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/dbcommands.h"
33 : #include "commands/progress.h"
34 : #include "commands/tablecmds.h"
35 : #include "commands/vacuum.h"
36 : #include "common/pg_prng.h"
37 : #include "executor/executor.h"
38 : #include "foreign/fdwapi.h"
39 : #include "miscadmin.h"
40 : #include "nodes/nodeFuncs.h"
41 : #include "parser/parse_oper.h"
42 : #include "parser/parse_relation.h"
43 : #include "pgstat.h"
44 : #include "statistics/extended_stats_internal.h"
45 : #include "statistics/statistics.h"
46 : #include "storage/bufmgr.h"
47 : #include "storage/procarray.h"
48 : #include "utils/attoptcache.h"
49 : #include "utils/datum.h"
50 : #include "utils/guc.h"
51 : #include "utils/lsyscache.h"
52 : #include "utils/memutils.h"
53 : #include "utils/pg_rusage.h"
54 : #include "utils/sampling.h"
55 : #include "utils/sortsupport.h"
56 : #include "utils/syscache.h"
57 : #include "utils/timestamp.h"
58 :
59 :
60 : /* Per-index data for ANALYZE */
61 : typedef struct AnlIndexData
62 : {
63 : IndexInfo *indexInfo; /* BuildIndexInfo result */
64 : double tupleFract; /* fraction of rows for partial index */
65 : VacAttrStats **vacattrstats; /* index attrs to analyze */
66 : int attr_cnt;
67 : } AnlIndexData;
68 :
69 :
70 : /* Default statistics target (GUC parameter) */
71 : int default_statistics_target = 100;
72 :
73 : /* A few variables that don't seem worth passing around as parameters */
74 : static MemoryContext anl_context = NULL;
75 : static BufferAccessStrategy vac_strategy;
76 :
77 :
78 : static void do_analyze_rel(Relation onerel,
79 : VacuumParams *params, List *va_cols,
80 : AcquireSampleRowsFunc acquirefunc, BlockNumber relpages,
81 : bool inh, bool in_outer_xact, int elevel);
82 : static void compute_index_stats(Relation onerel, double totalrows,
83 : AnlIndexData *indexdata, int nindexes,
84 : HeapTuple *rows, int numrows,
85 : MemoryContext col_context);
86 : static VacAttrStats *examine_attribute(Relation onerel, int attnum,
87 : Node *index_expr);
88 : static int acquire_sample_rows(Relation onerel, int elevel,
89 : HeapTuple *rows, int targrows,
90 : double *totalrows, double *totaldeadrows);
91 : static int compare_rows(const void *a, const void *b, void *arg);
92 : static int acquire_inherited_sample_rows(Relation onerel, int elevel,
93 : HeapTuple *rows, int targrows,
94 : double *totalrows, double *totaldeadrows);
95 : static void update_attstats(Oid relid, bool inh,
96 : int natts, VacAttrStats **vacattrstats);
97 : static Datum std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
98 : static Datum ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
99 :
100 :
101 : /*
102 : * analyze_rel() -- analyze one relation
103 : *
104 : * relid identifies the relation to analyze. If relation is supplied, use
105 : * the name therein for reporting any failure to open/lock the rel; do not
106 : * use it once we've successfully opened the rel, since it might be stale.
107 : */
108 : void
109 14488 : analyze_rel(Oid relid, RangeVar *relation,
110 : VacuumParams *params, List *va_cols, bool in_outer_xact,
111 : BufferAccessStrategy bstrategy)
112 : {
113 : Relation onerel;
114 : int elevel;
115 14488 : AcquireSampleRowsFunc acquirefunc = NULL;
116 14488 : BlockNumber relpages = 0;
117 :
118 : /* Select logging level */
119 14488 : if (params->options & VACOPT_VERBOSE)
120 0 : elevel = INFO;
121 : else
122 14488 : elevel = DEBUG2;
123 :
124 : /* Set up static variables */
125 14488 : vac_strategy = bstrategy;
126 :
127 : /*
128 : * Check for user-requested abort.
129 : */
130 14488 : CHECK_FOR_INTERRUPTS();
131 :
132 : /*
133 : * Open the relation, getting ShareUpdateExclusiveLock to ensure that two
134 : * ANALYZEs don't run on it concurrently. (This also locks out a
135 : * concurrent VACUUM, which doesn't matter much at the moment but might
136 : * matter if we ever try to accumulate stats on dead tuples.) If the rel
137 : * has been dropped since we last saw it, we don't need to process it.
138 : *
139 : * Make sure to generate only logs for ANALYZE in this case.
140 : */
141 14488 : onerel = vacuum_open_relation(relid, relation, params->options & ~(VACOPT_VACUUM),
142 14488 : params->log_min_duration >= 0,
143 : ShareUpdateExclusiveLock);
144 :
145 : /* leave if relation could not be opened or locked */
146 14488 : if (!onerel)
147 188 : return;
148 :
149 : /*
150 : * Check if relation needs to be skipped based on privileges. This check
151 : * happens also when building the relation list to analyze for a manual
152 : * operation, and needs to be done additionally here as ANALYZE could
153 : * happen across multiple transactions where privileges could have changed
154 : * in-between. Make sure to generate only logs for ANALYZE in this case.
155 : */
156 14480 : if (!vacuum_is_permitted_for_relation(RelationGetRelid(onerel),
157 : onerel->rd_rel,
158 14480 : params->options & ~VACOPT_VACUUM))
159 : {
160 36 : relation_close(onerel, ShareUpdateExclusiveLock);
161 36 : return;
162 : }
163 :
164 : /*
165 : * Silently ignore tables that are temp tables of other backends ---
166 : * trying to analyze these is rather pointless, since their contents are
167 : * probably not up-to-date on disk. (We don't throw a warning here; it
168 : * would just lead to chatter during a database-wide ANALYZE.)
169 : */
170 14444 : if (RELATION_IS_OTHER_TEMP(onerel))
171 : {
172 0 : relation_close(onerel, ShareUpdateExclusiveLock);
173 0 : return;
174 : }
175 :
176 : /*
177 : * We can ANALYZE any table except pg_statistic. See update_attstats
178 : */
179 14444 : if (RelationGetRelid(onerel) == StatisticRelationId)
180 : {
181 144 : relation_close(onerel, ShareUpdateExclusiveLock);
182 144 : return;
183 : }
184 :
185 : /*
186 : * Check that it's of an analyzable relkind, and set up appropriately.
187 : */
188 14300 : if (onerel->rd_rel->relkind == RELKIND_RELATION ||
189 794 : onerel->rd_rel->relkind == RELKIND_MATVIEW)
190 : {
191 : /* Regular table, so we'll use the regular row acquisition function */
192 13510 : acquirefunc = acquire_sample_rows;
193 : /* Also get regular table's size */
194 13510 : relpages = RelationGetNumberOfBlocks(onerel);
195 : }
196 790 : else if (onerel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
197 : {
198 : /*
199 : * For a foreign table, call the FDW's hook function to see whether it
200 : * supports analysis.
201 : */
202 : FdwRoutine *fdwroutine;
203 56 : bool ok = false;
204 :
205 56 : fdwroutine = GetFdwRoutineForRelation(onerel, false);
206 :
207 56 : if (fdwroutine->AnalyzeForeignTable != NULL)
208 56 : ok = fdwroutine->AnalyzeForeignTable(onerel,
209 : &acquirefunc,
210 : &relpages);
211 :
212 56 : if (!ok)
213 : {
214 0 : ereport(WARNING,
215 : (errmsg("skipping \"%s\" --- cannot analyze this foreign table",
216 : RelationGetRelationName(onerel))));
217 0 : relation_close(onerel, ShareUpdateExclusiveLock);
218 0 : return;
219 : }
220 : }
221 734 : else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
222 : {
223 : /*
224 : * For partitioned tables, we want to do the recursive ANALYZE below.
225 : */
226 : }
227 : else
228 : {
229 : /* No need for a WARNING if we already complained during VACUUM */
230 0 : if (!(params->options & VACOPT_VACUUM))
231 0 : ereport(WARNING,
232 : (errmsg("skipping \"%s\" --- cannot analyze non-tables or special system tables",
233 : RelationGetRelationName(onerel))));
234 0 : relation_close(onerel, ShareUpdateExclusiveLock);
235 0 : return;
236 : }
237 :
238 : /*
239 : * OK, let's do it. First, initialize progress reporting.
240 : */
241 14300 : pgstat_progress_start_command(PROGRESS_COMMAND_ANALYZE,
242 : RelationGetRelid(onerel));
243 :
244 : /*
245 : * Do the normal non-recursive ANALYZE. We can skip this for partitioned
246 : * tables, which don't contain any rows.
247 : */
248 14300 : if (onerel->rd_rel->relkind != RELKIND_PARTITIONED_TABLE)
249 13566 : do_analyze_rel(onerel, params, va_cols, acquirefunc,
250 : relpages, false, in_outer_xact, elevel);
251 :
252 : /*
253 : * If there are child tables, do recursive ANALYZE.
254 : */
255 14260 : if (onerel->rd_rel->relhassubclass)
256 842 : do_analyze_rel(onerel, params, va_cols, acquirefunc, relpages,
257 : true, in_outer_xact, elevel);
258 :
259 : /*
260 : * Close source relation now, but keep lock so that no one deletes it
261 : * before we commit. (If someone did, they'd fail to clean up the entries
262 : * we made in pg_statistic. Also, releasing the lock before commit would
263 : * expose us to concurrent-update failures in update_attstats.)
264 : */
265 14242 : relation_close(onerel, NoLock);
266 :
267 14242 : pgstat_progress_end_command();
268 : }
269 :
270 : /*
271 : * do_analyze_rel() -- analyze one relation, recursively or not
272 : *
273 : * Note that "acquirefunc" is only relevant for the non-inherited case.
274 : * For the inherited case, acquire_inherited_sample_rows() determines the
275 : * appropriate acquirefunc for each child table.
276 : */
277 : static void
278 14408 : do_analyze_rel(Relation onerel, VacuumParams *params,
279 : List *va_cols, AcquireSampleRowsFunc acquirefunc,
280 : BlockNumber relpages, bool inh, bool in_outer_xact,
281 : int elevel)
282 : {
283 : int attr_cnt,
284 : tcnt,
285 : i,
286 : ind;
287 : Relation *Irel;
288 : int nindexes;
289 : bool verbose,
290 : instrument,
291 : hasindex;
292 : VacAttrStats **vacattrstats;
293 : AnlIndexData *indexdata;
294 : int targrows,
295 : numrows,
296 : minrows;
297 : double totalrows,
298 : totaldeadrows;
299 : HeapTuple *rows;
300 : PGRUsage ru0;
301 14408 : TimestampTz starttime = 0;
302 : MemoryContext caller_context;
303 : Oid save_userid;
304 : int save_sec_context;
305 : int save_nestlevel;
306 14408 : WalUsage startwalusage = pgWalUsage;
307 14408 : BufferUsage startbufferusage = pgBufferUsage;
308 : BufferUsage bufferusage;
309 14408 : PgStat_Counter startreadtime = 0;
310 14408 : PgStat_Counter startwritetime = 0;
311 :
312 14408 : verbose = (params->options & VACOPT_VERBOSE) != 0;
313 14844 : instrument = (verbose || (AmAutoVacuumWorkerProcess() &&
314 436 : params->log_min_duration >= 0));
315 14408 : if (inh)
316 842 : ereport(elevel,
317 : (errmsg("analyzing \"%s.%s\" inheritance tree",
318 : get_namespace_name(RelationGetNamespace(onerel)),
319 : RelationGetRelationName(onerel))));
320 : else
321 13566 : ereport(elevel,
322 : (errmsg("analyzing \"%s.%s\"",
323 : get_namespace_name(RelationGetNamespace(onerel)),
324 : RelationGetRelationName(onerel))));
325 :
326 : /*
327 : * Set up a working context so that we can easily free whatever junk gets
328 : * created.
329 : */
330 14408 : anl_context = AllocSetContextCreate(CurrentMemoryContext,
331 : "Analyze",
332 : ALLOCSET_DEFAULT_SIZES);
333 14408 : caller_context = MemoryContextSwitchTo(anl_context);
334 :
335 : /*
336 : * Switch to the table owner's userid, so that any index functions are run
337 : * as that user. Also lock down security-restricted operations and
338 : * arrange to make GUC variable changes local to this command.
339 : */
340 14408 : GetUserIdAndSecContext(&save_userid, &save_sec_context);
341 14408 : SetUserIdAndSecContext(onerel->rd_rel->relowner,
342 : save_sec_context | SECURITY_RESTRICTED_OPERATION);
343 14408 : save_nestlevel = NewGUCNestLevel();
344 14408 : RestrictSearchPath();
345 :
346 : /*
347 : * When verbose or autovacuum logging is used, initialize a resource usage
348 : * snapshot and optionally track I/O timing.
349 : */
350 14408 : if (instrument)
351 : {
352 436 : if (track_io_timing)
353 : {
354 0 : startreadtime = pgStatBlockReadTime;
355 0 : startwritetime = pgStatBlockWriteTime;
356 : }
357 :
358 436 : pg_rusage_init(&ru0);
359 : }
360 :
361 : /* Used for instrumentation and stats report */
362 14408 : starttime = GetCurrentTimestamp();
363 :
364 : /*
365 : * Determine which columns to analyze
366 : *
367 : * Note that system attributes are never analyzed, so we just reject them
368 : * at the lookup stage. We also reject duplicate column mentions. (We
369 : * could alternatively ignore duplicates, but analyzing a column twice
370 : * won't work; we'd end up making a conflicting update in pg_statistic.)
371 : */
372 14408 : if (va_cols != NIL)
373 : {
374 100 : Bitmapset *unique_cols = NULL;
375 : ListCell *le;
376 :
377 100 : vacattrstats = (VacAttrStats **) palloc(list_length(va_cols) *
378 : sizeof(VacAttrStats *));
379 100 : tcnt = 0;
380 182 : foreach(le, va_cols)
381 : {
382 132 : char *col = strVal(lfirst(le));
383 :
384 132 : i = attnameAttNum(onerel, col, false);
385 132 : if (i == InvalidAttrNumber)
386 38 : ereport(ERROR,
387 : (errcode(ERRCODE_UNDEFINED_COLUMN),
388 : errmsg("column \"%s\" of relation \"%s\" does not exist",
389 : col, RelationGetRelationName(onerel))));
390 94 : if (bms_is_member(i, unique_cols))
391 12 : ereport(ERROR,
392 : (errcode(ERRCODE_DUPLICATE_COLUMN),
393 : errmsg("column \"%s\" of relation \"%s\" appears more than once",
394 : col, RelationGetRelationName(onerel))));
395 82 : unique_cols = bms_add_member(unique_cols, i);
396 :
397 82 : vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
398 82 : if (vacattrstats[tcnt] != NULL)
399 82 : tcnt++;
400 : }
401 50 : attr_cnt = tcnt;
402 : }
403 : else
404 : {
405 14308 : attr_cnt = onerel->rd_att->natts;
406 : vacattrstats = (VacAttrStats **)
407 14308 : palloc(attr_cnt * sizeof(VacAttrStats *));
408 14308 : tcnt = 0;
409 116150 : for (i = 1; i <= attr_cnt; i++)
410 : {
411 101842 : vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
412 101842 : if (vacattrstats[tcnt] != NULL)
413 101830 : tcnt++;
414 : }
415 14308 : attr_cnt = tcnt;
416 : }
417 :
418 : /*
419 : * Open all indexes of the relation, and see if there are any analyzable
420 : * columns in the indexes. We do not analyze index columns if there was
421 : * an explicit column list in the ANALYZE command, however.
