Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * analyze.c
4 : * the Postgres statistics generator
5 : *
6 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/commands/analyze.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 : #include "postgres.h"
16 :
17 : #include <math.h>
18 :
19 : #include "access/detoast.h"
20 : #include "access/genam.h"
21 : #include "access/multixact.h"
22 : #include "access/relation.h"
23 : #include "access/table.h"
24 : #include "access/tableam.h"
25 : #include "access/transam.h"
26 : #include "access/tupconvert.h"
27 : #include "access/visibilitymap.h"
28 : #include "access/xact.h"
29 : #include "catalog/index.h"
30 : #include "catalog/indexing.h"
31 : #include "catalog/pg_inherits.h"
32 : #include "commands/progress.h"
33 : #include "commands/tablecmds.h"
34 : #include "commands/vacuum.h"
35 : #include "common/pg_prng.h"
36 : #include "executor/executor.h"
37 : #include "foreign/fdwapi.h"
38 : #include "miscadmin.h"
39 : #include "nodes/nodeFuncs.h"
40 : #include "parser/parse_oper.h"
41 : #include "parser/parse_relation.h"
42 : #include "pgstat.h"
43 : #include "statistics/extended_stats_internal.h"
44 : #include "statistics/statistics.h"
45 : #include "storage/bufmgr.h"
46 : #include "storage/procarray.h"
47 : #include "utils/attoptcache.h"
48 : #include "utils/datum.h"
49 : #include "utils/guc.h"
50 : #include "utils/lsyscache.h"
51 : #include "utils/memutils.h"
52 : #include "utils/pg_rusage.h"
53 : #include "utils/sampling.h"
54 : #include "utils/sortsupport.h"
55 : #include "utils/syscache.h"
56 : #include "utils/timestamp.h"
57 :
58 :
59 : /* Per-index data for ANALYZE */
60 : typedef struct AnlIndexData
61 : {
62 : IndexInfo *indexInfo; /* BuildIndexInfo result */
63 : double tupleFract; /* fraction of rows for partial index */
64 : VacAttrStats **vacattrstats; /* index attrs to analyze */
65 : int attr_cnt;
66 : } AnlIndexData;
67 :
68 :
69 : /* Default statistics target (GUC parameter) */
70 : int default_statistics_target = 100;
71 :
72 : /* A few variables that don't seem worth passing around as parameters */
73 : static MemoryContext anl_context = NULL;
74 : static BufferAccessStrategy vac_strategy;
75 :
76 :
77 : static void do_analyze_rel(Relation onerel,
78 : const VacuumParams params, List *va_cols,
79 : AcquireSampleRowsFunc acquirefunc, BlockNumber relpages,
80 : bool inh, bool in_outer_xact, int elevel);
81 : static void compute_index_stats(Relation onerel, double totalrows,
82 : AnlIndexData *indexdata, int nindexes,
83 : HeapTuple *rows, int numrows,
84 : MemoryContext col_context);
85 : static VacAttrStats *examine_attribute(Relation onerel, int attnum,
86 : Node *index_expr);
87 : static int acquire_sample_rows(Relation onerel, int elevel,
88 : HeapTuple *rows, int targrows,
89 : double *totalrows, double *totaldeadrows);
90 : static int compare_rows(const void *a, const void *b, void *arg);
91 : static int acquire_inherited_sample_rows(Relation onerel, int elevel,
92 : HeapTuple *rows, int targrows,
93 : double *totalrows, double *totaldeadrows);
94 : static void update_attstats(Oid relid, bool inh,
95 : int natts, VacAttrStats **vacattrstats);
96 : static Datum std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
97 : static Datum ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
98 :
99 :
100 : /*
101 : * analyze_rel() -- analyze one relation
102 : *
103 : * relid identifies the relation to analyze. If relation is supplied, use
104 : * the name therein for reporting any failure to open/lock the rel; do not
105 : * use it once we've successfully opened the rel, since it might be stale.
106 : */
107 : void
108 16638 : analyze_rel(Oid relid, RangeVar *relation,
109 : const VacuumParams params, List *va_cols, bool in_outer_xact,
110 : BufferAccessStrategy bstrategy)
111 : {
112 : Relation onerel;
113 : int elevel;
114 16638 : AcquireSampleRowsFunc acquirefunc = NULL;
115 16638 : BlockNumber relpages = 0;
116 :
117 : /* Select logging level */
118 16638 : if (params.options & VACOPT_VERBOSE)
119 0 : elevel = INFO;
120 : else
121 16638 : elevel = DEBUG2;
122 :
123 : /* Set up static variables */
124 16638 : vac_strategy = bstrategy;
125 :
126 : /*
127 : * Check for user-requested abort.
128 : */
129 16638 : CHECK_FOR_INTERRUPTS();
130 :
131 : /*
132 : * Open the relation, getting ShareUpdateExclusiveLock to ensure that two
133 : * ANALYZEs don't run on it concurrently. (This also locks out a
134 : * concurrent VACUUM, which doesn't matter much at the moment but might
135 : * matter if we ever try to accumulate stats on dead tuples.) If the rel
136 : * has been dropped since we last saw it, we don't need to process it.
137 : *
138 : * Make sure to generate only logs for ANALYZE in this case.
139 : */
140 16638 : onerel = vacuum_open_relation(relid, relation, params.options & ~(VACOPT_VACUUM),
141 16638 : params.log_analyze_min_duration >= 0,
142 : ShareUpdateExclusiveLock);
143 :
144 : /* leave if relation could not be opened or locked */
145 16638 : if (!onerel)
146 216 : return;
147 :
148 : /*
149 : * Check if relation needs to be skipped based on privileges. This check
150 : * happens also when building the relation list to analyze for a manual
151 : * operation, and needs to be done additionally here as ANALYZE could
152 : * happen across multiple transactions where privileges could have changed
153 : * in-between. Make sure to generate only logs for ANALYZE in this case.
154 : */
155 16626 : if (!vacuum_is_permitted_for_relation(RelationGetRelid(onerel),
156 : onerel->rd_rel,
157 16626 : params.options & ~VACOPT_VACUUM))
158 : {
159 36 : relation_close(onerel, ShareUpdateExclusiveLock);
160 36 : return;
161 : }
162 :
163 : /*
164 : * Silently ignore tables that are temp tables of other backends ---
165 : * trying to analyze these is rather pointless, since their contents are
166 : * probably not up-to-date on disk. (We don't throw a warning here; it
167 : * would just lead to chatter during a database-wide ANALYZE.)
168 : */
169 16590 : if (RELATION_IS_OTHER_TEMP(onerel))
170 : {
171 0 : relation_close(onerel, ShareUpdateExclusiveLock);
172 0 : return;
173 : }
174 :
175 : /*
176 : * We can ANALYZE any table except pg_statistic. See update_attstats
177 : */
178 16590 : if (RelationGetRelid(onerel) == StatisticRelationId)
179 : {
180 168 : relation_close(onerel, ShareUpdateExclusiveLock);
181 168 : return;
182 : }
183 :
184 : /*
185 : * Check that it's of an analyzable relkind, and set up appropriately.
186 : */
187 16422 : if (onerel->rd_rel->relkind == RELKIND_RELATION ||
188 836 : onerel->rd_rel->relkind == RELKIND_MATVIEW)
189 : {
190 : /* Regular table, so we'll use the regular row acquisition function */
191 15588 : acquirefunc = acquire_sample_rows;
192 : /* Also get regular table's size */
193 15588 : relpages = RelationGetNumberOfBlocks(onerel);
194 : }
195 834 : else if (onerel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
196 : {
197 : /*
198 : * For a foreign table, call the FDW's hook function to see whether it
199 : * supports analysis.
200 : */
201 : FdwRoutine *fdwroutine;
202 66 : bool ok = false;
203 :
204 66 : fdwroutine = GetFdwRoutineForRelation(onerel, false);
205 :
206 66 : if (fdwroutine->AnalyzeForeignTable != NULL)
207 66 : ok = fdwroutine->AnalyzeForeignTable(onerel,
208 : &acquirefunc,
209 : &relpages);
210 :
211 66 : if (!ok)
212 : {
213 0 : ereport(WARNING,
214 : (errmsg("skipping \"%s\" --- cannot analyze this foreign table",
215 : RelationGetRelationName(onerel))));
216 0 : relation_close(onerel, ShareUpdateExclusiveLock);
217 0 : return;
218 : }
219 : }
220 768 : else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
221 : {
222 : /*
223 : * For partitioned tables, we want to do the recursive ANALYZE below.
224 : */
225 : }
226 : else
227 : {
228 : /* No need for a WARNING if we already complained during VACUUM */
229 0 : if (!(params.options & VACOPT_VACUUM))
230 0 : ereport(WARNING,
231 : (errmsg("skipping \"%s\" --- cannot analyze non-tables or special system tables",
232 : RelationGetRelationName(onerel))));
233 0 : relation_close(onerel, ShareUpdateExclusiveLock);
234 0 : return;
235 : }
236 :
237 : /*
238 : * OK, let's do it. First, initialize progress reporting.
239 : */
240 16422 : pgstat_progress_start_command(PROGRESS_COMMAND_ANALYZE,
241 : RelationGetRelid(onerel));
242 :
243 : /*
244 : * Do the normal non-recursive ANALYZE. We can skip this for partitioned
245 : * tables, which don't contain any rows.
246 : */
247 16422 : if (onerel->rd_rel->relkind != RELKIND_PARTITIONED_TABLE)
248 15654 : do_analyze_rel(onerel, params, va_cols, acquirefunc,
249 : relpages, false, in_outer_xact, elevel);
250 :
251 : /*
252 : * If there are child tables, do recursive ANALYZE.
253 : */
254 16382 : if (onerel->rd_rel->relhassubclass)
255 896 : do_analyze_rel(onerel, params, va_cols, acquirefunc, relpages,
256 : true, in_outer_xact, elevel);
257 :
258 : /*
259 : * Close source relation now, but keep lock so that no one deletes it
260 : * before we commit. (If someone did, they'd fail to clean up the entries
261 : * we made in pg_statistic. Also, releasing the lock before commit would
262 : * expose us to concurrent-update failures in update_attstats.)
263 : */
264 16364 : relation_close(onerel, NoLock);
265 :
266 16364 : pgstat_progress_end_command();
267 : }
268 :
269 : /*
270 : * do_analyze_rel() -- analyze one relation, recursively or not
271 : *
272 : * Note that "acquirefunc" is only relevant for the non-inherited case.
273 : * For the inherited case, acquire_inherited_sample_rows() determines the
274 : * appropriate acquirefunc for each child table.
275 : */
276 : static void
277 16550 : do_analyze_rel(Relation onerel, const VacuumParams params,
278 : List *va_cols, AcquireSampleRowsFunc acquirefunc,
279 : BlockNumber relpages, bool inh, bool in_outer_xact,
280 : int elevel)
281 : {
282 : int attr_cnt,
283 : tcnt,
284 : i,
285 : ind;
286 : Relation *Irel;
287 : int nindexes;
288 : bool verbose,
289 : instrument,
290 : hasindex;
291 : VacAttrStats **vacattrstats;
292 : AnlIndexData *indexdata;
293 : int targrows,
294 : numrows,
295 : minrows;
296 : double totalrows,
297 : totaldeadrows;
298 : HeapTuple *rows;
299 : PGRUsage ru0;
300 16550 : TimestampTz starttime = 0;
301 : MemoryContext caller_context;
302 : Oid save_userid;
303 : int save_sec_context;
304 : int save_nestlevel;
305 16550 : WalUsage startwalusage = pgWalUsage;
306 16550 : BufferUsage startbufferusage = pgBufferUsage;
307 : BufferUsage bufferusage;
308 16550 : PgStat_Counter startreadtime = 0;
309 16550 : PgStat_Counter startwritetime = 0;
310 :
311 16550 : verbose = (params.options & VACOPT_VERBOSE) != 0;
312 17196 : instrument = (verbose || (AmAutoVacuumWorkerProcess() &&
313 646 : params.log_analyze_min_duration >= 0));
314 16550 : if (inh)
315 896 : ereport(elevel,
316 : (errmsg("analyzing \"%s.%s\" inheritance tree",
317 : get_namespace_name(RelationGetNamespace(onerel)),
318 : RelationGetRelationName(onerel))));
319 : else
320 15654 : ereport(elevel,
321 : (errmsg("analyzing \"%s.%s\"",
322 : get_namespace_name(RelationGetNamespace(onerel)),
323 : RelationGetRelationName(onerel))));
324 :
325 : /*
326 : * Set up a working context so that we can easily free whatever junk gets
327 : * created.
328 : */
329 16550 : anl_context = AllocSetContextCreate(CurrentMemoryContext,
330 : "Analyze",
331 : ALLOCSET_DEFAULT_SIZES);
332 16550 : caller_context = MemoryContextSwitchTo(anl_context);
333 :
334 : /*
335 : * Switch to the table owner's userid, so that any index functions are run
336 : * as that user. Also lock down security-restricted operations and
337 : * arrange to make GUC variable changes local to this command.
338 : */
339 16550 : GetUserIdAndSecContext(&save_userid, &save_sec_context);
340 16550 : SetUserIdAndSecContext(onerel->rd_rel->relowner,
341 : save_sec_context | SECURITY_RESTRICTED_OPERATION);
342 16550 : save_nestlevel = NewGUCNestLevel();
343 16550 : RestrictSearchPath();
344 :
345 : /*
346 : * When verbose or autovacuum logging is used, initialize a resource usage
347 : * snapshot and optionally track I/O timing.
348 : */
349 16550 : if (instrument)
350 : {
351 646 : if (track_io_timing)
352 : {
353 0 : startreadtime = pgStatBlockReadTime;
354 0 : startwritetime = pgStatBlockWriteTime;
355 : }
356 :
357 646 : pg_rusage_init(&ru0);
358 : }
359 :
360 : /* Used for instrumentation and stats report */
361 16550 : starttime = GetCurrentTimestamp();
362 :
363 : /*
364 : * Determine which columns to analyze
365 : *
366 : * Note that system attributes are never analyzed, so we just reject them
367 : * at the lookup stage. We also reject duplicate column mentions. (We
368 : * could alternatively ignore duplicates, but analyzing a column twice
369 : * won't work; we'd end up making a conflicting update in pg_statistic.)
370 : */
371 16550 : if (va_cols != NIL)
372 : {
373 100 : Bitmapset *unique_cols = NULL;
374 : ListCell *le;
375 :
376 100 : vacattrstats = (VacAttrStats **) palloc(list_length(va_cols) *
377 : sizeof(VacAttrStats *));
378 100 : tcnt = 0;
379 182 : foreach(le, va_cols)
380 : {
381 132 : char *col = strVal(lfirst(le));
382 :
383 132 : i = attnameAttNum(onerel, col, false);
384 132 : if (i == InvalidAttrNumber)
385 38 : ereport(ERROR,
386 : (errcode(ERRCODE_UNDEFINED_COLUMN),
387 : errmsg("column \"%s\" of relation \"%s\" does not exist",
388 : col, RelationGetRelationName(onerel))));
389 94 : if (bms_is_member(i, unique_cols))
390 12 : ereport(ERROR,
391 : (errcode(ERRCODE_DUPLICATE_COLUMN),
392 : errmsg("column \"%s\" of relation \"%s\" appears more than once",
393 : col, RelationGetRelationName(onerel))));
394 82 : unique_cols = bms_add_member(unique_cols, i);
395 :
396 82 : vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
397 82 : if (vacattrstats[tcnt] != NULL)
398 82 : tcnt++;
399 : }
400 50 : attr_cnt = tcnt;
401 : }
402 : else
403 : {
404 16450 : attr_cnt = onerel->rd_att->natts;
405 : vacattrstats = (VacAttrStats **)
406 16450 : palloc(attr_cnt * sizeof(VacAttrStats *));
407 16450 : tcnt = 0;
408 135248 : for (i = 1; i <= attr_cnt; i++)
409 : {
410 118798 : vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
411 118798 : if (vacattrstats[tcnt] != NULL)
412 118730 : tcnt++;
413 : }
414 16450 : attr_cnt = tcnt;
415 : }
416 :
417 : /*
418 : * Open all indexes of the relation, and see if there are any analyzable
419 : * columns in the indexes. We do not analyze index columns if there was
420 : * an explicit column list in the ANALYZE command, however.
