support mark/restore.
3789 : */
3790 :
3791 930900 : cost_sort(&sort_path,
3792 : root,
3793 : innersortkeys,
3794 : inner_path->disabled_nodes,
3795 : inner_path->total_cost,
3796 : inner_path_rows,
3797 930900 : inner_path->pathtarget->width,
3798 : 0.0,
3799 : work_mem,
3800 : -1.0);
3801 930900 : disabled_nodes += sort_path.disabled_nodes;
3802 930900 : startup_cost += sort_path.startup_cost;
3803 930900 : startup_cost += (sort_path.total_cost - sort_path.startup_cost)
3804 930900 : * innerstartsel;
3805 930900 : inner_run_cost = (sort_path.total_cost - sort_path.startup_cost)
3806 930900 : * (innerendsel - innerstartsel);
3807 : }
3808 : else
3809 : {
3810 253438 : disabled_nodes += inner_path->disabled_nodes;
3811 253438 : startup_cost += inner_path->startup_cost;
3812 253438 : startup_cost += (inner_path->total_cost - inner_path->startup_cost)
3813 253438 : * innerstartsel;
3814 253438 : inner_run_cost = (inner_path->total_cost - inner_path->startup_cost)
3815 253438 : * (innerendsel - innerstartsel);
3816 : }
3817 :
3818 : /*
3819 : * We can't yet determine whether rescanning occurs, or whether
3820 : * materialization of the inner input should be done. The minimum
3821 : * possible inner input cost, regardless of rescan and materialization
3822 : * considerations, is inner_run_cost. We include that in
3823 : * workspace->total_cost, but not yet in run_cost.
3824 : */
3825 :
3826 : /* CPU costs left for later */
3827 :
3828 : /* Public result fields */
3829 1184338 : workspace->disabled_nodes = disabled_nodes;
3830 1184338 : workspace->startup_cost = startup_cost;
3831 1184338 : workspace->total_cost = startup_cost + run_cost + inner_run_cost;
3832 : /* Save private data for final_cost_mergejoin */
3833 1184338 : workspace->run_cost = run_cost;
3834 1184338 : workspace->inner_run_cost = inner_run_cost;
3835 1184338 : workspace->outer_rows = outer_rows;
3836 1184338 : workspace->inner_rows = inner_rows;
3837 1184338 : workspace->outer_skip_rows = outer_skip_rows;
3838 1184338 : workspace->inner_skip_rows = inner_skip_rows;
3839 1184338 : }
3840 :
3841 : /*
3842 : * final_cost_mergejoin
3843 : * Final estimate of the cost and result size of a mergejoin path.
3844 : *
3845 : * Unlike other costsize functions, this routine makes two actual decisions:
3846 : * whether the executor will need to do mark/restore, and whether we should
3847 : * materialize the inner path. It would be logically cleaner to build
3848 : * separate paths testing these alternatives, but that would require repeating
3849 : * most of the cost calculations, which are not all that cheap. Since the
3850 : * choice will not affect output pathkeys or startup cost, only total cost,
3851 : * there is no possibility of wanting to keep more than one path. So it seems
3852 : * best to make the decisions here and record them in the path's
3853 : * skip_mark_restore and materialize_inner fields.
3854 : *
3855 : * Mark/restore overhead is usually required, but can be skipped if we know
3856 : * that the executor need find only one match per outer tuple, and that the
3857 : * mergeclauses are sufficient to identify a match.
3858 : *
3859 : * We materialize the inner path if we need mark/restore and either the inner
3860 : * path can't support mark/restore, or it's cheaper to use an interposed
3861 : * Material node to handle mark/restore.
3862 : *
3863 : * 'path' is already filled in except for the rows and cost fields and
3864 : * skip_mark_restore and materialize_inner
3865 : * 'workspace' is the result from initial_cost_mergejoin
3866 : * 'extra' contains miscellaneous information about the join
3867 : */
3868 : void
3869 311058 : final_cost_mergejoin(PlannerInfo *root, MergePath *path,
3870 : JoinCostWorkspace *workspace,
3871 : JoinPathExtraData *extra)
3872 : {
3873 311058 : Path *outer_path = path->jpath.outerjoinpath;
3874 311058 : Path *inner_path = path->jpath.innerjoinpath;
3875 311058 : double inner_path_rows = inner_path->rows;
3876 311058 : List *mergeclauses = path->path_mergeclauses;
3877 311058 : List *innersortkeys = path->innersortkeys;
3878 311058 : Cost startup_cost = workspace->startup_cost;
3879 311058 : Cost run_cost = workspace->run_cost;
3880 311058 : Cost inner_run_cost = workspace->inner_run_cost;
3881 311058 : double outer_rows = workspace->outer_rows;
3882 311058 : double inner_rows = workspace->inner_rows;
3883 311058 : double outer_skip_rows = workspace->outer_skip_rows;
3884 311058 : double inner_skip_rows = workspace->inner_skip_rows;
3885 : Cost cpu_per_tuple,
3886 : bare_inner_cost,
3887 : mat_inner_cost;
3888 : QualCost merge_qual_cost;
3889 : QualCost qp_qual_cost;
3890 : double mergejointuples,
3891 : rescannedtuples;
3892 : double rescanratio;
3893 :
3894 : /* Set the number of disabled nodes. */
3895 311058 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
3896 :
3897 : /* Protect some assumptions below that rowcounts aren't zero */
3898 311058 : if (inner_path_rows <= 0)
3899 90 : inner_path_rows = 1;
3900 :
3901 : /* Mark the path with the correct row estimate */
3902 311058 : if (path->jpath.path.param_info)
3903 898 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3904 : else
3905 310160 : path->jpath.path.rows = path->jpath.path.parent->rows;
3906 :
3907 : /* For partial paths, scale row estimate. */
3908 311058 : if (path->jpath.path.parallel_workers > 0)
3909 : {
3910 10098 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3911 :
3912 10098 : path->jpath.path.rows =
3913 10098 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
3914 : }
3915 :
3916 : /*
3917 : * Compute cost of the mergequals and qpquals (other restriction clauses)
3918 : * separately.
3919 : */
3920 311058 : cost_qual_eval(&merge_qual_cost, mergeclauses, root);
3921 311058 : cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
3922 311058 : qp_qual_cost.startup -= merge_qual_cost.startup;
3923 311058 : qp_qual_cost.per_tuple -= merge_qual_cost.per_tuple;
3924 :
3925 : /*
3926 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
3927 : * executor will stop scanning for matches after the first match. When
3928 : * all the joinclauses are merge clauses, this means we don't ever need to
3929 : * back up the merge, and so we can skip mark/restore overhead.
3930 : */
3931 311058 : if ((path->jpath.jointype == JOIN_SEMI ||
3932 303980 : path->jpath.jointype == JOIN_ANTI ||
3933 438002 : extra->inner_unique) &&
3934 140970 : (list_length(path->jpath.joinrestrictinfo) ==
3935 140970 : list_length(path->path_mergeclauses)))
3936 119296 : path->skip_mark_restore = true;
3937 : else
3938 191762 : path->skip_mark_restore = false;
3939 :
3940 : /*
3941 : * Get approx # tuples passing the mergequals. We use approx_tuple_count
3942 : * here because we need an estimate done with JOIN_INNER semantics.
3943 : */
3944 311058 : mergejointuples = approx_tuple_count(root, &path->jpath, mergeclauses);
3945 :
3946 : /*
3947 : * When there are equal merge keys in the outer relation, the mergejoin
3948 : * must rescan any matching tuples in the inner relation. This means
3949 : * re-fetching inner tuples; we have to estimate how often that happens.
3950 : *
3951 : * For regular inner and outer joins, the number of re-fetches can be
3952 : * estimated approximately as size of merge join output minus size of
3953 : * inner relation. Assume that the distinct key values are 1, 2, ..., and
3954 : * denote the number of values of each key in the outer relation as m1,
3955 : * m2, ...; in the inner relation, n1, n2, ... Then we have
3956 : *
3957 : * size of join = m1 * n1 + m2 * n2 + ...
3958 : *
3959 : * number of rescanned tuples = (m1 - 1) * n1 + (m2 - 1) * n2 + ... = m1 *
3960 : * n1 + m2 * n2 + ... - (n1 + n2 + ...) = size of join - size of inner
3961 : * relation
3962 : *
3963 : * This equation works correctly for outer tuples having no inner match
3964 : * (nk = 0), but not for inner tuples having no outer match (mk = 0); we
3965 : * are effectively subtracting those from the number of rescanned tuples,
3966 : * when we should not. Can we do better without expensive selectivity
3967 : * computations?
3968 : *
3969 : * The whole issue is moot if we know we don't need to mark/restore at
3970 : * all, or if we are working from a unique-ified outer input.
3971 : */
3972 311058 : if (path->skip_mark_restore ||
3973 191762 : RELATION_WAS_MADE_UNIQUE(outer_path->parent, extra->sjinfo,
3974 : path->jpath.jointype))
3975 123484 : rescannedtuples = 0;
3976 : else
3977 : {
3978 187574 : rescannedtuples = mergejointuples - inner_path_rows;
3979 : /* Must clamp because of possible underestimate */
3980 187574 : if (rescannedtuples < 0)
3981 75618 : rescannedtuples = 0;
3982 : }
3983 :
3984 : /*
3985 : * We'll inflate various costs this much to account for rescanning. Note
3986 : * that this is to be multiplied by something involving inner_rows, or
3987 : * another number related to the portion of the inner rel we'll scan.
3988 : */
3989 311058 : rescanratio = 1.0 + (rescannedtuples / inner_rows);
3990 :
3991 : /*
3992 : * Decide whether we want to materialize the inner input to shield it from
3993 : * mark/restore and performing re-fetches. Our cost model for regular
3994 : * re-fetches is that a re-fetch costs the same as an original fetch,
3995 : * which is probably an overestimate; but on the other hand we ignore the
3996 : * bookkeeping costs of mark/restore. Not clear if it's worth developing
3997 : * a more refined model. So we just need to inflate the inner run cost by
3998 : * rescanratio.
3999 : */
4000 311058 : bare_inner_cost = inner_run_cost * rescanratio;
4001 :
4002 : /*
4003 : * When we interpose a Material node the re-fetch cost is assumed to be
4004 : * just cpu_operator_cost per tuple, independently of the underlying
4005 : * plan's cost; and we charge an extra cpu_operator_cost per original
4006 : * fetch as well. Note that we're assuming the materialize node will
4007 : * never spill to disk, since it only has to remember tuples back to the
4008 : * last mark. (If there are a huge number of duplicates, our other cost
4009 : * factors will make the path so expensive that it probably won't get
4010 : * chosen anyway.) So we don't use cost_rescan here.
4011 : *
4012 : * Note: keep this estimate in sync with create_mergejoin_plan's labeling
4013 : * of the generated Material node.
4014 : */
4015 311058 : mat_inner_cost = inner_run_cost +
4016 311058 : cpu_operator_cost * inner_rows * rescanratio;
4017 :
4018 : /*
4019 : * If we don't need mark/restore at all, we don't need materialization.
4020 : */
4021 311058 : if (path->skip_mark_restore)
4022 119296 : path->materialize_inner = false;
4023 :
4024 : /*
4025 : * Prefer materializing if it looks cheaper, unless the user has asked to
4026 : * suppress materialization.
4027 : */
4028 191762 : else if (enable_material && mat_inner_cost < bare_inner_cost)
4029 2870 : path->materialize_inner = true;
4030 :
4031 : /*
4032 : * Even if materializing doesn't look cheaper, we *must* do it if the
4033 : * inner path is to be used directly (without sorting) and it doesn't
4034 : * support mark/restore.
4035 : *
4036 : * Since the inner side must be ordered, and only Sorts and IndexScans can
4037 : * create order to begin with, and they both support mark/restore, you
4038 : * might think there's no problem --- but you'd be wrong. Nestloop and
4039 : * merge joins can *preserve* the order of their inputs, so they can be
4040 : * selected as the input of a mergejoin, and they don't support
4041 : * mark/restore at present.
4042 : *
4043 : * We don't test the value of enable_material here, because
4044 : * materialization is required for correctness in this case, and turning
4045 : * it off does not entitle us to deliver an invalid plan.
4046 : */
4047 188892 : else if (innersortkeys == NIL &&
4048 7526 : !ExecSupportsMarkRestore(inner_path))
4049 1846 : path->materialize_inner = true;
4050 :
4051 : /*
4052 : * Also, force materializing if the inner path is to be sorted and the
4053 : * sort is expected to spill to disk. This is because the final merge
4054 : * pass can be done on-the-fly if it doesn't have to support mark/restore.
4055 : * We don't try to adjust the cost estimates for this consideration,
4056 : * though.
4057 : *
4058 : * Since materialization is a performance optimization in this case,
4059 : * rather than necessary for correctness, we skip it if enable_material is
4060 : * off.
4061 : */
4062 187046 : else if (enable_material && innersortkeys != NIL &&
4063 181318 : relation_byte_size(inner_path_rows,
4064 181318 : inner_path->pathtarget->width) >
4065 181318 : work_mem * (Size) 1024)
4066 268 : path->materialize_inner = true;
4067 : else
4068 186778 : path->materialize_inner = false;
4069 :
4070 : /* Charge the right incremental cost for the chosen case */
4071 311058 : if (path->materialize_inner)
4072 4984 : run_cost += mat_inner_cost;
4073 : else
4074 306074 : run_cost += bare_inner_cost;
4075 :
4076 : /* CPU costs */
4077 :
4078 : /*
4079 : * The number of tuple comparisons needed is approximately number of outer
4080 : * rows plus number of inner rows plus number of rescanned tuples (can we
4081 : * refine this?). At each one, we need to evaluate the mergejoin quals.
4082 : */
4083 311058 : startup_cost += merge_qual_cost.startup;
4084 311058 : startup_cost += merge_qual_cost.per_tuple *
4085 311058 : (outer_skip_rows + inner_skip_rows * rescanratio);
4086 311058 : run_cost += merge_qual_cost.per_tuple *
4087 311058 : ((outer_rows - outer_skip_rows) +
4088 311058 : (inner_rows - inner_skip_rows) * rescanratio);
4089 :
4090 : /*
4091 : * For each tuple that gets through the mergejoin proper, we charge
4092 : * cpu_tuple_cost plus the cost of evaluating additional restriction
4093 : * clauses that are to be applied at the join. (This is pessimistic since
4094 : * not all of the quals may get evaluated at each tuple.)
4095 : *
4096 : * Note: we could adjust for SEMI/ANTI joins skipping some qual
4097 : * evaluations here, but it's probably not worth the trouble.