422 : *
423 : * If we are doing a recursive scan, we don't want to touch the parent's
424 : * indexes at all. If we're processing a partitioned table, we need to
425 : * know if there are any indexes, but we don't want to process them.
426 : */
427 14358 : if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
428 : {
429 716 : List *idxs = RelationGetIndexList(onerel);
430 :
431 716 : Irel = NULL;
432 716 : nindexes = 0;
433 716 : hasindex = idxs != NIL;
434 716 : list_free(idxs);
435 : }
436 13642 : else if (!inh)
437 : {
438 13534 : vac_open_indexes(onerel, AccessShareLock, &nindexes, &Irel);
439 13534 : hasindex = nindexes > 0;
440 : }
441 : else
442 : {
443 108 : Irel = NULL;
444 108 : nindexes = 0;
445 108 : hasindex = false;
446 : }
447 14358 : indexdata = NULL;
448 14358 : if (nindexes > 0)
449 : {
450 10494 : indexdata = (AnlIndexData *) palloc0(nindexes * sizeof(AnlIndexData));
451 30182 : for (ind = 0; ind < nindexes; ind++)
452 : {
453 19688 : AnlIndexData *thisdata = &indexdata[ind];
454 : IndexInfo *indexInfo;
455 :
456 19688 : thisdata->indexInfo = indexInfo = BuildIndexInfo(Irel[ind]);
457 19688 : thisdata->tupleFract = 1.0; /* fix later if partial */
458 19688 : if (indexInfo->ii_Expressions != NIL && va_cols == NIL)
459 : {
460 80 : ListCell *indexpr_item = list_head(indexInfo->ii_Expressions);
461 :
462 80 : thisdata->vacattrstats = (VacAttrStats **)
463 80 : palloc(indexInfo->ii_NumIndexAttrs * sizeof(VacAttrStats *));
464 80 : tcnt = 0;
465 162 : for (i = 0; i < indexInfo->ii_NumIndexAttrs; i++)
466 : {
467 82 : int keycol = indexInfo->ii_IndexAttrNumbers[i];
468 :
469 82 : if (keycol == 0)
470 : {
471 : /* Found an index expression */
472 : Node *indexkey;
473 :
474 80 : if (indexpr_item == NULL) /* shouldn't happen */
475 0 : elog(ERROR, "too few entries in indexprs list");
476 80 : indexkey = (Node *) lfirst(indexpr_item);
477 80 : indexpr_item = lnext(indexInfo->ii_Expressions,
478 : indexpr_item);
479 160 : thisdata->vacattrstats[tcnt] =
480 80 : examine_attribute(Irel[ind], i + 1, indexkey);
481 80 : if (thisdata->vacattrstats[tcnt] != NULL)
482 80 : tcnt++;
483 : }
484 : }
485 80 : thisdata->attr_cnt = tcnt;
486 : }
487 : }
488 : }
489 :
490 : /*
491 : * Determine how many rows we need to sample, using the worst case from
492 : * all analyzable columns. We use a lower bound of 100 rows to avoid
493 : * possible overflow in Vitter's algorithm. (Note: that will also be the
494 : * target in the corner case where there are no analyzable columns.)
495 : */
496 14358 : targrows = 100;
497 116246 : for (i = 0; i < attr_cnt; i++)
498 : {
499 101888 : if (targrows < vacattrstats[i]->minrows)
500 14322 : targrows = vacattrstats[i]->minrows;
501 : }
502 34046 : for (ind = 0; ind < nindexes; ind++)
503 : {
504 19688 : AnlIndexData *thisdata = &indexdata[ind];
505 :
506 19768 : for (i = 0; i < thisdata->attr_cnt; i++)
507 : {
508 80 : if (targrows < thisdata->vacattrstats[i]->minrows)
509 0 : targrows = thisdata->vacattrstats[i]->minrows;
510 : }
511 : }
512 :
513 : /*
514 : * Look at extended statistics objects too, as those may define custom
515 : * statistics target. So we may need to sample more rows and then build
516 : * the statistics with enough detail.
517 : */
518 14358 : minrows = ComputeExtStatisticsRows(onerel, attr_cnt, vacattrstats);
519 :
520 14358 : if (targrows < minrows)
521 0 : targrows = minrows;
522 :
523 : /*
524 : * Acquire the sample rows
525 : */
526 14358 : rows = (HeapTuple *) palloc(targrows * sizeof(HeapTuple));
527 14358 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
528 : inh ? PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS_INH :
529 : PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS);
530 14358 : if (inh)
531 824 : numrows = acquire_inherited_sample_rows(onerel, elevel,
532 : rows, targrows,
533 : &totalrows, &totaldeadrows);
534 : else
535 13534 : numrows = (*acquirefunc) (onerel, elevel,
536 : rows, targrows,
537 : &totalrows, &totaldeadrows);
538 :
539 : /*
540 : * Compute the statistics. Temporary results during the calculations for
541 : * each column are stored in a child context. The calc routines are
542 : * responsible to make sure that whatever they store into the VacAttrStats
543 : * structure is allocated in anl_context.
544 : */
545 14356 : if (numrows > 0)
546 : {
547 : MemoryContext col_context,
548 : old_context;
549 :
550 9822 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
551 : PROGRESS_ANALYZE_PHASE_COMPUTE_STATS);
552 :
553 9822 : col_context = AllocSetContextCreate(anl_context,
554 : "Analyze Column",
555 : ALLOCSET_DEFAULT_SIZES);
556 9822 : old_context = MemoryContextSwitchTo(col_context);
557 :
558 85184 : for (i = 0; i < attr_cnt; i++)
559 : {
560 75362 : VacAttrStats *stats = vacattrstats[i];
561 : AttributeOpts *aopt;
562 :
563 75362 : stats->rows = rows;
564 75362 : stats->tupDesc = onerel->rd_att;
565 75362 : stats->compute_stats(stats,
566 : std_fetch_func,
567 : numrows,
568 : totalrows);
569 :
570 : /*
571 : * If the appropriate flavor of the n_distinct option is
572 : * specified, override with the corresponding value.
573 : */
574 75362 : aopt = get_attribute_options(onerel->rd_id, stats->tupattnum);
575 75362 : if (aopt != NULL)
576 : {
577 : float8 n_distinct;
578 :
579 6 : n_distinct = inh ? aopt->n_distinct_inherited : aopt->n_distinct;
580 6 : if (n_distinct != 0.0)
581 6 : stats->stadistinct = n_distinct;
582 : }
583 :
584 75362 : MemoryContextReset(col_context);
585 : }
586 :
587 9822 : if (nindexes > 0)
588 6238 : compute_index_stats(onerel, totalrows,
589 : indexdata, nindexes,
590 : rows, numrows,
591 : col_context);
592 :
593 9816 : MemoryContextSwitchTo(old_context);
594 9816 : MemoryContextDelete(col_context);
595 :
596 : /*
597 : * Emit the completed stats rows into pg_statistic, replacing any
598 : * previous statistics for the target columns. (If there are stats in
599 : * pg_statistic for columns we didn't process, we leave them alone.)
600 : */
601 9816 : update_attstats(RelationGetRelid(onerel), inh,
602 : attr_cnt, vacattrstats);
603 :
604 21902 : for (ind = 0; ind < nindexes; ind++)
605 : {
606 12086 : AnlIndexData *thisdata = &indexdata[ind];
607 :
608 12086 : update_attstats(RelationGetRelid(Irel[ind]), false,
609 : thisdata->attr_cnt, thisdata->vacattrstats);
610 : }
611 :
612 : /* Build extended statistics (if there are any). */
613 9816 : BuildRelationExtStatistics(onerel, inh, totalrows, numrows, rows,
614 : attr_cnt, vacattrstats);
615 : }
616 :
617 14350 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
618 : PROGRESS_ANALYZE_PHASE_FINALIZE_ANALYZE);
619 :
620 : /*
621 : * Update pages/tuples stats in pg_class ... but not if we're doing
622 : * inherited stats.
623 : *
624 : * We assume that VACUUM hasn't set pg_class.reltuples already, even
625 : * during a VACUUM ANALYZE. Although VACUUM often updates pg_class,
626 : * exceptions exist. A "VACUUM (ANALYZE, INDEX_CLEANUP OFF)" command will
627 : * never update pg_class entries for index relations. It's also possible
628 : * that an individual index's pg_class entry won't be updated during
629 : * VACUUM if the index AM returns NULL from its amvacuumcleanup() routine.
630 : */
631 14350 : if (!inh)
632 : {
633 : BlockNumber relallvisible;
634 :
635 13526 : if (RELKIND_HAS_STORAGE(onerel->rd_rel->relkind))
636 13472 : visibilitymap_count(onerel, &relallvisible, NULL);
637 : else
638 54 : relallvisible = 0;
639 :
640 : /*
641 : * Update pg_class for table relation. CCI first, in case acquirefunc
642 : * updated pg_class.
643 : */
644 13526 : CommandCounterIncrement();
645 13526 : vac_update_relstats(onerel,
646 : relpages,
647 : totalrows,
648 : relallvisible,
649 : hasindex,
650 : InvalidTransactionId,
651 : InvalidMultiXactId,
652 : NULL, NULL,
653 : in_outer_xact);
654 :
655 : /* Same for indexes */
656 33202 : for (ind = 0; ind < nindexes; ind++)
657 : {
658 19676 : AnlIndexData *thisdata = &indexdata[ind];
659 : double totalindexrows;
660 :
661 19676 : totalindexrows = ceil(thisdata->tupleFract * totalrows);
662 19676 : vac_update_relstats(Irel[ind],
663 19676 : RelationGetNumberOfBlocks(Irel[ind]),
664 : totalindexrows,
665 : 0,
666 : false,
667 : InvalidTransactionId,
668 : InvalidMultiXactId,
669 : NULL, NULL,
670 : in_outer_xact);
671 : }
672 : }
673 824 : else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
674 : {
675 : /*
676 : * Partitioned tables don't have storage, so we don't set any fields
677 : * in their pg_class entries except for reltuples and relhasindex.
678 : */
679 716 : CommandCounterIncrement();
680 716 : vac_update_relstats(onerel, -1, totalrows,
681 : 0, hasindex, InvalidTransactionId,
682 : InvalidMultiXactId,
683 : NULL, NULL,
684 : in_outer_xact);
685 : }
686 :
687 : /*
688 : * Now report ANALYZE to the cumulative stats system. For regular tables,
689 : * we do it only if not doing inherited stats. For partitioned tables, we
690 : * only do it for inherited stats. (We're never called for not-inherited
691 : * stats on partitioned tables anyway.)
692 : *
693 : * Reset the changes_since_analyze counter only if we analyzed all
694 : * columns; otherwise, there is still work for auto-analyze to do.
695 : */
696 14350 : if (!inh)
697 13526 : pgstat_report_analyze(onerel, totalrows, totaldeadrows,
698 : (va_cols == NIL), starttime);
699 824 : else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
700 716 : pgstat_report_analyze(onerel, 0, 0, (va_cols == NIL), starttime);
701 :
702 : /*
703 : * If this isn't part of VACUUM ANALYZE, let index AMs do cleanup.
704 : *
705 : * Note that most index AMs perform a no-op as a matter of policy for
706 : * amvacuumcleanup() when called in ANALYZE-only mode. The only exception
707 : * among core index AMs is GIN/ginvacuumcleanup().
708 : */
709 14350 : if (!(params->options & VACOPT_VACUUM))
710 : {
711 30190 : for (ind = 0; ind < nindexes; ind++)
712 : {
713 : IndexBulkDeleteResult *stats;
714 : IndexVacuumInfo ivinfo;
715 :
716 17412 : ivinfo.index = Irel[ind];
717 17412 : ivinfo.heaprel = onerel;
718 17412 : ivinfo.analyze_only = true;
719 17412 : ivinfo.estimated_count = true;
720 17412 : ivinfo.message_level = elevel;
721 17412 : ivinfo.num_heap_tuples = onerel->rd_rel->reltuples;
722 17412 : ivinfo.strategy = vac_strategy;
723 :
724 17412 : stats = index_vacuum_cleanup(&ivinfo, NULL);
725 :
726 17412 : if (stats)
727 0 : pfree(stats);
728 : }
729 : }
730 :
731 : /* Done with indexes */
732 14350 : vac_close_indexes(nindexes, Irel, NoLock);
733 :
734 : /* Log the action if appropriate */
735 14350 : if (instrument)
736 : {
737 436 : TimestampTz endtime = GetCurrentTimestamp();
738 :
739 564 : if (verbose || params->log_min_duration == 0 ||
740 128 : TimestampDifferenceExceeds(starttime, endtime,
741 : params->log_min_duration))
742 : {
743 : long delay_in_ms;
744 : WalUsage walusage;
745 308 : double read_rate = 0;
746 308 : double write_rate = 0;
747 : char *msgfmt;
748 : StringInfoData buf;
749 : int64 total_blks_hit;
750 : int64 total_blks_read;
751 : int64 total_blks_dirtied;
752 :
753 308 : memset(&bufferusage, 0, sizeof(BufferUsage));
754 308 : BufferUsageAccumDiff(&bufferusage, &pgBufferUsage, &startbufferusage);
755 308 : memset(&walusage, 0, sizeof(WalUsage));
756 308 : WalUsageAccumDiff(&walusage, &pgWalUsage, &startwalusage);
757 :
758 308 : total_blks_hit = bufferusage.shared_blks_hit +
759 308 : bufferusage.local_blks_hit;
760 308 : total_blks_read = bufferusage.shared_blks_read +
761 308 : bufferusage.local_blks_read;
762 308 : total_blks_dirtied = bufferusage.shared_blks_dirtied +
763 308 : bufferusage.local_blks_dirtied;
764 :
765 : /*
766 : * We do not expect an analyze to take > 25 days and it simplifies
767 : * things a bit to use TimestampDifferenceMilliseconds.