421 : *
422 : * If we are doing a recursive scan, we don't want to touch the parent's
423 : * indexes at all. If we're processing a partitioned table, we need to
424 : * know if there are any indexes, but we don't want to process them.
425 : */
426 16500 : if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
427 : {
428 750 : List *idxs = RelationGetIndexList(onerel);
429 :
430 750 : Irel = NULL;
431 750 : nindexes = 0;
432 750 : hasindex = idxs != NIL;
433 750 : list_free(idxs);
434 : }
435 15750 : else if (!inh)
436 : {
437 15622 : vac_open_indexes(onerel, AccessShareLock, &nindexes, &Irel);
438 15622 : hasindex = nindexes > 0;
439 : }
440 : else
441 : {
442 128 : Irel = NULL;
443 128 : nindexes = 0;
444 128 : hasindex = false;
445 : }
446 16500 : indexdata = NULL;
447 16500 : if (nindexes > 0)
448 : {
449 12196 : indexdata = (AnlIndexData *) palloc0(nindexes * sizeof(AnlIndexData));
450 35132 : for (ind = 0; ind < nindexes; ind++)
451 : {
452 22936 : AnlIndexData *thisdata = &indexdata[ind];
453 : IndexInfo *indexInfo;
454 :
455 22936 : thisdata->indexInfo = indexInfo = BuildIndexInfo(Irel[ind]);
456 22936 : thisdata->tupleFract = 1.0; /* fix later if partial */
457 22936 : if (indexInfo->ii_Expressions != NIL && va_cols == NIL)
458 : {
459 108 : ListCell *indexpr_item = list_head(indexInfo->ii_Expressions);
460 :
461 108 : thisdata->vacattrstats = (VacAttrStats **)
462 108 : palloc(indexInfo->ii_NumIndexAttrs * sizeof(VacAttrStats *));
463 108 : tcnt = 0;
464 220 : for (i = 0; i < indexInfo->ii_NumIndexAttrs; i++)
465 : {
466 112 : int keycol = indexInfo->ii_IndexAttrNumbers[i];
467 :
468 112 : if (keycol == 0)
469 : {
470 : /* Found an index expression */
471 : Node *indexkey;
472 :
473 108 : if (indexpr_item == NULL) /* shouldn't happen */
474 0 : elog(ERROR, "too few entries in indexprs list");
475 108 : indexkey = (Node *) lfirst(indexpr_item);
476 108 : indexpr_item = lnext(indexInfo->ii_Expressions,
477 : indexpr_item);
478 216 : thisdata->vacattrstats[tcnt] =
479 108 : examine_attribute(Irel[ind], i + 1, indexkey);
480 108 : if (thisdata->vacattrstats[tcnt] != NULL)
481 108 : tcnt++;
482 : }
483 : }
484 108 : thisdata->attr_cnt = tcnt;
485 : }
486 : }
487 : }
488 :
489 : /*
490 : * Determine how many rows we need to sample, using the worst case from
491 : * all analyzable columns. We use a lower bound of 100 rows to avoid
492 : * possible overflow in Vitter's algorithm. (Note: that will also be the
493 : * target in the corner case where there are no analyzable columns.)
494 : */
495 16500 : targrows = 100;
496 135288 : for (i = 0; i < attr_cnt; i++)
497 : {
498 118788 : if (targrows < vacattrstats[i]->minrows)
499 16428 : targrows = vacattrstats[i]->minrows;
500 : }
501 39436 : for (ind = 0; ind < nindexes; ind++)
502 : {
503 22936 : AnlIndexData *thisdata = &indexdata[ind];
504 :
505 23044 : for (i = 0; i < thisdata->attr_cnt; i++)
506 : {
507 108 : if (targrows < thisdata->vacattrstats[i]->minrows)
508 12 : targrows = thisdata->vacattrstats[i]->minrows;
509 : }
510 : }
511 :
512 : /*
513 : * Look at extended statistics objects too, as those may define custom
514 : * statistics target. So we may need to sample more rows and then build
515 : * the statistics with enough detail.
516 : */
517 16500 : minrows = ComputeExtStatisticsRows(onerel, attr_cnt, vacattrstats);
518 :
519 16500 : if (targrows < minrows)
520 0 : targrows = minrows;
521 :
522 : /*
523 : * Acquire the sample rows
524 : */
525 16500 : rows = (HeapTuple *) palloc(targrows * sizeof(HeapTuple));
526 16500 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
527 : inh ? PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS_INH :
528 : PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS);
529 16500 : if (inh)
530 878 : numrows = acquire_inherited_sample_rows(onerel, elevel,
531 : rows, targrows,
532 : &totalrows, &totaldeadrows);
533 : else
534 15622 : numrows = (*acquirefunc) (onerel, elevel,
535 : rows, targrows,
536 : &totalrows, &totaldeadrows);
537 :
538 : /*
539 : * Compute the statistics. Temporary results during the calculations for
540 : * each column are stored in a child context. The calc routines are
541 : * responsible to make sure that whatever they store into the VacAttrStats
542 : * structure is allocated in anl_context.
543 : */
544 16498 : if (numrows > 0)
545 : {
546 : MemoryContext col_context,
547 : old_context;
548 :
549 11230 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
550 : PROGRESS_ANALYZE_PHASE_COMPUTE_STATS);
551 :
552 11230 : col_context = AllocSetContextCreate(anl_context,
553 : "Analyze Column",
554 : ALLOCSET_DEFAULT_SIZES);
555 11230 : old_context = MemoryContextSwitchTo(col_context);
556 :
557 98452 : for (i = 0; i < attr_cnt; i++)
558 : {
559 87222 : VacAttrStats *stats = vacattrstats[i];
560 : AttributeOpts *aopt;
561 :
562 87222 : stats->rows = rows;
563 87222 : stats->tupDesc = onerel->rd_att;
564 87222 : stats->compute_stats(stats,
565 : std_fetch_func,
566 : numrows,
567 : totalrows);
568 :
569 : /*
570 : * If the appropriate flavor of the n_distinct option is
571 : * specified, override with the corresponding value.
572 : */
573 87222 : aopt = get_attribute_options(onerel->rd_id, stats->tupattnum);
574 87222 : if (aopt != NULL)
575 : {
576 : float8 n_distinct;
577 :
578 6 : n_distinct = inh ? aopt->n_distinct_inherited : aopt->n_distinct;
579 6 : if (n_distinct != 0.0)
580 6 : stats->stadistinct = n_distinct;
581 : }
582 :
583 87222 : MemoryContextReset(col_context);
584 : }
585 :
586 11230 : if (nindexes > 0)
587 7222 : compute_index_stats(onerel, totalrows,
588 : indexdata, nindexes,
589 : rows, numrows,
590 : col_context);
591 :
592 11224 : MemoryContextSwitchTo(old_context);
593 11224 : MemoryContextDelete(col_context);
594 :
595 : /*
596 : * Emit the completed stats rows into pg_statistic, replacing any
597 : * previous statistics for the target columns. (If there are stats in
598 : * pg_statistic for columns we didn't process, we leave them alone.)
599 : */
600 11224 : update_attstats(RelationGetRelid(onerel), inh,
601 : attr_cnt, vacattrstats);
602 :
603 25286 : for (ind = 0; ind < nindexes; ind++)
604 : {
605 14062 : AnlIndexData *thisdata = &indexdata[ind];
606 :
607 14062 : update_attstats(RelationGetRelid(Irel[ind]), false,
608 : thisdata->attr_cnt, thisdata->vacattrstats);
609 : }
610 :
611 : /* Build extended statistics (if there are any). */
612 11224 : BuildRelationExtStatistics(onerel, inh, totalrows, numrows, rows,
613 : attr_cnt, vacattrstats);
614 : }
615 :
616 16492 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
617 : PROGRESS_ANALYZE_PHASE_FINALIZE_ANALYZE);
618 :
619 : /*
620 : * Update pages/tuples stats in pg_class ... but not if we're doing
621 : * inherited stats.
622 : *
623 : * We assume that VACUUM hasn't set pg_class.reltuples already, even
624 : * during a VACUUM ANALYZE. Although VACUUM often updates pg_class,
625 : * exceptions exist. A "VACUUM (ANALYZE, INDEX_CLEANUP OFF)" command will
626 : * never update pg_class entries for index relations. It's also possible
627 : * that an individual index's pg_class entry won't be updated during
628 : * VACUUM if the index AM returns NULL from its amvacuumcleanup() routine.
629 : */
630 16492 : if (!inh)
631 : {
632 15614 : BlockNumber relallvisible = 0;
633 15614 : BlockNumber relallfrozen = 0;
634 :
635 15614 : if (RELKIND_HAS_STORAGE(onerel->rd_rel->relkind))
636 15550 : visibilitymap_count(onerel, &relallvisible, &relallfrozen);
637 :
638 : /*
639 : * Update pg_class for table relation. CCI first, in case acquirefunc
640 : * updated pg_class.
641 : */
642 15614 : CommandCounterIncrement();
643 15614 : vac_update_relstats(onerel,
644 : relpages,
645 : totalrows,
646 : relallvisible,
647 : relallfrozen,
648 : hasindex,
649 : InvalidTransactionId,
650 : InvalidMultiXactId,
651 : NULL, NULL,
652 : in_outer_xact);
653 :
654 : /* Same for indexes */
655 38538 : for (ind = 0; ind < nindexes; ind++)
656 : {
657 22924 : AnlIndexData *thisdata = &indexdata[ind];
658 : double totalindexrows;
659 :
660 22924 : totalindexrows = ceil(thisdata->tupleFract * totalrows);
661 22924 : vac_update_relstats(Irel[ind],
662 22924 : RelationGetNumberOfBlocks(Irel[ind]),
663 : totalindexrows,
664 : 0, 0,
665 : false,
666 : InvalidTransactionId,
667 : InvalidMultiXactId,
668 : NULL, NULL,
669 : in_outer_xact);
670 : }
671 : }
672 878 : else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
673 : {
674 : /*
675 : * Partitioned tables don't have storage, so we don't set any fields
676 : * in their pg_class entries except for reltuples and relhasindex.
677 : */
678 750 : CommandCounterIncrement();
679 750 : vac_update_relstats(onerel, -1, totalrows,
680 : 0, 0, hasindex, InvalidTransactionId,
681 : InvalidMultiXactId,
682 : NULL, NULL,
683 : in_outer_xact);
684 : }
685 :
686 : /*
687 : * Now report ANALYZE to the cumulative stats system. For regular tables,
688 : * we do it only if not doing inherited stats. For partitioned tables, we
689 : * only do it for inherited stats. (We're never called for not-inherited
690 : * stats on partitioned tables anyway.)
691 : *
692 : * Reset the mod_since_analyze counter only if we analyzed all columns;
693 : * otherwise, there is still work for auto-analyze to do.
694 : */
695 16492 : if (!inh)
696 15614 : pgstat_report_analyze(onerel, totalrows, totaldeadrows,
697 : (va_cols == NIL), starttime);
698 878 : else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
699 750 : pgstat_report_analyze(onerel, 0, 0, (va_cols == NIL), starttime);
700 :
701 : /*
702 : * If this isn't part of VACUUM ANALYZE, let index AMs do cleanup.
703 : *
704 : * Note that most index AMs perform a no-op as a matter of policy for
705 : * amvacuumcleanup() when called in ANALYZE-only mode. The only exception
706 : * among core index AMs is GIN/ginvacuumcleanup().
707 : */
708 16492 : if (!(params.options & VACOPT_VACUUM))
709 : {
710 31456 : for (ind = 0; ind < nindexes; ind++)
711 : {
712 : IndexBulkDeleteResult *stats;
713 : IndexVacuumInfo ivinfo;
714 :
715 18000 : ivinfo.index = Irel[ind];
716 18000 : ivinfo.heaprel = onerel;
717 18000 : ivinfo.analyze_only = true;
718 18000 : ivinfo.estimated_count = true;
719 18000 : ivinfo.message_level = elevel;
720 18000 : ivinfo.num_heap_tuples = onerel->rd_rel->reltuples;
721 18000 : ivinfo.strategy = vac_strategy;
722 :
723 18000 : stats = index_vacuum_cleanup(&ivinfo, NULL);
724 :
725 18000 : if (stats)
726 0 : pfree(stats);
727 : }
728 : }
729 :
730 : /* Done with indexes */
731 16492 : vac_close_indexes(nindexes, Irel, NoLock);
732 :
733 : /* Log the action if appropriate */
734 16492 : if (instrument)
735 : {
736 646 : TimestampTz endtime = GetCurrentTimestamp();
737 :
738 1236 : if (verbose || params.log_analyze_min_duration == 0 ||
739 590 : TimestampDifferenceExceeds(starttime, endtime,
740 590 : params.log_analyze_min_duration))
741 : {
742 : long delay_in_ms;
743 : WalUsage walusage;
744 56 : double read_rate = 0;
745 56 : double write_rate = 0;
746 : char *msgfmt;
747 : StringInfoData buf;
748 : int64 total_blks_hit;
749 : int64 total_blks_read;
750 : int64 total_blks_dirtied;
751 :
752 56 : memset(&bufferusage, 0, sizeof(BufferUsage));
753 56 : BufferUsageAccumDiff(&bufferusage, &pgBufferUsage, &startbufferusage);
754 56 : memset(&walusage, 0, sizeof(WalUsage));
755 56 : WalUsageAccumDiff(&walusage, &pgWalUsage, &startwalusage);
756 :
757 56 : total_blks_hit = bufferusage.shared_blks_hit +
758 56 : bufferusage.local_blks_hit;
759 56 : total_blks_read = bufferusage.shared_blks_read +
760 56 : bufferusage.local_blks_read;
761 56 : total_blks_dirtied = bufferusage.shared_blks_dirtied +
762 56 : bufferusage.local_blks_dirtied;
763 :
764 : /*
765 : * We do not expect an analyze to take > 25 days and it simplifies
766 : * things a bit to use TimestampDifferenceMilliseconds.
767 : */
768 56 : delay_in_ms = TimestampDifferenceMilliseconds(starttime, endtime);
769 :
770 : /*
771 : * Note that we are reporting these read/write rates in the same
772 : * manner as VACUUM does, which means that while the 'average read
773 : * rate' here actually corresponds to page misses and resulting
774 : * reads which are also picked up by track_io_timing, if enabled,
775 : * the 'average write rate' is actually talking about the rate of
776 : * pages being dirtied, not being written out, so it's typical to
777 : * have a non-zero 'avg write rate' while I/O timings only reports
778 : * reads.
779 : *
780 : * It's not clear that an ANALYZE will ever result in
781 : * FlushBuffer() being called, but we track and support reporting
782 : * on I/O write time in case that changes as it's practically free
783 : * to do so anyway.
784 : */
785 :
786 56 : if (delay_in_ms > 0)
787 : {
788 56 : read_rate = (double) BLCKSZ * total_blks_read /
789 56 : (1024 * 1024) / (delay_in_ms / 1000.0);
790 56 : write_rate = (double) BLCKSZ * total_blks_dirtied /
791 56 : (1024 * 1024) / (delay_in_ms / 1000.0);
792 : }
793 :
794 : /*
795 : * We split this up so we don't emit empty I/O timing values when
796 : * track_io_timing isn't enabled.