4098 : */
4099 311058 : startup_cost += qp_qual_cost.startup;
4100 311058 : cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
4101 311058 : run_cost += cpu_per_tuple * mergejointuples;
4102 :
4103 : /* tlist eval costs are paid per output row, not per tuple scanned */
4104 311058 : startup_cost += path->jpath.path.pathtarget->cost.startup;
4105 311058 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
4106 :
4107 311058 : path->jpath.path.startup_cost = startup_cost;
4108 311058 : path->jpath.path.total_cost = startup_cost + run_cost;
4109 311058 : }
4110 :
4111 : /*
4112 : * run mergejoinscansel() with caching
4113 : */
4114 : static MergeScanSelCache *
4115 1178198 : cached_scansel(PlannerInfo *root, RestrictInfo *rinfo, PathKey *pathkey)
4116 : {
4117 : MergeScanSelCache *cache;
4118 : ListCell *lc;
4119 : Selectivity leftstartsel,
4120 : leftendsel,
4121 : rightstartsel,
4122 : rightendsel;
4123 : MemoryContext oldcontext;
4124 :
4125 : /* Do we have this result already? */
4126 1178276 : foreach(lc, rinfo->scansel_cache)
4127 : {
4128 1064034 : cache = (MergeScanSelCache *) lfirst(lc);
4129 1064034 : if (cache->opfamily == pathkey->pk_opfamily &&
4130 1064034 : cache->collation == pathkey->pk_eclass->ec_collation &&
4131 1064034 : cache->cmptype == pathkey->pk_cmptype &&
4132 1063956 : cache->nulls_first == pathkey->pk_nulls_first)
4133 1063956 : return cache;
4134 : }
4135 :
4136 : /* Nope, do the computation */
4137 114242 : mergejoinscansel(root,
4138 114242 : (Node *) rinfo->clause,
4139 : pathkey->pk_opfamily,
4140 : pathkey->pk_cmptype,
4141 114242 : pathkey->pk_nulls_first,
4142 : &leftstartsel,
4143 : &leftendsel,
4144 : &rightstartsel,
4145 : &rightendsel);
4146 :
4147 : /* Cache the result in suitably long-lived workspace */
4148 114242 : oldcontext = MemoryContextSwitchTo(root->planner_cxt);
4149 :
4150 114242 : cache = (MergeScanSelCache *) palloc(sizeof(MergeScanSelCache));
4151 114242 : cache->opfamily = pathkey->pk_opfamily;
4152 114242 : cache->collation = pathkey->pk_eclass->ec_collation;
4153 114242 : cache->cmptype = pathkey->pk_cmptype;
4154 114242 : cache->nulls_first = pathkey->pk_nulls_first;
4155 114242 : cache->leftstartsel = leftstartsel;
4156 114242 : cache->leftendsel = leftendsel;
4157 114242 : cache->rightstartsel = rightstartsel;
4158 114242 : cache->rightendsel = rightendsel;
4159 :
4160 114242 : rinfo->scansel_cache = lappend(rinfo->scansel_cache, cache);
4161 :
4162 114242 : MemoryContextSwitchTo(oldcontext);
4163 :
4164 114242 : return cache;
4165 : }
4166 :
4167 : /*
4168 : * initial_cost_hashjoin
4169 : * Preliminary estimate of the cost of a hashjoin path.
4170 : *
4171 : * This must quickly produce lower-bound estimates of the path's startup and
4172 : * total costs. If we are unable to eliminate the proposed path from
4173 : * consideration using the lower bounds, final_cost_hashjoin will be called
4174 : * to obtain the final estimates.
4175 : *
4176 : * The exact division of labor between this function and final_cost_hashjoin
4177 : * is private to them, and represents a tradeoff between speed of the initial
4178 : * estimate and getting a tight lower bound. We choose to not examine the
4179 : * join quals here (other than by counting the number of hash clauses),
4180 : * so we can't do much with CPU costs. We do assume that
4181 : * ExecChooseHashTableSize is cheap enough to use here.
4182 : *
4183 : * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
4184 : * other data to be used by final_cost_hashjoin
4185 : * 'jointype' is the type of join to be performed
4186 : * 'hashclauses' is the list of joinclauses to be used as hash clauses
4187 : * 'outer_path' is the outer input to the join
4188 : * 'inner_path' is the inner input to the join
4189 : * 'extra' contains miscellaneous information about the join
4190 : * 'parallel_hash' indicates that inner_path is partial and that a shared
4191 : * hash table will be built in parallel
4192 : */
4193 : void
4194 640258 : initial_cost_hashjoin(PlannerInfo *root, JoinCostWorkspace *workspace,
4195 : JoinType jointype,
4196 : List *hashclauses,
4197 : Path *outer_path, Path *inner_path,
4198 : JoinPathExtraData *extra,
4199 : bool parallel_hash)
4200 : {
4201 : int disabled_nodes;
4202 640258 : Cost startup_cost = 0;
4203 640258 : Cost run_cost = 0;
4204 640258 : double outer_path_rows = outer_path->rows;
4205 640258 : double inner_path_rows = inner_path->rows;
4206 640258 : double inner_path_rows_total = inner_path_rows;
4207 640258 : int num_hashclauses = list_length(hashclauses);
4208 : int numbuckets;
4209 : int numbatches;
4210 : int num_skew_mcvs;
4211 : size_t space_allowed; /* unused */
4212 :
4213 : /* Count up disabled nodes. */
4214 640258 : disabled_nodes = enable_hashjoin ? 0 : 1;
4215 640258 : disabled_nodes += inner_path->disabled_nodes;
4216 640258 : disabled_nodes += outer_path->disabled_nodes;
4217 :
4218 : /* cost of source data */
4219 640258 : startup_cost += outer_path->startup_cost;
4220 640258 : run_cost += outer_path->total_cost - outer_path->startup_cost;
4221 640258 : startup_cost += inner_path->total_cost;
4222 :
4223 : /*
4224 : * Cost of computing hash function: must do it once per input tuple. We
4225 : * charge one cpu_operator_cost for each column's hash function. Also,
4226 : * tack on one cpu_tuple_cost per inner row, to model the costs of
4227 : * inserting the row into the hashtable.
4228 : *
4229 : * XXX when a hashclause is more complex than a single operator, we really
4230 : * should charge the extra eval costs of the left or right side, as
4231 : * appropriate, here. This seems more work than it's worth at the moment.
4232 : */
4233 640258 : startup_cost += (cpu_operator_cost * num_hashclauses + cpu_tuple_cost)
4234 640258 : * inner_path_rows;
4235 640258 : run_cost += cpu_operator_cost * num_hashclauses * outer_path_rows;
4236 :
4237 : /*
4238 : * If this is a parallel hash build, then the value we have for
4239 : * inner_rows_total currently refers only to the rows returned by each
4240 : * participant. For shared hash table size estimation, we need the total
4241 : * number, so we need to undo the division.
4242 : */
4243 640258 : if (parallel_hash)
4244 12744 : inner_path_rows_total *= get_parallel_divisor(inner_path);
4245 :
4246 : /*
4247 : * Get hash table size that executor would use for inner relation.
4248 : *
4249 : * XXX for the moment, always assume that skew optimization will be
4250 : * performed. As long as SKEW_HASH_MEM_PERCENT is small, it's not worth
4251 : * trying to determine that for sure.
4252 : *
4253 : * XXX at some point it might be interesting to try to account for skew
4254 : * optimization in the cost estimate, but for now, we don't.
4255 : */
4256 640258 : ExecChooseHashTableSize(inner_path_rows_total,
4257 640258 : inner_path->pathtarget->width,
4258 : true, /* useskew */
4259 : parallel_hash, /* try_combined_hash_mem */
4260 : outer_path->parallel_workers,
4261 : &space_allowed,
4262 : &numbuckets,
4263 : &numbatches,
4264 : &num_skew_mcvs);
4265 :
4266 : /*
4267 : * If inner relation is too big then we will need to "batch" the join,
4268 : * which implies writing and reading most of the tuples to disk an extra
4269 : * time. Charge seq_page_cost per page, since the I/O should be nice and
4270 : * sequential. Writing the inner rel counts as startup cost, all the rest
4271 : * as run cost.
4272 : */
4273 640258 : if (numbatches > 1)
4274 : {
4275 4674 : double outerpages = page_size(outer_path_rows,
4276 4674 : outer_path->pathtarget->width);
4277 4674 : double innerpages = page_size(inner_path_rows,
4278 4674 : inner_path->pathtarget->width);
4279 :
4280 4674 : startup_cost += seq_page_cost * innerpages;
4281 4674 : run_cost += seq_page_cost * (innerpages + 2 * outerpages);
4282 : }
4283 :
4284 : /* CPU costs left for later */
4285 :
4286 : /* Public result fields */
4287 640258 : workspace->disabled_nodes = disabled_nodes;
4288 640258 : workspace->startup_cost = startup_cost;
4289 640258 : workspace->total_cost = startup_cost + run_cost;
4290 : /* Save private data for final_cost_hashjoin */
4291 640258 : workspace->run_cost = run_cost;
4292 640258 : workspace->numbuckets = numbuckets;
4293 640258 : workspace->numbatches = numbatches;
4294 640258 : workspace->inner_rows_total = inner_path_rows_total;
4295 640258 : }
4296 :
4297 : /*
4298 : * final_cost_hashjoin
4299 : * Final estimate of the cost and result size of a hashjoin path.
4300 : *
4301 : * Note: the numbatches estimate is also saved into 'path' for use later
4302 : *
4303 : * 'path' is already filled in except for the rows and cost fields and
4304 : * num_batches
4305 : * 'workspace' is the result from initial_cost_hashjoin
4306 : * 'extra' contains miscellaneous information about the join
4307 : */
4308 : void
4309 279926 : final_cost_hashjoin(PlannerInfo *root, HashPath *path,
4310 : JoinCostWorkspace *workspace,
4311 : JoinPathExtraData *extra)
4312 : {
4313 279926 : Path *outer_path = path->jpath.outerjoinpath;
4314 279926 : Path *inner_path = path->jpath.innerjoinpath;
4315 279926 : double outer_path_rows = outer_path->rows;
4316 279926 : double inner_path_rows = inner_path->rows;
4317 279926 : double inner_path_rows_total = workspace->inner_rows_total;
4318 279926 : List *hashclauses = path->path_hashclauses;
4319 279926 : Cost startup_cost = workspace->startup_cost;
4320 279926 : Cost run_cost = workspace->run_cost;
4321 279926 : int numbuckets = workspace->numbuckets;
4322 279926 : int numbatches = workspace->numbatches;
4323 : Cost cpu_per_tuple;
4324 : QualCost hash_qual_cost;
4325 : QualCost qp_qual_cost;
4326 : double hashjointuples;
4327 : double virtualbuckets;
4328 : Selectivity innerbucketsize;
4329 : Selectivity innermcvfreq;
4330 : ListCell *hcl;
4331 :
4332 : /* Set the number of disabled nodes. */
4333 279926 : path->jpath.path.disabled_nodes = workspace->disabled_nodes;
4334 :
4335 : /* Mark the path with the correct row estimate */
4336 279926 : if (path->jpath.path.param_info)
4337 2102 : path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
4338 : else
4339 277824 : path->jpath.path.rows = path->jpath.path.parent->rows;
4340 :
4341 : /* For partial paths, scale row estimate. */
4342 279926 : if (path->jpath.path.parallel_workers > 0)
4343 : {
4344 11778 : double parallel_divisor = get_parallel_divisor(&path->jpath.path);
4345 :
4346 11778 : path->jpath.path.rows =
4347 11778 : clamp_row_est(path->jpath.path.rows / parallel_divisor);
4348 : }
4349 :
4350 : /* mark the path with estimated # of batches */
4351 279926 : path->num_batches = numbatches;
4352 :
4353 : /* store the total number of tuples (sum of partial row estimates) */
4354 279926 : path->inner_rows_total = inner_path_rows_total;
4355 :
4356 : /* and compute the number of "virtual" buckets in the whole join */
4357 279926 : virtualbuckets = (double) numbuckets * (double) numbatches;
4358 :
4359 : /*
4360 : * Determine bucketsize fraction and MCV frequency for the inner relation.
4361 : * We use the smallest bucketsize or MCV frequency estimated for any
4362 : * individual hashclause; this is undoubtedly conservative.
4363 : *
4364 : * BUT: if inner relation has been unique-ified, we can assume it's good
4365 : * for hashing. This is important both because it's the right answer, and
4366 : * because we avoid contaminating the cache with a value that's wrong for
4367 : * non-unique-ified paths.
4368 : */
4369 279926 : if (RELATION_WAS_MADE_UNIQUE(inner_path->parent, extra->sjinfo,
4370 : path->jpath.jointype))
4371 : {
4372 4216 : innerbucketsize = 1.0 / virtualbuckets;
4373 4216 : innermcvfreq = 0.0;
4374 : }
4375 : else
4376 : {
4377 : List *otherclauses;
4378 :
4379 275710 : innerbucketsize = 1.0;
4380 275710 : innermcvfreq = 1.0;
4381 :
4382 : /* At first, try to estimate bucket size using extended statistics. */
4383 275710 : otherclauses = estimate_multivariate_bucketsize(root,
4384 : inner_path->parent,
4385 : hashclauses,
4386 : &innerbucketsize);
4387 :
4388 : /* Pass through the remaining clauses */
4389 583638 : foreach(hcl, otherclauses)
4390 : {
4391 307928 : RestrictInfo *restrictinfo = lfirst_node(RestrictInfo, hcl);
4392 : Selectivity thisbucketsize;
4393 : Selectivity thismcvfreq;
4394 :
4395 : /*
4396 : * First we have to figure out which side of the hashjoin clause
4397 : * is the inner side.
4398 : *
4399 : * Since we tend to visit the same clauses over and over when
4400 : * planning a large query, we cache the bucket stats estimates in
4401 : * the RestrictInfo node to avoid repeated lookups of statistics.
4402 : */
4403 307928 : if (bms_is_subset(restrictinfo->right_relids,
4404 307928 : inner_path->parent->relids))
4405 : {
4406 : /* righthand side is inner */
4407 161380 : thisbucketsize = restrictinfo->right_bucketsize;
4408 161380 : if (thisbucketsize < 0)
4409 : {
4410 : /* not cached yet */
4411 87936 : estimate_hash_bucket_stats(root,
4412 87936 : get_rightop(restrictinfo->clause),
4413 : virtualbuckets,
4414 : &restrictinfo->right_mcvfreq,
4415 : &restrictinfo->right_bucketsize);
4416 87936 : thisbucketsize = restrictinfo->right_bucketsize;
4417 : }
4418 161380 : thismcvfreq = restrictinfo->right_mcvfreq;
4419 : }
4420 : else
4421 : {
4422 : Assert(bms_is_subset(restrictinfo->left_relids,
4423 : inner_path->parent->relids));
4424 : /* lefthand side is inner */
4425 146548 : thisbucketsize = restrictinfo->left_bucketsize;
4426 146548 : if (thisbucketsize < 0)
4427 : {
4428 : /* not cached yet */
4429 76156 : estimate_hash_bucket_stats(root,
4430 76156 : get_leftop(restrictinfo->clause),
4431 : virtualbuckets,
4432 : &restrictinfo->left_mcvfreq,
4433 : &restrictinfo->left_bucketsize);
4434 76156 : thisbucketsize = restrictinfo->left_bucketsize;
4435 : }
4436 146548 : thismcvfreq = restrictinfo->left_mcvfreq;
4437 : }
4438 :
4439 307928 : if (innerbucketsize > thisbucketsize)
4440 194522 : innerbucketsize = thisbucketsize;
4441 307928 : if (innermcvfreq > thismcvfreq)
4442 279476 : innermcvfreq = thismcvfreq;
4443 : }
4444 : }
4445 :
4446 : /*
4447 : * If the bucket holding the inner MCV would exceed hash_mem, we don't
4448 : * want to hash unless there is really no other alternative, so apply
4449 : * disable_cost. (The executor normally copes with excessive memory usage
4450 : * by splitting batches, but obviously it cannot separate equal values
4451 : * that way, so it will be unable to drive the batch size below hash_mem
4452 : * when this is true.)