768 : */
769 308 : delay_in_ms = TimestampDifferenceMilliseconds(starttime, endtime);
770 :
771 : /*
772 : * Note that we are reporting these read/write rates in the same
773 : * manner as VACUUM does, which means that while the 'average read
774 : * rate' here actually corresponds to page misses and resulting
775 : * reads which are also picked up by track_io_timing, if enabled,
776 : * the 'average write rate' is actually talking about the rate of
777 : * pages being dirtied, not being written out, so it's typical to
778 : * have a non-zero 'avg write rate' while I/O timings only reports
779 : * reads.
780 : *
781 : * It's not clear that an ANALYZE will ever result in
782 : * FlushBuffer() being called, but we track and support reporting
783 : * on I/O write time in case that changes as it's practically free
784 : * to do so anyway.
785 : */
786 :
787 308 : if (delay_in_ms > 0)
788 : {
789 308 : read_rate = (double) BLCKSZ * total_blks_read /
790 308 : (1024 * 1024) / (delay_in_ms / 1000.0);
791 308 : write_rate = (double) BLCKSZ * total_blks_dirtied /
792 308 : (1024 * 1024) / (delay_in_ms / 1000.0);
793 : }
794 :
795 : /*
796 : * We split this up so we don't emit empty I/O timing values when
797 : * track_io_timing isn't enabled.
798 : */
799 :
800 308 : initStringInfo(&buf);
801 :
802 308 : if (AmAutoVacuumWorkerProcess())
803 308 : msgfmt = _("automatic analyze of table \"%s.%s.%s\"\n");
804 : else
805 0 : msgfmt = _("finished analyzing table \"%s.%s.%s\"\n");
806 :
807 308 : appendStringInfo(&buf, msgfmt,
808 : get_database_name(MyDatabaseId),
809 308 : get_namespace_name(RelationGetNamespace(onerel)),
810 308 : RelationGetRelationName(onerel));
811 308 : if (track_cost_delay_timing)
812 : {
813 : /*
814 : * We bypass the changecount mechanism because this value is
815 : * only updated by the calling process.
816 : */
817 0 : appendStringInfo(&buf, _("delay time: %.3f ms\n"),
818 0 : (double) MyBEEntry->st_progress_param[PROGRESS_ANALYZE_DELAY_TIME] / 1000000.0);
819 : }
820 308 : if (track_io_timing)
821 : {
822 0 : double read_ms = (double) (pgStatBlockReadTime - startreadtime) / 1000;
823 0 : double write_ms = (double) (pgStatBlockWriteTime - startwritetime) / 1000;
824 :
825 0 : appendStringInfo(&buf, _("I/O timings: read: %.3f ms, write: %.3f ms\n"),
826 : read_ms, write_ms);
827 : }
828 308 : appendStringInfo(&buf, _("avg read rate: %.3f MB/s, avg write rate: %.3f MB/s\n"),
829 : read_rate, write_rate);
830 308 : appendStringInfo(&buf, _("buffer usage: %lld hits, %lld reads, %lld dirtied\n"),
831 : (long long) total_blks_hit,
832 : (long long) total_blks_read,
833 : (long long) total_blks_dirtied);
834 308 : appendStringInfo(&buf,
835 308 : _("WAL usage: %lld records, %lld full page images, %llu bytes, %lld buffers full\n"),
836 308 : (long long) walusage.wal_records,
837 308 : (long long) walusage.wal_fpi,
838 308 : (unsigned long long) walusage.wal_bytes,
839 308 : (long long) walusage.wal_buffers_full);
840 308 : appendStringInfo(&buf, _("system usage: %s"), pg_rusage_show(&ru0));
841 :
842 308 : ereport(verbose ? INFO : LOG,
843 : (errmsg_internal("%s", buf.data)));
844 :
845 308 : pfree(buf.data);
846 : }
847 : }
848 :
849 : /* Roll back any GUC changes executed by index functions */
850 14350 : AtEOXact_GUC(false, save_nestlevel);
851 :
852 : /* Restore userid and security context */
853 14350 : SetUserIdAndSecContext(save_userid, save_sec_context);
854 :
855 : /* Restore current context and release memory */
856 14350 : MemoryContextSwitchTo(caller_context);
857 14350 : MemoryContextDelete(anl_context);
858 14350 : anl_context = NULL;
859 14350 : }
860 :
861 : /*
862 : * Compute statistics about indexes of a relation
863 : */
864 : static void
865 6238 : 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 6238 : ind_context = AllocSetContextCreate(anl_context,
878 : "Analyze Index",
879 : ALLOCSET_DEFAULT_SIZES);
880 6238 : old_context = MemoryContextSwitchTo(ind_context);
881 :
882 18330 : for (ind = 0; ind < nindexes; ind++)
883 : {
884 12098 : AnlIndexData *thisdata = &indexdata[ind];
885 12098 : IndexInfo *indexInfo = thisdata->indexInfo;
886 12098 : 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 12098 : if (attr_cnt == 0 && indexInfo->ii_Predicate == NIL)
900 11978 : 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 : */
907 120 : estate = CreateExecutorState();
908 120 : econtext = GetPerTupleExprContext(estate);
909 : /* Need a slot to hold the current heap tuple, too */
910 120 : slot = MakeSingleTupleTableSlot(RelationGetDescr(onerel),
911 : &TTSOpsHeapTuple);
912 :
913 : /* Arrange for econtext's scan tuple to be the tuple under test */
914 120 : econtext->ecxt_scantuple = slot;
915 :
916 : /* Set up execution state for predicate. */
917 120 : predicate = ExecPrepareQual(indexInfo->ii_Predicate, estate);
918 :
919 : /* Compute and save index expression values */
920 120 : exprvals = (Datum *) palloc(numrows * attr_cnt * sizeof(Datum));
921 120 : exprnulls = (bool *) palloc(numrows * attr_cnt * sizeof(bool));
922 120 : numindexrows = 0;
923 120 : tcnt = 0;
924 155070 : for (rowno = 0; rowno < numrows; rowno++)
925 : {
926 154956 : HeapTuple heapTuple = rows[rowno];
927 :
928 154956 : 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 154956 : ResetExprContext(econtext);
935 :
936 : /* Set up for predicate or expression evaluation */
937 154956 : ExecStoreHeapTuple(heapTuple, slot, false);
938 :
939 : /* If index is partial, check predicate */
940 154956 : if (predicate != NULL)
941 : {
942 40066 : if (!ExecQual(predicate, econtext))
943 21330 : continue;
944 : }
945 133626 : numindexrows++;
946 :
947 133626 : 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 114890 : 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 229768 : for (i = 0; i < attr_cnt; i++)
964 : {
965 114884 : VacAttrStats *stats = thisdata->vacattrstats[i];
966 114884 : int attnum = stats->tupattnum;
967 :
968 114884 : if (isnull[attnum - 1])
969 : {
970 6 : exprvals[tcnt] = (Datum) 0;
971 6 : exprnulls[tcnt] = true;
972 : }
973 : else
974 : {
975 229756 : exprvals[tcnt] = datumCopy(values[attnum - 1],
976 114878 : stats->attrtype->typbyval,
977 114878 : stats->attrtype->typlen);
978 114878 : exprnulls[tcnt] = false;
979 : }
980 114884 : 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 : */
989 114 : thisdata->tupleFract = (double) numindexrows / (double) numrows;
990 114 : totalindexrows = ceil(thisdata->tupleFract * totalrows);
991 :
992 : /*
993 : * Now we can compute the statistics for the expression columns.
994 : */
995 114 : if (numindexrows > 0)
996 : {
997 106 : MemoryContextSwitchTo(col_context);
998 174 : for (i = 0; i < attr_cnt; i++)
999 : {
1000 68 : VacAttrStats *stats = thisdata->vacattrstats[i];
1001 :
1002 68 : stats->exprvals = exprvals + i;
1003 68 : stats->exprnulls = exprnulls + i;
1004 68 : stats->rowstride = attr_cnt;
1005 68 : stats->compute_stats(stats,
1006 : ind_fetch_func,
1007 : numindexrows,
1008 : totalindexrows);
1009 :
1010 68 : MemoryContextReset(col_context);
1011 : }
1012 : }
1013 :
1014 : /* And clean up */
1015 114 : MemoryContextSwitchTo(ind_context);
1016 :
1017 114 : ExecDropSingleTupleTableSlot(slot);
1018 114 : FreeExecutorState(estate);
1019 114 : MemoryContextReset(ind_context);
1020 : }
1021 :
1022 6232 : MemoryContextSwitchTo(old_context);
1023 6232 : MemoryContextDelete(ind_context);
1024 6232 : }
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 102004 : examine_attribute(Relation onerel, int attnum, Node *index_expr)
1037 : {
1038 102004 : 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 102004 : if (attr->attisdropped)
1050 6 : return NULL;
1051 :
1052 : /* Don't analyze virtual generated columns */
1053 101998 : if (attr->attgenerated == ATTRIBUTE_GENERATED_VIRTUAL)
1054 0 : 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 101998 : atttuple = SearchSysCache2(ATTNUM, ObjectIdGetDatum(RelationGetRelid(onerel)), Int16GetDatum(attnum));
1062 101998 : if (!HeapTupleIsValid(atttuple))
1063 0 : elog(ERROR, "cache lookup failed for attribute %d of relation %u",
1064 : attnum, RelationGetRelid(onerel));
1065 101998 : dat = SysCacheGetAttr(ATTNUM, atttuple, Anum_pg_attribute_attstattarget, &isnull);
1066 101998 : attstattarget = isnull ? -1 : DatumGetInt16(dat);
1067 101998 : ReleaseSysCache(atttuple);
1068 :
1069 : /* Don't analyze column if user has specified not to */
1070 101998 : if (attstattarget == 0)
1071 6 : return NULL;
1072 :
1073 : /*
1074 : * Create the VacAttrStats struct.
1075 : */
1076 101992 : stats = (VacAttrStats *) palloc0(sizeof(VacAttrStats));
1077 101992 : 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 101992 : if (index_expr)
1089 : {
1090 80 : stats->attrtypid = exprType(index_expr);
1091 80 : 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 80 : if (OidIsValid(onerel->rd_indcollation[attnum - 1]))
1099 12 : stats->attrcollid = onerel->rd_indcollation[attnum - 1];
1100 : else
1101 68 : stats->attrcollid = exprCollation(index_expr);
1102 : }
1103 : else
1104 : {
1105 101912 : stats->attrtypid = attr->atttypid;
1106 101912 : stats->attrtypmod = attr->atttypmod;
1107 101912 : stats->attrcollid = attr->attcollation;
1108 : }
1109 :
1110 101992 : typtuple = SearchSysCacheCopy1(TYPEOID,
1111 : ObjectIdGetDatum(stats->attrtypid));
1112 101992 : if (!HeapTupleIsValid(typtuple))
1113 0 : elog(ERROR, "cache lookup failed for type %u", stats->attrtypid);
1114 101992 : stats->attrtype = (Form_pg_type) GETSTRUCT(typtuple);
1115 101992 : stats->anl_context = anl_context;
1116 101992 : 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 611952 : for (i = 0; i < STATISTIC_NUM_SLOTS; i++)
1124 : {
1125 509960 : stats->statypid[i] = stats->attrtypid;
1126 509960 : stats->statyplen[i] = stats->attrtype->typlen;
1127 509960 : stats->statypbyval[i] = stats->attrtype->typbyval;
1128 509960 : 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 101992 : if (OidIsValid(stats->attrtype->typanalyze))
1136 6572 : ok = DatumGetBool(OidFunctionCall1(stats->attrtype->typanalyze,
1137 : PointerGetDatum(stats)));
1138 : else
1139 95420 : ok = std_typanalyze(stats);
1140 :
1141 101992 : 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 101992 : 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 140834 : block_sampling_read_stream_next(ReadStream *stream,
1157 : void *callback_private_data,
1158 : void *per_buffer_data)
1159 : {
1160 140834 : BlockSamplerData *bs = callback_private_data;
1161 :
1162 140834 : 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 15412 : acquire_sample_rows(Relation onerel, int elevel,
1200 : HeapTuple *rows, int targrows,
1201 : double *totalrows, double *totaldeadrows)
1202 : {
1203 15412 : int numrows = 0; /* # rows now in reservoir */
1204 15412 : double samplerows = 0; /* total # rows collected */
1205 15412 : double liverows = 0; /* # live rows seen */
1206 15412 : double deadrows = 0; /* # dead rows seen */
1207 15412 : 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 15412 : BlockNumber blksdone = 0;
1217 : ReadStream *stream;
1218 :
1219 : Assert(targrows > 0);
1220 :
1221 15412 : totalblocks = RelationGetNumberOfBlocks(onerel);
1222 :
1223 : /* Need a cutoff xmin for HeapTupleSatisfiesVacuum */
1224 15412 : OldestXmin = GetOldestNonRemovableTransactionId(onerel);
1225 :
1226 : /* Prepare for sampling block numbers */
1227 15412 : randseed = pg_prng_uint32(&pg_global_prng_state);
1228 15412 : nblocks = BlockSampler_Init(&bs, totalblocks, targrows, randseed);
1229 :
1230 : /* Report sampling block numbers */
1231 15412 : pgstat_progress_update_param(PROGRESS_ANALYZE_BLOCKS_TOTAL,
1232 : nblocks);
1233 :
1234 : /* Prepare for sampling rows */
1235 15412 : reservoir_init_selection_state(&rstate, targrows);
1236 :
1237 15412 : scan = table_beginscan_analyze(onerel);
1238 15412 : slot = table_slot_create(onerel, NULL);
1239 :
1240 15412 : stream = read_stream_begin_relation(READ_STREAM_MAINTENANCE,
1241 : vac_strategy,
1242 : scan->rs_rd,
1243 : MAIN_FORKNUM,
1244 : block_sampling_read_stream_next,
1245 : &bs,
1246 : 0);
1247 :
1248 : /* Outer loop over blocks to sample */
1249 140834 : while (table_scan_analyze_next_block(scan, stream))
1250 : {
1251 125422 : vacuum_delay_point(true);
1252 :
1253 9950510 : while (table_scan_analyze_next_tuple(scan, OldestXmin, &liverows, &deadrows, slot))
1254 : {
1255 : /*
1256 : * The first targrows sample rows are simply copied into the
1257 : * reservoir. Then we start replacing tuples in the sample until
1258 : * we reach the end of the relation. This algorithm is from Jeff
1259 : * Vitter's paper (see full citation in utils/misc/sampling.c). It
1260 : * works by repeatedly computing the number of tuples to skip
1261 : * before selecting a tuple, which replaces a randomly chosen
1262 : * element of the reservoir (current set of tuples). At all times
1263 : * the reservoir is a true random sample of the tuples we've
1264 : * passed over so far, so when we fall off the end of the relation
1265 : * we're done.