797 : */
798 :
799 56 : initStringInfo(&buf);
800 :
801 56 : if (AmAutoVacuumWorkerProcess())
802 56 : msgfmt = _("automatic analyze of table \"%s.%s.%s\"\n");
803 : else
804 0 : msgfmt = _("finished analyzing table \"%s.%s.%s\"\n");
805 :
806 56 : appendStringInfo(&buf, msgfmt,
807 : get_database_name(MyDatabaseId),
808 56 : get_namespace_name(RelationGetNamespace(onerel)),
809 56 : RelationGetRelationName(onerel));
810 56 : if (track_cost_delay_timing)
811 : {
812 : /*
813 : * We bypass the changecount mechanism because this value is
814 : * only updated by the calling process.
815 : */
816 0 : appendStringInfo(&buf, _("delay time: %.3f ms\n"),
817 0 : (double) MyBEEntry->st_progress_param[PROGRESS_ANALYZE_DELAY_TIME] / 1000000.0);
818 : }
819 56 : if (track_io_timing)
820 : {
821 0 : double read_ms = (double) (pgStatBlockReadTime - startreadtime) / 1000;
822 0 : double write_ms = (double) (pgStatBlockWriteTime - startwritetime) / 1000;
823 :
824 0 : appendStringInfo(&buf, _("I/O timings: read: %.3f ms, write: %.3f ms\n"),
825 : read_ms, write_ms);
826 : }
827 56 : appendStringInfo(&buf, _("avg read rate: %.3f MB/s, avg write rate: %.3f MB/s\n"),
828 : read_rate, write_rate);
829 56 : appendStringInfo(&buf, _("buffer usage: %" PRId64 " hits, %" PRId64 " reads, %" PRId64 " dirtied\n"),
830 : total_blks_hit,
831 : total_blks_read,
832 : total_blks_dirtied);
833 56 : appendStringInfo(&buf,
834 56 : _("WAL usage: %" PRId64 " records, %" PRId64 " full page images, %" PRIu64 " bytes, %" PRIu64 " full page image bytes, %" PRId64 " buffers full\n"),
835 : walusage.wal_records,
836 : walusage.wal_fpi,
837 : walusage.wal_bytes,
838 : walusage.wal_fpi_bytes,
839 : walusage.wal_buffers_full);
840 56 : appendStringInfo(&buf, _("system usage: %s"), pg_rusage_show(&ru0));
841 :
842 56 : ereport(verbose ? INFO : LOG,
843 : (errmsg_internal("%s", buf.data)));
844 :
845 56 : pfree(buf.data);
846 : }
847 : }
848 :
849 : /* Roll back any GUC changes executed by index functions */
850 16492 : AtEOXact_GUC(false, save_nestlevel);
851 :
852 : /* Restore userid and security context */
853 16492 : SetUserIdAndSecContext(save_userid, save_sec_context);
854 :
855 : /* Restore current context and release memory */
856 16492 : MemoryContextSwitchTo(caller_context);
857 16492 : MemoryContextDelete(anl_context);
858 16492 : anl_context = NULL;
859 16492 : }
860 :
861 : /*
862 : * Compute statistics about indexes of a relation
863 : */
864 : static void
865 7222 : compute_index_stats(Relation onerel, double totalrows,
866 : AnlIndexData *indexdata, int nindexes,
867 : HeapTuple *rows, int numrows,
868 : MemoryContext col_context)
869 : {
870 : MemoryContext ind_context,
871 : old_context;
872 : Datum values[INDEX_MAX_KEYS];
873 : bool isnull[INDEX_MAX_KEYS];
874 : int ind,
875 : i;
876 :
877 7222 : ind_context = AllocSetContextCreate(anl_context,
878 : "Analyze Index",
879 : ALLOCSET_DEFAULT_SIZES);
880 7222 : old_context = MemoryContextSwitchTo(ind_context);
881 :
882 21290 : for (ind = 0; ind < nindexes; ind++)
883 : {
884 14074 : AnlIndexData *thisdata = &indexdata[ind];
885 14074 : IndexInfo *indexInfo = thisdata->indexInfo;
886 14074 : int attr_cnt = thisdata->attr_cnt;
887 : TupleTableSlot *slot;
888 : EState *estate;
889 : ExprContext *econtext;
890 : ExprState *predicate;
891 : Datum *exprvals;
892 : bool *exprnulls;
893 : int numindexrows,
894 : tcnt,
895 : rowno;
896 : double totalindexrows;
897 :
898 : /* Ignore index if no columns to analyze and not partial */
899 14074 : if (attr_cnt == 0 && indexInfo->ii_Predicate == NIL)
900 13938 : continue;
901 :
902 : /*
903 : * Need an EState for evaluation of index expressions and
904 : * partial-index predicates. Create it in the per-index context to be
905 : * sure it gets cleaned up at the bottom of the loop.
906 : */
907 136 : estate = CreateExecutorState();
908 136 : econtext = GetPerTupleExprContext(estate);
909 : /* Need a slot to hold the current heap tuple, too */
910 136 : slot = MakeSingleTupleTableSlot(RelationGetDescr(onerel),
911 : &TTSOpsHeapTuple);
912 :
913 : /* Arrange for econtext's scan tuple to be the tuple under test */
914 136 : econtext->ecxt_scantuple = slot;
915 :
916 : /* Set up execution state for predicate. */
917 136 : predicate = ExecPrepareQual(indexInfo->ii_Predicate, estate);
918 :
919 : /* Compute and save index expression values */
920 136 : exprvals = (Datum *) palloc(numrows * attr_cnt * sizeof(Datum));
921 136 : exprnulls = (bool *) palloc(numrows * attr_cnt * sizeof(bool));
922 136 : numindexrows = 0;
923 136 : tcnt = 0;
924 155224 : for (rowno = 0; rowno < numrows; rowno++)
925 : {
926 155094 : HeapTuple heapTuple = rows[rowno];
927 :
928 155094 : vacuum_delay_point(true);
929 :
930 : /*
931 : * Reset the per-tuple context each time, to reclaim any cruft
932 : * left behind by evaluating the predicate or index expressions.
933 : */
934 155094 : ResetExprContext(econtext);
935 :
936 : /* Set up for predicate or expression evaluation */
937 155094 : ExecStoreHeapTuple(heapTuple, slot, false);
938 :
939 : /* If index is partial, check predicate */
940 155094 : if (predicate != NULL)
941 : {
942 40066 : if (!ExecQual(predicate, econtext))
943 21330 : continue;
944 : }
945 133764 : numindexrows++;
946 :
947 133764 : if (attr_cnt > 0)
948 : {
949 : /*
950 : * Evaluate the index row to compute expression values. We
951 : * could do this by hand, but FormIndexDatum is convenient.
952 : */
953 115028 : FormIndexDatum(indexInfo,
954 : slot,
955 : estate,
956 : values,
957 : isnull);
958 :
959 : /*
960 : * Save just the columns we care about. We copy the values
961 : * into ind_context from the estate's per-tuple context.
962 : */
963 230044 : for (i = 0; i < attr_cnt; i++)
964 : {
965 115022 : VacAttrStats *stats = thisdata->vacattrstats[i];
966 115022 : int attnum = stats->tupattnum;
967 :
968 115022 : if (isnull[attnum - 1])
969 : {
970 6 : exprvals[tcnt] = (Datum) 0;
971 6 : exprnulls[tcnt] = true;
972 : }
973 : else
974 : {
975 230032 : exprvals[tcnt] = datumCopy(values[attnum - 1],
976 115016 : stats->attrtype->typbyval,
977 115016 : stats->attrtype->typlen);
978 115016 : exprnulls[tcnt] = false;
979 : }
980 115022 : tcnt++;
981 : }
982 : }
983 : }
984 :
985 : /*
986 : * Having counted the number of rows that pass the predicate in the
987 : * sample, we can estimate the total number of rows in the index.
988 : */
989 130 : thisdata->tupleFract = (double) numindexrows / (double) numrows;
990 130 : totalindexrows = ceil(thisdata->tupleFract * totalrows);
991 :
992 : /*
993 : * Now we can compute the statistics for the expression columns.
994 : */
995 130 : if (numindexrows > 0)
996 : {
997 122 : MemoryContextSwitchTo(col_context);
998 206 : for (i = 0; i < attr_cnt; i++)
999 : {
1000 84 : VacAttrStats *stats = thisdata->vacattrstats[i];
1001 :
1002 84 : stats->exprvals = exprvals + i;
1003 84 : stats->exprnulls = exprnulls + i;
1004 84 : stats->rowstride = attr_cnt;
1005 84 : stats->compute_stats(stats,
1006 : ind_fetch_func,
1007 : numindexrows,
1008 : totalindexrows);
1009 :
1010 84 : MemoryContextReset(col_context);
1011 : }
1012 : }
1013 :
1014 : /* And clean up */
1015 130 : MemoryContextSwitchTo(ind_context);
1016 :
1017 130 : ExecDropSingleTupleTableSlot(slot);
1018 130 : FreeExecutorState(estate);
1019 130 : MemoryContextReset(ind_context);
1020 : }
1021 :
1022 7216 : MemoryContextSwitchTo(old_context);
1023 7216 : MemoryContextDelete(ind_context);
1024 7216 : }
1025 :
1026 : /*
1027 : * examine_attribute -- pre-analysis of a single column
1028 : *
1029 : * Determine whether the column is analyzable; if so, create and initialize
1030 : * a VacAttrStats struct for it. If not, return NULL.
1031 : *
1032 : * If index_expr isn't NULL, then we're trying to analyze an expression index,
1033 : * and index_expr is the expression tree representing the column's data.
1034 : */
1035 : static VacAttrStats *
1036 118988 : examine_attribute(Relation onerel, int attnum, Node *index_expr)
1037 : {
1038 118988 : Form_pg_attribute attr = TupleDescAttr(onerel->rd_att, attnum - 1);
1039 : int attstattarget;
1040 : HeapTuple atttuple;
1041 : Datum dat;
1042 : bool isnull;
1043 : HeapTuple typtuple;
1044 : VacAttrStats *stats;
1045 : int i;
1046 : bool ok;
1047 :
1048 : /* Never analyze dropped columns */
1049 118988 : if (attr->attisdropped)
1050 6 : return NULL;
1051 :
1052 : /* Don't analyze virtual generated columns */
1053 118982 : if (attr->attgenerated == ATTRIBUTE_GENERATED_VIRTUAL)
1054 56 : return NULL;
1055 :
1056 : /*
1057 : * Get attstattarget value. Set to -1 if null. (Analyze functions expect
1058 : * -1 to mean use default_statistics_target; see for example
1059 : * std_typanalyze.)
1060 : */
1061 118926 : atttuple = SearchSysCache2(ATTNUM, ObjectIdGetDatum(RelationGetRelid(onerel)), Int16GetDatum(attnum));
1062 118926 : if (!HeapTupleIsValid(atttuple))
1063 0 : elog(ERROR, "cache lookup failed for attribute %d of relation %u",
1064 : attnum, RelationGetRelid(onerel));
1065 118926 : dat = SysCacheGetAttr(ATTNUM, atttuple, Anum_pg_attribute_attstattarget, &isnull);
1066 118926 : attstattarget = isnull ? -1 : DatumGetInt16(dat);
1067 118926 : ReleaseSysCache(atttuple);
1068 :
1069 : /* Don't analyze column if user has specified not to */
1070 118926 : if (attstattarget == 0)
1071 6 : return NULL;
1072 :
1073 : /*
1074 : * Create the VacAttrStats struct.
1075 : */
1076 118920 : stats = (VacAttrStats *) palloc0(sizeof(VacAttrStats));
1077 118920 : stats->attstattarget = attstattarget;
1078 :
1079 : /*
1080 : * When analyzing an expression index, believe the expression tree's type
1081 : * not the column datatype --- the latter might be the opckeytype storage
1082 : * type of the opclass, which is not interesting for our purposes. (Note:
1083 : * if we did anything with non-expression index columns, we'd need to
1084 : * figure out where to get the correct type info from, but for now that's
1085 : * not a problem.) It's not clear whether anyone will care about the
1086 : * typmod, but we store that too just in case.
1087 : */
1088 118920 : if (index_expr)
1089 : {
1090 108 : stats->attrtypid = exprType(index_expr);
1091 108 : stats->attrtypmod = exprTypmod(index_expr);
1092 :
1093 : /*
1094 : * If a collation has been specified for the index column, use that in
1095 : * preference to anything else; but if not, fall back to whatever we
1096 : * can get from the expression.
1097 : */
1098 108 : if (OidIsValid(onerel->rd_indcollation[attnum - 1]))
1099 14 : stats->attrcollid = onerel->rd_indcollation[attnum - 1];
1100 : else
1101 94 : stats->attrcollid = exprCollation(index_expr);
1102 : }
1103 : else
1104 : {
1105 118812 : stats->attrtypid = attr->atttypid;
1106 118812 : stats->attrtypmod = attr->atttypmod;
1107 118812 : stats->attrcollid = attr->attcollation;
1108 : }
1109 :
1110 118920 : typtuple = SearchSysCacheCopy1(TYPEOID,
1111 : ObjectIdGetDatum(stats->attrtypid));
1112 118920 : if (!HeapTupleIsValid(typtuple))
1113 0 : elog(ERROR, "cache lookup failed for type %u", stats->attrtypid);
1114 118920 : stats->attrtype = (Form_pg_type) GETSTRUCT(typtuple);
1115 118920 : stats->anl_context = anl_context;
1116 118920 : stats->tupattnum = attnum;
1117 :
1118 : /*
1119 : * The fields describing the stats->stavalues[n] element types default to
1120 : * the type of the data being analyzed, but the type-specific typanalyze
1121 : * function can change them if it wants to store something else.
1122 : */
1123 713520 : for (i = 0; i < STATISTIC_NUM_SLOTS; i++)
1124 : {
1125 594600 : stats->statypid[i] = stats->attrtypid;
1126 594600 : stats->statyplen[i] = stats->attrtype->typlen;
1127 594600 : stats->statypbyval[i] = stats->attrtype->typbyval;
1128 594600 : stats->statypalign[i] = stats->attrtype->typalign;
1129 : }
1130 :
1131 : /*
1132 : * Call the type-specific typanalyze function. If none is specified, use
1133 : * std_typanalyze().
1134 : */
1135 118920 : if (OidIsValid(stats->attrtype->typanalyze))
1136 7648 : ok = DatumGetBool(OidFunctionCall1(stats->attrtype->typanalyze,
1137 : PointerGetDatum(stats)));
1138 : else
1139 111272 : ok = std_typanalyze(stats);
1140 :
1141 118920 : if (!ok || stats->compute_stats == NULL || stats->minrows <= 0)
1142 : {
1143 0 : heap_freetuple(typtuple);
1144 0 : pfree(stats);
1145 0 : return NULL;
1146 : }
1147 :
1148 118920 : return stats;
1149 : }
1150 :
1151 : /*
1152 : * Read stream callback returning the next BlockNumber as chosen by the
1153 : * BlockSampling algorithm.
1154 : */
1155 : static BlockNumber
1156 164684 : block_sampling_read_stream_next(ReadStream *stream,
1157 : void *callback_private_data,
1158 : void *per_buffer_data)
1159 : {
1160 164684 : BlockSamplerData *bs = callback_private_data;
1161 :
1162 164684 : return BlockSampler_HasMore(bs) ? BlockSampler_Next(bs) : InvalidBlockNumber;
1163 : }
1164 :
1165 : /*
1166 : * acquire_sample_rows -- acquire a random sample of rows from the table
1167 : *
1168 : * Selected rows are returned in the caller-allocated array rows[], which
1169 : * must have at least targrows entries.