4453 : */
4454 279926 : if (relation_byte_size(clamp_row_est(inner_path_rows * innermcvfreq),
4455 559852 : inner_path->pathtarget->width) > get_hash_memory_limit())
4456 6 : startup_cost += disable_cost;
4457 :
4458 : /*
4459 : * Compute cost of the hashquals and qpquals (other restriction clauses)
4460 : * separately.
4461 : */
4462 279926 : cost_qual_eval(&hash_qual_cost, hashclauses, root);
4463 279926 : cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
4464 279926 : qp_qual_cost.startup -= hash_qual_cost.startup;
4465 279926 : qp_qual_cost.per_tuple -= hash_qual_cost.per_tuple;
4466 :
4467 : /* CPU costs */
4468 :
4469 279926 : if (path->jpath.jointype == JOIN_SEMI ||
4470 273736 : path->jpath.jointype == JOIN_ANTI ||
4471 269234 : extra->inner_unique)
4472 117272 : {
4473 : double outer_matched_rows;
4474 : Selectivity inner_scan_frac;
4475 :
4476 : /*
4477 : * With a SEMI or ANTI join, or if the innerrel is known unique, the
4478 : * executor will stop after the first match.
4479 : *
4480 : * For an outer-rel row that has at least one match, we can expect the
4481 : * bucket scan to stop after a fraction 1/(match_count+1) of the
4482 : * bucket's rows, if the matches are evenly distributed. Since they
4483 : * probably aren't quite evenly distributed, we apply a fuzz factor of
4484 : * 2.0 to that fraction. (If we used a larger fuzz factor, we'd have
4485 : * to clamp inner_scan_frac to at most 1.0; but since match_count is
4486 : * at least 1, no such clamp is needed now.)
4487 : */
4488 117272 : outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
4489 117272 : inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
4490 :
4491 117272 : startup_cost += hash_qual_cost.startup;
4492 234544 : run_cost += hash_qual_cost.per_tuple * outer_matched_rows *
4493 117272 : clamp_row_est(inner_path_rows * innerbucketsize * inner_scan_frac) * 0.5;
4494 :
4495 : /*
4496 : * For unmatched outer-rel rows, the picture is quite a lot different.
4497 : * In the first place, there is no reason to assume that these rows
4498 : * preferentially hit heavily-populated buckets; instead assume they
4499 : * are uncorrelated with the inner distribution and so they see an
4500 : * average bucket size of inner_path_rows / virtualbuckets. In the
4501 : * second place, it seems likely that they will have few if any exact
4502 : * hash-code matches and so very few of the tuples in the bucket will
4503 : * actually require eval of the hash quals. We don't have any good
4504 : * way to estimate how many will, but for the moment assume that the
4505 : * effective cost per bucket entry is one-tenth what it is for
4506 : * matchable tuples.
4507 : */
4508 234544 : run_cost += hash_qual_cost.per_tuple *
4509 234544 : (outer_path_rows - outer_matched_rows) *
4510 117272 : clamp_row_est(inner_path_rows / virtualbuckets) * 0.05;
4511 :
4512 : /* Get # of tuples that will pass the basic join */
4513 117272 : if (path->jpath.jointype == JOIN_ANTI)
4514 4502 : hashjointuples = outer_path_rows - outer_matched_rows;
4515 : else
4516 112770 : hashjointuples = outer_matched_rows;
4517 : }
4518 : else
4519 : {
4520 : /*
4521 : * The number of tuple comparisons needed is the number of outer
4522 : * tuples times the typical number of tuples in a hash bucket, which
4523 : * is the inner relation size times its bucketsize fraction. At each
4524 : * one, we need to evaluate the hashjoin quals. But actually,
4525 : * charging the full qual eval cost at each tuple is pessimistic,
4526 : * since we don't evaluate the quals unless the hash values match
4527 : * exactly. For lack of a better idea, halve the cost estimate to
4528 : * allow for that.
4529 : */
4530 162654 : startup_cost += hash_qual_cost.startup;
4531 325308 : run_cost += hash_qual_cost.per_tuple * outer_path_rows *
4532 162654 : clamp_row_est(inner_path_rows * innerbucketsize) * 0.5;
4533 :
4534 : /*
4535 : * Get approx # tuples passing the hashquals. We use
4536 : * approx_tuple_count here because we need an estimate done with
4537 : * JOIN_INNER semantics.
4538 : */
4539 162654 : hashjointuples = approx_tuple_count(root, &path->jpath, hashclauses);
4540 : }
4541 :
4542 : /*
4543 : * For each tuple that gets through the hashjoin proper, we charge
4544 : * cpu_tuple_cost plus the cost of evaluating additional restriction
4545 : * clauses that are to be applied at the join. (This is pessimistic since
4546 : * not all of the quals may get evaluated at each tuple.)
4547 : */
4548 279926 : startup_cost += qp_qual_cost.startup;
4549 279926 : cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
4550 279926 : run_cost += cpu_per_tuple * hashjointuples;
4551 :
4552 : /* tlist eval costs are paid per output row, not per tuple scanned */
4553 279926 : startup_cost += path->jpath.path.pathtarget->cost.startup;
4554 279926 : run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
4555 :
4556 279926 : path->jpath.path.startup_cost = startup_cost;
4557 279926 : path->jpath.path.total_cost = startup_cost + run_cost;
4558 279926 : }
4559 :
4560 :
4561 : /*
4562 : * cost_subplan
4563 : * Figure the costs for a SubPlan (or initplan).
4564 : *
4565 : * Note: we could dig the subplan's Plan out of the root list, but in practice
4566 : * all callers have it handy already, so we make them pass it.
4567 : */
4568 : void
4569 43310 : cost_subplan(PlannerInfo *root, SubPlan *subplan, Plan *plan)
4570 : {
4571 : QualCost sp_cost;
4572 :
4573 : /* Figure any cost for evaluating the testexpr */
4574 43310 : cost_qual_eval(&sp_cost,
4575 43310 : make_ands_implicit((Expr *) subplan->testexpr),
4576 : root);
4577 :
4578 43310 : if (subplan->useHashTable)
4579 : {
4580 : /*
4581 : * If we are using a hash table for the subquery outputs, then the
4582 : * cost of evaluating the query is a one-time cost. We charge one
4583 : * cpu_operator_cost per tuple for the work of loading the hashtable,
4584 : * too.
4585 : */
4586 2090 : sp_cost.startup += plan->total_cost +
4587 2090 : cpu_operator_cost * plan->plan_rows;
4588 :
4589 : /*
4590 : * The per-tuple costs include the cost of evaluating the lefthand
4591 : * expressions, plus the cost of probing the hashtable. We already
4592 : * accounted for the lefthand expressions as part of the testexpr, and
4593 : * will also have counted one cpu_operator_cost for each comparison
4594 : * operator. That is probably too low for the probing cost, but it's
4595 : * hard to make a better estimate, so live with it for now.
4596 : */
4597 : }
4598 : else
4599 : {
4600 : /*
4601 : * Otherwise we will be rescanning the subplan output on each
4602 : * evaluation. We need to estimate how much of the output we will
4603 : * actually need to scan. NOTE: this logic should agree with the
4604 : * tuple_fraction estimates used by make_subplan() in
4605 : * plan/subselect.c.
4606 : */
4607 41220 : Cost plan_run_cost = plan->total_cost - plan->startup_cost;
4608 :
4609 41220 : if (subplan->subLinkType == EXISTS_SUBLINK)
4610 : {
4611 : /* we only need to fetch 1 tuple; clamp to avoid zero divide */
4612 2484 : sp_cost.per_tuple += plan_run_cost / clamp_row_est(plan->plan_rows);
4613 : }
4614 38736 : else if (subplan->subLinkType == ALL_SUBLINK ||
4615 38718 : subplan->subLinkType == ANY_SUBLINK)
4616 : {
4617 : /* assume we need 50% of the tuples */
4618 144 : sp_cost.per_tuple += 0.50 * plan_run_cost;
4619 : /* also charge a cpu_operator_cost per row examined */
4620 144 : sp_cost.per_tuple += 0.50 * plan->plan_rows * cpu_operator_cost;
4621 : }
4622 : else
4623 : {
4624 : /* assume we need all tuples */
4625 38592 : sp_cost.per_tuple += plan_run_cost;
4626 : }
4627 :
4628 : /*
4629 : * Also account for subplan's startup cost. If the subplan is
4630 : * uncorrelated or undirect correlated, AND its topmost node is one
4631 : * that materializes its output, assume that we'll only need to pay
4632 : * its startup cost once; otherwise assume we pay the startup cost
4633 : * every time.
4634 : */
4635 54226 : if (subplan->parParam == NIL &&
4636 13006 : ExecMaterializesOutput(nodeTag(plan)))
4637 720 : sp_cost.startup += plan->startup_cost;
4638 : else
4639 40500 : sp_cost.per_tuple += plan->startup_cost;
4640 : }
4641 :
4642 43310 : subplan->startup_cost = sp_cost.startup;
4643 43310 : subplan->per_call_cost = sp_cost.per_tuple;
4644 43310 : }
4645 :
4646 :
4647 : /*
4648 : * cost_rescan
4649 : * Given a finished Path, estimate the costs of rescanning it after
4650 : * having done so the first time. For some Path types a rescan is
4651 : * cheaper than an original scan (if no parameters change), and this
4652 : * function embodies knowledge about that. The default is to return
4653 : * the same costs stored in the Path. (Note that the cost estimates
4654 : * actually stored in Paths are always for first scans.)
4655 : *
4656 : * This function is not currently intended to model effects such as rescans
4657 : * being cheaper due to disk block caching; what we are concerned with is
4658 : * plan types wherein the executor caches results explicitly, or doesn't
4659 : * redo startup calculations, etc.
4660 : */
4661 : static void
4662 2717354 : cost_rescan(PlannerInfo *root, Path *path,
4663 : Cost *rescan_startup_cost, /* output parameters */
4664 : Cost *rescan_total_cost)
4665 : {
4666 2717354 : switch (path->pathtype)
4667 : {
4668 49114 : case T_FunctionScan:
4669 :
4670 : /*
4671 : * Currently, nodeFunctionscan.c always executes the function to
4672 : * completion before returning any rows, and caches the results in
4673 : * a tuplestore. So the function eval cost is all startup cost
4674 : * and isn't paid over again on rescans. However, all run costs
4675 : * will be paid over again.
4676 : */
4677 49114 : *rescan_startup_cost = 0;
4678 49114 : *rescan_total_cost = path->total_cost - path->startup_cost;
4679 49114 : break;
4680 122154 : case T_HashJoin:
4681 :
4682 : /*
4683 : * If it's a single-batch join, we don't need to rebuild the hash
4684 : * table during a rescan.
4685 : */
4686 122154 : if (((HashPath *) path)->num_batches == 1)
4687 : {
4688 : /* Startup cost is exactly the cost of hash table building */
4689 122154 : *rescan_startup_cost = 0;
4690 122154 : *rescan_total_cost = path->total_cost - path->startup_cost;
4691 : }
4692 : else
4693 : {
4694 : /* Otherwise, no special treatment */
4695 0 : *rescan_startup_cost = path->startup_cost;
4696 0 : *rescan_total_cost = path->total_cost;
4697 : }
4698 122154 : break;
4699 7926 : case T_CteScan:
4700 : case T_WorkTableScan:
4701 : {
4702 : /*
4703 : * These plan types materialize their final result in a
4704 : * tuplestore or tuplesort object. So the rescan cost is only
4705 : * cpu_tuple_cost per tuple, unless the result is large enough
4706 : * to spill to disk.
4707 : */
4708 7926 : Cost run_cost = cpu_tuple_cost * path->rows;
4709 7926 : double nbytes = relation_byte_size(path->rows,
4710 7926 : path->pathtarget->width);
4711 7926 : double work_mem_bytes = work_mem * (Size) 1024;
4712 :
4713 7926 : if (nbytes > work_mem_bytes)
4714 : {
4715 : /* It will spill, so account for re-read cost */
4716 352 : double npages = ceil(nbytes / BLCKSZ);
4717 :
4718 352 : run_cost += seq_page_cost * npages;
4719 : }
4720 7926 : *rescan_startup_cost = 0;
4721 7926 : *rescan_total_cost = run_cost;
4722 : }
4723 7926 : break;
4724 934500 : case T_Material:
4725 : case T_Sort:
4726 : {
4727 : /*
4728 : * These plan types not only materialize their results, but do
4729 : * not implement qual filtering or projection. So they are
4730 : * even cheaper to rescan than the ones above. We charge only
4731 : * cpu_operator_cost per tuple. (Note: keep that in sync with
4732 : * the run_cost charge in cost_sort, and also see comments in
4733 : * cost_material before you change it.)
4734 : */
4735 934500 : Cost run_cost = cpu_operator_cost * path->rows;
4736 934500 : double nbytes = relation_byte_size(path->rows,
4737 934500 : path->pathtarget->width);
4738 934500 : double work_mem_bytes = work_mem * (Size) 1024;
4739 :
4740 934500 : if (nbytes > work_mem_bytes)
4741 : {
4742 : /* It will spill, so account for re-read cost */
4743 9790 : double npages = ceil(nbytes / BLCKSZ);
4744 :
4745 9790 : run_cost += seq_page_cost * npages;
4746 : }
4747 934500 : *rescan_startup_cost = 0;
4748 934500 : *rescan_total_cost = run_cost;
4749 : }
4750 934500 : break;
4751 274376 : case T_Memoize:
4752 : /* All the hard work is done by cost_memoize_rescan */
4753 274376 : cost_memoize_rescan(root, (MemoizePath *) path,
4754 : rescan_startup_cost, rescan_total_cost);
4755 274376 : break;
4756 1329284 : default:
4757 1329284 : *rescan_startup_cost = path->startup_cost;
4758 1329284 : *rescan_total_cost = path->total_cost;
4759 1329284 : break;
4760 : }
4761 2717354 : }
4762 :
4763 :
4764 : /*
4765 : * cost_qual_eval
4766 : * Estimate the CPU costs of evaluating a WHERE clause.
4767 : * The input can be either an implicitly-ANDed list of boolean
4768 : * expressions, or a list of RestrictInfo nodes. (The latter is
4769 : * preferred since it allows caching of the results.)
4770 : * The result includes both a one-time (startup) component,
4771 : * and a per-evaluation component.
4772 : *
4773 : * Note: in some code paths root can be passed as NULL, resulting in
4774 : * slightly worse estimates.
4775 : */
4776 : void
4777 3783600 : cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
4778 : {
4779 : cost_qual_eval_context context;
4780 : ListCell *l;
4781 :
4782 3783600 : context.root = root;
4783 3783600 : context.total.startup = 0;
4784 3783600 : context.total.per_tuple = 0;
4785 :
4786 : /* We don't charge any cost for the implicit ANDing at top level ... */
4787 :
4788 7127852 : foreach(l, quals)
4789 : {
4790 3344252 : Node *qual = (Node *) lfirst(l);
4791 :
4792 3344252 : cost_qual_eval_walker(qual, &context);
4793 : }
4794 :
4795 3783600 : *cost = context.total;
4796 3783600 : }
4797 :
4798 : /*
4799 : * cost_qual_eval_node
4800 : * As above, for a single RestrictInfo or expression.