1266 : */
1267 9825088 : if (numrows < targrows)
1268 9574136 : rows[numrows++] = ExecCopySlotHeapTuple(slot);
1269 : else
1270 : {
1271 : /*
1272 : * t in Vitter's paper is the number of records already
1273 : * processed. If we need to compute a new S value, we must
1274 : * use the not-yet-incremented value of samplerows as t.
1275 : */
1276 250952 : if (rowstoskip < 0)
1277 115358 : rowstoskip = reservoir_get_next_S(&rstate, samplerows, targrows);
1278 :
1279 250952 : if (rowstoskip <= 0)
1280 : {
1281 : /*
1282 : * Found a suitable tuple, so save it, replacing one old
1283 : * tuple at random
1284 : */
1285 115312 : int k = (int) (targrows * sampler_random_fract(&rstate.randstate));
1286 :
1287 : Assert(k >= 0 && k < targrows);
1288 115312 : heap_freetuple(rows[k]);
1289 115312 : rows[k] = ExecCopySlotHeapTuple(slot);
1290 : }
1291 :
1292 250952 : rowstoskip -= 1;
1293 : }
1294 :
1295 9825088 : samplerows += 1;
1296 : }
1297 :
1298 125422 : pgstat_progress_update_param(PROGRESS_ANALYZE_BLOCKS_DONE,
1299 : ++blksdone);
1300 : }
1301 :
1302 15412 : read_stream_end(stream);
1303 :
1304 15412 : ExecDropSingleTupleTableSlot(slot);
1305 15412 : table_endscan(scan);
1306 :
1307 : /*
1308 : * If we didn't find as many tuples as we wanted then we're done. No sort
1309 : * is needed, since they're already in order.
1310 : *
1311 : * Otherwise we need to sort the collected tuples by position
1312 : * (itempointer). It's not worth worrying about corner cases where the
1313 : * tuples are already sorted.
1314 : */
1315 15412 : if (numrows == targrows)
1316 162 : qsort_interruptible(rows, numrows, sizeof(HeapTuple),
1317 : compare_rows, NULL);
1318 :
1319 : /*
1320 : * Estimate total numbers of live and dead rows in relation, extrapolating
1321 : * on the assumption that the average tuple density in pages we didn't
1322 : * scan is the same as in the pages we did scan. Since what we scanned is
1323 : * a random sample of the pages in the relation, this should be a good
1324 : * assumption.
1325 : */
1326 15412 : if (bs.m > 0)
1327 : {
1328 10968 : *totalrows = floor((liverows / bs.m) * totalblocks + 0.5);
1329 10968 : *totaldeadrows = floor((deadrows / bs.m) * totalblocks + 0.5);
1330 : }
1331 : else
1332 : {
1333 4444 : *totalrows = 0.0;
1334 4444 : *totaldeadrows = 0.0;
1335 : }
1336 :
1337 : /*
1338 : * Emit some interesting relation info
1339 : */
1340 15412 : ereport(elevel,
1341 : (errmsg("\"%s\": scanned %d of %u pages, "
1342 : "containing %.0f live rows and %.0f dead rows; "
1343 : "%d rows in sample, %.0f estimated total rows",
1344 : RelationGetRelationName(onerel),
1345 : bs.m, totalblocks,
1346 : liverows, deadrows,
1347 : numrows, *totalrows)));
1348 :
1349 15412 : return numrows;
1350 : }
1351 :
1352 : /*
1353 : * Comparator for sorting rows[] array
1354 : */
1355 : static int
1356 4023782 : compare_rows(const void *a, const void *b, void *arg)
1357 : {
1358 4023782 : HeapTuple ha = *(const HeapTuple *) a;
1359 4023782 : HeapTuple hb = *(const HeapTuple *) b;
1360 4023782 : BlockNumber ba = ItemPointerGetBlockNumber(&ha->t_self);
1361 4023782 : OffsetNumber oa = ItemPointerGetOffsetNumber(&ha->t_self);
1362 4023782 : BlockNumber bb = ItemPointerGetBlockNumber(&hb->t_self);
1363 4023782 : OffsetNumber ob = ItemPointerGetOffsetNumber(&hb->t_self);
1364 :
1365 4023782 : if (ba < bb)
1366 889988 : return -1;
1367 3133794 : if (ba > bb)
1368 860026 : return 1;
1369 2273768 : if (oa < ob)
1370 1529874 : return -1;
1371 743894 : if (oa > ob)
1372 743894 : return 1;
1373 0 : return 0;
1374 : }
1375 :
1376 :
1377 : /*
1378 : * acquire_inherited_sample_rows -- acquire sample rows from inheritance tree
1379 : *
1380 : * This has the same API as acquire_sample_rows, except that rows are
1381 : * collected from all inheritance children as well as the specified table.
1382 : * We fail and return zero if there are no inheritance children, or if all
1383 : * children are foreign tables that don't support ANALYZE.
1384 : */
1385 : static int
1386 824 : acquire_inherited_sample_rows(Relation onerel, int elevel,
1387 : HeapTuple *rows, int targrows,
1388 : double *totalrows, double *totaldeadrows)
1389 : {
1390 : List *tableOIDs;
1391 : Relation *rels;
1392 : AcquireSampleRowsFunc *acquirefuncs;
1393 : double *relblocks;
1394 : double totalblocks;
1395 : int numrows,
1396 : nrels,
1397 : i;
1398 : ListCell *lc;
1399 : bool has_child;
1400 :
1401 : /* Initialize output parameters to zero now, in case we exit early */
1402 824 : *totalrows = 0;
1403 824 : *totaldeadrows = 0;
1404 :
1405 : /*
1406 : * Find all members of inheritance set. We only need AccessShareLock on
1407 : * the children.
1408 : */
1409 : tableOIDs =
1410 824 : find_all_inheritors(RelationGetRelid(onerel), AccessShareLock, NULL);
1411 :
1412 : /*
1413 : * Check that there's at least one descendant, else fail. This could
1414 : * happen despite analyze_rel's relhassubclass check, if table once had a
1415 : * child but no longer does. In that case, we can clear the
1416 : * relhassubclass field so as not to make the same mistake again later.
1417 : * (This is safe because we hold ShareUpdateExclusiveLock.)
1418 : */
1419 824 : if (list_length(tableOIDs) < 2)
1420 : {
1421 : /* CCI because we already updated the pg_class row in this command */
1422 14 : CommandCounterIncrement();
1423 14 : SetRelationHasSubclass(RelationGetRelid(onerel), false);
1424 14 : ereport(elevel,
1425 : (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no child tables",
1426 : get_namespace_name(RelationGetNamespace(onerel)),
1427 : RelationGetRelationName(onerel))));
1428 14 : return 0;
1429 : }
1430 :
1431 : /*
1432 : * Identify acquirefuncs to use, and count blocks in all the relations.
1433 : * The result could overflow BlockNumber, so we use double arithmetic.
1434 : */
1435 810 : rels = (Relation *) palloc(list_length(tableOIDs) * sizeof(Relation));
1436 : acquirefuncs = (AcquireSampleRowsFunc *)
1437 810 : palloc(list_length(tableOIDs) * sizeof(AcquireSampleRowsFunc));
1438 810 : relblocks = (double *) palloc(list_length(tableOIDs) * sizeof(double));
1439 810 : totalblocks = 0;
1440 810 : nrels = 0;
1441 810 : has_child = false;
1442 3728 : foreach(lc, tableOIDs)
1443 : {
1444 2918 : Oid childOID = lfirst_oid(lc);
1445 : Relation childrel;
1446 2918 : AcquireSampleRowsFunc acquirefunc = NULL;
1447 2918 : BlockNumber relpages = 0;
1448 :
1449 : /* We already got the needed lock */
1450 2918 : childrel = table_open(childOID, NoLock);
1451 :
1452 : /* Ignore if temp table of another backend */
1453 2918 : if (RELATION_IS_OTHER_TEMP(childrel))
1454 : {
1455 : /* ... but release the lock on it */
1456 : Assert(childrel != onerel);
1457 0 : table_close(childrel, AccessShareLock);
1458 784 : continue;
1459 : }
1460 :
1461 : /* Check table type (MATVIEW can't happen, but might as well allow) */
1462 2918 : if (childrel->rd_rel->relkind == RELKIND_RELATION ||
1463 814 : childrel->rd_rel->relkind == RELKIND_MATVIEW)
1464 : {
1465 : /* Regular table, so use the regular row acquisition function */
1466 2104 : acquirefunc = acquire_sample_rows;
1467 2104 : relpages = RelationGetNumberOfBlocks(childrel);
1468 : }
1469 814 : else if (childrel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
1470 : {
1471 : /*
1472 : * For a foreign table, call the FDW's hook function to see
1473 : * whether it supports analysis.
1474 : */
1475 : FdwRoutine *fdwroutine;
1476 30 : bool ok = false;
1477 :
1478 30 : fdwroutine = GetFdwRoutineForRelation(childrel, false);
1479 :
1480 30 : if (fdwroutine->AnalyzeForeignTable != NULL)
1481 30 : ok = fdwroutine->AnalyzeForeignTable(childrel,
1482 : &acquirefunc,
1483 : &relpages);
1484 :
1485 30 : if (!ok)
1486 : {
1487 : /* ignore, but release the lock on it */
1488 : Assert(childrel != onerel);
1489 0 : table_close(childrel, AccessShareLock);
1490 0 : continue;
1491 : }
1492 : }
1493 : else
1494 : {
1495 : /*
1496 : * ignore, but release the lock on it. don't try to unlock the
1497 : * passed-in relation
1498 : */
1499 : Assert(childrel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE);
1500 784 : if (childrel != onerel)
1501 74 : table_close(childrel, AccessShareLock);
1502 : else
1503 710 : table_close(childrel, NoLock);
1504 784 : continue;
1505 : }
1506 :
1507 : /* OK, we'll process this child */
1508 2134 : has_child = true;
1509 2134 : rels[nrels] = childrel;
1510 2134 : acquirefuncs[nrels] = acquirefunc;
1511 2134 : relblocks[nrels] = (double) relpages;
1512 2134 : totalblocks += (double) relpages;
1513 2134 : nrels++;
1514 : }
1515 :
1516 : /*
1517 : * If we don't have at least one child table to consider, fail. If the
1518 : * relation is a partitioned table, it's not counted as a child table.
1519 : */
1520 810 : if (!has_child)
1521 : {
1522 0 : ereport(elevel,
1523 : (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no analyzable child tables",
1524 : get_namespace_name(RelationGetNamespace(onerel)),
1525 : RelationGetRelationName(onerel))));
1526 0 : return 0;
1527 : }
1528 :
1529 : /*
1530 : * Now sample rows from each relation, proportionally to its fraction of
1531 : * the total block count. (This might be less than desirable if the child
1532 : * rels have radically different free-space percentages, but it's not
1533 : * clear that it's worth working harder.)
1534 : */
1535 810 : pgstat_progress_update_param(PROGRESS_ANALYZE_CHILD_TABLES_TOTAL,
1536 : nrels);
1537 810 : numrows = 0;
1538 2944 : for (i = 0; i < nrels; i++)
1539 : {
1540 2134 : Relation childrel = rels[i];
1541 2134 : AcquireSampleRowsFunc acquirefunc = acquirefuncs[i];
1542 2134 : double childblocks = relblocks[i];
1543 :
1544 : /*
1545 : * Report progress. The sampling function will normally report blocks
1546 : * done/total, but we need to reset them to 0 here, so that they don't
1547 : * show an old value until that.
1548 : */
1549 : {
1550 2134 : const int progress_index[] = {
1551 : PROGRESS_ANALYZE_CURRENT_CHILD_TABLE_RELID,
1552 : PROGRESS_ANALYZE_BLOCKS_DONE,
1553 : PROGRESS_ANALYZE_BLOCKS_TOTAL
1554 : };
1555 2134 : const int64 progress_vals[] = {
1556 2134 : RelationGetRelid(childrel),
1557 : 0,
1558 : 0,
1559 : };
1560 :
1561 2134 : pgstat_progress_update_multi_param(3, progress_index, progress_vals);
1562 : }
1563 :
1564 2134 : if (childblocks > 0)
1565 : {
1566 : int childtargrows;
1567 :
1568 1964 : childtargrows = (int) rint(targrows * childblocks / totalblocks);
1569 : /* Make sure we don't overrun due to roundoff error */
1570 1964 : childtargrows = Min(childtargrows, targrows - numrows);
1571 1964 : if (childtargrows > 0)
1572 : {
1573 : int childrows;
1574 : double trows,
1575 : tdrows;
1576 :
1577 : /* Fetch a random sample of the child's rows */
1578 1964 : childrows = (*acquirefunc) (childrel, elevel,
1579 1964 : rows + numrows, childtargrows,
1580 : &trows, &tdrows);
1581 :
1582 : /* We may need to convert from child's rowtype to parent's */
1583 1964 : if (childrows > 0 &&
1584 1964 : !equalRowTypes(RelationGetDescr(childrel),
1585 : RelationGetDescr(onerel)))
1586 : {
1587 : TupleConversionMap *map;
1588 :
1589 1882 : map = convert_tuples_by_name(RelationGetDescr(childrel),
1590 : RelationGetDescr(onerel));
1591 1882 : if (map != NULL)
1592 : {
1593 : int j;
1594 :
1595 106604 : for (j = 0; j < childrows; j++)
1596 : {
1597 : HeapTuple newtup;
1598 :
1599 106472 : newtup = execute_attr_map_tuple(rows[numrows + j], map);
1600 106472 : heap_freetuple(rows[numrows + j]);
1601 106472 : rows[numrows + j] = newtup;
1602 : }
1603 132 : free_conversion_map(map);
1604 : }
1605 : }
1606 :
1607 : /* And add to counts */
1608 1964 : numrows += childrows;
1609 1964 : *totalrows += trows;
1610 1964 : *totaldeadrows += tdrows;
1611 : }
1612 : }
1613 :
1614 : /*
1615 : * Note: we cannot release the child-table locks, since we may have
1616 : * pointers to their TOAST tables in the sampled rows.
1617 : */
1618 2134 : table_close(childrel, NoLock);
1619 2134 : pgstat_progress_update_param(PROGRESS_ANALYZE_CHILD_TABLES_DONE,
1620 2134 : i + 1);
1621 : }
1622 :
1623 810 : return numrows;
1624 : }
1625 :
1626 :
1627 : /*
1628 : * update_attstats() -- update attribute statistics for one relation
1629 : *
1630 : * Statistics are stored in several places: the pg_class row for the
1631 : * relation has stats about the whole relation, and there is a
1632 : * pg_statistic row for each (non-system) attribute that has ever
1633 : * been analyzed. The pg_class values are updated by VACUUM, not here.