1170 : * The actual number of rows selected is returned as the function result.
1171 : * We also estimate the total numbers of live and dead rows in the table,
1172 : * and return them into *totalrows and *totaldeadrows, respectively.
1173 : *
1174 : * The returned list of tuples is in order by physical position in the table.
1175 : * (We will rely on this later to derive correlation estimates.)
1176 : *
1177 : * As of May 2004 we use a new two-stage method: Stage one selects up
1178 : * to targrows random blocks (or all blocks, if there aren't so many).
1179 : * Stage two scans these blocks and uses the Vitter algorithm to create
1180 : * a random sample of targrows rows (or less, if there are less in the
1181 : * sample of blocks). The two stages are executed simultaneously: each
1182 : * block is processed as soon as stage one returns its number and while
1183 : * the rows are read stage two controls which ones are to be inserted
1184 : * into the sample.
1185 : *
1186 : * Although every row has an equal chance of ending up in the final
1187 : * sample, this sampling method is not perfect: not every possible
1188 : * sample has an equal chance of being selected. For large relations
1189 : * the number of different blocks represented by the sample tends to be
1190 : * too small. We can live with that for now. Improvements are welcome.
1191 : *
1192 : * An important property of this sampling method is that because we do
1193 : * look at a statistically unbiased set of blocks, we should get
1194 : * unbiased estimates of the average numbers of live and dead rows per
1195 : * block. The previous sampling method put too much credence in the row
1196 : * density near the start of the table.
1197 : */
1198 : static int
1199 17606 : acquire_sample_rows(Relation onerel, int elevel,
1200 : HeapTuple *rows, int targrows,
1201 : double *totalrows, double *totaldeadrows)
1202 : {
1203 17606 : int numrows = 0; /* # rows now in reservoir */
1204 17606 : double samplerows = 0; /* total # rows collected */
1205 17606 : double liverows = 0; /* # live rows seen */
1206 17606 : double deadrows = 0; /* # dead rows seen */
1207 17606 : double rowstoskip = -1; /* -1 means not set yet */
1208 : uint32 randseed; /* Seed for block sampler(s) */
1209 : BlockNumber totalblocks;
1210 : TransactionId OldestXmin;
1211 : BlockSamplerData bs;
1212 : ReservoirStateData rstate;
1213 : TupleTableSlot *slot;
1214 : TableScanDesc scan;
1215 : BlockNumber nblocks;
1216 17606 : BlockNumber blksdone = 0;
1217 : ReadStream *stream;
1218 :
1219 : Assert(targrows > 0);
1220 :
1221 17606 : totalblocks = RelationGetNumberOfBlocks(onerel);
1222 :
1223 : /* Need a cutoff xmin for HeapTupleSatisfiesVacuum */
1224 17606 : OldestXmin = GetOldestNonRemovableTransactionId(onerel);
1225 :
1226 : /* Prepare for sampling block numbers */
1227 17606 : randseed = pg_prng_uint32(&pg_global_prng_state);
1228 17606 : nblocks = BlockSampler_Init(&bs, totalblocks, targrows, randseed);
1229 :
1230 : /* Report sampling block numbers */
1231 17606 : pgstat_progress_update_param(PROGRESS_ANALYZE_BLOCKS_TOTAL,
1232 : nblocks);
1233 :
1234 : /* Prepare for sampling rows */
1235 17606 : reservoir_init_selection_state(&rstate, targrows);
1236 :
1237 17606 : scan = table_beginscan_analyze(onerel);
1238 17606 : slot = table_slot_create(onerel, NULL);
1239 :
1240 : /*
1241 : * It is safe to use batching, as block_sampling_read_stream_next never
1242 : * blocks.
1243 : */
1244 17606 : stream = read_stream_begin_relation(READ_STREAM_MAINTENANCE |
1245 : READ_STREAM_USE_BATCHING,
1246 : vac_strategy,
1247 : scan->rs_rd,
1248 : MAIN_FORKNUM,
1249 : block_sampling_read_stream_next,
1250 : &bs,
1251 : 0);
1252 :
1253 : /* Outer loop over blocks to sample */
1254 164684 : while (table_scan_analyze_next_block(scan, stream))
1255 : {
1256 147078 : vacuum_delay_point(true);
1257 :
1258 11351620 : while (table_scan_analyze_next_tuple(scan, OldestXmin, &liverows, &deadrows, slot))
1259 : {
1260 : /*
1261 : * The first targrows sample rows are simply copied into the
1262 : * reservoir. Then we start replacing tuples in the sample until
1263 : * we reach the end of the relation. This algorithm is from Jeff
1264 : * Vitter's paper (see full citation in utils/misc/sampling.c). It
1265 : * works by repeatedly computing the number of tuples to skip
1266 : * before selecting a tuple, which replaces a randomly chosen
1267 : * element of the reservoir (current set of tuples). At all times
1268 : * the reservoir is a true random sample of the tuples we've
1269 : * passed over so far, so when we fall off the end of the relation
1270 : * we're done.
1271 : */
1272 11204542 : if (numrows < targrows)
1273 10951580 : rows[numrows++] = ExecCopySlotHeapTuple(slot);
1274 : else
1275 : {
1276 : /*
1277 : * t in Vitter's paper is the number of records already
1278 : * processed. If we need to compute a new S value, we must
1279 : * use the not-yet-incremented value of samplerows as t.
1280 : */
1281 252962 : if (rowstoskip < 0)
1282 115180 : rowstoskip = reservoir_get_next_S(&rstate, samplerows, targrows);
1283 :
1284 252962 : if (rowstoskip <= 0)
1285 : {
1286 : /*
1287 : * Found a suitable tuple, so save it, replacing one old
1288 : * tuple at random
1289 : */
1290 115132 : int k = (int) (targrows * sampler_random_fract(&rstate.randstate));
1291 :
1292 : Assert(k >= 0 && k < targrows);
1293 115132 : heap_freetuple(rows[k]);
1294 115132 : rows[k] = ExecCopySlotHeapTuple(slot);
1295 : }
1296 :
1297 252962 : rowstoskip -= 1;
1298 : }
1299 :
1300 11204542 : samplerows += 1;
1301 : }
1302 :
1303 147078 : pgstat_progress_update_param(PROGRESS_ANALYZE_BLOCKS_DONE,
1304 : ++blksdone);
1305 : }
1306 :
1307 17606 : read_stream_end(stream);
1308 :
1309 17606 : ExecDropSingleTupleTableSlot(slot);
1310 17606 : table_endscan(scan);
1311 :
1312 : /*
1313 : * If we didn't find as many tuples as we wanted then we're done. No sort
1314 : * is needed, since they're already in order.
1315 : *
1316 : * Otherwise we need to sort the collected tuples by position
1317 : * (itempointer). It's not worth worrying about corner cases where the
1318 : * tuples are already sorted.
1319 : */
1320 17606 : if (numrows == targrows)
1321 162 : qsort_interruptible(rows, numrows, sizeof(HeapTuple),
1322 : compare_rows, NULL);
1323 :
1324 : /*
1325 : * Estimate total numbers of live and dead rows in relation, extrapolating
1326 : * on the assumption that the average tuple density in pages we didn't
1327 : * scan is the same as in the pages we did scan. Since what we scanned is
1328 : * a random sample of the pages in the relation, this should be a good
1329 : * assumption.
1330 : */
1331 17606 : if (bs.m > 0)
1332 : {
1333 12436 : *totalrows = floor((liverows / bs.m) * totalblocks + 0.5);
1334 12436 : *totaldeadrows = floor((deadrows / bs.m) * totalblocks + 0.5);
1335 : }
1336 : else
1337 : {
1338 5170 : *totalrows = 0.0;
1339 5170 : *totaldeadrows = 0.0;
1340 : }
1341 :
1342 : /*
1343 : * Emit some interesting relation info
1344 : */
1345 17606 : ereport(elevel,
1346 : (errmsg("\"%s\": scanned %d of %u pages, "
1347 : "containing %.0f live rows and %.0f dead rows; "
1348 : "%d rows in sample, %.0f estimated total rows",
1349 : RelationGetRelationName(onerel),
1350 : bs.m, totalblocks,
1351 : liverows, deadrows,
1352 : numrows, *totalrows)));
1353 :
1354 17606 : return numrows;
1355 : }
1356 :
1357 : /*
1358 : * Comparator for sorting rows[] array
1359 : */
1360 : static int
1361 3976420 : compare_rows(const void *a, const void *b, void *arg)
1362 : {
1363 3976420 : HeapTuple ha = *(const HeapTuple *) a;
1364 3976420 : HeapTuple hb = *(const HeapTuple *) b;
1365 3976420 : BlockNumber ba = ItemPointerGetBlockNumber(&ha->t_self);
1366 3976420 : OffsetNumber oa = ItemPointerGetOffsetNumber(&ha->t_self);
1367 3976420 : BlockNumber bb = ItemPointerGetBlockNumber(&hb->t_self);
1368 3976420 : OffsetNumber ob = ItemPointerGetOffsetNumber(&hb->t_self);
1369 :
1370 3976420 : if (ba < bb)
1371 868608 : return -1;
1372 3107812 : if (ba > bb)
1373 843810 : return 1;
1374 2264002 : if (oa < ob)
1375 1529520 : return -1;
1376 734482 : if (oa > ob)
1377 734482 : return 1;
1378 0 : return 0;
1379 : }
1380 :
1381 :
1382 : /*
1383 : * acquire_inherited_sample_rows -- acquire sample rows from inheritance tree
1384 : *
1385 : * This has the same API as acquire_sample_rows, except that rows are
1386 : * collected from all inheritance children as well as the specified table.
1387 : * We fail and return zero if there are no inheritance children, or if all
1388 : * children are foreign tables that don't support ANALYZE.
1389 : */
1390 : static int
1391 878 : acquire_inherited_sample_rows(Relation onerel, int elevel,
1392 : HeapTuple *rows, int targrows,
1393 : double *totalrows, double *totaldeadrows)
1394 : {
1395 : List *tableOIDs;
1396 : Relation *rels;
1397 : AcquireSampleRowsFunc *acquirefuncs;
1398 : double *relblocks;
1399 : double totalblocks;
1400 : int numrows,
1401 : nrels,
1402 : i;
1403 : ListCell *lc;
1404 : bool has_child;
1405 :
1406 : /* Initialize output parameters to zero now, in case we exit early */
1407 878 : *totalrows = 0;
1408 878 : *totaldeadrows = 0;
1409 :
1410 : /*
1411 : * Find all members of inheritance set. We only need AccessShareLock on
1412 : * the children.
1413 : */
1414 : tableOIDs =
1415 878 : find_all_inheritors(RelationGetRelid(onerel), AccessShareLock, NULL);
1416 :
1417 : /*
1418 : * Check that there's at least one descendant, else fail. This could
1419 : * happen despite analyze_rel's relhassubclass check, if table once had a
1420 : * child but no longer does. In that case, we can clear the
1421 : * relhassubclass field so as not to make the same mistake again later.
1422 : * (This is safe because we hold ShareUpdateExclusiveLock.)
1423 : */
1424 878 : if (list_length(tableOIDs) < 2)
1425 : {
1426 : /* CCI because we already updated the pg_class row in this command */
1427 20 : CommandCounterIncrement();
1428 20 : SetRelationHasSubclass(RelationGetRelid(onerel), false);
1429 20 : ereport(elevel,
1430 : (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no child tables",
1431 : get_namespace_name(RelationGetNamespace(onerel)),
1432 : RelationGetRelationName(onerel))));
1433 20 : return 0;
1434 : }
1435 :
1436 : /*
1437 : * Identify acquirefuncs to use, and count blocks in all the relations.
1438 : * The result could overflow BlockNumber, so we use double arithmetic.
1439 : */
1440 858 : rels = (Relation *) palloc(list_length(tableOIDs) * sizeof(Relation));
1441 : acquirefuncs = (AcquireSampleRowsFunc *)
1442 858 : palloc(list_length(tableOIDs) * sizeof(AcquireSampleRowsFunc));
1443 858 : relblocks = (double *) palloc(list_length(tableOIDs) * sizeof(double));
1444 858 : totalblocks = 0;
1445 858 : nrels = 0;
1446 858 : has_child = false;
1447 3926 : foreach(lc, tableOIDs)
1448 : {
1449 3068 : Oid childOID = lfirst_oid(lc);
1450 : Relation childrel;
1451 3068 : AcquireSampleRowsFunc acquirefunc = NULL;
1452 3068 : BlockNumber relpages = 0;
1453 :
1454 : /* We already got the needed lock */
1455 3068 : childrel = table_open(childOID, NoLock);
1456 :
1457 : /* Ignore if temp table of another backend */
1458 3068 : if (RELATION_IS_OTHER_TEMP(childrel))
1459 : {
1460 : /* ... but release the lock on it */
1461 : Assert(childrel != onerel);
1462 0 : table_close(childrel, AccessShareLock);
1463 822 : continue;
1464 : }
1465 :
1466 : /* Check table type (MATVIEW can't happen, but might as well allow) */
1467 3068 : if (childrel->rd_rel->relkind == RELKIND_RELATION ||
1468 852 : childrel->rd_rel->relkind == RELKIND_MATVIEW)
1469 : {
1470 : /* Regular table, so use the regular row acquisition function */
1471 2216 : acquirefunc = acquire_sample_rows;
1472 2216 : relpages = RelationGetNumberOfBlocks(childrel);
1473 : }
1474 852 : else if (childrel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
1475 : {
1476 : /*
1477 : * For a foreign table, call the FDW's hook function to see
1478 : * whether it supports analysis.
1479 : */
1480 : FdwRoutine *fdwroutine;
1481 30 : bool ok = false;
1482 :
1483 30 : fdwroutine = GetFdwRoutineForRelation(childrel, false);
1484 :
1485 30 : if (fdwroutine->AnalyzeForeignTable != NULL)
1486 30 : ok = fdwroutine->AnalyzeForeignTable(childrel,
1487 : &acquirefunc,
1488 : &relpages);
1489 :
1490 30 : if (!ok)
1491 : {
1492 : /* ignore, but release the lock on it */
1493 : Assert(childrel != onerel);
1494 0 : table_close(childrel, AccessShareLock);
1495 0 : continue;
1496 : }
1497 : }
1498 : else
1499 : {
1500 : /*
1501 : * ignore, but release the lock on it. don't try to unlock the
1502 : * passed-in relation
1503 : */
1504 : Assert(childrel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE);
1505 822 : if (childrel != onerel)
1506 78 : table_close(childrel, AccessShareLock);
1507 : else
1508 744 : table_close(childrel, NoLock);
1509 822 : continue;
1510 : }
1511 :
1512 : /* OK, we'll process this child */
1513 2246 : has_child = true;
1514 2246 : rels[nrels] = childrel;
1515 2246 : acquirefuncs[nrels] = acquirefunc;
1516 2246 : relblocks[nrels] = (double) relpages;
1517 2246 : totalblocks += (double) relpages;
1518 2246 : nrels++;
1519 : }
1520 :
1521 : /*
1522 : * If we don't have at least one child table to consider, fail. If the
1523 : * relation is a partitioned table, it's not counted as a child table.
1524 : */
1525 858 : if (!has_child)
1526 : {
1527 0 : ereport(elevel,
1528 : (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no analyzable child tables",
1529 : get_namespace_name(RelationGetNamespace(onerel)),
1530 : RelationGetRelationName(onerel))));
1531 0 : return 0;
1532 : }
1533 :
1534 : /*
1535 : * Now sample rows from each relation, proportionally to its fraction of
1536 : * the total block count. (This might be less than desirable if the child
1537 : * rels have radically different free-space percentages, but it's not
1538 : * clear that it's worth working harder.)