4801 : */
4802 : void
4803 1740998 : cost_qual_eval_node(QualCost *cost, Node *qual, PlannerInfo *root)
4804 : {
4805 : cost_qual_eval_context context;
4806 :
4807 1740998 : context.root = root;
4808 1740998 : context.total.startup = 0;
4809 1740998 : context.total.per_tuple = 0;
4810 :
4811 1740998 : cost_qual_eval_walker(qual, &context);
4812 :
4813 1740998 : *cost = context.total;
4814 1740998 : }
4815 :
4816 : static bool
4817 8416450 : cost_qual_eval_walker(Node *node, cost_qual_eval_context *context)
4818 : {
4819 8416450 : if (node == NULL)
4820 85140 : return false;
4821 :
4822 : /*
4823 : * RestrictInfo nodes contain an eval_cost field reserved for this
4824 : * routine's use, so that it's not necessary to evaluate the qual clause's
4825 : * cost more than once. If the clause's cost hasn't been computed yet,
4826 : * the field's startup value will contain -1.
4827 : */
4828 8331310 : if (IsA(node, RestrictInfo))
4829 : {
4830 3522534 : RestrictInfo *rinfo = (RestrictInfo *) node;
4831 :
4832 3522534 : if (rinfo->eval_cost.startup < 0)
4833 : {
4834 : cost_qual_eval_context locContext;
4835 :
4836 563060 : locContext.root = context->root;
4837 563060 : locContext.total.startup = 0;
4838 563060 : locContext.total.per_tuple = 0;
4839 :
4840 : /*
4841 : * For an OR clause, recurse into the marked-up tree so that we
4842 : * set the eval_cost for contained RestrictInfos too.
4843 : */
4844 563060 : if (rinfo->orclause)
4845 9480 : cost_qual_eval_walker((Node *) rinfo->orclause, &locContext);
4846 : else
4847 553580 : cost_qual_eval_walker((Node *) rinfo->clause, &locContext);
4848 :
4849 : /*
4850 : * If the RestrictInfo is marked pseudoconstant, it will be tested
4851 : * only once, so treat its cost as all startup cost.
4852 : */
4853 563060 : if (rinfo->pseudoconstant)
4854 : {
4855 : /* count one execution during startup */
4856 9894 : locContext.total.startup += locContext.total.per_tuple;
4857 9894 : locContext.total.per_tuple = 0;
4858 : }
4859 563060 : rinfo->eval_cost = locContext.total;
4860 : }
4861 3522534 : context->total.startup += rinfo->eval_cost.startup;
4862 3522534 : context->total.per_tuple += rinfo->eval_cost.per_tuple;
4863 : /* do NOT recurse into children */
4864 3522534 : return false;
4865 : }
4866 :
4867 : /*
4868 : * For each operator or function node in the given tree, we charge the
4869 : * estimated execution cost given by pg_proc.procost (remember to multiply
4870 : * this by cpu_operator_cost).
4871 : *
4872 : * Vars and Consts are charged zero, and so are boolean operators (AND,
4873 : * OR, NOT). Simplistic, but a lot better than no model at all.
4874 : *
4875 : * Should we try to account for the possibility of short-circuit
4876 : * evaluation of AND/OR? Probably *not*, because that would make the
4877 : * results depend on the clause ordering, and we are not in any position
4878 : * to expect that the current ordering of the clauses is the one that's
4879 : * going to end up being used. The above per-RestrictInfo caching would
4880 : * not mix well with trying to re-order clauses anyway.
4881 : *
4882 : * Another issue that is entirely ignored here is that if a set-returning
4883 : * function is below top level in the tree, the functions/operators above
4884 : * it will need to be evaluated multiple times. In practical use, such
4885 : * cases arise so seldom as to not be worth the added complexity needed;
4886 : * moreover, since our rowcount estimates for functions tend to be pretty
4887 : * phony, the results would also be pretty phony.
4888 : */
4889 4808776 : if (IsA(node, FuncExpr))
4890 : {
4891 337968 : add_function_cost(context->root, ((FuncExpr *) node)->funcid, node,
4892 : &context->total);
4893 : }
4894 4470808 : else if (IsA(node, OpExpr) ||
4895 3852110 : IsA(node, DistinctExpr) ||
4896 3851052 : IsA(node, NullIfExpr))
4897 : {
4898 : /* rely on struct equivalence to treat these all alike */
4899 619880 : set_opfuncid((OpExpr *) node);
4900 619880 : add_function_cost(context->root, ((OpExpr *) node)->opfuncid, node,
4901 : &context->total);
4902 : }
4903 3850928 : else if (IsA(node, ScalarArrayOpExpr))
4904 : {
4905 43666 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) node;
4906 43666 : Node *arraynode = (Node *) lsecond(saop->args);
4907 : QualCost sacosts;
4908 : QualCost hcosts;
4909 43666 : double estarraylen = estimate_array_length(context->root, arraynode);
4910 :
4911 43666 : set_sa_opfuncid(saop);
4912 43666 : sacosts.startup = sacosts.per_tuple = 0;
4913 43666 : add_function_cost(context->root, saop->opfuncid, NULL,
4914 : &sacosts);
4915 :
4916 43666 : if (OidIsValid(saop->hashfuncid))
4917 : {
4918 : /* Handle costs for hashed ScalarArrayOpExpr */
4919 428 : hcosts.startup = hcosts.per_tuple = 0;
4920 :
4921 428 : add_function_cost(context->root, saop->hashfuncid, NULL, &hcosts);
4922 428 : context->total.startup += sacosts.startup + hcosts.startup;
4923 :
4924 : /* Estimate the cost of building the hashtable. */
4925 428 : context->total.startup += estarraylen * hcosts.per_tuple;
4926 :
4927 : /*
4928 : * XXX should we charge a little bit for sacosts.per_tuple when
4929 : * building the table, or is it ok to assume there will be zero
4930 : * hash collision?
4931 : */
4932 :
4933 : /*
4934 : * Charge for hashtable lookups. Charge a single hash and a
4935 : * single comparison.
4936 : */
4937 428 : context->total.per_tuple += hcosts.per_tuple + sacosts.per_tuple;
4938 : }
4939 : else
4940 : {
4941 : /*
4942 : * Estimate that the operator will be applied to about half of the
4943 : * array elements before the answer is determined.
4944 : */
4945 43238 : context->total.startup += sacosts.startup;
4946 86476 : context->total.per_tuple += sacosts.per_tuple *
4947 43238 : estimate_array_length(context->root, arraynode) * 0.5;
4948 : }
4949 : }
4950 3807262 : else if (IsA(node, Aggref) ||
4951 3752256 : IsA(node, WindowFunc))
4952 : {
4953 : /*
4954 : * Aggref and WindowFunc nodes are (and should be) treated like Vars,
4955 : * ie, zero execution cost in the current model, because they behave
4956 : * essentially like Vars at execution. We disregard the costs of
4957 : * their input expressions for the same reason. The actual execution
4958 : * costs of the aggregate/window functions and their arguments have to
4959 : * be factored into plan-node-specific costing of the Agg or WindowAgg
4960 : * plan node.
4961 : */
4962 58530 : return false; /* don't recurse into children */
4963 : }
4964 3748732 : else if (IsA(node, GroupingFunc))
4965 : {
4966 : /* Treat this as having cost 1 */
4967 422 : context->total.per_tuple += cpu_operator_cost;
4968 422 : return false; /* don't recurse into children */
4969 : }
4970 3748310 : else if (IsA(node, CoerceViaIO))
4971 : {
4972 22146 : CoerceViaIO *iocoerce = (CoerceViaIO *) node;
4973 : Oid iofunc;
4974 : Oid typioparam;
4975 : bool typisvarlena;
4976 :
4977 : /* check the result type's input function */
4978 22146 : getTypeInputInfo(iocoerce->resulttype,
4979 : &iofunc, &typioparam);
4980 22146 : add_function_cost(context->root, iofunc, NULL,
4981 : &context->total);
4982 : /* check the input type's output function */
4983 22146 : getTypeOutputInfo(exprType((Node *) iocoerce->arg),
4984 : &iofunc, &typisvarlena);
4985 22146 : add_function_cost(context->root, iofunc, NULL,
4986 : &context->total);
4987 : }
4988 3726164 : else if (IsA(node, ArrayCoerceExpr))
4989 : {
4990 5076 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
4991 : QualCost perelemcost;
4992 :
4993 5076 : cost_qual_eval_node(&perelemcost, (Node *) acoerce->elemexpr,
4994 : context->root);
4995 5076 : context->total.startup += perelemcost.startup;
4996 5076 : if (perelemcost.per_tuple > 0)
4997 66 : context->total.per_tuple += perelemcost.per_tuple *
4998 66 : estimate_array_length(context->root, (Node *) acoerce->arg);
4999 : }
5000 3721088 : else if (IsA(node, RowCompareExpr))
5001 : {
5002 : /* Conservatively assume we will check all the columns */
5003 252 : RowCompareExpr *rcexpr = (RowCompareExpr *) node;
5004 : ListCell *lc;
5005 :
5006 810 : foreach(lc, rcexpr->opnos)
5007 : {
5008 558 : Oid opid = lfirst_oid(lc);
5009 :
5010 558 : add_function_cost(context->root, get_opcode(opid), NULL,
5011 : &context->total);
5012 : }
5013 : }
5014 3720836 : else if (IsA(node, MinMaxExpr) ||
5015 3720576 : IsA(node, SQLValueFunction) ||
5016 3715852 : IsA(node, XmlExpr) ||
5017 3715150 : IsA(node, CoerceToDomain) ||
5018 3705432 : IsA(node, NextValueExpr) ||
5019 3705072 : IsA(node, JsonExpr))
5020 : {
5021 : /* Treat all these as having cost 1 */
5022 18312 : context->total.per_tuple += cpu_operator_cost;
5023 : }
5024 3702524 : else if (IsA(node, SubLink))
5025 : {
5026 : /* This routine should not be applied to un-planned expressions */
5027 0 : elog(ERROR, "cannot handle unplanned sub-select");
5028 : }
5029 3702524 : else if (IsA(node, SubPlan))
5030 : {
5031 : /*
5032 : * A subplan node in an expression typically indicates that the
5033 : * subplan will be executed on each evaluation, so charge accordingly.
5034 : * (Sub-selects that can be executed as InitPlans have already been
5035 : * removed from the expression.)
5036 : */
5037 42884 : SubPlan *subplan = (SubPlan *) node;
5038 :
5039 42884 : context->total.startup += subplan->startup_cost;
5040 42884 : context->total.per_tuple += subplan->per_call_cost;
5041 :
5042 : /*
5043 : * We don't want to recurse into the testexpr, because it was already
5044 : * counted in the SubPlan node's costs. So we're done.
5045 : */
5046 42884 : return false;
5047 : }
5048 3659640 : else if (IsA(node, AlternativeSubPlan))
5049 : {
5050 : /*
5051 : * Arbitrarily use the first alternative plan for costing. (We should
5052 : * certainly only include one alternative, and we don't yet have
5053 : * enough information to know which one the executor is most likely to
5054 : * use.)
5055 : */
5056 1828 : AlternativeSubPlan *asplan = (AlternativeSubPlan *) node;
5057 :
5058 1828 : return cost_qual_eval_walker((Node *) linitial(asplan->subplans),
5059 : context);
5060 : }
5061 3657812 : else if (IsA(node, PlaceHolderVar))
5062 : {
5063 : /*
5064 : * A PlaceHolderVar should be given cost zero when considering general
5065 : * expression evaluation costs. The expense of doing the contained
5066 : * expression is charged as part of the tlist eval costs of the scan
5067 : * or join where the PHV is first computed (see set_rel_width and
5068 : * add_placeholders_to_joinrel). If we charged it again here, we'd be
5069 : * double-counting the cost for each level of plan that the PHV
5070 : * bubbles up through. Hence, return without recursing into the
5071 : * phexpr.
5072 : */
5073 5160 : return false;
5074 : }
5075 :
5076 : /* recurse into children */
5077 4699952 : return expression_tree_walker(node, cost_qual_eval_walker, context);
5078 : }
5079 :
5080 : /*
5081 : * get_restriction_qual_cost
5082 : * Compute evaluation costs of a baserel's restriction quals, plus any
5083 : * movable join quals that have been pushed down to the scan.
5084 : * Results are returned into *qpqual_cost.
5085 : *
5086 : * This is a convenience subroutine that works for seqscans and other cases
5087 : * where all the given quals will be evaluated the hard way. It's not useful
5088 : * for cost_index(), for example, where the index machinery takes care of
5089 : * some of the quals. We assume baserestrictcost was previously set by
5090 : * set_baserel_size_estimates().
5091 : */
5092 : static void
5093 1028922 : get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel,
5094 : ParamPathInfo *param_info,
5095 : QualCost *qpqual_cost)
5096 : {
5097 1028922 : if (param_info)
5098 : {
5099 : /* Include costs of pushed-down clauses */
5100 223518 : cost_qual_eval(qpqual_cost, param_info->ppi_clauses, root);
5101 :
5102 223518 : qpqual_cost->startup += baserel->baserestrictcost.startup;
5103 223518 : qpqual_cost->per_tuple += baserel->baserestrictcost.per_tuple;
5104 : }
5105 : else
5106 805404 : *qpqual_cost = baserel->baserestrictcost;
5107 1028922 : }
5108 :
5109 :
5110 : /*
5111 : * compute_semi_anti_join_factors
5112 : * Estimate how much of the inner input a SEMI, ANTI, or inner_unique join
5113 : * can be expected to scan.
5114 : *
5115 : * In a hash or nestloop SEMI/ANTI join, the executor will stop scanning
5116 : * inner rows as soon as it finds a match to the current outer row.
5117 : * The same happens if we have detected the inner rel is unique.
5118 : * We should therefore adjust some of the cost components for this effect.
5119 : * This function computes some estimates needed for these adjustments.
5120 : * These estimates will be the same regardless of the particular paths used
5121 : * for the outer and inner relation, so we compute these once and then pass
5122 : * them to all the join cost estimation functions.
5123 : *
5124 : * Input parameters:
5125 : * joinrel: join relation under consideration
5126 : * outerrel: outer relation under consideration
5127 : * innerrel: inner relation under consideration
5128 : * jointype: if not JOIN_SEMI or JOIN_ANTI, we assume it's inner_unique
5129 : * sjinfo: SpecialJoinInfo relevant to this join
5130 : * restrictlist: join quals
5131 : * Output parameters:
5132 : * *semifactors is filled in (see pathnodes.h for field definitions)
5133 : */
5134 : void
5135 199218 : compute_semi_anti_join_factors(PlannerInfo *root,
5136 : RelOptInfo *joinrel,
5137 : RelOptInfo *outerrel,
5138 : RelOptInfo *innerrel,
5139 : JoinType jointype,
5140 : SpecialJoinInfo *sjinfo,
5141 : List *restrictlist,
5142 : SemiAntiJoinFactors *semifactors)
5143 : {
5144 : Selectivity jselec;
5145 : Selectivity nselec;
5146 : Selectivity avgmatch;
5147 : SpecialJoinInfo norm_sjinfo;
5148 : List *joinquals;
5149 : ListCell *l;
5150 :
5151 : /*
5152 : * In an ANTI join, we must ignore clauses that are "pushed down", since
5153 : * those won't affect the match logic. In a SEMI join, we do not
5154 : * distinguish joinquals from "pushed down" quals, so just use the whole
5155 : * restrictinfo list. For other outer join types, we should consider only
5156 : * non-pushed-down quals, so that this devolves to an IS_OUTER_JOIN check.