1634 : *
1635 : * pg_statistic rows are just added or updated normally. This means
1636 : * that pg_statistic will probably contain some deleted rows at the
1637 : * completion of a vacuum cycle, unless it happens to get vacuumed last.
1638 : *
1639 : * To keep things simple, we punt for pg_statistic, and don't try
1640 : * to compute or store rows for pg_statistic itself in pg_statistic.
1641 : * This could possibly be made to work, but it's not worth the trouble.
1642 : * Note analyze_rel() has seen to it that we won't come here when
1643 : * vacuuming pg_statistic itself.
1644 : *
1645 : * Note: there would be a race condition here if two backends could
1646 : * ANALYZE the same table concurrently. Presently, we lock that out
1647 : * by taking a self-exclusive lock on the relation in analyze_rel().
1648 : */
1649 : static void
1650 21902 : update_attstats(Oid relid, bool inh, int natts, VacAttrStats **vacattrstats)
1651 : {
1652 : Relation sd;
1653 : int attno;
1654 21902 : CatalogIndexState indstate = NULL;
1655 :
1656 21902 : if (natts <= 0)
1657 12030 : return; /* nothing to do */
1658 :
1659 9872 : sd = table_open(StatisticRelationId, RowExclusiveLock);
1660 :
1661 85302 : for (attno = 0; attno < natts; attno++)
1662 : {
1663 75430 : VacAttrStats *stats = vacattrstats[attno];
1664 : HeapTuple stup,
1665 : oldtup;
1666 : int i,
1667 : k,
1668 : n;
1669 : Datum values[Natts_pg_statistic];
1670 : bool nulls[Natts_pg_statistic];
1671 : bool replaces[Natts_pg_statistic];
1672 :
1673 : /* Ignore attr if we weren't able to collect stats */
1674 75430 : if (!stats->stats_valid)
1675 6 : continue;
1676 :
1677 : /*
1678 : * Construct a new pg_statistic tuple
1679 : */
1680 2413568 : for (i = 0; i < Natts_pg_statistic; ++i)
1681 : {
1682 2338144 : nulls[i] = false;
1683 2338144 : replaces[i] = true;
1684 : }
1685 :
1686 75424 : values[Anum_pg_statistic_starelid - 1] = ObjectIdGetDatum(relid);
1687 75424 : values[Anum_pg_statistic_staattnum - 1] = Int16GetDatum(stats->tupattnum);
1688 75424 : values[Anum_pg_statistic_stainherit - 1] = BoolGetDatum(inh);
1689 75424 : values[Anum_pg_statistic_stanullfrac - 1] = Float4GetDatum(stats->stanullfrac);
1690 75424 : values[Anum_pg_statistic_stawidth - 1] = Int32GetDatum(stats->stawidth);
1691 75424 : values[Anum_pg_statistic_stadistinct - 1] = Float4GetDatum(stats->stadistinct);
1692 75424 : i = Anum_pg_statistic_stakind1 - 1;
1693 452544 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1694 : {
1695 377120 : values[i++] = Int16GetDatum(stats->stakind[k]); /* stakindN */
1696 : }
1697 75424 : i = Anum_pg_statistic_staop1 - 1;
1698 452544 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1699 : {
1700 377120 : values[i++] = ObjectIdGetDatum(stats->staop[k]); /* staopN */
1701 : }
1702 75424 : i = Anum_pg_statistic_stacoll1 - 1;
1703 452544 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1704 : {
1705 377120 : values[i++] = ObjectIdGetDatum(stats->stacoll[k]); /* stacollN */
1706 : }
1707 75424 : i = Anum_pg_statistic_stanumbers1 - 1;
1708 452544 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1709 : {
1710 377120 : int nnum = stats->numnumbers[k];
1711 :
1712 377120 : if (nnum > 0)
1713 : {
1714 117338 : Datum *numdatums = (Datum *) palloc(nnum * sizeof(Datum));
1715 : ArrayType *arry;
1716 :
1717 969596 : for (n = 0; n < nnum; n++)
1718 852258 : numdatums[n] = Float4GetDatum(stats->stanumbers[k][n]);
1719 117338 : arry = construct_array_builtin(numdatums, nnum, FLOAT4OID);
1720 117338 : values[i++] = PointerGetDatum(arry); /* stanumbersN */
1721 : }
1722 : else
1723 : {
1724 259782 : nulls[i] = true;
1725 259782 : values[i++] = (Datum) 0;
1726 : }
1727 : }
1728 75424 : i = Anum_pg_statistic_stavalues1 - 1;
1729 452544 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1730 : {
1731 377120 : if (stats->numvalues[k] > 0)
1732 : {
1733 : ArrayType *arry;
1734 :
1735 82694 : arry = construct_array(stats->stavalues[k],
1736 : stats->numvalues[k],
1737 : stats->statypid[k],
1738 82694 : stats->statyplen[k],
1739 82694 : stats->statypbyval[k],
1740 82694 : stats->statypalign[k]);
1741 82694 : values[i++] = PointerGetDatum(arry); /* stavaluesN */
1742 : }
1743 : else
1744 : {
1745 294426 : nulls[i] = true;
1746 294426 : values[i++] = (Datum) 0;
1747 : }
1748 : }
1749 :
1750 : /* Is there already a pg_statistic tuple for this attribute? */
1751 150848 : oldtup = SearchSysCache3(STATRELATTINH,
1752 : ObjectIdGetDatum(relid),
1753 75424 : Int16GetDatum(stats->tupattnum),
1754 : BoolGetDatum(inh));
1755 :
1756 : /* Open index information when we know we need it */
1757 75424 : if (indstate == NULL)
1758 9866 : indstate = CatalogOpenIndexes(sd);
1759 :
1760 75424 : if (HeapTupleIsValid(oldtup))
1761 : {
1762 : /* Yes, replace it */
1763 32490 : stup = heap_modify_tuple(oldtup,
1764 : RelationGetDescr(sd),
1765 : values,
1766 : nulls,
1767 : replaces);
1768 32490 : ReleaseSysCache(oldtup);
1769 32490 : CatalogTupleUpdateWithInfo(sd, &stup->t_self, stup, indstate);
1770 : }
1771 : else
1772 : {
1773 : /* No, insert new tuple */
1774 42934 : stup = heap_form_tuple(RelationGetDescr(sd), values, nulls);
1775 42934 : CatalogTupleInsertWithInfo(sd, stup, indstate);
1776 : }
1777 :
1778 75424 : heap_freetuple(stup);
1779 : }
1780 :
1781 9872 : if (indstate != NULL)
1782 9866 : CatalogCloseIndexes(indstate);
1783 9872 : table_close(sd, RowExclusiveLock);
1784 : }
1785 :
1786 : /*
1787 : * Standard fetch function for use by compute_stats subroutines.
1788 : *
1789 : * This exists to provide some insulation between compute_stats routines
1790 : * and the actual storage of the sample data.
1791 : */
1792 : static Datum
1793 71876178 : std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1794 : {
1795 71876178 : int attnum = stats->tupattnum;
1796 71876178 : HeapTuple tuple = stats->rows[rownum];
1797 71876178 : TupleDesc tupDesc = stats->tupDesc;
1798 :
1799 71876178 : return heap_getattr(tuple, attnum, tupDesc, isNull);
1800 : }
1801 :
1802 : /*
1803 : * Fetch function for analyzing index expressions.
1804 : *
1805 : * We have not bothered to construct index tuples, instead the data is
1806 : * just in Datum arrays.
1807 : */
1808 : static Datum
1809 114884 : ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1810 : {
1811 : int i;
1812 :
1813 : /* exprvals and exprnulls are already offset for proper column */
1814 114884 : i = rownum * stats->rowstride;
1815 114884 : *isNull = stats->exprnulls[i];
1816 114884 : return stats->exprvals[i];
1817 : }
1818 :
1819 :
1820 : /*==========================================================================
1821 : *
1822 : * Code below this point represents the "standard" type-specific statistics
1823 : * analysis algorithms. This code can be replaced on a per-data-type basis
1824 : * by setting a nonzero value in pg_type.typanalyze.
1825 : *
1826 : *==========================================================================
1827 : */
1828 :
1829 :
1830 : /*
1831 : * To avoid consuming too much memory during analysis and/or too much space
1832 : * in the resulting pg_statistic rows, we ignore varlena datums that are wider
1833 : * than WIDTH_THRESHOLD (after detoasting!). This is legitimate for MCV
1834 : * and distinct-value calculations since a wide value is unlikely to be
1835 : * duplicated at all, much less be a most-common value. For the same reason,
1836 : * ignoring wide values will not affect our estimates of histogram bin
1837 : * boundaries very much.
1838 : */
1839 : #define WIDTH_THRESHOLD 1024
1840 :
1841 : #define swapInt(a,b) do {int _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1842 : #define swapDatum(a,b) do {Datum _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1843 :
1844 : /*
1845 : * Extra information used by the default analysis routines
1846 : */
1847 : typedef struct
1848 : {
1849 : int count; /* # of duplicates */
1850 : int first; /* values[] index of first occurrence */
1851 : } ScalarMCVItem;
1852 :
1853 : typedef struct
1854 : {
1855 : SortSupport ssup;
1856 : int *tupnoLink;
1857 : } CompareScalarsContext;
1858 :
1859 :
1860 : static void compute_trivial_stats(VacAttrStatsP stats,
1861 : AnalyzeAttrFetchFunc fetchfunc,
1862 : int samplerows,
1863 : double totalrows);
1864 : static void compute_distinct_stats(VacAttrStatsP stats,
1865 : AnalyzeAttrFetchFunc fetchfunc,
1866 : int samplerows,
1867 : double totalrows);
1868 : static void compute_scalar_stats(VacAttrStatsP stats,
1869 : AnalyzeAttrFetchFunc fetchfunc,
1870 : int samplerows,
1871 : double totalrows);
1872 : static int compare_scalars(const void *a, const void *b, void *arg);
1873 : static int compare_mcvs(const void *a, const void *b, void *arg);
1874 : static int analyze_mcv_list(int *mcv_counts,
1875 : int num_mcv,
1876 : double stadistinct,
1877 : double stanullfrac,
1878 : int samplerows,
1879 : double totalrows);
1880 :
1881 :
1882 : /*
1883 : * std_typanalyze -- the default type-specific typanalyze function
1884 : */
1885 : bool
1886 103150 : std_typanalyze(VacAttrStats *stats)
1887 : {
1888 : Oid ltopr;
1889 : Oid eqopr;
1890 : StdAnalyzeData *mystats;
1891 :
1892 : /* If the attstattarget column is negative, use the default value */
1893 103150 : if (stats->attstattarget < 0)
1894 102544 : stats->attstattarget = default_statistics_target;
1895 :
1896 : /* Look for default "<" and "=" operators for column's type */
1897 103150 : get_sort_group_operators(stats->attrtypid,
1898 : false, false, false,
1899 : <opr, &eqopr, NULL,
1900 : NULL);
1901 :
1902 : /* Save the operator info for compute_stats routines */
1903 103150 : mystats = (StdAnalyzeData *) palloc(sizeof(StdAnalyzeData));
1904 103150 : mystats->eqopr = eqopr;
1905 103150 : mystats->eqfunc = OidIsValid(eqopr) ? get_opcode(eqopr) : InvalidOid;
1906 103150 : mystats->ltopr = ltopr;
1907 103150 : stats->extra_data = mystats;
1908 :
1909 : /*
1910 : * Determine which standard statistics algorithm to use
1911 : */
1912 103150 : if (OidIsValid(eqopr) && OidIsValid(ltopr))
1913 : {
1914 : /* Seems to be a scalar datatype */
1915 99962 : stats->compute_stats = compute_scalar_stats;
1916 : /*--------------------
1917 : * The following choice of minrows is based on the paper
1918 : * "Random sampling for histogram construction: how much is enough?"
1919 : * by Surajit Chaudhuri, Rajeev Motwani and Vivek Narasayya, in
1920 : * Proceedings of ACM SIGMOD International Conference on Management
1921 : * of Data, 1998, Pages 436-447. Their Corollary 1 to Theorem 5
1922 : * says that for table size n, histogram size k, maximum relative
1923 : * error in bin size f, and error probability gamma, the minimum
1924 : * random sample size is
1925 : * r = 4 * k * ln(2*n/gamma) / f^2
1926 : * Taking f = 0.5, gamma = 0.01, n = 10^6 rows, we obtain
1927 : * r = 305.82 * k
1928 : * Note that because of the log function, the dependence on n is
1929 : * quite weak; even at n = 10^12, a 300*k sample gives <= 0.66
1930 : * bin size error with probability 0.99. So there's no real need to
1931 : * scale for n, which is a good thing because we don't necessarily
1932 : * know it at this point.
1933 : *--------------------
1934 : */
1935 99962 : stats->minrows = 300 * stats->attstattarget;
1936 : }
1937 3188 : else if (OidIsValid(eqopr))
1938 : {
1939 : /* We can still recognize distinct values */
1940 2728 : stats->compute_stats = compute_distinct_stats;
1941 : /* Might as well use the same minrows as above */
1942 2728 : stats->minrows = 300 * stats->attstattarget;
1943 : }
1944 : else
1945 : {
1946 : /* Can't do much but the trivial stuff */
1947 460 : stats->compute_stats = compute_trivial_stats;
1948 : /* Might as well use the same minrows as above */
1949 460 : stats->minrows = 300 * stats->attstattarget;
1950 : }
1951 :
1952 103150 : return true;
1953 : }
1954 :
1955 :
1956 : /*
1957 : * compute_trivial_stats() -- compute very basic column statistics
1958 : *
1959 : * We use this when we cannot find a hash "=" operator for the datatype.
1960 : *
1961 : * We determine the fraction of non-null rows and the average datum width.
1962 : */
1963 : static void
1964 316 : compute_trivial_stats(VacAttrStatsP stats,
1965 : AnalyzeAttrFetchFunc fetchfunc,
1966 : int samplerows,
1967 : double totalrows)
1968 : {
1969 : int i;
1970 316 : int null_cnt = 0;
1971 316 : int nonnull_cnt = 0;
1972 316 : double total_width = 0;
1973 632 : bool is_varlena = (!stats->attrtype->typbyval &&
1974 316 : stats->attrtype->typlen == -1);
1975 632 : bool is_varwidth = (!stats->attrtype->typbyval &&
1976 316 : stats->attrtype->typlen < 0);
1977 :
1978 943614 : for (i = 0; i < samplerows; i++)
1979 : {
1980 : Datum value;
1981 : bool isnull;
1982 :
1983 943298 : vacuum_delay_point(true);
1984 :
1985 943298 : value = fetchfunc(stats, i, &isnull);
1986 :
1987 : /* Check for null/nonnull */
1988 943298 : if (isnull)
1989 : {
1990 529422 : null_cnt++;
1991 529422 : continue;
1992 : }
1993 413876 : nonnull_cnt++;
1994 :
1995 : /*
1996 : * If it's a variable-width field, add up widths for average width
1997 : * calculation. Note that if the value is toasted, we use the toasted
1998 : * width. We don't bother with this calculation if it's a fixed-width
1999 : * type.