1539 : */
1540 858 : pgstat_progress_update_param(PROGRESS_ANALYZE_CHILD_TABLES_TOTAL,
1541 : nrels);
1542 858 : numrows = 0;
1543 3104 : for (i = 0; i < nrels; i++)
1544 : {
1545 2246 : Relation childrel = rels[i];
1546 2246 : AcquireSampleRowsFunc acquirefunc = acquirefuncs[i];
1547 2246 : double childblocks = relblocks[i];
1548 :
1549 : /*
1550 : * Report progress. The sampling function will normally report blocks
1551 : * done/total, but we need to reset them to 0 here, so that they don't
1552 : * show an old value until that.
1553 : */
1554 : {
1555 2246 : const int progress_index[] = {
1556 : PROGRESS_ANALYZE_CURRENT_CHILD_TABLE_RELID,
1557 : PROGRESS_ANALYZE_BLOCKS_DONE,
1558 : PROGRESS_ANALYZE_BLOCKS_TOTAL
1559 : };
1560 2246 : const int64 progress_vals[] = {
1561 2246 : RelationGetRelid(childrel),
1562 : 0,
1563 : 0,
1564 : };
1565 :
1566 2246 : pgstat_progress_update_multi_param(3, progress_index, progress_vals);
1567 : }
1568 :
1569 2246 : if (childblocks > 0)
1570 : {
1571 : int childtargrows;
1572 :
1573 2080 : childtargrows = (int) rint(targrows * childblocks / totalblocks);
1574 : /* Make sure we don't overrun due to roundoff error */
1575 2080 : childtargrows = Min(childtargrows, targrows - numrows);
1576 2080 : if (childtargrows > 0)
1577 : {
1578 : int childrows;
1579 : double trows,
1580 : tdrows;
1581 :
1582 : /* Fetch a random sample of the child's rows */
1583 2080 : childrows = (*acquirefunc) (childrel, elevel,
1584 2080 : rows + numrows, childtargrows,
1585 : &trows, &tdrows);
1586 :
1587 : /* We may need to convert from child's rowtype to parent's */
1588 2080 : if (childrows > 0 &&
1589 2080 : !equalRowTypes(RelationGetDescr(childrel),
1590 : RelationGetDescr(onerel)))
1591 : {
1592 : TupleConversionMap *map;
1593 :
1594 1990 : map = convert_tuples_by_name(RelationGetDescr(childrel),
1595 : RelationGetDescr(onerel));
1596 1990 : if (map != NULL)
1597 : {
1598 : int j;
1599 :
1600 106604 : for (j = 0; j < childrows; j++)
1601 : {
1602 : HeapTuple newtup;
1603 :
1604 106472 : newtup = execute_attr_map_tuple(rows[numrows + j], map);
1605 106472 : heap_freetuple(rows[numrows + j]);
1606 106472 : rows[numrows + j] = newtup;
1607 : }
1608 132 : free_conversion_map(map);
1609 : }
1610 : }
1611 :
1612 : /* And add to counts */
1613 2080 : numrows += childrows;
1614 2080 : *totalrows += trows;
1615 2080 : *totaldeadrows += tdrows;
1616 : }
1617 : }
1618 :
1619 : /*
1620 : * Note: we cannot release the child-table locks, since we may have
1621 : * pointers to their TOAST tables in the sampled rows.
1622 : */
1623 2246 : table_close(childrel, NoLock);
1624 2246 : pgstat_progress_update_param(PROGRESS_ANALYZE_CHILD_TABLES_DONE,
1625 2246 : i + 1);
1626 : }
1627 :
1628 858 : return numrows;
1629 : }
1630 :
1631 :
1632 : /*
1633 : * update_attstats() -- update attribute statistics for one relation
1634 : *
1635 : * Statistics are stored in several places: the pg_class row for the
1636 : * relation has stats about the whole relation, and there is a
1637 : * pg_statistic row for each (non-system) attribute that has ever
1638 : * been analyzed. The pg_class values are updated by VACUUM, not here.
1639 : *
1640 : * pg_statistic rows are just added or updated normally. This means
1641 : * that pg_statistic will probably contain some deleted rows at the
1642 : * completion of a vacuum cycle, unless it happens to get vacuumed last.
1643 : *
1644 : * To keep things simple, we punt for pg_statistic, and don't try
1645 : * to compute or store rows for pg_statistic itself in pg_statistic.
1646 : * This could possibly be made to work, but it's not worth the trouble.
1647 : * Note analyze_rel() has seen to it that we won't come here when
1648 : * vacuuming pg_statistic itself.
1649 : *
1650 : * Note: there would be a race condition here if two backends could
1651 : * ANALYZE the same table concurrently. Presently, we lock that out
1652 : * by taking a self-exclusive lock on the relation in analyze_rel().
1653 : */
1654 : static void
1655 25286 : update_attstats(Oid relid, bool inh, int natts, VacAttrStats **vacattrstats)
1656 : {
1657 : Relation sd;
1658 : int attno;
1659 25286 : CatalogIndexState indstate = NULL;
1660 :
1661 25286 : if (natts <= 0)
1662 13990 : return; /* nothing to do */
1663 :
1664 11296 : sd = table_open(StatisticRelationId, RowExclusiveLock);
1665 :
1666 98602 : for (attno = 0; attno < natts; attno++)
1667 : {
1668 87306 : VacAttrStats *stats = vacattrstats[attno];
1669 : HeapTuple stup,
1670 : oldtup;
1671 : int i,
1672 : k,
1673 : n;
1674 : Datum values[Natts_pg_statistic];
1675 : bool nulls[Natts_pg_statistic];
1676 : bool replaces[Natts_pg_statistic];
1677 :
1678 : /* Ignore attr if we weren't able to collect stats */
1679 87306 : if (!stats->stats_valid)
1680 6 : continue;
1681 :
1682 : /*
1683 : * Construct a new pg_statistic tuple
1684 : */
1685 2793600 : for (i = 0; i < Natts_pg_statistic; ++i)
1686 : {
1687 2706300 : nulls[i] = false;
1688 2706300 : replaces[i] = true;
1689 : }
1690 :
1691 87300 : values[Anum_pg_statistic_starelid - 1] = ObjectIdGetDatum(relid);
1692 87300 : values[Anum_pg_statistic_staattnum - 1] = Int16GetDatum(stats->tupattnum);
1693 87300 : values[Anum_pg_statistic_stainherit - 1] = BoolGetDatum(inh);
1694 87300 : values[Anum_pg_statistic_stanullfrac - 1] = Float4GetDatum(stats->stanullfrac);
1695 87300 : values[Anum_pg_statistic_stawidth - 1] = Int32GetDatum(stats->stawidth);
1696 87300 : values[Anum_pg_statistic_stadistinct - 1] = Float4GetDatum(stats->stadistinct);
1697 87300 : i = Anum_pg_statistic_stakind1 - 1;
1698 523800 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1699 : {
1700 436500 : values[i++] = Int16GetDatum(stats->stakind[k]); /* stakindN */
1701 : }
1702 87300 : i = Anum_pg_statistic_staop1 - 1;
1703 523800 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1704 : {
1705 436500 : values[i++] = ObjectIdGetDatum(stats->staop[k]); /* staopN */
1706 : }
1707 87300 : i = Anum_pg_statistic_stacoll1 - 1;
1708 523800 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1709 : {
1710 436500 : values[i++] = ObjectIdGetDatum(stats->stacoll[k]); /* stacollN */
1711 : }
1712 87300 : i = Anum_pg_statistic_stanumbers1 - 1;
1713 523800 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1714 : {
1715 436500 : if (stats->stanumbers[k] != NULL)
1716 : {
1717 136650 : int nnum = stats->numnumbers[k];
1718 136650 : Datum *numdatums = (Datum *) palloc(nnum * sizeof(Datum));
1719 : ArrayType *arry;
1720 :
1721 1140666 : for (n = 0; n < nnum; n++)
1722 1004016 : numdatums[n] = Float4GetDatum(stats->stanumbers[k][n]);
1723 136650 : arry = construct_array_builtin(numdatums, nnum, FLOAT4OID);
1724 136650 : values[i++] = PointerGetDatum(arry); /* stanumbersN */
1725 : }
1726 : else
1727 : {
1728 299850 : nulls[i] = true;
1729 299850 : values[i++] = (Datum) 0;
1730 : }
1731 : }
1732 87300 : i = Anum_pg_statistic_stavalues1 - 1;
1733 523800 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1734 : {
1735 436500 : if (stats->stavalues[k] != NULL)
1736 : {
1737 : ArrayType *arry;
1738 :
1739 96294 : arry = construct_array(stats->stavalues[k],
1740 : stats->numvalues[k],
1741 : stats->statypid[k],
1742 96294 : stats->statyplen[k],
1743 96294 : stats->statypbyval[k],
1744 96294 : stats->statypalign[k]);
1745 96294 : values[i++] = PointerGetDatum(arry); /* stavaluesN */
1746 : }
1747 : else
1748 : {
1749 340206 : nulls[i] = true;
1750 340206 : values[i++] = (Datum) 0;
1751 : }
1752 : }
1753 :
1754 : /* Is there already a pg_statistic tuple for this attribute? */
1755 174600 : oldtup = SearchSysCache3(STATRELATTINH,
1756 : ObjectIdGetDatum(relid),
1757 87300 : Int16GetDatum(stats->tupattnum),
1758 : BoolGetDatum(inh));
1759 :
1760 : /* Open index information when we know we need it */
1761 87300 : if (indstate == NULL)
1762 11290 : indstate = CatalogOpenIndexes(sd);
1763 :
1764 87300 : if (HeapTupleIsValid(oldtup))
1765 : {
1766 : /* Yes, replace it */
1767 39302 : stup = heap_modify_tuple(oldtup,
1768 : RelationGetDescr(sd),
1769 : values,
1770 : nulls,
1771 : replaces);
1772 39302 : ReleaseSysCache(oldtup);
1773 39302 : CatalogTupleUpdateWithInfo(sd, &stup->t_self, stup, indstate);
1774 : }
1775 : else
1776 : {
1777 : /* No, insert new tuple */
1778 47998 : stup = heap_form_tuple(RelationGetDescr(sd), values, nulls);
1779 47998 : CatalogTupleInsertWithInfo(sd, stup, indstate);
1780 : }
1781 :
1782 87300 : heap_freetuple(stup);
1783 : }
1784 :
1785 11296 : if (indstate != NULL)
1786 11290 : CatalogCloseIndexes(indstate);
1787 11296 : table_close(sd, RowExclusiveLock);
1788 : }
1789 :
1790 : /*
1791 : * Standard fetch function for use by compute_stats subroutines.
1792 : *
1793 : * This exists to provide some insulation between compute_stats routines
1794 : * and the actual storage of the sample data.
1795 : */
1796 : static Datum
1797 85497530 : std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1798 : {
1799 85497530 : int attnum = stats->tupattnum;
1800 85497530 : HeapTuple tuple = stats->rows[rownum];
1801 85497530 : TupleDesc tupDesc = stats->tupDesc;
1802 :
1803 85497530 : return heap_getattr(tuple, attnum, tupDesc, isNull);
1804 : }
1805 :
1806 : /*
1807 : * Fetch function for analyzing index expressions.
1808 : *
1809 : * We have not bothered to construct index tuples, instead the data is
1810 : * just in Datum arrays.
1811 : */
1812 : static Datum
1813 115022 : ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1814 : {
1815 : int i;
1816 :
1817 : /* exprvals and exprnulls are already offset for proper column */
1818 115022 : i = rownum * stats->rowstride;
1819 115022 : *isNull = stats->exprnulls[i];
1820 115022 : return stats->exprvals[i];
1821 : }
1822 :
1823 :
1824 : /*==========================================================================
1825 : *
1826 : * Code below this point represents the "standard" type-specific statistics
1827 : * analysis algorithms. This code can be replaced on a per-data-type basis
1828 : * by setting a nonzero value in pg_type.typanalyze.
1829 : *
1830 : *==========================================================================
1831 : */
1832 :
1833 :
1834 : /*
1835 : * To avoid consuming too much memory during analysis and/or too much space
1836 : * in the resulting pg_statistic rows, we ignore varlena datums that are wider
1837 : * than WIDTH_THRESHOLD (after detoasting!). This is legitimate for MCV
1838 : * and distinct-value calculations since a wide value is unlikely to be
1839 : * duplicated at all, much less be a most-common value. For the same reason,
1840 : * ignoring wide values will not affect our estimates of histogram bin
1841 : * boundaries very much.
1842 : */
1843 : #define WIDTH_THRESHOLD 1024
1844 :
1845 : #define swapInt(a,b) do {int _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1846 : #define swapDatum(a,b) do {Datum _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1847 :
1848 : /*
1849 : * Extra information used by the default analysis routines
1850 : */
1851 : typedef struct
1852 : {
1853 : int count; /* # of duplicates */
1854 : int first; /* values[] index of first occurrence */
1855 : } ScalarMCVItem;
1856 :
1857 : typedef struct
1858 : {
1859 : SortSupport ssup;
1860 : int *tupnoLink;
1861 : } CompareScalarsContext;
1862 :
1863 :
1864 : static void compute_trivial_stats(VacAttrStatsP stats,
1865 : AnalyzeAttrFetchFunc fetchfunc,
1866 : int samplerows,
1867 : double totalrows);
1868 : static void compute_distinct_stats(VacAttrStatsP stats,
1869 : AnalyzeAttrFetchFunc fetchfunc,
1870 : int samplerows,
1871 : double totalrows);
1872 : static void compute_scalar_stats(VacAttrStatsP stats,
1873 : AnalyzeAttrFetchFunc fetchfunc,
1874 : int samplerows,
1875 : double totalrows);
1876 : static int compare_scalars(const void *a, const void *b, void *arg);
1877 : static int compare_mcvs(const void *a, const void *b, void *arg);
1878 : static int analyze_mcv_list(int *mcv_counts,
1879 : int num_mcv,
1880 : double stadistinct,
1881 : double stanullfrac,
1882 : int samplerows,
1883 : double totalrows);
1884 :
1885 :
1886 : /*
1887 : * std_typanalyze -- the default type-specific typanalyze function
1888 : */
1889 : bool
1890 120098 : std_typanalyze(VacAttrStats *stats)
1891 : {
1892 : Oid ltopr;
1893 : Oid eqopr;
1894 : StdAnalyzeData *mystats;
1895 :
1896 : /* If the attstattarget column is negative, use the default value */
1897 120098 : if (stats->attstattarget < 0)
1898 119480 : stats->attstattarget = default_statistics_target;
1899 :
1900 : /* Look for default "<" and "=" operators for column's type */
1901 120098 : get_sort_group_operators(stats->attrtypid,
1902 : false, false, false,
1903 : <opr, &eqopr, NULL,
1904 : NULL);
1905 :
1906 : /* Save the operator info for compute_stats routines */
1907 120098 : mystats = (StdAnalyzeData *) palloc(sizeof(StdAnalyzeData));
1908 120098 : mystats->eqopr = eqopr;
1909 120098 : mystats->eqfunc = OidIsValid(eqopr) ? get_opcode(eqopr) : InvalidOid;
1910 120098 : mystats->ltopr = ltopr;
1911 120098 : stats->extra_data = mystats;
1912 :
1913 : /*
1914 : * Determine which standard statistics algorithm to use
1915 : */
1916 120098 : if (OidIsValid(eqopr) && OidIsValid(ltopr))
1917 : {
1918 : /* Seems to be a scalar datatype */
1919 116400 : stats->compute_stats = compute_scalar_stats;
1920 : /*--------------------
1921 : * The following choice of minrows is based on the paper
1922 : * "Random sampling for histogram construction: how much is enough?"