5157 : */
5158 199218 : if (IS_OUTER_JOIN(jointype))
5159 : {
5160 74012 : joinquals = NIL;
5161 162024 : foreach(l, restrictlist)
5162 : {
5163 88012 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5164 :
5165 88012 : if (!RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5166 83226 : joinquals = lappend(joinquals, rinfo);
5167 : }
5168 : }
5169 : else
5170 125206 : joinquals = restrictlist;
5171 :
5172 : /*
5173 : * Get the JOIN_SEMI or JOIN_ANTI selectivity of the join clauses.
5174 : */
5175 199218 : jselec = clauselist_selectivity(root,
5176 : joinquals,
5177 : 0,
5178 : (jointype == JOIN_ANTI) ? JOIN_ANTI : JOIN_SEMI,
5179 : sjinfo);
5180 :
5181 : /*
5182 : * Also get the normal inner-join selectivity of the join clauses.
5183 : */
5184 199218 : init_dummy_sjinfo(&norm_sjinfo, outerrel->relids, innerrel->relids);
5185 :
5186 199218 : nselec = clauselist_selectivity(root,
5187 : joinquals,
5188 : 0,
5189 : JOIN_INNER,
5190 : &norm_sjinfo);
5191 :
5192 : /* Avoid leaking a lot of ListCells */
5193 199218 : if (IS_OUTER_JOIN(jointype))
5194 74012 : list_free(joinquals);
5195 :
5196 : /*
5197 : * jselec can be interpreted as the fraction of outer-rel rows that have
5198 : * any matches (this is true for both SEMI and ANTI cases). And nselec is
5199 : * the fraction of the Cartesian product that matches. So, the average
5200 : * number of matches for each outer-rel row that has at least one match is
5201 : * nselec * inner_rows / jselec.
5202 : *
5203 : * Note: it is correct to use the inner rel's "rows" count here, even
5204 : * though we might later be considering a parameterized inner path with
5205 : * fewer rows. This is because we have included all the join clauses in
5206 : * the selectivity estimate.
5207 : */
5208 199218 : if (jselec > 0) /* protect against zero divide */
5209 : {
5210 198874 : avgmatch = nselec * innerrel->rows / jselec;
5211 : /* Clamp to sane range */
5212 198874 : avgmatch = Max(1.0, avgmatch);
5213 : }
5214 : else
5215 344 : avgmatch = 1.0;
5216 :
5217 199218 : semifactors->outer_match_frac = jselec;
5218 199218 : semifactors->match_count = avgmatch;
5219 199218 : }
5220 :
5221 : /*
5222 : * has_indexed_join_quals
5223 : * Check whether all the joinquals of a nestloop join are used as
5224 : * inner index quals.
5225 : *
5226 : * If the inner path of a SEMI/ANTI join is an indexscan (including bitmap
5227 : * indexscan) that uses all the joinquals as indexquals, we can assume that an
5228 : * unmatched outer tuple is cheap to process, whereas otherwise it's probably
5229 : * expensive.
5230 : */
5231 : static bool
5232 853894 : has_indexed_join_quals(NestPath *path)
5233 : {
5234 853894 : JoinPath *joinpath = &path->jpath;
5235 853894 : Relids joinrelids = joinpath->path.parent->relids;
5236 853894 : Path *innerpath = joinpath->innerjoinpath;
5237 : List *indexclauses;
5238 : bool found_one;
5239 : ListCell *lc;
5240 :
5241 : /* If join still has quals to evaluate, it's not fast */
5242 853894 : if (joinpath->joinrestrictinfo != NIL)
5243 607048 : return false;
5244 : /* Nor if the inner path isn't parameterized at all */
5245 246846 : if (innerpath->param_info == NULL)
5246 3300 : return false;
5247 :
5248 : /* Find the indexclauses list for the inner scan */
5249 243546 : switch (innerpath->pathtype)
5250 : {
5251 146954 : case T_IndexScan:
5252 : case T_IndexOnlyScan:
5253 146954 : indexclauses = ((IndexPath *) innerpath)->indexclauses;
5254 146954 : break;
5255 270 : case T_BitmapHeapScan:
5256 : {
5257 : /* Accept only a simple bitmap scan, not AND/OR cases */
5258 270 : Path *bmqual = ((BitmapHeapPath *) innerpath)->bitmapqual;
5259 :
5260 270 : if (IsA(bmqual, IndexPath))
5261 222 : indexclauses = ((IndexPath *) bmqual)->indexclauses;
5262 : else
5263 48 : return false;
5264 222 : break;
5265 : }
5266 96322 : default:
5267 :
5268 : /*
5269 : * If it's not a simple indexscan, it probably doesn't run quickly
5270 : * for zero rows out, even if it's a parameterized path using all
5271 : * the joinquals.
5272 : */
5273 96322 : return false;
5274 : }
5275 :
5276 : /*
5277 : * Examine the inner path's param clauses. Any that are from the outer
5278 : * path must be found in the indexclauses list, either exactly or in an
5279 : * equivalent form generated by equivclass.c. Also, we must find at least
5280 : * one such clause, else it's a clauseless join which isn't fast.
5281 : */
5282 147176 : found_one = false;
5283 293244 : foreach(lc, innerpath->param_info->ppi_clauses)
5284 : {
5285 150408 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
5286 :
5287 150408 : if (join_clause_is_movable_into(rinfo,
5288 150408 : innerpath->parent->relids,
5289 : joinrelids))
5290 : {
5291 149856 : if (!is_redundant_with_indexclauses(rinfo, indexclauses))
5292 4340 : return false;
5293 145516 : found_one = true;
5294 : }
5295 : }
5296 142836 : return found_one;
5297 : }
5298 :
5299 :
5300 : /*
5301 : * approx_tuple_count
5302 : * Quick-and-dirty estimation of the number of join rows passing
5303 : * a set of qual conditions.
5304 : *
5305 : * The quals can be either an implicitly-ANDed list of boolean expressions,
5306 : * or a list of RestrictInfo nodes (typically the latter).
5307 : *
5308 : * We intentionally compute the selectivity under JOIN_INNER rules, even
5309 : * if it's some type of outer join. This is appropriate because we are
5310 : * trying to figure out how many tuples pass the initial merge or hash
5311 : * join step.
5312 : *
5313 : * This is quick-and-dirty because we bypass clauselist_selectivity, and
5314 : * simply multiply the independent clause selectivities together. Now
5315 : * clauselist_selectivity often can't do any better than that anyhow, but
5316 : * for some situations (such as range constraints) it is smarter. However,
5317 : * we can't effectively cache the results of clauselist_selectivity, whereas
5318 : * the individual clause selectivities can be and are cached.
5319 : *
5320 : * Since we are only using the results to estimate how many potential
5321 : * output tuples are generated and passed through qpqual checking, it
5322 : * seems OK to live with the approximation.
5323 : */
5324 : static double
5325 473712 : approx_tuple_count(PlannerInfo *root, JoinPath *path, List *quals)
5326 : {
5327 : double tuples;
5328 473712 : double outer_tuples = path->outerjoinpath->rows;
5329 473712 : double inner_tuples = path->innerjoinpath->rows;
5330 : SpecialJoinInfo sjinfo;
5331 473712 : Selectivity selec = 1.0;
5332 : ListCell *l;
5333 :
5334 : /*
5335 : * Make up a SpecialJoinInfo for JOIN_INNER semantics.
5336 : */
5337 473712 : init_dummy_sjinfo(&sjinfo, path->outerjoinpath->parent->relids,
5338 473712 : path->innerjoinpath->parent->relids);
5339 :
5340 : /* Get the approximate selectivity */
5341 1021802 : foreach(l, quals)
5342 : {
5343 548090 : Node *qual = (Node *) lfirst(l);
5344 :
5345 : /* Note that clause_selectivity will be able to cache its result */
5346 548090 : selec *= clause_selectivity(root, qual, 0, JOIN_INNER, &sjinfo);
5347 : }
5348 :
5349 : /* Apply it to the input relation sizes */
5350 473712 : tuples = selec * outer_tuples * inner_tuples;
5351 :
5352 473712 : return clamp_row_est(tuples);
5353 : }
5354 :
5355 :
5356 : /*
5357 : * set_baserel_size_estimates
5358 : * Set the size estimates for the given base relation.
5359 : *
5360 : * The rel's targetlist and restrictinfo list must have been constructed
5361 : * already, and rel->tuples must be set.
5362 : *
5363 : * We set the following fields of the rel node:
5364 : * rows: the estimated number of output tuples (after applying
5365 : * restriction clauses).
5366 : * width: the estimated average output tuple width in bytes.
5367 : * baserestrictcost: estimated cost of evaluating baserestrictinfo clauses.
5368 : */
5369 : void
5370 488772 : set_baserel_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5371 : {
5372 : double nrows;
5373 :
5374 : /* Should only be applied to base relations */
5375 : Assert(rel->relid > 0);
5376 :
5377 977514 : nrows = rel->tuples *
5378 488772 : clauselist_selectivity(root,
5379 : rel->baserestrictinfo,
5380 : 0,
5381 : JOIN_INNER,
5382 : NULL);
5383 :
5384 488742 : rel->rows = clamp_row_est(nrows);
5385 :
5386 488742 : cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo, root);
5387 :
5388 488742 : set_rel_width(root, rel);
5389 488742 : }
5390 :
5391 : /*
5392 : * get_parameterized_baserel_size
5393 : * Make a size estimate for a parameterized scan of a base relation.
5394 : *
5395 : * 'param_clauses' lists the additional join clauses to be used.
5396 : *
5397 : * set_baserel_size_estimates must have been applied already.
5398 : */
5399 : double
5400 143492 : get_parameterized_baserel_size(PlannerInfo *root, RelOptInfo *rel,
5401 : List *param_clauses)
5402 : {
5403 : List *allclauses;
5404 : double nrows;
5405 :
5406 : /*
5407 : * Estimate the number of rows returned by the parameterized scan, knowing
5408 : * that it will apply all the extra join clauses as well as the rel's own
5409 : * restriction clauses. Note that we force the clauses to be treated as
5410 : * non-join clauses during selectivity estimation.
5411 : */
5412 143492 : allclauses = list_concat_copy(param_clauses, rel->baserestrictinfo);
5413 286984 : nrows = rel->tuples *
5414 143492 : clauselist_selectivity(root,
5415 : allclauses,
5416 143492 : rel->relid, /* do not use 0! */
5417 : JOIN_INNER,
5418 : NULL);
5419 143492 : nrows = clamp_row_est(nrows);
5420 : /* For safety, make sure result is not more than the base estimate */
5421 143492 : if (nrows > rel->rows)
5422 0 : nrows = rel->rows;
5423 143492 : return nrows;
5424 : }
5425 :
5426 : /*
5427 : * set_joinrel_size_estimates
5428 : * Set the size estimates for the given join relation.
5429 : *
5430 : * The rel's targetlist must have been constructed already, and a
5431 : * restriction clause list that matches the given component rels must
5432 : * be provided.
5433 : *
5434 : * Since there is more than one way to make a joinrel for more than two
5435 : * base relations, the results we get here could depend on which component
5436 : * rel pair is provided. In theory we should get the same answers no matter
5437 : * which pair is provided; in practice, since the selectivity estimation
5438 : * routines don't handle all cases equally well, we might not. But there's
5439 : * not much to be done about it. (Would it make sense to repeat the
5440 : * calculations for each pair of input rels that's encountered, and somehow
5441 : * average the results? Probably way more trouble than it's worth, and
5442 : * anyway we must keep the rowcount estimate the same for all paths for the
5443 : * joinrel.)
5444 : *
5445 : * We set only the rows field here. The reltarget field was already set by
5446 : * build_joinrel_tlist, and baserestrictcost is not used for join rels.
5447 : */
5448 : void
5449 206938 : set_joinrel_size_estimates(PlannerInfo *root, RelOptInfo *rel,
5450 : RelOptInfo *outer_rel,
5451 : RelOptInfo *inner_rel,
5452 : SpecialJoinInfo *sjinfo,
5453 : List *restrictlist)
5454 : {
5455 206938 : rel->rows = calc_joinrel_size_estimate(root,
5456 : rel,
5457 : outer_rel,
5458 : inner_rel,
5459 : outer_rel->rows,
5460 : inner_rel->rows,
5461 : sjinfo,
5462 : restrictlist);
5463 206938 : }
5464 :
5465 : /*
5466 : * get_parameterized_joinrel_size
5467 : * Make a size estimate for a parameterized scan of a join relation.
5468 : *
5469 : * 'rel' is the joinrel under consideration.
5470 : * 'outer_path', 'inner_path' are (probably also parameterized) Paths that
5471 : * produce the relations being joined.
5472 : * 'sjinfo' is any SpecialJoinInfo relevant to this join.
5473 : * 'restrict_clauses' lists the join clauses that need to be applied at the
5474 : * join node (including any movable clauses that were moved down to this join,
5475 : * and not including any movable clauses that were pushed down into the
5476 : * child paths).
5477 : *
5478 : * set_joinrel_size_estimates must have been applied already.
5479 : */
5480 : double
5481 8472 : get_parameterized_joinrel_size(PlannerInfo *root, RelOptInfo *rel,
5482 : Path *outer_path,
5483 : Path *inner_path,
5484 : SpecialJoinInfo *sjinfo,
5485 : List *restrict_clauses)
5486 : {
5487 : double nrows;
5488 :
5489 : /*
5490 : * Estimate the number of rows returned by the parameterized join as the
5491 : * sizes of the input paths times the selectivity of the clauses that have
5492 : * ended up at this join node.
5493 : *
5494 : * As with set_joinrel_size_estimates, the rowcount estimate could depend
5495 : * on the pair of input paths provided, though ideally we'd get the same
5496 : * estimate for any pair with the same parameterization.
5497 : */
5498 8472 : nrows = calc_joinrel_size_estimate(root,
5499 : rel,
5500 : outer_path->parent,
5501 : inner_path->parent,
5502 : outer_path->rows,
5503 : inner_path->rows,
5504 : sjinfo,
5505 : restrict_clauses);
5506 : /* For safety, make sure result is not more than the base estimate */
5507 8472 : if (nrows > rel->rows)
5508 12 : nrows = rel->rows;
5509 8472 : return nrows;
5510 : }
5511 :
5512 : /*
5513 : * calc_joinrel_size_estimate
5514 : * Workhorse for set_joinrel_size_estimates and
5515 : * get_parameterized_joinrel_size.
5516 : *
5517 : * outer_rel/inner_rel are the relations being joined, but they should be
5518 : * assumed to have sizes outer_rows/inner_rows; those numbers might be less
5519 : * than what rel->rows says, when we are considering parameterized paths.