2000 : */
2001 413876 : if (is_varlena)
2002 : {
2003 79884 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2004 : }
2005 333992 : else if (is_varwidth)
2006 : {
2007 : /* must be cstring */
2008 0 : total_width += strlen(DatumGetCString(value)) + 1;
2009 : }
2010 : }
2011 :
2012 : /* We can only compute average width if we found some non-null values. */
2013 316 : if (nonnull_cnt > 0)
2014 : {
2015 150 : stats->stats_valid = true;
2016 : /* Do the simple null-frac and width stats */
2017 150 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2018 150 : if (is_varwidth)
2019 68 : stats->stawidth = total_width / (double) nonnull_cnt;
2020 : else
2021 82 : stats->stawidth = stats->attrtype->typlen;
2022 150 : stats->stadistinct = 0.0; /* "unknown" */
2023 : }
2024 166 : else if (null_cnt > 0)
2025 : {
2026 : /* We found only nulls; assume the column is entirely null */
2027 166 : stats->stats_valid = true;
2028 166 : stats->stanullfrac = 1.0;
2029 166 : if (is_varwidth)
2030 166 : stats->stawidth = 0; /* "unknown" */
2031 : else
2032 0 : stats->stawidth = stats->attrtype->typlen;
2033 166 : stats->stadistinct = 0.0; /* "unknown" */
2034 : }
2035 316 : }
2036 :
2037 :
2038 : /*
2039 : * compute_distinct_stats() -- compute column statistics including ndistinct
2040 : *
2041 : * We use this when we can find only an "=" operator for the datatype.
2042 : *
2043 : * We determine the fraction of non-null rows, the average width, the
2044 : * most common values, and the (estimated) number of distinct values.
2045 : *
2046 : * The most common values are determined by brute force: we keep a list
2047 : * of previously seen values, ordered by number of times seen, as we scan
2048 : * the samples. A newly seen value is inserted just after the last
2049 : * multiply-seen value, causing the bottommost (oldest) singly-seen value
2050 : * to drop off the list. The accuracy of this method, and also its cost,
2051 : * depend mainly on the length of the list we are willing to keep.
2052 : */
2053 : static void
2054 2004 : compute_distinct_stats(VacAttrStatsP stats,
2055 : AnalyzeAttrFetchFunc fetchfunc,
2056 : int samplerows,
2057 : double totalrows)
2058 : {
2059 : int i;
2060 2004 : int null_cnt = 0;
2061 2004 : int nonnull_cnt = 0;
2062 2004 : int toowide_cnt = 0;
2063 2004 : double total_width = 0;
2064 3388 : bool is_varlena = (!stats->attrtype->typbyval &&
2065 1384 : stats->attrtype->typlen == -1);
2066 3388 : bool is_varwidth = (!stats->attrtype->typbyval &&
2067 1384 : stats->attrtype->typlen < 0);
2068 : FmgrInfo f_cmpeq;
2069 : typedef struct
2070 : {
2071 : Datum value;
2072 : int count;
2073 : } TrackItem;
2074 : TrackItem *track;
2075 : int track_cnt,
2076 : track_max;
2077 2004 : int num_mcv = stats->attstattarget;
2078 2004 : StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2079 :
2080 : /*
2081 : * We track up to 2*n values for an n-element MCV list; but at least 10
2082 : */
2083 2004 : track_max = 2 * num_mcv;
2084 2004 : if (track_max < 10)
2085 78 : track_max = 10;
2086 2004 : track = (TrackItem *) palloc(track_max * sizeof(TrackItem));
2087 2004 : track_cnt = 0;
2088 :
2089 2004 : fmgr_info(mystats->eqfunc, &f_cmpeq);
2090 :
2091 1356744 : for (i = 0; i < samplerows; i++)
2092 : {
2093 : Datum value;
2094 : bool isnull;
2095 : bool match;
2096 : int firstcount1,
2097 : j;
2098 :
2099 1354740 : vacuum_delay_point(true);
2100 :
2101 1354740 : value = fetchfunc(stats, i, &isnull);
2102 :
2103 : /* Check for null/nonnull */
2104 1354740 : if (isnull)
2105 : {
2106 1129694 : null_cnt++;
2107 1129694 : continue;
2108 : }
2109 225046 : nonnull_cnt++;
2110 :
2111 : /*
2112 : * If it's a variable-width field, add up widths for average width
2113 : * calculation. Note that if the value is toasted, we use the toasted
2114 : * width. We don't bother with this calculation if it's a fixed-width
2115 : * type.
2116 : */
2117 225046 : if (is_varlena)
2118 : {
2119 82350 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2120 :
2121 : /*
2122 : * If the value is toasted, we want to detoast it just once to
2123 : * avoid repeated detoastings and resultant excess memory usage
2124 : * during the comparisons. Also, check to see if the value is
2125 : * excessively wide, and if so don't detoast at all --- just
2126 : * ignore the value.
2127 : */
2128 82350 : if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
2129 : {
2130 0 : toowide_cnt++;
2131 0 : continue;
2132 : }
2133 82350 : value = PointerGetDatum(PG_DETOAST_DATUM(value));
2134 : }
2135 142696 : else if (is_varwidth)
2136 : {
2137 : /* must be cstring */
2138 0 : total_width += strlen(DatumGetCString(value)) + 1;
2139 : }
2140 :
2141 : /*
2142 : * See if the value matches anything we're already tracking.
2143 : */
2144 225046 : match = false;
2145 225046 : firstcount1 = track_cnt;
2146 426974 : for (j = 0; j < track_cnt; j++)
2147 : {
2148 421364 : if (DatumGetBool(FunctionCall2Coll(&f_cmpeq,
2149 : stats->attrcollid,
2150 421364 : value, track[j].value)))
2151 : {
2152 219436 : match = true;
2153 219436 : break;
2154 : }
2155 201928 : if (j < firstcount1 && track[j].count == 1)
2156 3356 : firstcount1 = j;
2157 : }
2158 :
2159 225046 : if (match)
2160 : {
2161 : /* Found a match */
2162 219436 : track[j].count++;
2163 : /* This value may now need to "bubble up" in the track list */
2164 223990 : while (j > 0 && track[j].count > track[j - 1].count)
2165 : {
2166 4554 : swapDatum(track[j].value, track[j - 1].value);
2167 4554 : swapInt(track[j].count, track[j - 1].count);
2168 4554 : j--;
2169 : }
2170 : }
2171 : else
2172 : {
2173 : /* No match. Insert at head of count-1 list */
2174 5610 : if (track_cnt < track_max)
2175 5312 : track_cnt++;
2176 78234 : for (j = track_cnt - 1; j > firstcount1; j--)
2177 : {
2178 72624 : track[j].value = track[j - 1].value;
2179 72624 : track[j].count = track[j - 1].count;
2180 : }
2181 5610 : if (firstcount1 < track_cnt)
2182 : {
2183 5610 : track[firstcount1].value = value;
2184 5610 : track[firstcount1].count = 1;
2185 : }
2186 : }
2187 : }
2188 :
2189 : /* We can only compute real stats if we found some non-null values. */
2190 2004 : if (nonnull_cnt > 0)
2191 : {
2192 : int nmultiple,
2193 : summultiple;
2194 :
2195 1466 : stats->stats_valid = true;
2196 : /* Do the simple null-frac and width stats */
2197 1466 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2198 1466 : if (is_varwidth)
2199 846 : stats->stawidth = total_width / (double) nonnull_cnt;
2200 : else
2201 620 : stats->stawidth = stats->attrtype->typlen;
2202 :
2203 : /* Count the number of values we found multiple times */
2204 1466 : summultiple = 0;
2205 5346 : for (nmultiple = 0; nmultiple < track_cnt; nmultiple++)
2206 : {
2207 4622 : if (track[nmultiple].count == 1)
2208 742 : break;
2209 3880 : summultiple += track[nmultiple].count;
2210 : }
2211 :
2212 1466 : if (nmultiple == 0)
2213 : {
2214 : /*
2215 : * If we found no repeated non-null values, assume it's a unique
2216 : * column; but be sure to discount for any nulls we found.
2217 : */
2218 174 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2219 : }
2220 1292 : else if (track_cnt < track_max && toowide_cnt == 0 &&
2221 : nmultiple == track_cnt)
2222 : {
2223 : /*
2224 : * Our track list includes every value in the sample, and every
2225 : * value appeared more than once. Assume the column has just
2226 : * these values. (This case is meant to address columns with
2227 : * small, fixed sets of possible values, such as boolean or enum
2228 : * columns. If there are any values that appear just once in the
2229 : * sample, including too-wide values, we should assume that that's
2230 : * not what we're dealing with.)
2231 : */
2232 724 : stats->stadistinct = track_cnt;
2233 : }
2234 : else
2235 : {
2236 : /*----------
2237 : * Estimate the number of distinct values using the estimator
2238 : * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2239 : * n*d / (n - f1 + f1*n/N)
2240 : * where f1 is the number of distinct values that occurred
2241 : * exactly once in our sample of n rows (from a total of N),
2242 : * and d is the total number of distinct values in the sample.
2243 : * This is their Duj1 estimator; the other estimators they
2244 : * recommend are considerably more complex, and are numerically
2245 : * very unstable when n is much smaller than N.
2246 : *
2247 : * In this calculation, we consider only non-nulls. We used to
2248 : * include rows with null values in the n and N counts, but that
2249 : * leads to inaccurate answers in columns with many nulls, and
2250 : * it's intuitively bogus anyway considering the desired result is
2251 : * the number of distinct non-null values.
2252 : *
2253 : * We assume (not very reliably!) that all the multiply-occurring
2254 : * values are reflected in the final track[] list, and the other
2255 : * nonnull values all appeared but once. (XXX this usually
2256 : * results in a drastic overestimate of ndistinct. Can we do
2257 : * any better?)
2258 : *----------
2259 : */
2260 568 : int f1 = nonnull_cnt - summultiple;
2261 568 : int d = f1 + nmultiple;
2262 568 : double n = samplerows - null_cnt;
2263 568 : double N = totalrows * (1.0 - stats->stanullfrac);
2264 : double stadistinct;
2265 :
2266 : /* N == 0 shouldn't happen, but just in case ... */
2267 568 : if (N > 0)
2268 568 : stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2269 : else
2270 0 : stadistinct = 0;
2271 :
2272 : /* Clamp to sane range in case of roundoff error */
2273 568 : if (stadistinct < d)
2274 156 : stadistinct = d;
2275 568 : if (stadistinct > N)
2276 0 : stadistinct = N;
2277 : /* And round to integer */
2278 568 : stats->stadistinct = floor(stadistinct + 0.5);
2279 : }
2280 :
2281 : /*
2282 : * If we estimated the number of distinct values at more than 10% of
2283 : * the total row count (a very arbitrary limit), then assume that
2284 : * stadistinct should scale with the row count rather than be a fixed
2285 : * value.
2286 : */
2287 1466 : if (stats->stadistinct > 0.1 * totalrows)
2288 320 : stats->stadistinct = -(stats->stadistinct / totalrows);
2289 :
2290 : /*
2291 : * Decide how many values are worth storing as most-common values. If
2292 : * we are able to generate a complete MCV list (all the values in the
2293 : * sample will fit, and we think these are all the ones in the table),
2294 : * then do so. Otherwise, store only those values that are
2295 : * significantly more common than the values not in the list.
2296 : *
2297 : * Note: the first of these cases is meant to address columns with
2298 : * small, fixed sets of possible values, such as boolean or enum
2299 : * columns. If we can *completely* represent the column population by
2300 : * an MCV list that will fit into the stats target, then we should do
2301 : * so and thus provide the planner with complete information. But if
2302 : * the MCV list is not complete, it's generally worth being more
2303 : * selective, and not just filling it all the way up to the stats
2304 : * target.
2305 : */
2306 1466 : if (track_cnt < track_max && toowide_cnt == 0 &&
2307 1458 : stats->stadistinct > 0 &&
2308 : track_cnt <= num_mcv)
2309 : {
2310 : /* Track list includes all values seen, and all will fit */
2311 942 : num_mcv = track_cnt;
2312 : }
2313 : else
2314 : {
2315 : int *mcv_counts;
2316 :
2317 : /* Incomplete list; decide how many values are worth keeping */
2318 524 : if (num_mcv > track_cnt)
2319 466 : num_mcv = track_cnt;
2320 :
2321 524 : if (num_mcv > 0)
2322 : {
2323 524 : mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2324 1396 : for (i = 0; i < num_mcv; i++)
2325 872 : mcv_counts[i] = track[i].count;
2326 :
2327 524 : num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2328 524 : stats->stadistinct,
2329 524 : stats->stanullfrac,
2330 : samplerows, totalrows);
2331 : }
2332 : }
2333 :
2334 : /* Generate MCV slot entry */
2335 1466 : if (num_mcv > 0)
2336 : {
2337 : MemoryContext old_context;
2338 : Datum *mcv_values;
2339 : float4 *mcv_freqs;
2340 :
2341 : /* Must copy the target values into anl_context */
2342 1458 : old_context = MemoryContextSwitchTo(stats->anl_context);
2343 1458 : mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2344 1458 : mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2345 6458 : for (i = 0; i < num_mcv; i++)
2346 : {
2347 10000 : mcv_values[i] = datumCopy(track[i].value,
2348 5000 : stats->attrtype->typbyval,
2349 5000 : stats->attrtype->typlen);
2350 5000 : mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2351 : }
2352 1458 : MemoryContextSwitchTo(old_context);
2353 :
2354 1458 : stats->stakind[0] = STATISTIC_KIND_MCV;
2355 1458 : stats->staop[0] = mystats->eqopr;
2356 1458 : stats->stacoll[0] = stats->attrcollid;
2357 1458 : stats->stanumbers[0] = mcv_freqs;
2358 1458 : stats->numnumbers[0] = num_mcv;
2359 1458 : stats->stavalues[0] = mcv_values;
2360 1458 : stats->numvalues[0] = num_mcv;
2361 :
2362 : /*
2363 : * Accept the defaults for stats->statypid and others. They have
2364 : * been set before we were called (see vacuum.h)
2365 : */
2366 : }
2367 : }
2368 538 : else if (null_cnt > 0)
2369 : {
2370 : /* We found only nulls; assume the column is entirely null */
2371 538 : stats->stats_valid = true;
2372 538 : stats->stanullfrac = 1.0;
2373 538 : if (is_varwidth)
2374 538 : stats->stawidth = 0; /* "unknown" */
2375 : else
2376 0 : stats->stawidth = stats->attrtype->typlen;
2377 538 : stats->stadistinct = 0.0; /* "unknown" */
2378 : }
2379 :
2380 : /* We don't need to bother cleaning up any of our temporary palloc's */
2381 2004 : }
2382 :
2383 :
2384 : /*
2385 : * compute_scalar_stats() -- compute column statistics
2386 : *
2387 : * We use this when we can find "=" and "<" operators for the datatype.