1923 : * by Surajit Chaudhuri, Rajeev Motwani and Vivek Narasayya, in
1924 : * Proceedings of ACM SIGMOD International Conference on Management
1925 : * of Data, 1998, Pages 436-447. Their Corollary 1 to Theorem 5
1926 : * says that for table size n, histogram size k, maximum relative
1927 : * error in bin size f, and error probability gamma, the minimum
1928 : * random sample size is
1929 : * r = 4 * k * ln(2*n/gamma) / f^2
1930 : * Taking f = 0.5, gamma = 0.01, n = 10^6 rows, we obtain
1931 : * r = 305.82 * k
1932 : * Note that because of the log function, the dependence on n is
1933 : * quite weak; even at n = 10^12, a 300*k sample gives <= 0.66
1934 : * bin size error with probability 0.99. So there's no real need to
1935 : * scale for n, which is a good thing because we don't necessarily
1936 : * know it at this point.
1937 : *--------------------
1938 : */
1939 116400 : stats->minrows = 300 * stats->attstattarget;
1940 : }
1941 3698 : else if (OidIsValid(eqopr))
1942 : {
1943 : /* We can still recognize distinct values */
1944 3162 : stats->compute_stats = compute_distinct_stats;
1945 : /* Might as well use the same minrows as above */
1946 3162 : stats->minrows = 300 * stats->attstattarget;
1947 : }
1948 : else
1949 : {
1950 : /* Can't do much but the trivial stuff */
1951 536 : stats->compute_stats = compute_trivial_stats;
1952 : /* Might as well use the same minrows as above */
1953 536 : stats->minrows = 300 * stats->attstattarget;
1954 : }
1955 :
1956 120098 : return true;
1957 : }
1958 :
1959 :
1960 : /*
1961 : * compute_trivial_stats() -- compute very basic column statistics
1962 : *
1963 : * We use this when we cannot find a hash "=" operator for the datatype.
1964 : *
1965 : * We determine the fraction of non-null rows and the average datum width.
1966 : */
1967 : static void
1968 370 : compute_trivial_stats(VacAttrStatsP stats,
1969 : AnalyzeAttrFetchFunc fetchfunc,
1970 : int samplerows,
1971 : double totalrows)
1972 : {
1973 : int i;
1974 370 : int null_cnt = 0;
1975 370 : int nonnull_cnt = 0;
1976 370 : double total_width = 0;
1977 740 : bool is_varlena = (!stats->attrtype->typbyval &&
1978 370 : stats->attrtype->typlen == -1);
1979 740 : bool is_varwidth = (!stats->attrtype->typbyval &&
1980 370 : stats->attrtype->typlen < 0);
1981 :
1982 1246368 : for (i = 0; i < samplerows; i++)
1983 : {
1984 : Datum value;
1985 : bool isnull;
1986 :
1987 1245998 : vacuum_delay_point(true);
1988 :
1989 1245998 : value = fetchfunc(stats, i, &isnull);
1990 :
1991 : /* Check for null/nonnull */
1992 1245998 : if (isnull)
1993 : {
1994 671736 : null_cnt++;
1995 671736 : continue;
1996 : }
1997 574262 : nonnull_cnt++;
1998 :
1999 : /*
2000 : * If it's a variable-width field, add up widths for average width
2001 : * calculation. Note that if the value is toasted, we use the toasted
2002 : * width. We don't bother with this calculation if it's a fixed-width
2003 : * type.
2004 : */
2005 574262 : if (is_varlena)
2006 : {
2007 103856 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2008 : }
2009 470406 : else if (is_varwidth)
2010 : {
2011 : /* must be cstring */
2012 0 : total_width += strlen(DatumGetCString(value)) + 1;
2013 : }
2014 : }
2015 :
2016 : /* We can only compute average width if we found some non-null values. */
2017 370 : if (nonnull_cnt > 0)
2018 : {
2019 182 : stats->stats_valid = true;
2020 : /* Do the simple null-frac and width stats */
2021 182 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2022 182 : if (is_varwidth)
2023 82 : stats->stawidth = total_width / (double) nonnull_cnt;
2024 : else
2025 100 : stats->stawidth = stats->attrtype->typlen;
2026 182 : stats->stadistinct = 0.0; /* "unknown" */
2027 : }
2028 188 : else if (null_cnt > 0)
2029 : {
2030 : /* We found only nulls; assume the column is entirely null */
2031 188 : stats->stats_valid = true;
2032 188 : stats->stanullfrac = 1.0;
2033 188 : if (is_varwidth)
2034 188 : stats->stawidth = 0; /* "unknown" */
2035 : else
2036 0 : stats->stawidth = stats->attrtype->typlen;
2037 188 : stats->stadistinct = 0.0; /* "unknown" */
2038 : }
2039 370 : }
2040 :
2041 :
2042 : /*
2043 : * compute_distinct_stats() -- compute column statistics including ndistinct
2044 : *
2045 : * We use this when we can find only an "=" operator for the datatype.
2046 : *
2047 : * We determine the fraction of non-null rows, the average width, the
2048 : * most common values, and the (estimated) number of distinct values.
2049 : *
2050 : * The most common values are determined by brute force: we keep a list
2051 : * of previously seen values, ordered by number of times seen, as we scan
2052 : * the samples. A newly seen value is inserted just after the last
2053 : * multiply-seen value, causing the bottommost (oldest) singly-seen value
2054 : * to drop off the list. The accuracy of this method, and also its cost,
2055 : * depend mainly on the length of the list we are willing to keep.
2056 : */
2057 : static void
2058 2318 : compute_distinct_stats(VacAttrStatsP stats,
2059 : AnalyzeAttrFetchFunc fetchfunc,
2060 : int samplerows,
2061 : double totalrows)
2062 : {
2063 : int i;
2064 2318 : int null_cnt = 0;
2065 2318 : int nonnull_cnt = 0;
2066 2318 : int toowide_cnt = 0;
2067 2318 : double total_width = 0;
2068 3920 : bool is_varlena = (!stats->attrtype->typbyval &&
2069 1602 : stats->attrtype->typlen == -1);
2070 3920 : bool is_varwidth = (!stats->attrtype->typbyval &&
2071 1602 : stats->attrtype->typlen < 0);
2072 : FmgrInfo f_cmpeq;
2073 : typedef struct
2074 : {
2075 : Datum value;
2076 : int count;
2077 : } TrackItem;
2078 : TrackItem *track;
2079 : int track_cnt,
2080 : track_max;
2081 2318 : int num_mcv = stats->attstattarget;
2082 2318 : StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2083 :
2084 : /*
2085 : * We track up to 2*n values for an n-element MCV list; but at least 10
2086 : */
2087 2318 : track_max = 2 * num_mcv;
2088 2318 : if (track_max < 10)
2089 78 : track_max = 10;
2090 2318 : track = (TrackItem *) palloc(track_max * sizeof(TrackItem));
2091 2318 : track_cnt = 0;
2092 :
2093 2318 : fmgr_info(mystats->eqfunc, &f_cmpeq);
2094 :
2095 1686074 : for (i = 0; i < samplerows; i++)
2096 : {
2097 : Datum value;
2098 : bool isnull;
2099 : bool match;
2100 : int firstcount1,
2101 : j;
2102 :
2103 1683756 : vacuum_delay_point(true);
2104 :
2105 1683756 : value = fetchfunc(stats, i, &isnull);
2106 :
2107 : /* Check for null/nonnull */
2108 1683756 : if (isnull)
2109 : {
2110 1407174 : null_cnt++;
2111 1407174 : continue;
2112 : }
2113 276582 : nonnull_cnt++;
2114 :
2115 : /*
2116 : * If it's a variable-width field, add up widths for average width
2117 : * calculation. Note that if the value is toasted, we use the toasted
2118 : * width. We don't bother with this calculation if it's a fixed-width
2119 : * type.
2120 : */
2121 276582 : if (is_varlena)
2122 : {
2123 97218 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2124 :
2125 : /*
2126 : * If the value is toasted, we want to detoast it just once to
2127 : * avoid repeated detoastings and resultant excess memory usage
2128 : * during the comparisons. Also, check to see if the value is
2129 : * excessively wide, and if so don't detoast at all --- just
2130 : * ignore the value.
2131 : */
2132 97218 : if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
2133 : {
2134 0 : toowide_cnt++;
2135 0 : continue;
2136 : }
2137 97218 : value = PointerGetDatum(PG_DETOAST_DATUM(value));
2138 : }
2139 179364 : else if (is_varwidth)
2140 : {
2141 : /* must be cstring */
2142 0 : total_width += strlen(DatumGetCString(value)) + 1;
2143 : }
2144 :
2145 : /*
2146 : * See if the value matches anything we're already tracking.
2147 : */
2148 276582 : match = false;
2149 276582 : firstcount1 = track_cnt;
2150 721918 : for (j = 0; j < track_cnt; j++)
2151 : {
2152 713788 : if (DatumGetBool(FunctionCall2Coll(&f_cmpeq,
2153 : stats->attrcollid,
2154 713788 : value, track[j].value)))
2155 : {
2156 268452 : match = true;
2157 268452 : break;
2158 : }
2159 445336 : if (j < firstcount1 && track[j].count == 1)
2160 5872 : firstcount1 = j;
2161 : }
2162 :
2163 276582 : if (match)
2164 : {
2165 : /* Found a match */
2166 268452 : track[j].count++;
2167 : /* This value may now need to "bubble up" in the track list */
2168 278834 : while (j > 0 && track[j].count > track[j - 1].count)
2169 : {
2170 10382 : swapDatum(track[j].value, track[j - 1].value);
2171 10382 : swapInt(track[j].count, track[j - 1].count);
2172 10382 : j--;
2173 : }
2174 : }
2175 : else
2176 : {
2177 : /* No match. Insert at head of count-1 list */
2178 8130 : if (track_cnt < track_max)
2179 7556 : track_cnt++;
2180 242654 : for (j = track_cnt - 1; j > firstcount1; j--)
2181 : {
2182 234524 : track[j].value = track[j - 1].value;
2183 234524 : track[j].count = track[j - 1].count;
2184 : }
2185 8130 : if (firstcount1 < track_cnt)
2186 : {
2187 8130 : track[firstcount1].value = value;
2188 8130 : track[firstcount1].count = 1;
2189 : }
2190 : }
2191 : }
2192 :
2193 : /* We can only compute real stats if we found some non-null values. */
2194 2318 : if (nonnull_cnt > 0)
2195 : {
2196 : int nmultiple,
2197 : summultiple;
2198 :
2199 1696 : stats->stats_valid = true;
2200 : /* Do the simple null-frac and width stats */
2201 1696 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2202 1696 : if (is_varwidth)
2203 980 : stats->stawidth = total_width / (double) nonnull_cnt;
2204 : else
2205 716 : stats->stawidth = stats->attrtype->typlen;
2206 :
2207 : /* Count the number of values we found multiple times */
2208 1696 : summultiple = 0;
2209 6590 : for (nmultiple = 0; nmultiple < track_cnt; nmultiple++)
2210 : {
2211 5758 : if (track[nmultiple].count == 1)
2212 864 : break;
2213 4894 : summultiple += track[nmultiple].count;
2214 : }
2215 :
2216 1696 : if (nmultiple == 0)
2217 : {
2218 : /*
2219 : * If we found no repeated non-null values, assume it's a unique
2220 : * column; but be sure to discount for any nulls we found.
2221 : */
2222 196 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2223 : }
2224 1500 : else if (track_cnt < track_max && toowide_cnt == 0 &&
2225 : nmultiple == track_cnt)
2226 : {
2227 : /*
2228 : * Our track list includes every value in the sample, and every
2229 : * value appeared more than once. Assume the column has just
2230 : * these values. (This case is meant to address columns with
2231 : * small, fixed sets of possible values, such as boolean or enum
2232 : * columns. If there are any values that appear just once in the
2233 : * sample, including too-wide values, we should assume that that's
2234 : * not what we're dealing with.)
2235 : */
2236 832 : stats->stadistinct = track_cnt;
2237 : }
2238 : else
2239 : {
2240 : /*----------
2241 : * Estimate the number of distinct values using the estimator
2242 : * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2243 : * n*d / (n - f1 + f1*n/N)
2244 : * where f1 is the number of distinct values that occurred
2245 : * exactly once in our sample of n rows (from a total of N),
2246 : * and d is the total number of distinct values in the sample.
2247 : * This is their Duj1 estimator; the other estimators they
2248 : * recommend are considerably more complex, and are numerically
2249 : * very unstable when n is much smaller than N.
2250 : *
2251 : * In this calculation, we consider only non-nulls. We used to
2252 : * include rows with null values in the n and N counts, but that
2253 : * leads to inaccurate answers in columns with many nulls, and
2254 : * it's intuitively bogus anyway considering the desired result is
2255 : * the number of distinct non-null values.
2256 : *
2257 : * We assume (not very reliably!) that all the multiply-occurring
2258 : * values are reflected in the final track[] list, and the other
2259 : * nonnull values all appeared but once. (XXX this usually
2260 : * results in a drastic overestimate of ndistinct. Can we do
2261 : * any better?)
2262 : *----------
2263 : */
2264 668 : int f1 = nonnull_cnt - summultiple;
2265 668 : int d = f1 + nmultiple;
2266 668 : double n = samplerows - null_cnt;
2267 668 : double N = totalrows * (1.0 - stats->stanullfrac);
2268 : double stadistinct;
2269 :
2270 : /* N == 0 shouldn't happen, but just in case ... */
2271 668 : if (N > 0)
2272 668 : stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2273 : else
2274 0 : stadistinct = 0;
2275 :
2276 : /* Clamp to sane range in case of roundoff error */
2277 668 : if (stadistinct < d)
2278 182 : stadistinct = d;
2279 668 : if (stadistinct > N)
2280 0 : stadistinct = N;
2281 : /* And round to integer */
2282 668 : stats->stadistinct = floor(stadistinct + 0.5);
2283 : }
2284 :
2285 : /*
2286 : * If we estimated the number of distinct values at more than 10% of
2287 : * the total row count (a very arbitrary limit), then assume that
2288 : * stadistinct should scale with the row count rather than be a fixed
2289 : * value.
2290 : */
2291 1696 : if (stats->stadistinct > 0.1 * totalrows)
2292 392 : stats->stadistinct = -(stats->stadistinct / totalrows);
2293 :
2294 : /*
2295 : * Decide how many values are worth storing as most-common values. If
2296 : * we are able to generate a complete MCV list (all the values in the
2297 : * sample will fit, and we think these are all the ones in the table),
2298 : * then do so. Otherwise, store only those values that are
2299 : * significantly more common than the values not in the list.
2300 : *
2301 : * Note: the first of these cases is meant to address columns with
2302 : * small, fixed sets of possible values, such as boolean or enum
2303 : * columns. If we can *completely* represent the column population by
2304 : * an MCV list that will fit into the stats target, then we should do
2305 : * so and thus provide the planner with complete information. But if
2306 : * the MCV list is not complete, it's generally worth being more
2307 : * selective, and not just filling it all the way up to the stats
2308 : * target.