5520 : */
5521 : static double
5522 215410 : calc_joinrel_size_estimate(PlannerInfo *root,
5523 : RelOptInfo *joinrel,
5524 : RelOptInfo *outer_rel,
5525 : RelOptInfo *inner_rel,
5526 : double outer_rows,
5527 : double inner_rows,
5528 : SpecialJoinInfo *sjinfo,
5529 : List *restrictlist)
5530 : {
5531 215410 : JoinType jointype = sjinfo->jointype;
5532 : Selectivity fkselec;
5533 : Selectivity jselec;
5534 : Selectivity pselec;
5535 : double nrows;
5536 :
5537 : /*
5538 : * Compute joinclause selectivity. Note that we are only considering
5539 : * clauses that become restriction clauses at this join level; we are not
5540 : * double-counting them because they were not considered in estimating the
5541 : * sizes of the component rels.
5542 : *
5543 : * First, see whether any of the joinclauses can be matched to known FK
5544 : * constraints. If so, drop those clauses from the restrictlist, and
5545 : * instead estimate their selectivity using FK semantics. (We do this
5546 : * without regard to whether said clauses are local or "pushed down".
5547 : * Probably, an FK-matching clause could never be seen as pushed down at
5548 : * an outer join, since it would be strict and hence would be grounds for
5549 : * join strength reduction.) fkselec gets the net selectivity for
5550 : * FK-matching clauses, or 1.0 if there are none.
5551 : */
5552 215410 : fkselec = get_foreign_key_join_selectivity(root,
5553 : outer_rel->relids,
5554 : inner_rel->relids,
5555 : sjinfo,
5556 : &restrictlist);
5557 :
5558 : /*
5559 : * For an outer join, we have to distinguish the selectivity of the join's
5560 : * own clauses (JOIN/ON conditions) from any clauses that were "pushed
5561 : * down". For inner joins we just count them all as joinclauses.
5562 : */
5563 215410 : if (IS_OUTER_JOIN(jointype))
5564 : {
5565 79394 : List *joinquals = NIL;
5566 79394 : List *pushedquals = NIL;
5567 : ListCell *l;
5568 :
5569 : /* Grovel through the clauses to separate into two lists */
5570 178842 : foreach(l, restrictlist)
5571 : {
5572 99448 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5573 :
5574 99448 : if (RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5575 4248 : pushedquals = lappend(pushedquals, rinfo);
5576 : else
5577 95200 : joinquals = lappend(joinquals, rinfo);
5578 : }
5579 :
5580 : /* Get the separate selectivities */
5581 79394 : jselec = clauselist_selectivity(root,
5582 : joinquals,
5583 : 0,
5584 : jointype,
5585 : sjinfo);
5586 79394 : pselec = clauselist_selectivity(root,
5587 : pushedquals,
5588 : 0,
5589 : jointype,
5590 : sjinfo);
5591 :
5592 : /* Avoid leaking a lot of ListCells */
5593 79394 : list_free(joinquals);
5594 79394 : list_free(pushedquals);
5595 : }
5596 : else
5597 : {
5598 136016 : jselec = clauselist_selectivity(root,
5599 : restrictlist,
5600 : 0,
5601 : jointype,
5602 : sjinfo);
5603 136016 : pselec = 0.0; /* not used, keep compiler quiet */
5604 : }
5605 :
5606 : /*
5607 : * Basically, we multiply size of Cartesian product by selectivity.
5608 : *
5609 : * If we are doing an outer join, take that into account: the joinqual
5610 : * selectivity has to be clamped using the knowledge that the output must
5611 : * be at least as large as the non-nullable input. However, any
5612 : * pushed-down quals are applied after the outer join, so their
5613 : * selectivity applies fully.
5614 : *
5615 : * For JOIN_SEMI and JOIN_ANTI, the selectivity is defined as the fraction
5616 : * of LHS rows that have matches, and we apply that straightforwardly.
5617 : */
5618 215410 : switch (jointype)
5619 : {
5620 127918 : case JOIN_INNER:
5621 127918 : nrows = outer_rows * inner_rows * fkselec * jselec;
5622 : /* pselec not used */
5623 127918 : break;
5624 72706 : case JOIN_LEFT:
5625 72706 : nrows = outer_rows * inner_rows * fkselec * jselec;
5626 72706 : if (nrows < outer_rows)
5627 28422 : nrows = outer_rows;
5628 72706 : nrows *= pselec;
5629 72706 : break;
5630 1714 : case JOIN_FULL:
5631 1714 : nrows = outer_rows * inner_rows * fkselec * jselec;
5632 1714 : if (nrows < outer_rows)
5633 1136 : nrows = outer_rows;
5634 1714 : if (nrows < inner_rows)
5635 120 : nrows = inner_rows;
5636 1714 : nrows *= pselec;
5637 1714 : break;
5638 8098 : case JOIN_SEMI:
5639 8098 : nrows = outer_rows * fkselec * jselec;
5640 : /* pselec not used */
5641 8098 : break;
5642 4974 : case JOIN_ANTI:
5643 4974 : nrows = outer_rows * (1.0 - fkselec * jselec);
5644 4974 : nrows *= pselec;
5645 4974 : break;
5646 0 : default:
5647 : /* other values not expected here */
5648 0 : elog(ERROR, "unrecognized join type: %d", (int) jointype);
5649 : nrows = 0; /* keep compiler quiet */
5650 : break;
5651 : }
5652 :
5653 215410 : return clamp_row_est(nrows);
5654 : }
5655 :
5656 : /*
5657 : * get_foreign_key_join_selectivity
5658 : * Estimate join selectivity for foreign-key-related clauses.
5659 : *
5660 : * Remove any clauses that can be matched to FK constraints from *restrictlist,
5661 : * and return a substitute estimate of their selectivity. 1.0 is returned
5662 : * when there are no such clauses.
5663 : *
5664 : * The reason for treating such clauses specially is that we can get better
5665 : * estimates this way than by relying on clauselist_selectivity(), especially
5666 : * for multi-column FKs where that function's assumption that the clauses are
5667 : * independent falls down badly. But even with single-column FKs, we may be
5668 : * able to get a better answer when the pg_statistic stats are missing or out
5669 : * of date.
5670 : */
5671 : static Selectivity
5672 215410 : get_foreign_key_join_selectivity(PlannerInfo *root,
5673 : Relids outer_relids,
5674 : Relids inner_relids,
5675 : SpecialJoinInfo *sjinfo,
5676 : List **restrictlist)
5677 : {
5678 215410 : Selectivity fkselec = 1.0;
5679 215410 : JoinType jointype = sjinfo->jointype;
5680 215410 : List *worklist = *restrictlist;
5681 : ListCell *lc;
5682 :
5683 : /* Consider each FK constraint that is known to match the query */
5684 217376 : foreach(lc, root->fkey_list)
5685 : {
5686 1966 : ForeignKeyOptInfo *fkinfo = (ForeignKeyOptInfo *) lfirst(lc);
5687 : bool ref_is_outer;
5688 : List *removedlist;
5689 : ListCell *cell;
5690 :
5691 : /*
5692 : * This FK is not relevant unless it connects a baserel on one side of
5693 : * this join to a baserel on the other side.
5694 : */
5695 3580 : if (bms_is_member(fkinfo->con_relid, outer_relids) &&
5696 1614 : bms_is_member(fkinfo->ref_relid, inner_relids))
5697 1440 : ref_is_outer = false;
5698 866 : else if (bms_is_member(fkinfo->ref_relid, outer_relids) &&
5699 340 : bms_is_member(fkinfo->con_relid, inner_relids))
5700 130 : ref_is_outer = true;
5701 : else
5702 396 : continue;
5703 :
5704 : /*
5705 : * If we're dealing with a semi/anti join, and the FK's referenced
5706 : * relation is on the outside, then knowledge of the FK doesn't help
5707 : * us figure out what we need to know (which is the fraction of outer
5708 : * rows that have matches). On the other hand, if the referenced rel
5709 : * is on the inside, then all outer rows must have matches in the
5710 : * referenced table (ignoring nulls). But any restriction or join
5711 : * clauses that filter that table will reduce the fraction of matches.
5712 : * We can account for restriction clauses, but it's too hard to guess
5713 : * how many table rows would get through a join that's inside the RHS.
5714 : * Hence, if either case applies, punt and ignore the FK.
5715 : */
5716 1570 : if ((jointype == JOIN_SEMI || jointype == JOIN_ANTI) &&
5717 1048 : (ref_is_outer || bms_membership(inner_relids) != BMS_SINGLETON))
5718 12 : continue;
5719 :
5720 : /*
5721 : * Modify the restrictlist by removing clauses that match the FK (and
5722 : * putting them into removedlist instead). It seems unsafe to modify
5723 : * the originally-passed List structure, so we make a shallow copy the
5724 : * first time through.
5725 : */
5726 1558 : if (worklist == *restrictlist)
5727 1334 : worklist = list_copy(worklist);
5728 :
5729 1558 : removedlist = NIL;
5730 3252 : foreach(cell, worklist)
5731 : {
5732 1694 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
5733 1694 : bool remove_it = false;
5734 : int i;
5735 :
5736 : /* Drop this clause if it matches any column of the FK */
5737 2140 : for (i = 0; i < fkinfo->nkeys; i++)
5738 : {
5739 2110 : if (rinfo->parent_ec)
5740 : {
5741 : /*
5742 : * EC-derived clauses can only match by EC. It is okay to
5743 : * consider any clause derived from the same EC as
5744 : * matching the FK: even if equivclass.c chose to generate
5745 : * a clause equating some other pair of Vars, it could
5746 : * have generated one equating the FK's Vars. So for
5747 : * purposes of estimation, we can act as though it did so.
5748 : *
5749 : * Note: checking parent_ec is a bit of a cheat because
5750 : * there are EC-derived clauses that don't have parent_ec
5751 : * set; but such clauses must compare expressions that
5752 : * aren't just Vars, so they cannot match the FK anyway.
5753 : */
5754 304 : if (fkinfo->eclass[i] == rinfo->parent_ec)
5755 : {
5756 298 : remove_it = true;
5757 298 : break;
5758 : }
5759 : }
5760 : else
5761 : {
5762 : /*
5763 : * Otherwise, see if rinfo was previously matched to FK as
5764 : * a "loose" clause.
5765 : */
5766 1806 : if (list_member_ptr(fkinfo->rinfos[i], rinfo))
5767 : {
5768 1366 : remove_it = true;
5769 1366 : break;
5770 : }
5771 : }
5772 : }
5773 1694 : if (remove_it)
5774 : {
5775 1664 : worklist = foreach_delete_current(worklist, cell);
5776 1664 : removedlist = lappend(removedlist, rinfo);
5777 : }
5778 : }
5779 :
5780 : /*
5781 : * If we failed to remove all the matching clauses we expected to
5782 : * find, chicken out and ignore this FK; applying its selectivity
5783 : * might result in double-counting. Put any clauses we did manage to
5784 : * remove back into the worklist.
5785 : *
5786 : * Since the matching clauses are known not outerjoin-delayed, they
5787 : * would normally have appeared in the initial joinclause list. If we
5788 : * didn't find them, there are two possibilities:
5789 : *
5790 : * 1. If the FK match is based on an EC that is ec_has_const, it won't
5791 : * have generated any join clauses at all. We discount such ECs while
5792 : * checking to see if we have "all" the clauses. (Below, we'll adjust
5793 : * the selectivity estimate for this case.)
5794 : *
5795 : * 2. The clauses were matched to some other FK in a previous
5796 : * iteration of this loop, and thus removed from worklist. (A likely
5797 : * case is that two FKs are matched to the same EC; there will be only
5798 : * one EC-derived clause in the initial list, so the first FK will
5799 : * consume it.) Applying both FKs' selectivity independently risks
5800 : * underestimating the join size; in particular, this would undo one
5801 : * of the main things that ECs were invented for, namely to avoid
5802 : * double-counting the selectivity of redundant equality conditions.
5803 : * Later we might think of a reasonable way to combine the estimates,
5804 : * but for now, just punt, since this is a fairly uncommon situation.
5805 : */
5806 1558 : if (removedlist == NIL ||
5807 1272 : list_length(removedlist) !=
5808 1272 : (fkinfo->nmatched_ec - fkinfo->nconst_ec + fkinfo->nmatched_ri))
5809 : {
5810 286 : worklist = list_concat(worklist, removedlist);
5811 286 : continue;
5812 : }
5813 :
5814 : /*
5815 : * Finally we get to the payoff: estimate selectivity using the
5816 : * knowledge that each referencing row will match exactly one row in
5817 : * the referenced table.
5818 : *
5819 : * XXX that's not true in the presence of nulls in the referencing
5820 : * column(s), so in principle we should derate the estimate for those.
5821 : * However (1) if there are any strict restriction clauses for the
5822 : * referencing column(s) elsewhere in the query, derating here would
5823 : * be double-counting the null fraction, and (2) it's not very clear
5824 : * how to combine null fractions for multiple referencing columns. So
5825 : * we do nothing for now about correcting for nulls.
5826 : *
5827 : * XXX another point here is that if either side of an FK constraint
5828 : * is an inheritance parent, we estimate as though the constraint
5829 : * covers all its children as well. This is not an unreasonable
5830 : * assumption for a referencing table, ie the user probably applied
5831 : * identical constraints to all child tables (though perhaps we ought
5832 : * to check that). But it's not possible to have done that for a
5833 : * referenced table. Fortunately, precisely because that doesn't
5834 : * work, it is uncommon in practice to have an FK referencing a parent
5835 : * table. So, at least for now, disregard inheritance here.
5836 : */
5837 1272 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
5838 824 : {
5839 : /*
5840 : * For JOIN_SEMI and JOIN_ANTI, we only get here when the FK's
5841 : * referenced table is exactly the inside of the join. The join
5842 : * selectivity is defined as the fraction of LHS rows that have
5843 : * matches. The FK implies that every LHS row has a match *in the
5844 : * referenced table*; but any restriction clauses on it will
5845 : * reduce the number of matches. Hence we take the join
5846 : * selectivity as equal to the selectivity of the table's
5847 : * restriction clauses, which is rows / tuples; but we must guard
5848 : * against tuples == 0.
5849 : */
5850 824 : RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5851 824 : double ref_tuples = Max(ref_rel->tuples, 1.0);
5852 :
5853 824 : fkselec *= ref_rel->rows / ref_tuples;
5854 : }
5855 : else
5856 : {
5857 : /*
5858 : * Otherwise, selectivity is exactly 1/referenced-table-size; but
5859 : * guard against tuples == 0. Note we should use the raw table
5860 : * tuple count, not any estimate of its filtered or joined size.
5861 : */
5862 448 : RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5863 448 : double ref_tuples = Max(ref_rel->tuples, 1.0);
5864 :
5865 448 : fkselec *= 1.0 / ref_tuples;
5866 : }
5867 :
5868 : /*
5869 : * If any of the FK columns participated in ec_has_const ECs, then
5870 : * equivclass.c will have generated "var = const" restrictions for
5871 : * each side of the join, thus reducing the sizes of both input
5872 : * relations. Taking the fkselec at face value would amount to
5873 : * double-counting the selectivity of the constant restriction for the
5874 : * referencing Var. Hence, look for the restriction clause(s) that
5875 : * were applied to the referencing Var(s), and divide out their
5876 : * selectivity to correct for this.