2388 : *
2389 : * We determine the fraction of non-null rows, the average width, the
2390 : * most common values, the (estimated) number of distinct values, the
2391 : * distribution histogram, and the correlation of physical to logical order.
2392 : *
2393 : * The desired stats can be determined fairly easily after sorting the
2394 : * data values into order.
2395 : */
2396 : static void
2397 73368 : compute_scalar_stats(VacAttrStatsP stats,
2398 : AnalyzeAttrFetchFunc fetchfunc,
2399 : int samplerows,
2400 : double totalrows)
2401 : {
2402 : int i;
2403 73368 : int null_cnt = 0;
2404 73368 : int nonnull_cnt = 0;
2405 73368 : int toowide_cnt = 0;
2406 73368 : double total_width = 0;
2407 91462 : bool is_varlena = (!stats->attrtype->typbyval &&
2408 18094 : stats->attrtype->typlen == -1);
2409 91462 : bool is_varwidth = (!stats->attrtype->typbyval &&
2410 18094 : stats->attrtype->typlen < 0);
2411 : double corr_xysum;
2412 : SortSupportData ssup;
2413 : ScalarItem *values;
2414 73368 : int values_cnt = 0;
2415 : int *tupnoLink;
2416 : ScalarMCVItem *track;
2417 73368 : int track_cnt = 0;
2418 73368 : int num_mcv = stats->attstattarget;
2419 73368 : int num_bins = stats->attstattarget;
2420 73368 : StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2421 :
2422 73368 : values = (ScalarItem *) palloc(samplerows * sizeof(ScalarItem));
2423 73368 : tupnoLink = (int *) palloc(samplerows * sizeof(int));
2424 73368 : track = (ScalarMCVItem *) palloc(num_mcv * sizeof(ScalarMCVItem));
2425 :
2426 73368 : memset(&ssup, 0, sizeof(ssup));
2427 73368 : ssup.ssup_cxt = CurrentMemoryContext;
2428 73368 : ssup.ssup_collation = stats->attrcollid;
2429 73368 : ssup.ssup_nulls_first = false;
2430 :
2431 : /*
2432 : * For now, don't perform abbreviated key conversion, because full values
2433 : * are required for MCV slot generation. Supporting that optimization
2434 : * would necessitate teaching compare_scalars() to call a tie-breaker.
2435 : */
2436 73368 : ssup.abbreviate = false;
2437 :
2438 73368 : PrepareSortSupportFromOrderingOp(mystats->ltopr, &ssup);
2439 :
2440 : /* Initial scan to find sortable values */
2441 65704012 : for (i = 0; i < samplerows; i++)
2442 : {
2443 : Datum value;
2444 : bool isnull;
2445 :
2446 65630644 : vacuum_delay_point(true);
2447 :
2448 65630644 : value = fetchfunc(stats, i, &isnull);
2449 :
2450 : /* Check for null/nonnull */
2451 65630644 : if (isnull)
2452 : {
2453 8660496 : null_cnt++;
2454 8692720 : continue;
2455 : }
2456 56970148 : nonnull_cnt++;
2457 :
2458 : /*
2459 : * If it's a variable-width field, add up widths for average width
2460 : * calculation. Note that if the value is toasted, we use the toasted
2461 : * width. We don't bother with this calculation if it's a fixed-width
2462 : * type.
2463 : */
2464 56970148 : if (is_varlena)
2465 : {
2466 6648848 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2467 :
2468 : /*
2469 : * If the value is toasted, we want to detoast it just once to
2470 : * avoid repeated detoastings and resultant excess memory usage
2471 : * during the comparisons. Also, check to see if the value is
2472 : * excessively wide, and if so don't detoast at all --- just
2473 : * ignore the value.
2474 : */
2475 6648848 : if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
2476 : {
2477 32224 : toowide_cnt++;
2478 32224 : continue;
2479 : }
2480 6616624 : value = PointerGetDatum(PG_DETOAST_DATUM(value));
2481 : }
2482 50321300 : else if (is_varwidth)
2483 : {
2484 : /* must be cstring */
2485 0 : total_width += strlen(DatumGetCString(value)) + 1;
2486 : }
2487 :
2488 : /* Add it to the list to be sorted */
2489 56937924 : values[values_cnt].value = value;
2490 56937924 : values[values_cnt].tupno = values_cnt;
2491 56937924 : tupnoLink[values_cnt] = values_cnt;
2492 56937924 : values_cnt++;
2493 : }
2494 :
2495 : /* We can only compute real stats if we found some sortable values. */
2496 73368 : if (values_cnt > 0)
2497 : {
2498 : int ndistinct, /* # distinct values in sample */
2499 : nmultiple, /* # that appear multiple times */
2500 : num_hist,
2501 : dups_cnt;
2502 68430 : int slot_idx = 0;
2503 : CompareScalarsContext cxt;
2504 :
2505 : /* Sort the collected values */
2506 68430 : cxt.ssup = &ssup;
2507 68430 : cxt.tupnoLink = tupnoLink;
2508 68430 : qsort_interruptible(values, values_cnt, sizeof(ScalarItem),
2509 : compare_scalars, &cxt);
2510 :
2511 : /*
2512 : * Now scan the values in order, find the most common ones, and also
2513 : * accumulate ordering-correlation statistics.
2514 : *
2515 : * To determine which are most common, we first have to count the
2516 : * number of duplicates of each value. The duplicates are adjacent in
2517 : * the sorted list, so a brute-force approach is to compare successive
2518 : * datum values until we find two that are not equal. However, that
2519 : * requires N-1 invocations of the datum comparison routine, which are
2520 : * completely redundant with work that was done during the sort. (The
2521 : * sort algorithm must at some point have compared each pair of items
2522 : * that are adjacent in the sorted order; otherwise it could not know
2523 : * that it's ordered the pair correctly.) We exploit this by having
2524 : * compare_scalars remember the highest tupno index that each
2525 : * ScalarItem has been found equal to. At the end of the sort, a
2526 : * ScalarItem's tupnoLink will still point to itself if and only if it
2527 : * is the last item of its group of duplicates (since the group will
2528 : * be ordered by tupno).
2529 : */
2530 68430 : corr_xysum = 0;
2531 68430 : ndistinct = 0;
2532 68430 : nmultiple = 0;
2533 68430 : dups_cnt = 0;
2534 57006354 : for (i = 0; i < values_cnt; i++)
2535 : {
2536 56937924 : int tupno = values[i].tupno;
2537 :
2538 56937924 : corr_xysum += ((double) i) * ((double) tupno);
2539 56937924 : dups_cnt++;
2540 56937924 : if (tupnoLink[tupno] == tupno)
2541 : {
2542 : /* Reached end of duplicates of this value */
2543 12324180 : ndistinct++;
2544 12324180 : if (dups_cnt > 1)
2545 : {
2546 1033914 : nmultiple++;
2547 1033914 : if (track_cnt < num_mcv ||
2548 436830 : dups_cnt > track[track_cnt - 1].count)
2549 : {
2550 : /*
2551 : * Found a new item for the mcv list; find its
2552 : * position, bubbling down old items if needed. Loop
2553 : * invariant is that j points at an empty/ replaceable
2554 : * slot.
2555 : */
2556 : int j;
2557 :
2558 684902 : if (track_cnt < num_mcv)
2559 597084 : track_cnt++;
2560 8984684 : for (j = track_cnt - 1; j > 0; j--)
2561 : {
2562 8908828 : if (dups_cnt <= track[j - 1].count)
2563 609046 : break;
2564 8299782 : track[j].count = track[j - 1].count;
2565 8299782 : track[j].first = track[j - 1].first;
2566 : }
2567 684902 : track[j].count = dups_cnt;
2568 684902 : track[j].first = i + 1 - dups_cnt;
2569 : }
2570 : }
2571 12324180 : dups_cnt = 0;
2572 : }
2573 : }
2574 :
2575 68430 : stats->stats_valid = true;
2576 : /* Do the simple null-frac and width stats */
2577 68430 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2578 68430 : if (is_varwidth)
2579 10060 : stats->stawidth = total_width / (double) nonnull_cnt;
2580 : else
2581 58370 : stats->stawidth = stats->attrtype->typlen;
2582 :
2583 68430 : if (nmultiple == 0)
2584 : {
2585 : /*
2586 : * If we found no repeated non-null values, assume it's a unique
2587 : * column; but be sure to discount for any nulls we found.
2588 : */
2589 18096 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2590 : }
2591 50334 : else if (toowide_cnt == 0 && nmultiple == ndistinct)
2592 : {
2593 : /*
2594 : * Every value in the sample appeared more than once. Assume the
2595 : * column has just these values. (This case is meant to address
2596 : * columns with small, fixed sets of possible values, such as
2597 : * boolean or enum columns. If there are any values that appear
2598 : * just once in the sample, including too-wide values, we should
2599 : * assume that that's not what we're dealing with.)
2600 : */
2601 30920 : stats->stadistinct = ndistinct;
2602 : }
2603 : else
2604 : {
2605 : /*----------
2606 : * Estimate the number of distinct values using the estimator
2607 : * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2608 : * n*d / (n - f1 + f1*n/N)
2609 : * where f1 is the number of distinct values that occurred
2610 : * exactly once in our sample of n rows (from a total of N),
2611 : * and d is the total number of distinct values in the sample.
2612 : * This is their Duj1 estimator; the other estimators they
2613 : * recommend are considerably more complex, and are numerically
2614 : * very unstable when n is much smaller than N.
2615 : *
2616 : * In this calculation, we consider only non-nulls. We used to
2617 : * include rows with null values in the n and N counts, but that
2618 : * leads to inaccurate answers in columns with many nulls, and
2619 : * it's intuitively bogus anyway considering the desired result is
2620 : * the number of distinct non-null values.
2621 : *
2622 : * Overwidth values are assumed to have been distinct.
2623 : *----------
2624 : */
2625 19414 : int f1 = ndistinct - nmultiple + toowide_cnt;
2626 19414 : int d = f1 + nmultiple;
2627 19414 : double n = samplerows - null_cnt;
2628 19414 : double N = totalrows * (1.0 - stats->stanullfrac);
2629 : double stadistinct;
2630 :
2631 : /* N == 0 shouldn't happen, but just in case ... */
2632 19414 : if (N > 0)
2633 19414 : stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2634 : else
2635 0 : stadistinct = 0;
2636 :
2637 : /* Clamp to sane range in case of roundoff error */
2638 19414 : if (stadistinct < d)
2639 964 : stadistinct = d;
2640 19414 : if (stadistinct > N)
2641 0 : stadistinct = N;
2642 : /* And round to integer */
2643 19414 : stats->stadistinct = floor(stadistinct + 0.5);
2644 : }
2645 :
2646 : /*
2647 : * If we estimated the number of distinct values at more than 10% of
2648 : * the total row count (a very arbitrary limit), then assume that
2649 : * stadistinct should scale with the row count rather than be a fixed
2650 : * value.
2651 : */
2652 68430 : if (stats->stadistinct > 0.1 * totalrows)
2653 14440 : stats->stadistinct = -(stats->stadistinct / totalrows);
2654 :
2655 : /*
2656 : * Decide how many values are worth storing as most-common values. If
2657 : * we are able to generate a complete MCV list (all the values in the
2658 : * sample will fit, and we think these are all the ones in the table),
2659 : * then do so. Otherwise, store only those values that are
2660 : * significantly more common than the values not in the list.
2661 : *
2662 : * Note: the first of these cases is meant to address columns with
2663 : * small, fixed sets of possible values, such as boolean or enum
2664 : * columns. If we can *completely* represent the column population by
2665 : * an MCV list that will fit into the stats target, then we should do
2666 : * so and thus provide the planner with complete information. But if
2667 : * the MCV list is not complete, it's generally worth being more
2668 : * selective, and not just filling it all the way up to the stats
2669 : * target.
2670 : */
2671 68430 : if (track_cnt == ndistinct && toowide_cnt == 0 &&
2672 30240 : stats->stadistinct > 0 &&
2673 : track_cnt <= num_mcv)
2674 : {
2675 : /* Track list includes all values seen, and all will fit */
2676 27022 : num_mcv = track_cnt;
2677 : }
2678 : else
2679 : {
2680 : int *mcv_counts;
2681 :
2682 : /* Incomplete list; decide how many values are worth keeping */
2683 41408 : if (num_mcv > track_cnt)
2684 37726 : num_mcv = track_cnt;
2685 :
2686 41408 : if (num_mcv > 0)
2687 : {
2688 23312 : mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2689 474006 : for (i = 0; i < num_mcv; i++)
2690 450694 : mcv_counts[i] = track[i].count;
2691 :
2692 23312 : num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2693 23312 : stats->stadistinct,
2694 23312 : stats->stanullfrac,
2695 : samplerows, totalrows);
2696 : }
2697 : }
2698 :
2699 : /* Generate MCV slot entry */
2700 68430 : if (num_mcv > 0)
2701 : {
2702 : MemoryContext old_context;
2703 : Datum *mcv_values;
2704 : float4 *mcv_freqs;
2705 :
2706 : /* Must copy the target values into anl_context */
2707 50284 : old_context = MemoryContextSwitchTo(stats->anl_context);
2708 50284 : mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2709 50284 : mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2710 647150 : for (i = 0; i < num_mcv; i++)
2711 : {
2712 1193732 : mcv_values[i] = datumCopy(values[track[i].first].value,
2713 596866 : stats->attrtype->typbyval,
2714 596866 : stats->attrtype->typlen);
2715 596866 : mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2716 : }
2717 50284 : MemoryContextSwitchTo(old_context);
2718 :
2719 50284 : stats->stakind[slot_idx] = STATISTIC_KIND_MCV;
2720 50284 : stats->staop[slot_idx] = mystats->eqopr;
2721 50284 : stats->stacoll[slot_idx] = stats->attrcollid;
2722 50284 : stats->stanumbers[slot_idx] = mcv_freqs;
2723 50284 : stats->numnumbers[slot_idx] = num_mcv;
2724 50284 : stats->stavalues[slot_idx] = mcv_values;
2725 50284 : stats->numvalues[slot_idx] = num_mcv;
2726 :
2727 : /*
2728 : * Accept the defaults for stats->statypid and others. They have
2729 : * been set before we were called (see vacuum.h)
2730 : */
2731 50284 : slot_idx++;
2732 : }
2733 :
2734 : /*
2735 : * Generate a histogram slot entry if there are at least two distinct
2736 : * values not accounted for in the MCV list. (This ensures the
2737 : * histogram won't collapse to empty or a singleton.)