2309 : */
2310 1696 : if (track_cnt < track_max && toowide_cnt == 0 &&
2311 1682 : stats->stadistinct > 0 &&
2312 : track_cnt <= num_mcv)
2313 : {
2314 : /* Track list includes all values seen, and all will fit */
2315 1078 : num_mcv = track_cnt;
2316 : }
2317 : else
2318 : {
2319 : int *mcv_counts;
2320 :
2321 : /* Incomplete list; decide how many values are worth keeping */
2322 618 : if (num_mcv > track_cnt)
2323 552 : num_mcv = track_cnt;
2324 :
2325 618 : if (num_mcv > 0)
2326 : {
2327 618 : mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2328 2414 : for (i = 0; i < num_mcv; i++)
2329 1796 : mcv_counts[i] = track[i].count;
2330 :
2331 618 : num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2332 618 : stats->stadistinct,
2333 618 : stats->stanullfrac,
2334 : samplerows, totalrows);
2335 : }
2336 : }
2337 :
2338 : /* Generate MCV slot entry */
2339 1696 : if (num_mcv > 0)
2340 : {
2341 : MemoryContext old_context;
2342 : Datum *mcv_values;
2343 : float4 *mcv_freqs;
2344 :
2345 : /* Must copy the target values into anl_context */
2346 1688 : old_context = MemoryContextSwitchTo(stats->anl_context);
2347 1688 : mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2348 1688 : mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2349 8274 : for (i = 0; i < num_mcv; i++)
2350 : {
2351 13172 : mcv_values[i] = datumCopy(track[i].value,
2352 6586 : stats->attrtype->typbyval,
2353 6586 : stats->attrtype->typlen);
2354 6586 : mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2355 : }
2356 1688 : MemoryContextSwitchTo(old_context);
2357 :
2358 1688 : stats->stakind[0] = STATISTIC_KIND_MCV;
2359 1688 : stats->staop[0] = mystats->eqopr;
2360 1688 : stats->stacoll[0] = stats->attrcollid;
2361 1688 : stats->stanumbers[0] = mcv_freqs;
2362 1688 : stats->numnumbers[0] = num_mcv;
2363 1688 : stats->stavalues[0] = mcv_values;
2364 1688 : stats->numvalues[0] = num_mcv;
2365 :
2366 : /*
2367 : * Accept the defaults for stats->statypid and others. They have
2368 : * been set before we were called (see vacuum.h)
2369 : */
2370 : }
2371 : }
2372 622 : else if (null_cnt > 0)
2373 : {
2374 : /* We found only nulls; assume the column is entirely null */
2375 622 : stats->stats_valid = true;
2376 622 : stats->stanullfrac = 1.0;
2377 622 : if (is_varwidth)
2378 622 : stats->stawidth = 0; /* "unknown" */
2379 : else
2380 0 : stats->stawidth = stats->attrtype->typlen;
2381 622 : stats->stadistinct = 0.0; /* "unknown" */
2382 : }
2383 :
2384 : /* We don't need to bother cleaning up any of our temporary palloc's */
2385 2318 : }
2386 :
2387 :
2388 : /*
2389 : * compute_scalar_stats() -- compute column statistics
2390 : *
2391 : * We use this when we can find "=" and "<" operators for the datatype.
2392 : *
2393 : * We determine the fraction of non-null rows, the average width, the
2394 : * most common values, the (estimated) number of distinct values, the
2395 : * distribution histogram, and the correlation of physical to logical order.
2396 : *
2397 : * The desired stats can be determined fairly easily after sorting the
2398 : * data values into order.
2399 : */
2400 : static void
2401 84878 : compute_scalar_stats(VacAttrStatsP stats,
2402 : AnalyzeAttrFetchFunc fetchfunc,
2403 : int samplerows,
2404 : double totalrows)
2405 : {
2406 : int i;
2407 84878 : int null_cnt = 0;
2408 84878 : int nonnull_cnt = 0;
2409 84878 : int toowide_cnt = 0;
2410 84878 : double total_width = 0;
2411 105604 : bool is_varlena = (!stats->attrtype->typbyval &&
2412 20726 : stats->attrtype->typlen == -1);
2413 105604 : bool is_varwidth = (!stats->attrtype->typbyval &&
2414 20726 : stats->attrtype->typlen < 0);
2415 : double corr_xysum;
2416 : SortSupportData ssup;
2417 : ScalarItem *values;
2418 84878 : int values_cnt = 0;
2419 : int *tupnoLink;
2420 : ScalarMCVItem *track;
2421 84878 : int track_cnt = 0;
2422 84878 : int num_mcv = stats->attstattarget;
2423 84878 : int num_bins = stats->attstattarget;
2424 84878 : StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2425 :
2426 84878 : values = (ScalarItem *) palloc(samplerows * sizeof(ScalarItem));
2427 84878 : tupnoLink = (int *) palloc(samplerows * sizeof(int));
2428 84878 : track = (ScalarMCVItem *) palloc(num_mcv * sizeof(ScalarMCVItem));
2429 :
2430 84878 : memset(&ssup, 0, sizeof(ssup));
2431 84878 : ssup.ssup_cxt = CurrentMemoryContext;
2432 84878 : ssup.ssup_collation = stats->attrcollid;
2433 84878 : ssup.ssup_nulls_first = false;
2434 :
2435 : /*
2436 : * For now, don't perform abbreviated key conversion, because full values
2437 : * are required for MCV slot generation. Supporting that optimization
2438 : * would necessitate teaching compare_scalars() to call a tie-breaker.
2439 : */
2440 84878 : ssup.abbreviate = false;
2441 :
2442 84878 : PrepareSortSupportFromOrderingOp(mystats->ltopr, &ssup);
2443 :
2444 : /* Initial scan to find sortable values */
2445 77872764 : for (i = 0; i < samplerows; i++)
2446 : {
2447 : Datum value;
2448 : bool isnull;
2449 :
2450 77787886 : vacuum_delay_point(true);
2451 :
2452 77787886 : value = fetchfunc(stats, i, &isnull);
2453 :
2454 : /* Check for null/nonnull */
2455 77787886 : if (isnull)
2456 : {
2457 10078898 : null_cnt++;
2458 10117174 : continue;
2459 : }
2460 67708988 : nonnull_cnt++;
2461 :
2462 : /*
2463 : * If it's a variable-width field, add up widths for average width
2464 : * calculation. Note that if the value is toasted, we use the toasted
2465 : * width. We don't bother with this calculation if it's a fixed-width
2466 : * type.
2467 : */
2468 67708988 : if (is_varlena)
2469 : {
2470 8193932 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2471 :
2472 : /*
2473 : * If the value is toasted, we want to detoast it just once to
2474 : * avoid repeated detoastings and resultant excess memory usage
2475 : * during the comparisons. Also, check to see if the value is
2476 : * excessively wide, and if so don't detoast at all --- just
2477 : * ignore the value.
2478 : */
2479 8193932 : if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
2480 : {
2481 38276 : toowide_cnt++;
2482 38276 : continue;
2483 : }
2484 8155656 : value = PointerGetDatum(PG_DETOAST_DATUM(value));
2485 : }
2486 59515056 : else if (is_varwidth)
2487 : {
2488 : /* must be cstring */
2489 0 : total_width += strlen(DatumGetCString(value)) + 1;
2490 : }
2491 :
2492 : /* Add it to the list to be sorted */
2493 67670712 : values[values_cnt].value = value;
2494 67670712 : values[values_cnt].tupno = values_cnt;
2495 67670712 : tupnoLink[values_cnt] = values_cnt;
2496 67670712 : values_cnt++;
2497 : }
2498 :
2499 : /* We can only compute real stats if we found some sortable values. */
2500 84878 : if (values_cnt > 0)
2501 : {
2502 : int ndistinct, /* # distinct values in sample */
2503 : nmultiple, /* # that appear multiple times */
2504 : num_hist,
2505 : dups_cnt;
2506 79320 : int slot_idx = 0;
2507 : CompareScalarsContext cxt;
2508 :
2509 : /* Sort the collected values */
2510 79320 : cxt.ssup = &ssup;
2511 79320 : cxt.tupnoLink = tupnoLink;
2512 79320 : qsort_interruptible(values, values_cnt, sizeof(ScalarItem),
2513 : compare_scalars, &cxt);
2514 :
2515 : /*
2516 : * Now scan the values in order, find the most common ones, and also
2517 : * accumulate ordering-correlation statistics.
2518 : *
2519 : * To determine which are most common, we first have to count the
2520 : * number of duplicates of each value. The duplicates are adjacent in
2521 : * the sorted list, so a brute-force approach is to compare successive
2522 : * datum values until we find two that are not equal. However, that
2523 : * requires N-1 invocations of the datum comparison routine, which are
2524 : * completely redundant with work that was done during the sort. (The
2525 : * sort algorithm must at some point have compared each pair of items
2526 : * that are adjacent in the sorted order; otherwise it could not know
2527 : * that it's ordered the pair correctly.) We exploit this by having
2528 : * compare_scalars remember the highest tupno index that each
2529 : * ScalarItem has been found equal to. At the end of the sort, a
2530 : * ScalarItem's tupnoLink will still point to itself if and only if it
2531 : * is the last item of its group of duplicates (since the group will
2532 : * be ordered by tupno).
2533 : */
2534 79320 : corr_xysum = 0;
2535 79320 : ndistinct = 0;
2536 79320 : nmultiple = 0;
2537 79320 : dups_cnt = 0;
2538 67750032 : for (i = 0; i < values_cnt; i++)
2539 : {
2540 67670712 : int tupno = values[i].tupno;
2541 :
2542 67670712 : corr_xysum += ((double) i) * ((double) tupno);
2543 67670712 : dups_cnt++;
2544 67670712 : if (tupnoLink[tupno] == tupno)
2545 : {
2546 : /* Reached end of duplicates of this value */
2547 14126896 : ndistinct++;
2548 14126896 : if (dups_cnt > 1)
2549 : {
2550 1203848 : nmultiple++;
2551 1203848 : if (track_cnt < num_mcv ||
2552 503974 : dups_cnt > track[track_cnt - 1].count)
2553 : {
2554 : /*
2555 : * Found a new item for the mcv list; find its
2556 : * position, bubbling down old items if needed. Loop
2557 : * invariant is that j points at an empty/ replaceable
2558 : * slot.
2559 : */
2560 : int j;
2561 :
2562 806228 : if (track_cnt < num_mcv)
2563 699874 : track_cnt++;
2564 10742838 : for (j = track_cnt - 1; j > 0; j--)
2565 : {
2566 10653912 : if (dups_cnt <= track[j - 1].count)
2567 717302 : break;
2568 9936610 : track[j].count = track[j - 1].count;
2569 9936610 : track[j].first = track[j - 1].first;
2570 : }
2571 806228 : track[j].count = dups_cnt;
2572 806228 : track[j].first = i + 1 - dups_cnt;
2573 : }
2574 : }
2575 14126896 : dups_cnt = 0;
2576 : }
2577 : }
2578 :
2579 79320 : stats->stats_valid = true;
2580 : /* Do the simple null-frac and width stats */
2581 79320 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2582 79320 : if (is_varwidth)
2583 11542 : stats->stawidth = total_width / (double) nonnull_cnt;
2584 : else
2585 67778 : stats->stawidth = stats->attrtype->typlen;
2586 :
2587 79320 : if (nmultiple == 0)
2588 : {
2589 : /*
2590 : * If we found no repeated non-null values, assume it's a unique
2591 : * column; but be sure to discount for any nulls we found.
2592 : */
2593 20532 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2594 : }
2595 58788 : else if (toowide_cnt == 0 && nmultiple == ndistinct)
2596 : {
2597 : /*
2598 : * Every value in the sample appeared more than once. Assume the
2599 : * column has just these values. (This case is meant to address
2600 : * columns with small, fixed sets of possible values, such as
2601 : * boolean or enum columns. If there are any values that appear
2602 : * just once in the sample, including too-wide values, we should
2603 : * assume that that's not what we're dealing with.)
2604 : */
2605 35930 : stats->stadistinct = ndistinct;
2606 : }
2607 : else
2608 : {
2609 : /*----------
2610 : * Estimate the number of distinct values using the estimator
2611 : * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2612 : * n*d / (n - f1 + f1*n/N)
2613 : * where f1 is the number of distinct values that occurred
2614 : * exactly once in our sample of n rows (from a total of N),
2615 : * and d is the total number of distinct values in the sample.
2616 : * This is their Duj1 estimator; the other estimators they
2617 : * recommend are considerably more complex, and are numerically
2618 : * very unstable when n is much smaller than N.
2619 : *
2620 : * In this calculation, we consider only non-nulls. We used to
2621 : * include rows with null values in the n and N counts, but that
2622 : * leads to inaccurate answers in columns with many nulls, and
2623 : * it's intuitively bogus anyway considering the desired result is
2624 : * the number of distinct non-null values.
2625 : *
2626 : * Overwidth values are assumed to have been distinct.
2627 : *----------
2628 : */
2629 22858 : int f1 = ndistinct - nmultiple + toowide_cnt;
2630 22858 : int d = f1 + nmultiple;
2631 22858 : double n = samplerows - null_cnt;
2632 22858 : double N = totalrows * (1.0 - stats->stanullfrac);
2633 : double stadistinct;
2634 :
2635 : /* N == 0 shouldn't happen, but just in case ... */
2636 22858 : if (N > 0)
2637 22858 : stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2638 : else
2639 0 : stadistinct = 0;
2640 :
2641 : /* Clamp to sane range in case of roundoff error */
2642 22858 : if (stadistinct < d)
2643 978 : stadistinct = d;
2644 22858 : if (stadistinct > N)
2645 0 : stadistinct = N;
2646 : /* And round to integer */
2647 22858 : stats->stadistinct = floor(stadistinct + 0.5);
2648 : }
2649 :
2650 : /*
2651 : * If we estimated the number of distinct values at more than 10% of
2652 : * the total row count (a very arbitrary limit), then assume that
2653 : * stadistinct should scale with the row count rather than be a fixed
2654 : * value.
2655 : */
2656 79320 : if (stats->stadistinct > 0.1 * totalrows)
2657 16884 : stats->stadistinct = -(stats->stadistinct / totalrows);
2658 :
2659 : /*
2660 : * Decide how many values are worth storing as most-common values. If
2661 : * we are able to generate a complete MCV list (all the values in the
2662 : * sample will fit, and we think these are all the ones in the table),
2663 : * then do so. Otherwise, store only those values that are
2664 : * significantly more common than the values not in the list.
2665 : *
2666 : * Note: the first of these cases is meant to address columns with
2667 : * small, fixed sets of possible values, such as boolean or enum
2668 : * columns. If we can *completely* represent the column population by
2669 : * an MCV list that will fit into the stats target, then we should do
2670 : * so and thus provide the planner with complete information. But if
2671 : * the MCV list is not complete, it's generally worth being more
2672 : * selective, and not just filling it all the way up to the stats
2673 : * target.
2674 : */
2675 79320 : if (track_cnt == ndistinct && toowide_cnt == 0 &&
2676 35242 : stats->stadistinct > 0 &&
2677 : track_cnt <= num_mcv)
2678 : {
2679 : /* Track list includes all values seen, and all will fit */
2680 31580 : num_mcv = track_cnt;
2681 : }
2682 : else
2683 : {
2684 : int *mcv_counts;
2685 :
2686 : /* Incomplete list; decide how many values are worth keeping */
2687 47740 : if (num_mcv > track_cnt)
2688 43720 : num_mcv = track_cnt;
2689 :
2690 47740 : if (num_mcv > 0)
2691 : {
2692 27208 : mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2693 558690 : for (i = 0; i < num_mcv; i++)
2694 531482 : mcv_counts[i] = track[i].count;
2695 :
2696 27208 : num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2697 27208 : stats->stadistinct,
2698 27208 : stats->stanullfrac,
2699 : samplerows, totalrows);
2700 : }
2701 : }
2702 :
2703 : /* Generate MCV slot entry */
2704 79320 : if (num_mcv > 0)
2705 : {
2706 : MemoryContext old_context;
2707 : Datum *mcv_values;
2708 : float4 *mcv_freqs;
2709 :
2710 : /* Must copy the target values into anl_context */
2711 58746 : old_context = MemoryContextSwitchTo(stats->anl_context);
2712 58746 : mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2713 58746 : mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2714 758410 : for (i = 0; i < num_mcv; i++)
2715 : {
2716 1399328 : mcv_values[i] = datumCopy(values[track[i].first].value,
2717 699664 : stats->attrtype->typbyval,
2718 699664 : stats->attrtype->typlen);
2719 699664 : mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2720 : }
2721 58746 : MemoryContextSwitchTo(old_context);
2722 :
2723 58746 : stats->stakind[slot_idx] = STATISTIC_KIND_MCV;
2724 58746 : stats->staop[slot_idx] = mystats->eqopr;
2725 58746 : stats->stacoll[slot_idx] = stats->attrcollid;
2726 58746 : stats->stanumbers[slot_idx] = mcv_freqs;
2727 58746 : stats->numnumbers[slot_idx] = num_mcv;
2728 58746 : stats->stavalues[slot_idx] = mcv_values;
2729 58746 : stats->numvalues[slot_idx] = num_mcv;
2730 :
2731 : /*
2732 : * Accept the defaults for stats->statypid and others. They have
2733 : * been set before we were called (see vacuum.h)
2734 : */
2735 58746 : slot_idx++;
2736 : }
2737 :
2738 : /*
2739 : * Generate a histogram slot entry if there are at least two distinct
2740 : * values not accounted for in the MCV list. (This ensures the
2741 : * histogram won't collapse to empty or a singleton.)