5877 : */
5878 1272 : if (fkinfo->nconst_ec > 0)
5879 : {
5880 24 : for (int i = 0; i < fkinfo->nkeys; i++)
5881 : {
5882 18 : EquivalenceClass *ec = fkinfo->eclass[i];
5883 :
5884 18 : if (ec && ec->ec_has_const)
5885 : {
5886 6 : EquivalenceMember *em = fkinfo->fk_eclass_member[i];
5887 6 : RestrictInfo *rinfo = find_derived_clause_for_ec_member(root,
5888 : ec,
5889 : em);
5890 :
5891 6 : if (rinfo)
5892 : {
5893 : Selectivity s0;
5894 :
5895 6 : s0 = clause_selectivity(root,
5896 : (Node *) rinfo,
5897 : 0,
5898 : jointype,
5899 : sjinfo);
5900 6 : if (s0 > 0)
5901 6 : fkselec /= s0;
5902 : }
5903 : }
5904 : }
5905 : }
5906 : }
5907 :
5908 215410 : *restrictlist = worklist;
5909 215410 : CLAMP_PROBABILITY(fkselec);
5910 215410 : return fkselec;
5911 : }
5912 :
5913 : /*
5914 : * set_subquery_size_estimates
5915 : * Set the size estimates for a base relation that is a subquery.
5916 : *
5917 : * The rel's targetlist and restrictinfo list must have been constructed
5918 : * already, and the Paths for the subquery must have been completed.
5919 : * We look at the subquery's PlannerInfo to extract data.
5920 : *
5921 : * We set the same fields as set_baserel_size_estimates.
5922 : */
5923 : void
5924 27504 : set_subquery_size_estimates(PlannerInfo *root, RelOptInfo *rel)
5925 : {
5926 27504 : PlannerInfo *subroot = rel->subroot;
5927 : RelOptInfo *sub_final_rel;
5928 : ListCell *lc;
5929 :
5930 : /* Should only be applied to base relations that are subqueries */
5931 : Assert(rel->relid > 0);
5932 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_SUBQUERY);
5933 :
5934 : /*
5935 : * Copy raw number of output rows from subquery. All of its paths should
5936 : * have the same output rowcount, so just look at cheapest-total.
5937 : */
5938 27504 : sub_final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
5939 27504 : rel->tuples = sub_final_rel->cheapest_total_path->rows;
5940 :
5941 : /*
5942 : * Compute per-output-column width estimates by examining the subquery's
5943 : * targetlist. For any output that is a plain Var, get the width estimate
5944 : * that was made while planning the subquery. Otherwise, we leave it to
5945 : * set_rel_width to fill in a datatype-based default estimate.
5946 : */
5947 111450 : foreach(lc, subroot->parse->targetList)
5948 : {
5949 83946 : TargetEntry *te = lfirst_node(TargetEntry, lc);
5950 83946 : Node *texpr = (Node *) te->expr;
5951 83946 : int32 item_width = 0;
5952 :
5953 : /* junk columns aren't visible to upper query */
5954 83946 : if (te->resjunk)
5955 1286 : continue;
5956 :
5957 : /*
5958 : * The subquery could be an expansion of a view that's had columns
5959 : * added to it since the current query was parsed, so that there are
5960 : * non-junk tlist columns in it that don't correspond to any column
5961 : * visible at our query level. Ignore such columns.
5962 : */
5963 82660 : if (te->resno < rel->min_attr || te->resno > rel->max_attr)
5964 0 : continue;
5965 :
5966 : /*
5967 : * XXX This currently doesn't work for subqueries containing set
5968 : * operations, because the Vars in their tlists are bogus references
5969 : * to the first leaf subquery, which wouldn't give the right answer
5970 : * even if we could still get to its PlannerInfo.
5971 : *
5972 : * Also, the subquery could be an appendrel for which all branches are
5973 : * known empty due to constraint exclusion, in which case
5974 : * set_append_rel_pathlist will have left the attr_widths set to zero.
5975 : *
5976 : * In either case, we just leave the width estimate zero until
5977 : * set_rel_width fixes it.
5978 : */
5979 82660 : if (IsA(texpr, Var) &&
5980 36766 : subroot->parse->setOperations == NULL)
5981 : {
5982 34990 : Var *var = (Var *) texpr;
5983 34990 : RelOptInfo *subrel = find_base_rel(subroot, var->varno);
5984 :
5985 34990 : item_width = subrel->attr_widths[var->varattno - subrel->min_attr];
5986 : }
5987 82660 : rel->attr_widths[te->resno - rel->min_attr] = item_width;
5988 : }
5989 :
5990 : /* Now estimate number of output rows, etc */
5991 27504 : set_baserel_size_estimates(root, rel);
5992 27504 : }
5993 :
5994 : /*
5995 : * set_function_size_estimates
5996 : * Set the size estimates for a base relation that is a function call.
5997 : *
5998 : * The rel's targetlist and restrictinfo list must have been constructed
5999 : * already.
6000 : *
6001 : * We set the same fields as set_baserel_size_estimates.
6002 : */
6003 : void
6004 51118 : set_function_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6005 : {
6006 : RangeTblEntry *rte;
6007 : ListCell *lc;
6008 :
6009 : /* Should only be applied to base relations that are functions */
6010 : Assert(rel->relid > 0);
6011 51118 : rte = planner_rt_fetch(rel->relid, root);
6012 : Assert(rte->rtekind == RTE_FUNCTION);
6013 :
6014 : /*
6015 : * Estimate number of rows the functions will return. The rowcount of the
6016 : * node is that of the largest function result.
6017 : */
6018 51118 : rel->tuples = 0;
6019 102734 : foreach(lc, rte->functions)
6020 : {
6021 51616 : RangeTblFunction *rtfunc = (RangeTblFunction *) lfirst(lc);
6022 51616 : double ntup = expression_returns_set_rows(root, rtfunc->funcexpr);
6023 :
6024 51616 : if (ntup > rel->tuples)
6025 51142 : rel->tuples = ntup;
6026 : }
6027 :
6028 : /* Now estimate number of output rows, etc */
6029 51118 : set_baserel_size_estimates(root, rel);
6030 51118 : }
6031 :
6032 : /*
6033 : * set_function_size_estimates
6034 : * Set the size estimates for a base relation that is a function call.
6035 : *
6036 : * The rel's targetlist and restrictinfo list must have been constructed
6037 : * already.
6038 : *
6039 : * We set the same fields as set_tablefunc_size_estimates.
6040 : */
6041 : void
6042 626 : set_tablefunc_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6043 : {
6044 : /* Should only be applied to base relations that are functions */
6045 : Assert(rel->relid > 0);
6046 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_TABLEFUNC);
6047 :
6048 626 : rel->tuples = 100;
6049 :
6050 : /* Now estimate number of output rows, etc */
6051 626 : set_baserel_size_estimates(root, rel);
6052 626 : }
6053 :
6054 : /*
6055 : * set_values_size_estimates
6056 : * Set the size estimates for a base relation that is a values list.
6057 : *
6058 : * The rel's targetlist and restrictinfo list must have been constructed
6059 : * already.
6060 : *
6061 : * We set the same fields as set_baserel_size_estimates.
6062 : */
6063 : void
6064 8218 : set_values_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6065 : {
6066 : RangeTblEntry *rte;
6067 :
6068 : /* Should only be applied to base relations that are values lists */
6069 : Assert(rel->relid > 0);
6070 8218 : rte = planner_rt_fetch(rel->relid, root);
6071 : Assert(rte->rtekind == RTE_VALUES);
6072 :
6073 : /*
6074 : * Estimate number of rows the values list will return. We know this
6075 : * precisely based on the list length (well, barring set-returning
6076 : * functions in list items, but that's a refinement not catered for
6077 : * anywhere else either).
6078 : */
6079 8218 : rel->tuples = list_length(rte->values_lists);
6080 :
6081 : /* Now estimate number of output rows, etc */
6082 8218 : set_baserel_size_estimates(root, rel);
6083 8218 : }
6084 :
6085 : /*
6086 : * set_cte_size_estimates
6087 : * Set the size estimates for a base relation that is a CTE reference.
6088 : *
6089 : * The rel's targetlist and restrictinfo list must have been constructed
6090 : * already, and we need an estimate of the number of rows returned by the CTE
6091 : * (if a regular CTE) or the non-recursive term (if a self-reference).
6092 : *
6093 : * We set the same fields as set_baserel_size_estimates.
6094 : */
6095 : void
6096 5176 : set_cte_size_estimates(PlannerInfo *root, RelOptInfo *rel, double cte_rows)
6097 : {
6098 : RangeTblEntry *rte;
6099 :
6100 : /* Should only be applied to base relations that are CTE references */
6101 : Assert(rel->relid > 0);
6102 5176 : rte = planner_rt_fetch(rel->relid, root);
6103 : Assert(rte->rtekind == RTE_CTE);
6104 :
6105 5176 : if (rte->self_reference)
6106 : {
6107 : /*
6108 : * In a self-reference, we assume the average worktable size is a
6109 : * multiple of the nonrecursive term's size. The best multiplier will
6110 : * vary depending on query "fan-out", so make its value adjustable.
6111 : */
6112 926 : rel->tuples = clamp_row_est(recursive_worktable_factor * cte_rows);
6113 : }
6114 : else
6115 : {
6116 : /* Otherwise just believe the CTE's rowcount estimate */
6117 4250 : rel->tuples = cte_rows;
6118 : }
6119 :
6120 : /* Now estimate number of output rows, etc */
6121 5176 : set_baserel_size_estimates(root, rel);
6122 5176 : }
6123 :
6124 : /*
6125 : * set_namedtuplestore_size_estimates
6126 : * Set the size estimates for a base relation that is a tuplestore reference.
6127 : *
6128 : * The rel's targetlist and restrictinfo list must have been constructed
6129 : * already.
6130 : *
6131 : * We set the same fields as set_baserel_size_estimates.
6132 : */
6133 : void
6134 478 : set_namedtuplestore_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6135 : {
6136 : RangeTblEntry *rte;
6137 :
6138 : /* Should only be applied to base relations that are tuplestore references */
6139 : Assert(rel->relid > 0);
6140 478 : rte = planner_rt_fetch(rel->relid, root);
6141 : Assert(rte->rtekind == RTE_NAMEDTUPLESTORE);
6142 :
6143 : /*
6144 : * Use the estimate provided by the code which is generating the named
6145 : * tuplestore. In some cases, the actual number might be available; in
6146 : * others the same plan will be re-used, so a "typical" value might be
6147 : * estimated and used.
6148 : */
6149 478 : rel->tuples = rte->enrtuples;
6150 478 : if (rel->tuples < 0)
6151 0 : rel->tuples = 1000;
6152 :
6153 : /* Now estimate number of output rows, etc */
6154 478 : set_baserel_size_estimates(root, rel);
6155 478 : }
6156 :
6157 : /*
6158 : * set_result_size_estimates
6159 : * Set the size estimates for an RTE_RESULT base relation
6160 : *
6161 : * The rel's targetlist and restrictinfo list must have been constructed
6162 : * already.
6163 : *
6164 : * We set the same fields as set_baserel_size_estimates.
6165 : */
6166 : void
6167 4208 : set_result_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6168 : {
6169 : /* Should only be applied to RTE_RESULT base relations */
6170 : Assert(rel->relid > 0);
6171 : Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_RESULT);
6172 :
6173 : /* RTE_RESULT always generates a single row, natively */
6174 4208 : rel->tuples = 1;
6175 :
6176 : /* Now estimate number of output rows, etc */
6177 4208 : set_baserel_size_estimates(root, rel);
6178 4208 : }
6179 :
6180 : /*
6181 : * set_foreign_size_estimates
6182 : * Set the size estimates for a base relation that is a foreign table.
6183 : *
6184 : * There is not a whole lot that we can do here; the foreign-data wrapper
6185 : * is responsible for producing useful estimates. We can do a decent job
6186 : * of estimating baserestrictcost, so we set that, and we also set up width
6187 : * using what will be purely datatype-driven estimates from the targetlist.
6188 : * There is no way to do anything sane with the rows value, so we just put
6189 : * a default estimate and hope that the wrapper can improve on it. The
6190 : * wrapper's GetForeignRelSize function will be called momentarily.
6191 : *
6192 : * The rel's targetlist and restrictinfo list must have been constructed
6193 : * already.
6194 : */
6195 : void
6196 2440 : set_foreign_size_estimates(PlannerInfo *root, RelOptInfo *rel)
6197 : {
6198 : /* Should only be applied to base relations */
6199 : Assert(rel->relid > 0);
6200 :
6201 2440 : rel->rows = 1000; /* entirely bogus default estimate */
6202 :
6203 2440 : cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo, root);
6204 :
6205 2440 : set_rel_width(root, rel);
6206 2440 : }
6207 :
6208 :
6209 : /*
6210 : * set_rel_width
6211 : * Set the estimated output width of a base relation.
6212 : *
6213 : * The estimated output width is the sum of the per-attribute width estimates
6214 : * for the actually-referenced columns, plus any PHVs or other expressions
6215 : * that have to be calculated at this relation. This is the amount of data
6216 : * we'd need to pass upwards in case of a sort, hash, etc.
6217 : *
6218 : * This function also sets reltarget->cost, so it's a bit misnamed now.
6219 : *
6220 : * NB: this works best on plain relations because it prefers to look at
6221 : * real Vars. For subqueries, set_subquery_size_estimates will already have
6222 : * copied up whatever per-column estimates were made within the subquery,
6223 : * and for other types of rels there isn't much we can do anyway. We fall
6224 : * back on (fairly stupid) datatype-based width estimates if we can't get
6225 : * any better number.
6226 : *
6227 : * The per-attribute width estimates are cached for possible re-use while
6228 : * building join relations or post-scan/join pathtargets.
6229 : */
6230 : static void
6231 491182 : set_rel_width(PlannerInfo *root, RelOptInfo *rel)
6232 : {
6233 491182 : Oid reloid = planner_rt_fetch(rel->relid, root)->relid;
6234 491182 : int64 tuple_width = 0;
6235 491182 : bool have_wholerow_var = false;
6236 : ListCell *lc;
6237 :
6238 : /* Vars are assumed to have cost zero, but other exprs do not */
6239 491182 : rel->reltarget->cost.startup = 0;
6240 491182 : rel->reltarget->cost.per_tuple = 0;
6241 :
6242 1754044 : foreach(lc, rel->reltarget->exprs)
6243 : {
6244 1262862 : Node *node = (Node *) lfirst(lc);
6245 :
6246 : /*
6247 : * Ordinarily, a Var in a rel's targetlist must belong to that rel;
6248 : * but there are corner cases involving LATERAL references where that
6249 : * isn't so. If the Var has the wrong varno, fall through to the
6250 : * generic case (it doesn't seem worth the trouble to be any smarter).
6251 : */
6252 1262862 : if (IsA(node, Var) &&
6253 1238896 : ((Var *) node)->varno == rel->relid)
6254 321186 : {
6255 1238830 : Var *var = (Var *) node;
6256 : int ndx;
6257 : int32 item_width;
6258 :
6259 : Assert(var->varattno >= rel->min_attr);
6260 : Assert(var->varattno <= rel->max_attr);
6261 :
6262 1238830 : ndx = var->varattno - rel->min_attr;
6263 :
6264 : /*
6265 : * If it's a whole-row Var, we'll deal with it below after we have
6266 : * already cached as many attr widths as possible.