2738 : */
2739 68430 : num_hist = ndistinct - num_mcv;
2740 68430 : if (num_hist > num_bins)
2741 11470 : num_hist = num_bins + 1;
2742 68430 : if (num_hist >= 2)
2743 : {
2744 : MemoryContext old_context;
2745 : Datum *hist_values;
2746 : int nvals;
2747 : int pos,
2748 : posfrac,
2749 : delta,
2750 : deltafrac;
2751 :
2752 : /* Sort the MCV items into position order to speed next loop */
2753 30306 : qsort_interruptible(track, num_mcv, sizeof(ScalarMCVItem),
2754 : compare_mcvs, NULL);
2755 :
2756 : /*
2757 : * Collapse out the MCV items from the values[] array.
2758 : *
2759 : * Note we destroy the values[] array here... but we don't need it
2760 : * for anything more. We do, however, still need values_cnt.
2761 : * nvals will be the number of remaining entries in values[].
2762 : */
2763 30306 : if (num_mcv > 0)
2764 : {
2765 : int src,
2766 : dest;
2767 : int j;
2768 :
2769 16230 : src = dest = 0;
2770 16230 : j = 0; /* index of next interesting MCV item */
2771 581208 : while (src < values_cnt)
2772 : {
2773 : int ncopy;
2774 :
2775 564978 : if (j < num_mcv)
2776 : {
2777 552214 : int first = track[j].first;
2778 :
2779 552214 : if (src >= first)
2780 : {
2781 : /* advance past this MCV item */
2782 403748 : src = first + track[j].count;
2783 403748 : j++;
2784 403748 : continue;
2785 : }
2786 148466 : ncopy = first - src;
2787 : }
2788 : else
2789 12764 : ncopy = values_cnt - src;
2790 161230 : memmove(&values[dest], &values[src],
2791 : ncopy * sizeof(ScalarItem));
2792 161230 : src += ncopy;
2793 161230 : dest += ncopy;
2794 : }
2795 16230 : nvals = dest;
2796 : }
2797 : else
2798 14076 : nvals = values_cnt;
2799 : Assert(nvals >= num_hist);
2800 :
2801 : /* Must copy the target values into anl_context */
2802 30306 : old_context = MemoryContextSwitchTo(stats->anl_context);
2803 30306 : hist_values = (Datum *) palloc(num_hist * sizeof(Datum));
2804 :
2805 : /*
2806 : * The object of this loop is to copy the first and last values[]
2807 : * entries along with evenly-spaced values in between. So the
2808 : * i'th value is values[(i * (nvals - 1)) / (num_hist - 1)]. But
2809 : * computing that subscript directly risks integer overflow when
2810 : * the stats target is more than a couple thousand. Instead we
2811 : * add (nvals - 1) / (num_hist - 1) to pos at each step, tracking
2812 : * the integral and fractional parts of the sum separately.
2813 : */
2814 30306 : delta = (nvals - 1) / (num_hist - 1);
2815 30306 : deltafrac = (nvals - 1) % (num_hist - 1);
2816 30306 : pos = posfrac = 0;
2817 :
2818 1615422 : for (i = 0; i < num_hist; i++)
2819 : {
2820 3170232 : hist_values[i] = datumCopy(values[pos].value,
2821 1585116 : stats->attrtype->typbyval,
2822 1585116 : stats->attrtype->typlen);
2823 1585116 : pos += delta;
2824 1585116 : posfrac += deltafrac;
2825 1585116 : if (posfrac >= (num_hist - 1))
2826 : {
2827 : /* fractional part exceeds 1, carry to integer part */
2828 546578 : pos++;
2829 546578 : posfrac -= (num_hist - 1);
2830 : }
2831 : }
2832 :
2833 30306 : MemoryContextSwitchTo(old_context);
2834 :
2835 30306 : stats->stakind[slot_idx] = STATISTIC_KIND_HISTOGRAM;
2836 30306 : stats->staop[slot_idx] = mystats->ltopr;
2837 30306 : stats->stacoll[slot_idx] = stats->attrcollid;
2838 30306 : stats->stavalues[slot_idx] = hist_values;
2839 30306 : stats->numvalues[slot_idx] = num_hist;
2840 :
2841 : /*
2842 : * Accept the defaults for stats->statypid and others. They have
2843 : * been set before we were called (see vacuum.h)
2844 : */
2845 30306 : slot_idx++;
2846 : }
2847 :
2848 : /* Generate a correlation entry if there are multiple values */
2849 68430 : if (values_cnt > 1)
2850 : {
2851 : MemoryContext old_context;
2852 : float4 *corrs;
2853 : double corr_xsum,
2854 : corr_x2sum;
2855 :
2856 : /* Must copy the target values into anl_context */
2857 64360 : old_context = MemoryContextSwitchTo(stats->anl_context);
2858 64360 : corrs = (float4 *) palloc(sizeof(float4));
2859 64360 : MemoryContextSwitchTo(old_context);
2860 :
2861 : /*----------
2862 : * Since we know the x and y value sets are both
2863 : * 0, 1, ..., values_cnt-1
2864 : * we have sum(x) = sum(y) =
2865 : * (values_cnt-1)*values_cnt / 2
2866 : * and sum(x^2) = sum(y^2) =
2867 : * (values_cnt-1)*values_cnt*(2*values_cnt-1) / 6.
2868 : *----------
2869 : */
2870 64360 : corr_xsum = ((double) (values_cnt - 1)) *
2871 64360 : ((double) values_cnt) / 2.0;
2872 64360 : corr_x2sum = ((double) (values_cnt - 1)) *
2873 64360 : ((double) values_cnt) * (double) (2 * values_cnt - 1) / 6.0;
2874 :
2875 : /* And the correlation coefficient reduces to */
2876 64360 : corrs[0] = (values_cnt * corr_xysum - corr_xsum * corr_xsum) /
2877 64360 : (values_cnt * corr_x2sum - corr_xsum * corr_xsum);
2878 :
2879 64360 : stats->stakind[slot_idx] = STATISTIC_KIND_CORRELATION;
2880 64360 : stats->staop[slot_idx] = mystats->ltopr;
2881 64360 : stats->stacoll[slot_idx] = stats->attrcollid;
2882 64360 : stats->stanumbers[slot_idx] = corrs;
2883 64360 : stats->numnumbers[slot_idx] = 1;
2884 64360 : slot_idx++;
2885 : }
2886 : }
2887 4938 : else if (nonnull_cnt > 0)
2888 : {
2889 : /* We found some non-null values, but they were all too wide */
2890 : Assert(nonnull_cnt == toowide_cnt);
2891 308 : stats->stats_valid = true;
2892 : /* Do the simple null-frac and width stats */
2893 308 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2894 308 : if (is_varwidth)
2895 308 : stats->stawidth = total_width / (double) nonnull_cnt;
2896 : else
2897 0 : stats->stawidth = stats->attrtype->typlen;
2898 : /* Assume all too-wide values are distinct, so it's a unique column */
2899 308 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2900 : }
2901 4630 : else if (null_cnt > 0)
2902 : {
2903 : /* We found only nulls; assume the column is entirely null */
2904 4630 : stats->stats_valid = true;
2905 4630 : stats->stanullfrac = 1.0;
2906 4630 : if (is_varwidth)
2907 4024 : stats->stawidth = 0; /* "unknown" */
2908 : else
2909 606 : stats->stawidth = stats->attrtype->typlen;
2910 4630 : stats->stadistinct = 0.0; /* "unknown" */
2911 : }
2912 :
2913 : /* We don't need to bother cleaning up any of our temporary palloc's */
2914 73368 : }
2915 :
2916 : /*
2917 : * Comparator for sorting ScalarItems
2918 : *
2919 : * Aside from sorting the items, we update the tupnoLink[] array
2920 : * whenever two ScalarItems are found to contain equal datums. The array
2921 : * is indexed by tupno; for each ScalarItem, it contains the highest
2922 : * tupno that that item's datum has been found to be equal to. This allows
2923 : * us to avoid additional comparisons in compute_scalar_stats().
2924 : */
2925 : static int
2926 526763846 : compare_scalars(const void *a, const void *b, void *arg)
2927 : {
2928 526763846 : Datum da = ((const ScalarItem *) a)->value;
2929 526763846 : int ta = ((const ScalarItem *) a)->tupno;
2930 526763846 : Datum db = ((const ScalarItem *) b)->value;
2931 526763846 : int tb = ((const ScalarItem *) b)->tupno;
2932 526763846 : CompareScalarsContext *cxt = (CompareScalarsContext *) arg;
2933 : int compare;
2934 :
2935 526763846 : compare = ApplySortComparator(da, false, db, false, cxt->ssup);
2936 526763846 : if (compare != 0)
2937 205555314 : return compare;
2938 :
2939 : /*
2940 : * The two datums are equal, so update cxt->tupnoLink[].
2941 : */
2942 321208532 : if (cxt->tupnoLink[ta] < tb)
2943 46459806 : cxt->tupnoLink[ta] = tb;
2944 321208532 : if (cxt->tupnoLink[tb] < ta)
2945 3369258 : cxt->tupnoLink[tb] = ta;
2946 :
2947 : /*
2948 : * For equal datums, sort by tupno
2949 : */
2950 321208532 : return ta - tb;
2951 : }
2952 :
2953 : /*
2954 : * Comparator for sorting ScalarMCVItems by position
2955 : */
2956 : static int
2957 2068382 : compare_mcvs(const void *a, const void *b, void *arg)
2958 : {
2959 2068382 : int da = ((const ScalarMCVItem *) a)->first;
2960 2068382 : int db = ((const ScalarMCVItem *) b)->first;
2961 :
2962 2068382 : return da - db;
2963 : }
2964 :
2965 : /*
2966 : * Analyze the list of common values in the sample and decide how many are
2967 : * worth storing in the table's MCV list.
2968 : *
2969 : * mcv_counts is assumed to be a list of the counts of the most common values
2970 : * seen in the sample, starting with the most common. The return value is the
2971 : * number that are significantly more common than the values not in the list,
2972 : * and which are therefore deemed worth storing in the table's MCV list.
2973 : */
2974 : static int
2975 23836 : analyze_mcv_list(int *mcv_counts,
2976 : int num_mcv,
2977 : double stadistinct,
2978 : double stanullfrac,
2979 : int samplerows,
2980 : double totalrows)
2981 : {
2982 : double ndistinct_table;
2983 : double sumcount;
2984 : int i;
2985 :
2986 : /*
2987 : * If the entire table was sampled, keep the whole list. This also
2988 : * protects us against division by zero in the code below.
2989 : */
2990 23836 : if (samplerows == totalrows || totalrows <= 1.0)
2991 22996 : return num_mcv;
2992 :
2993 : /* Re-extract the estimated number of distinct nonnull values in table */
2994 840 : ndistinct_table = stadistinct;
2995 840 : if (ndistinct_table < 0)
2996 160 : ndistinct_table = -ndistinct_table * totalrows;
2997 :
2998 : /*
2999 : * Exclude the least common values from the MCV list, if they are not
3000 : * significantly more common than the estimated selectivity they would
3001 : * have if they weren't in the list. All non-MCV values are assumed to be
3002 : * equally common, after taking into account the frequencies of all the
3003 : * values in the MCV list and the number of nulls (c.f. eqsel()).
3004 : *
3005 : * Here sumcount tracks the total count of all but the last (least common)
3006 : * value in the MCV list, allowing us to determine the effect of excluding
3007 : * that value from the list.
3008 : *
3009 : * Note that we deliberately do this by removing values from the full
3010 : * list, rather than starting with an empty list and adding values,
3011 : * because the latter approach can fail to add any values if all the most
3012 : * common values have around the same frequency and make up the majority
3013 : * of the table, so that the overall average frequency of all values is
3014 : * roughly the same as that of the common values. This would lead to any
3015 : * uncommon values being significantly overestimated.
3016 : */
3017 840 : sumcount = 0.0;
3018 1786 : for (i = 0; i < num_mcv - 1; i++)
3019 946 : sumcount += mcv_counts[i];
3020 :
3021 1008 : while (num_mcv > 0)
3022 : {
3023 : double selec,
3024 : otherdistinct,
3025 : N,
3026 : n,
3027 : K,
3028 : variance,
3029 : stddev;
3030 :
3031 : /*
3032 : * Estimated selectivity the least common value would have if it
3033 : * wasn't in the MCV list (c.f. eqsel()).
3034 : */
3035 1008 : selec = 1.0 - sumcount / samplerows - stanullfrac;
3036 1008 : if (selec < 0.0)
3037 0 : selec = 0.0;
3038 1008 : if (selec > 1.0)
3039 0 : selec = 1.0;
3040 1008 : otherdistinct = ndistinct_table - (num_mcv - 1);
3041 1008 : if (otherdistinct > 1)
3042 1008 : selec /= otherdistinct;
3043 :
3044 : /*
3045 : * If the value is kept in the MCV list, its population frequency is
3046 : * assumed to equal its sample frequency. We use the lower end of a
3047 : * textbook continuity-corrected Wald-type confidence interval to
3048 : * determine if that is significantly more common than the non-MCV
3049 : * frequency --- specifically we assume the population frequency is
3050 : * highly likely to be within around 2 standard errors of the sample
3051 : * frequency, which equates to an interval of 2 standard deviations
3052 : * either side of the sample count, plus an additional 0.5 for the
3053 : * continuity correction. Since we are sampling without replacement,
3054 : * this is a hypergeometric distribution.
3055 : *
3056 : * XXX: Empirically, this approach seems to work quite well, but it
3057 : * may be worth considering more advanced techniques for estimating
3058 : * the confidence interval of the hypergeometric distribution.
3059 : */
3060 1008 : N = totalrows;
3061 1008 : n = samplerows;
3062 1008 : K = N * mcv_counts[num_mcv - 1] / n;
3063 1008 : variance = n * K * (N - K) * (N - n) / (N * N * (N - 1));
3064 1008 : stddev = sqrt(variance);
3065 :
3066 1008 : if (mcv_counts[num_mcv - 1] > selec * samplerows + 2 * stddev + 0.5)
3067 : {
3068 : /*
3069 : * The value is significantly more common than the non-MCV
3070 : * selectivity would suggest. Keep it, and all the other more
3071 : * common values in the list.
3072 : */
3073 782 : break;
3074 : }
3075 : else
3076 : {
3077 : /* Discard this value and consider the next least common value */
3078 226 : num_mcv--;
3079 226 : if (num_mcv == 0)
3080 58 : break;
3081 168 : sumcount -= mcv_counts[num_mcv - 1];
3082 : }
3083 : }
3084 840 : return num_mcv;
3085 : }
|