2742 : */
2743 79320 : num_hist = ndistinct - num_mcv;
2744 79320 : if (num_hist > num_bins)
2745 12882 : num_hist = num_bins + 1;
2746 79320 : if (num_hist >= 2)
2747 : {
2748 : MemoryContext old_context;
2749 : Datum *hist_values;
2750 : int nvals;
2751 : int pos,
2752 : posfrac,
2753 : delta,
2754 : deltafrac;
2755 :
2756 : /* Sort the MCV items into position order to speed next loop */
2757 35032 : qsort_interruptible(track, num_mcv, sizeof(ScalarMCVItem),
2758 : compare_mcvs, NULL);
2759 :
2760 : /*
2761 : * Collapse out the MCV items from the values[] array.
2762 : *
2763 : * Note we destroy the values[] array here... but we don't need it
2764 : * for anything more. We do, however, still need values_cnt.
2765 : * nvals will be the number of remaining entries in values[].
2766 : */
2767 35032 : if (num_mcv > 0)
2768 : {
2769 : int src,
2770 : dest;
2771 : int j;
2772 :
2773 19150 : src = dest = 0;
2774 19150 : j = 0; /* index of next interesting MCV item */
2775 692324 : while (src < values_cnt)
2776 : {
2777 : int ncopy;
2778 :
2779 673174 : if (j < num_mcv)
2780 : {
2781 658596 : int first = track[j].first;
2782 :
2783 658596 : if (src >= first)
2784 : {
2785 : /* advance past this MCV item */
2786 479890 : src = first + track[j].count;
2787 479890 : j++;
2788 479890 : continue;
2789 : }
2790 178706 : ncopy = first - src;
2791 : }
2792 : else
2793 14578 : ncopy = values_cnt - src;
2794 193284 : memmove(&values[dest], &values[src],
2795 : ncopy * sizeof(ScalarItem));
2796 193284 : src += ncopy;
2797 193284 : dest += ncopy;
2798 : }
2799 19150 : nvals = dest;
2800 : }
2801 : else
2802 15882 : nvals = values_cnt;
2803 : Assert(nvals >= num_hist);
2804 :
2805 : /* Must copy the target values into anl_context */
2806 35032 : old_context = MemoryContextSwitchTo(stats->anl_context);
2807 35032 : hist_values = (Datum *) palloc(num_hist * sizeof(Datum));
2808 :
2809 : /*
2810 : * The object of this loop is to copy the first and last values[]
2811 : * entries along with evenly-spaced values in between. So the
2812 : * i'th value is values[(i * (nvals - 1)) / (num_hist - 1)]. But
2813 : * computing that subscript directly risks integer overflow when
2814 : * the stats target is more than a couple thousand. Instead we
2815 : * add (nvals - 1) / (num_hist - 1) to pos at each step, tracking
2816 : * the integral and fractional parts of the sum separately.
2817 : */
2818 35032 : delta = (nvals - 1) / (num_hist - 1);
2819 35032 : deltafrac = (nvals - 1) % (num_hist - 1);
2820 35032 : pos = posfrac = 0;
2821 :
2822 1833814 : for (i = 0; i < num_hist; i++)
2823 : {
2824 3597564 : hist_values[i] = datumCopy(values[pos].value,
2825 1798782 : stats->attrtype->typbyval,
2826 1798782 : stats->attrtype->typlen);
2827 1798782 : pos += delta;
2828 1798782 : posfrac += deltafrac;
2829 1798782 : if (posfrac >= (num_hist - 1))
2830 : {
2831 : /* fractional part exceeds 1, carry to integer part */
2832 584138 : pos++;
2833 584138 : posfrac -= (num_hist - 1);
2834 : }
2835 : }
2836 :
2837 35032 : MemoryContextSwitchTo(old_context);
2838 :
2839 35032 : stats->stakind[slot_idx] = STATISTIC_KIND_HISTOGRAM;
2840 35032 : stats->staop[slot_idx] = mystats->ltopr;
2841 35032 : stats->stacoll[slot_idx] = stats->attrcollid;
2842 35032 : stats->stavalues[slot_idx] = hist_values;
2843 35032 : stats->numvalues[slot_idx] = num_hist;
2844 :
2845 : /*
2846 : * Accept the defaults for stats->statypid and others. They have
2847 : * been set before we were called (see vacuum.h)
2848 : */
2849 35032 : slot_idx++;
2850 : }
2851 :
2852 : /* Generate a correlation entry if there are multiple values */
2853 79320 : if (values_cnt > 1)
2854 : {
2855 : MemoryContext old_context;
2856 : float4 *corrs;
2857 : double corr_xsum,
2858 : corr_x2sum;
2859 :
2860 : /* Must copy the target values into anl_context */
2861 74628 : old_context = MemoryContextSwitchTo(stats->anl_context);
2862 74628 : corrs = (float4 *) palloc(sizeof(float4));
2863 74628 : MemoryContextSwitchTo(old_context);
2864 :
2865 : /*----------
2866 : * Since we know the x and y value sets are both
2867 : * 0, 1, ..., values_cnt-1
2868 : * we have sum(x) = sum(y) =
2869 : * (values_cnt-1)*values_cnt / 2
2870 : * and sum(x^2) = sum(y^2) =
2871 : * (values_cnt-1)*values_cnt*(2*values_cnt-1) / 6.
2872 : *----------
2873 : */
2874 74628 : corr_xsum = ((double) (values_cnt - 1)) *
2875 74628 : ((double) values_cnt) / 2.0;
2876 74628 : corr_x2sum = ((double) (values_cnt - 1)) *
2877 74628 : ((double) values_cnt) * (double) (2 * values_cnt - 1) / 6.0;
2878 :
2879 : /* And the correlation coefficient reduces to */
2880 74628 : corrs[0] = (values_cnt * corr_xysum - corr_xsum * corr_xsum) /
2881 74628 : (values_cnt * corr_x2sum - corr_xsum * corr_xsum);
2882 :
2883 74628 : stats->stakind[slot_idx] = STATISTIC_KIND_CORRELATION;
2884 74628 : stats->staop[slot_idx] = mystats->ltopr;
2885 74628 : stats->stacoll[slot_idx] = stats->attrcollid;
2886 74628 : stats->stanumbers[slot_idx] = corrs;
2887 74628 : stats->numnumbers[slot_idx] = 1;
2888 74628 : slot_idx++;
2889 : }
2890 : }
2891 5558 : else if (nonnull_cnt > 0)
2892 : {
2893 : /* We found some non-null values, but they were all too wide */
2894 : Assert(nonnull_cnt == toowide_cnt);
2895 354 : stats->stats_valid = true;
2896 : /* Do the simple null-frac and width stats */
2897 354 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2898 354 : if (is_varwidth)
2899 354 : stats->stawidth = total_width / (double) nonnull_cnt;
2900 : else
2901 0 : stats->stawidth = stats->attrtype->typlen;
2902 : /* Assume all too-wide values are distinct, so it's a unique column */
2903 354 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2904 : }
2905 5204 : else if (null_cnt > 0)
2906 : {
2907 : /* We found only nulls; assume the column is entirely null */
2908 5204 : stats->stats_valid = true;
2909 5204 : stats->stanullfrac = 1.0;
2910 5204 : if (is_varwidth)
2911 4520 : stats->stawidth = 0; /* "unknown" */
2912 : else
2913 684 : stats->stawidth = stats->attrtype->typlen;
2914 5204 : stats->stadistinct = 0.0; /* "unknown" */
2915 : }
2916 :
2917 : /* We don't need to bother cleaning up any of our temporary palloc's */
2918 84878 : }
2919 :
2920 : /*
2921 : * Comparator for sorting ScalarItems
2922 : *
2923 : * Aside from sorting the items, we update the tupnoLink[] array
2924 : * whenever two ScalarItems are found to contain equal datums. The array
2925 : * is indexed by tupno; for each ScalarItem, it contains the highest
2926 : * tupno that that item's datum has been found to be equal to. This allows
2927 : * us to avoid additional comparisons in compute_scalar_stats().
2928 : */
2929 : static int
2930 634265120 : compare_scalars(const void *a, const void *b, void *arg)
2931 : {
2932 634265120 : Datum da = ((const ScalarItem *) a)->value;
2933 634265120 : int ta = ((const ScalarItem *) a)->tupno;
2934 634265120 : Datum db = ((const ScalarItem *) b)->value;
2935 634265120 : int tb = ((const ScalarItem *) b)->tupno;
2936 634265120 : CompareScalarsContext *cxt = (CompareScalarsContext *) arg;
2937 : int compare;
2938 :
2939 634265120 : compare = ApplySortComparator(da, false, db, false, cxt->ssup);
2940 634265120 : if (compare != 0)
2941 239759852 : return compare;
2942 :
2943 : /*
2944 : * The two datums are equal, so update cxt->tupnoLink[].
2945 : */
2946 394505268 : if (cxt->tupnoLink[ta] < tb)
2947 56021438 : cxt->tupnoLink[ta] = tb;
2948 394505268 : if (cxt->tupnoLink[tb] < ta)
2949 3884570 : cxt->tupnoLink[tb] = ta;
2950 :
2951 : /*
2952 : * For equal datums, sort by tupno
2953 : */
2954 394505268 : return ta - tb;
2955 : }
2956 :
2957 : /*
2958 : * Comparator for sorting ScalarMCVItems by position
2959 : */
2960 : static int
2961 2505532 : compare_mcvs(const void *a, const void *b, void *arg)
2962 : {
2963 2505532 : int da = ((const ScalarMCVItem *) a)->first;
2964 2505532 : int db = ((const ScalarMCVItem *) b)->first;
2965 :
2966 2505532 : return da - db;
2967 : }
2968 :
2969 : /*
2970 : * Analyze the list of common values in the sample and decide how many are
2971 : * worth storing in the table's MCV list.
2972 : *
2973 : * mcv_counts is assumed to be a list of the counts of the most common values
2974 : * seen in the sample, starting with the most common. The return value is the
2975 : * number that are significantly more common than the values not in the list,
2976 : * and which are therefore deemed worth storing in the table's MCV list.
2977 : */
2978 : static int
2979 27826 : analyze_mcv_list(int *mcv_counts,
2980 : int num_mcv,
2981 : double stadistinct,
2982 : double stanullfrac,
2983 : int samplerows,
2984 : double totalrows)
2985 : {
2986 : double ndistinct_table;
2987 : double sumcount;
2988 : int i;
2989 :
2990 : /*
2991 : * If the entire table was sampled, keep the whole list. This also
2992 : * protects us against division by zero in the code below.
2993 : */
2994 27826 : if (samplerows == totalrows || totalrows <= 1.0)
2995 27002 : return num_mcv;
2996 :
2997 : /* Re-extract the estimated number of distinct nonnull values in table */
2998 824 : ndistinct_table = stadistinct;
2999 824 : if (ndistinct_table < 0)
3000 150 : ndistinct_table = -ndistinct_table * totalrows;
3001 :
3002 : /*
3003 : * Exclude the least common values from the MCV list, if they are not
3004 : * significantly more common than the estimated selectivity they would
3005 : * have if they weren't in the list. All non-MCV values are assumed to be
3006 : * equally common, after taking into account the frequencies of all the
3007 : * values in the MCV list and the number of nulls (c.f. eqsel()).
3008 : *
3009 : * Here sumcount tracks the total count of all but the last (least common)
3010 : * value in the MCV list, allowing us to determine the effect of excluding
3011 : * that value from the list.
3012 : *
3013 : * Note that we deliberately do this by removing values from the full
3014 : * list, rather than starting with an empty list and adding values,
3015 : * because the latter approach can fail to add any values if all the most
3016 : * common values have around the same frequency and make up the majority
3017 : * of the table, so that the overall average frequency of all values is
3018 : * roughly the same as that of the common values. This would lead to any
3019 : * uncommon values being significantly overestimated.
3020 : */
3021 824 : sumcount = 0.0;
3022 1734 : for (i = 0; i < num_mcv - 1; i++)
3023 910 : sumcount += mcv_counts[i];
3024 :
3025 992 : while (num_mcv > 0)
3026 : {
3027 : double selec,
3028 : otherdistinct,
3029 : N,
3030 : n,
3031 : K,
3032 : variance,
3033 : stddev;
3034 :
3035 : /*
3036 : * Estimated selectivity the least common value would have if it
3037 : * wasn't in the MCV list (c.f. eqsel()).
3038 : */
3039 992 : selec = 1.0 - sumcount / samplerows - stanullfrac;
3040 992 : if (selec < 0.0)
3041 0 : selec = 0.0;
3042 992 : if (selec > 1.0)
3043 0 : selec = 1.0;
3044 992 : otherdistinct = ndistinct_table - (num_mcv - 1);
3045 992 : if (otherdistinct > 1)
3046 992 : selec /= otherdistinct;
3047 :
3048 : /*
3049 : * If the value is kept in the MCV list, its population frequency is
3050 : * assumed to equal its sample frequency. We use the lower end of a
3051 : * textbook continuity-corrected Wald-type confidence interval to
3052 : * determine if that is significantly more common than the non-MCV
3053 : * frequency --- specifically we assume the population frequency is
3054 : * highly likely to be within around 2 standard errors of the sample
3055 : * frequency, which equates to an interval of 2 standard deviations
3056 : * either side of the sample count, plus an additional 0.5 for the
3057 : * continuity correction. Since we are sampling without replacement,
3058 : * this is a hypergeometric distribution.
3059 : *
3060 : * XXX: Empirically, this approach seems to work quite well, but it
3061 : * may be worth considering more advanced techniques for estimating
3062 : * the confidence interval of the hypergeometric distribution.
3063 : */
3064 992 : N = totalrows;
3065 992 : n = samplerows;
3066 992 : K = N * mcv_counts[num_mcv - 1] / n;
3067 992 : variance = n * K * (N - K) * (N - n) / (N * N * (N - 1));
3068 992 : stddev = sqrt(variance);
3069 :
3070 992 : if (mcv_counts[num_mcv - 1] > selec * samplerows + 2 * stddev + 0.5)
3071 : {
3072 : /*
3073 : * The value is significantly more common than the non-MCV
3074 : * selectivity would suggest. Keep it, and all the other more
3075 : * common values in the list.
3076 : */
3077 774 : break;
3078 : }
3079 : else
3080 : {
3081 : /* Discard this value and consider the next least common value */
3082 218 : num_mcv--;
3083 218 : if (num_mcv == 0)
3084 50 : break;
3085 168 : sumcount -= mcv_counts[num_mcv - 1];
3086 : }
3087 : }
3088 824 : return num_mcv;
3089 : }
|