6267 : */
6268 1238830 : if (var->varattno == 0)
6269 : {
6270 3018 : have_wholerow_var = true;
6271 3018 : continue;
6272 : }
6273 :
6274 : /*
6275 : * The width may have been cached already (especially if it's a
6276 : * subquery), so don't duplicate effort.
6277 : */
6278 1235812 : if (rel->attr_widths[ndx] > 0)
6279 : {
6280 227208 : tuple_width += rel->attr_widths[ndx];
6281 227208 : continue;
6282 : }
6283 :
6284 : /* Try to get column width from statistics */
6285 1008604 : if (reloid != InvalidOid && var->varattno > 0)
6286 : {
6287 797494 : item_width = get_attavgwidth(reloid, var->varattno);
6288 797494 : if (item_width > 0)
6289 : {
6290 687418 : rel->attr_widths[ndx] = item_width;
6291 687418 : tuple_width += item_width;
6292 687418 : continue;
6293 : }
6294 : }
6295 :
6296 : /*
6297 : * Not a plain relation, or can't find statistics for it. Estimate
6298 : * using just the type info.
6299 : */
6300 321186 : item_width = get_typavgwidth(var->vartype, var->vartypmod);
6301 : Assert(item_width > 0);
6302 321186 : rel->attr_widths[ndx] = item_width;
6303 321186 : tuple_width += item_width;
6304 : }
6305 24032 : else if (IsA(node, PlaceHolderVar))
6306 : {
6307 : /*
6308 : * We will need to evaluate the PHV's contained expression while
6309 : * scanning this rel, so be sure to include it in reltarget->cost.
6310 : */
6311 1978 : PlaceHolderVar *phv = (PlaceHolderVar *) node;
6312 1978 : PlaceHolderInfo *phinfo = find_placeholder_info(root, phv);
6313 : QualCost cost;
6314 :
6315 1978 : tuple_width += phinfo->ph_width;
6316 1978 : cost_qual_eval_node(&cost, (Node *) phv->phexpr, root);
6317 1978 : rel->reltarget->cost.startup += cost.startup;
6318 1978 : rel->reltarget->cost.per_tuple += cost.per_tuple;
6319 : }
6320 : else
6321 : {
6322 : /*
6323 : * We could be looking at an expression pulled up from a subquery,
6324 : * or a ROW() representing a whole-row child Var, etc. Do what we
6325 : * can using the expression type information.
6326 : */
6327 : int32 item_width;
6328 : QualCost cost;
6329 :
6330 22054 : item_width = get_typavgwidth(exprType(node), exprTypmod(node));
6331 : Assert(item_width > 0);
6332 22054 : tuple_width += item_width;
6333 : /* Not entirely clear if we need to account for cost, but do so */
6334 22054 : cost_qual_eval_node(&cost, node, root);
6335 22054 : rel->reltarget->cost.startup += cost.startup;
6336 22054 : rel->reltarget->cost.per_tuple += cost.per_tuple;
6337 : }
6338 : }
6339 :
6340 : /*
6341 : * If we have a whole-row reference, estimate its width as the sum of
6342 : * per-column widths plus heap tuple header overhead.
6343 : */
6344 491182 : if (have_wholerow_var)
6345 : {
6346 3018 : int64 wholerow_width = MAXALIGN(SizeofHeapTupleHeader);
6347 :
6348 3018 : if (reloid != InvalidOid)
6349 : {
6350 : /* Real relation, so estimate true tuple width */
6351 2360 : wholerow_width += get_relation_data_width(reloid,
6352 2360 : rel->attr_widths - rel->min_attr);
6353 : }
6354 : else
6355 : {
6356 : /* Do what we can with info for a phony rel */
6357 : AttrNumber i;
6358 :
6359 1794 : for (i = 1; i <= rel->max_attr; i++)
6360 1136 : wholerow_width += rel->attr_widths[i - rel->min_attr];
6361 : }
6362 :
6363 3018 : rel->attr_widths[0 - rel->min_attr] = clamp_width_est(wholerow_width);
6364 :
6365 : /*
6366 : * Include the whole-row Var as part of the output tuple. Yes, that
6367 : * really is what happens at runtime.
6368 : */
6369 3018 : tuple_width += wholerow_width;
6370 : }
6371 :
6372 491182 : rel->reltarget->width = clamp_width_est(tuple_width);
6373 491182 : }
6374 :
6375 : /*
6376 : * set_pathtarget_cost_width
6377 : * Set the estimated eval cost and output width of a PathTarget tlist.
6378 : *
6379 : * As a notational convenience, returns the same PathTarget pointer passed in.
6380 : *
6381 : * Most, though not quite all, uses of this function occur after we've run
6382 : * set_rel_width() for base relations; so we can usually obtain cached width
6383 : * estimates for Vars. If we can't, fall back on datatype-based width
6384 : * estimates. Present early-planning uses of PathTargets don't need accurate
6385 : * widths badly enough to justify going to the catalogs for better data.
6386 : */
6387 : PathTarget *
6388 593656 : set_pathtarget_cost_width(PlannerInfo *root, PathTarget *target)
6389 : {
6390 593656 : int64 tuple_width = 0;
6391 : ListCell *lc;
6392 :
6393 : /* Vars are assumed to have cost zero, but other exprs do not */
6394 593656 : target->cost.startup = 0;
6395 593656 : target->cost.per_tuple = 0;
6396 :
6397 2039952 : foreach(lc, target->exprs)
6398 : {
6399 1446296 : Node *node = (Node *) lfirst(lc);
6400 :
6401 1446296 : tuple_width += get_expr_width(root, node);
6402 :
6403 : /* For non-Vars, account for evaluation cost */
6404 1446296 : if (!IsA(node, Var))
6405 : {
6406 : QualCost cost;
6407 :
6408 614310 : cost_qual_eval_node(&cost, node, root);
6409 614310 : target->cost.startup += cost.startup;
6410 614310 : target->cost.per_tuple += cost.per_tuple;
6411 : }
6412 : }
6413 :
6414 593656 : target->width = clamp_width_est(tuple_width);
6415 :
6416 593656 : return target;
6417 : }
6418 :
6419 : /*
6420 : * get_expr_width
6421 : * Estimate the width of the given expr attempting to use the width
6422 : * cached in a Var's owning RelOptInfo, else fallback on the type's
6423 : * average width when unable to or when the given Node is not a Var.
6424 : */
6425 : static int32
6426 1758048 : get_expr_width(PlannerInfo *root, const Node *expr)
6427 : {
6428 : int32 width;
6429 :
6430 1758048 : if (IsA(expr, Var))
6431 : {
6432 1131096 : const Var *var = (const Var *) expr;
6433 :
6434 : /* We should not see any upper-level Vars here */
6435 : Assert(var->varlevelsup == 0);
6436 :
6437 : /* Try to get data from RelOptInfo cache */
6438 1131096 : if (!IS_SPECIAL_VARNO(var->varno) &&
6439 1125326 : var->varno < root->simple_rel_array_size)
6440 : {
6441 1125326 : RelOptInfo *rel = root->simple_rel_array[var->varno];
6442 :
6443 1125326 : if (rel != NULL &&
6444 1097238 : var->varattno >= rel->min_attr &&
6445 1097238 : var->varattno <= rel->max_attr)
6446 : {
6447 1097238 : int ndx = var->varattno - rel->min_attr;
6448 :
6449 1097238 : if (rel->attr_widths[ndx] > 0)
6450 1065716 : return rel->attr_widths[ndx];
6451 : }
6452 : }
6453 :
6454 : /*
6455 : * No cached data available, so estimate using just the type info.
6456 : */
6457 65380 : width = get_typavgwidth(var->vartype, var->vartypmod);
6458 : Assert(width > 0);
6459 :
6460 65380 : return width;
6461 : }
6462 :
6463 626952 : width = get_typavgwidth(exprType(expr), exprTypmod(expr));
6464 : Assert(width > 0);
6465 626952 : return width;
6466 : }
6467 :
6468 : /*
6469 : * relation_byte_size
6470 : * Estimate the storage space in bytes for a given number of tuples
6471 : * of a given width (size in bytes).
6472 : */
6473 : static double
6474 3849486 : relation_byte_size(double tuples, int width)
6475 : {
6476 3849486 : return tuples * (MAXALIGN(width) + MAXALIGN(SizeofHeapTupleHeader));
6477 : }
6478 :
6479 : /*
6480 : * page_size
6481 : * Returns an estimate of the number of pages covered by a given
6482 : * number of tuples of a given width (size in bytes).
6483 : */
6484 : static double
6485 9348 : page_size(double tuples, int width)
6486 : {
6487 9348 : return ceil(relation_byte_size(tuples, width) / BLCKSZ);
6488 : }
6489 :
6490 : /*
6491 : * Estimate the fraction of the work that each worker will do given the
6492 : * number of workers budgeted for the path.
6493 : */
6494 : static double
6495 189176 : get_parallel_divisor(Path *path)
6496 : {
6497 189176 : double parallel_divisor = path->parallel_workers;
6498 :
6499 : /*
6500 : * Early experience with parallel query suggests that when there is only
6501 : * one worker, the leader often makes a very substantial contribution to
6502 : * executing the parallel portion of the plan, but as more workers are
6503 : * added, it does less and less, because it's busy reading tuples from the
6504 : * workers and doing whatever non-parallel post-processing is needed. By
6505 : * the time we reach 4 workers, the leader no longer makes a meaningful
6506 : * contribution. Thus, for now, estimate that the leader spends 30% of
6507 : * its time servicing each worker, and the remainder executing the
6508 : * parallel plan.
6509 : */
6510 189176 : if (parallel_leader_participation)
6511 : {
6512 : double leader_contribution;
6513 :
6514 187874 : leader_contribution = 1.0 - (0.3 * path->parallel_workers);
6515 187874 : if (leader_contribution > 0)
6516 185558 : parallel_divisor += leader_contribution;
6517 : }
6518 :
6519 189176 : return parallel_divisor;
6520 : }
6521 :
6522 : /*
6523 : * compute_bitmap_pages
6524 : * Estimate number of pages fetched from heap in a bitmap heap scan.
6525 : *
6526 : * 'baserel' is the relation to be scanned
6527 : * 'bitmapqual' is a tree of IndexPaths, BitmapAndPaths, and BitmapOrPaths
6528 : * 'loop_count' is the number of repetitions of the indexscan to factor into
6529 : * estimates of caching behavior
6530 : *
6531 : * If cost_p isn't NULL, the indexTotalCost estimate is returned in *cost_p.
6532 : * If tuples_p isn't NULL, the tuples_fetched estimate is returned in *tuples_p.
6533 : */
6534 : double
6535 657080 : compute_bitmap_pages(PlannerInfo *root, RelOptInfo *baserel,
6536 : Path *bitmapqual, double loop_count,
6537 : Cost *cost_p, double *tuples_p)
6538 : {
6539 : Cost indexTotalCost;
6540 : Selectivity indexSelectivity;
6541 : double T;
6542 : double pages_fetched;
6543 : double tuples_fetched;
6544 : double heap_pages;
6545 : double maxentries;
6546 :
6547 : /*
6548 : * Fetch total cost of obtaining the bitmap, as well as its total
6549 : * selectivity.
6550 : */
6551 657080 : cost_bitmap_tree_node(bitmapqual, &indexTotalCost, &indexSelectivity);
6552 :
6553 : /*
6554 : * Estimate number of main-table pages fetched.
6555 : */
6556 657080 : tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
6557 :
6558 657080 : T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
6559 :
6560 : /*
6561 : * For a single scan, the number of heap pages that need to be fetched is
6562 : * the same as the Mackert and Lohman formula for the case T <= b (ie, no
6563 : * re-reads needed).
6564 : */
6565 657080 : pages_fetched = (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
6566 :
6567 : /*
6568 : * Calculate the number of pages fetched from the heap. Then based on
6569 : * current work_mem estimate get the estimated maxentries in the bitmap.
6570 : * (Note that we always do this calculation based on the number of pages
6571 : * that would be fetched in a single iteration, even if loop_count > 1.
6572 : * That's correct, because only that number of entries will be stored in
6573 : * the bitmap at one time.)
6574 : */
6575 657080 : heap_pages = Min(pages_fetched, baserel->pages);
6576 657080 : maxentries = tbm_calculate_entries(work_mem * (Size) 1024);
6577 :
6578 657080 : if (loop_count > 1)
6579 : {
6580 : /*
6581 : * For repeated bitmap scans, scale up the number of tuples fetched in
6582 : * the Mackert and Lohman formula by the number of scans, so that we
6583 : * estimate the number of pages fetched by all the scans. Then
6584 : * pro-rate for one scan.
6585 : */
6586 133444 : pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
6587 : baserel->pages,
6588 : get_indexpath_pages(bitmapqual),
6589 : root);
6590 133444 : pages_fetched /= loop_count;
6591 : }
6592 :
6593 657080 : if (pages_fetched >= T)
6594 65558 : pages_fetched = T;
6595 : else
6596 591522 : pages_fetched = ceil(pages_fetched);
6597 :
6598 657080 : if (maxentries < heap_pages)
6599 : {
6600 : double exact_pages;
6601 : double lossy_pages;
6602 :
6603 : /*
6604 : * Crude approximation of the number of lossy pages. Because of the
6605 : * way tbm_lossify() is coded, the number of lossy pages increases
6606 : * very sharply as soon as we run short of memory; this formula has
6607 : * that property and seems to perform adequately in testing, but it's
6608 : * possible we could do better somehow.
6609 : */
6610 18 : lossy_pages = Max(0, heap_pages - maxentries / 2);
6611 18 : exact_pages = heap_pages - lossy_pages;
6612 :
6613 : /*
6614 : * If there are lossy pages then recompute the number of tuples
6615 : * processed by the bitmap heap node. We assume here that the chance
6616 : * of a given tuple coming from an exact page is the same as the
6617 : * chance that a given page is exact. This might not be true, but
6618 : * it's not clear how we can do any better.
6619 : */
6620 18 : if (lossy_pages > 0)
6621 : tuples_fetched =
6622 18 : clamp_row_est(indexSelectivity *
6623 18 : (exact_pages / heap_pages) * baserel->tuples +
6624 18 : (lossy_pages / heap_pages) * baserel->tuples);
6625 : }
6626 :
6627 657080 : if (cost_p)
6628 516524 : *cost_p = indexTotalCost;
6629 657080 : if (tuples_p)
6630 516524 : *tuples_p = tuples_fetched;
6631 :
6632 657080 : return pages_fetched;
6633 : }
6634 :
6635 : /*
6636 : * compute_gather_rows
6637 : * Estimate number of rows for gather (merge) nodes.
6638 : *
6639 : * In a parallel plan, each worker's row estimate is determined by dividing the
6640 : * total number of rows by parallel_divisor, which accounts for the leader's
6641 : * contribution in addition to the number of workers. Accordingly, when
6642 : * estimating the number of rows for gather (merge) nodes, we multiply the rows
6643 : * per worker by the same parallel_divisor to undo the division.
6644 : */
6645 : double
6646 37038 : compute_gather_rows(Path *path)
6647 : {
6648 : Assert(path->parallel_workers > 0);
6649 :
6650 37038 : return clamp_row_est(path->rows * get_parallel_divisor(path));
6651 : }
|
