Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * analyze.c
4 : * the Postgres statistics generator
5 : *
6 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/commands/analyze.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 : #include "postgres.h"
16 :
17 : #include <math.h>
18 :
19 : #include "access/detoast.h"
20 : #include "access/genam.h"
21 : #include "access/multixact.h"
22 : #include "access/relation.h"
23 : #include "access/table.h"
24 : #include "access/tableam.h"
25 : #include "access/transam.h"
26 : #include "access/tupconvert.h"
27 : #include "access/visibilitymap.h"
28 : #include "access/xact.h"
29 : #include "catalog/index.h"
30 : #include "catalog/indexing.h"
31 : #include "catalog/pg_inherits.h"
32 : #include "commands/progress.h"
33 : #include "commands/tablecmds.h"
34 : #include "commands/vacuum.h"
35 : #include "common/pg_prng.h"
36 : #include "executor/executor.h"
37 : #include "executor/instrument.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 : const 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 void validate_va_cols_list(Relation onerel, List *va_cols);
87 : static VacAttrStats *examine_attribute(Relation onerel, int attnum,
88 : Node *index_expr);
89 : static int acquire_sample_rows(Relation onerel, int elevel,
90 : HeapTuple *rows, int targrows,
91 : double *totalrows, double *totaldeadrows);
92 : static int compare_rows(const void *a, const void *b, void *arg);
93 : static int acquire_inherited_sample_rows(Relation onerel, int elevel,
94 : HeapTuple *rows, int targrows,
95 : double *totalrows, double *totaldeadrows);
96 : static void update_attstats(Oid relid, bool inh,
97 : int natts, VacAttrStats **vacattrstats);
98 : static Datum std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
99 : static Datum ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
100 :
101 :
102 : /*
103 : * analyze_rel() -- analyze one relation
104 : *
105 : * relid identifies the relation to analyze. If relation is supplied, use
106 : * the name therein for reporting any failure to open/lock the rel; do not
107 : * use it once we've successfully opened the rel, since it might be stale.
108 : */
109 : void
110 10295 : analyze_rel(Oid relid, RangeVar *relation,
111 : const VacuumParams *params, List *va_cols, bool in_outer_xact,
112 : BufferAccessStrategy bstrategy)
113 : {
114 : Relation onerel;
115 : int elevel;
116 10295 : AcquireSampleRowsFunc acquirefunc = NULL;
117 10295 : BlockNumber relpages = 0;
118 10295 : bool stats_imported = false;
119 :
120 : /* Select logging level */
121 10295 : if (params->options & VACOPT_VERBOSE)
122 6 : elevel = INFO;
123 : else
124 10289 : elevel = DEBUG2;
125 :
126 : /* Set up static variables */
127 10295 : vac_strategy = bstrategy;
128 :
129 : /*
130 : * Check for user-requested abort.
131 : */
132 10295 : CHECK_FOR_INTERRUPTS();
133 :
134 : /*
135 : * Open the relation, getting ShareUpdateExclusiveLock to ensure that two
136 : * ANALYZEs don't run on it concurrently. (This also locks out a
137 : * concurrent VACUUM, which doesn't matter much at the moment but might
138 : * matter if we ever try to accumulate stats on dead tuples.) If the rel
139 : * has been dropped since we last saw it, we don't need to process it.
140 : *
141 : * Make sure to generate only logs for ANALYZE in this case.
142 : */
143 10295 : onerel = vacuum_open_relation(relid, relation, params->options & ~(VACOPT_VACUUM),
144 10295 : params->log_analyze_min_duration >= 0,
145 : ShareUpdateExclusiveLock);
146 :
147 : /* leave if relation could not be opened or locked */
148 10295 : if (!onerel)
149 124 : return;
150 :
151 : /*
152 : * Check if relation needs to be skipped based on privileges. This check
153 : * happens also when building the relation list to analyze for a manual
154 : * operation, and needs to be done additionally here as ANALYZE could
155 : * happen across multiple transactions where privileges could have changed
156 : * in-between. Make sure to generate only logs for ANALYZE in this case.
157 : */
158 10289 : if (!vacuum_is_permitted_for_relation(RelationGetRelid(onerel),
159 : onerel->rd_rel,
160 10289 : params->options & ~VACOPT_VACUUM))
161 : {
162 24 : relation_close(onerel, ShareUpdateExclusiveLock);
163 24 : return;
164 : }
165 :
166 : /*
167 : * Silently ignore tables that are temp tables of other backends ---
168 : * trying to analyze these is rather pointless, since their contents are
169 : * probably not up-to-date on disk. (We don't throw a warning here; it
170 : * would just lead to chatter during a database-wide ANALYZE.)
171 : */
172 10265 : if (RELATION_IS_OTHER_TEMP(onerel))
173 : {
174 0 : relation_close(onerel, ShareUpdateExclusiveLock);
175 0 : return;
176 : }
177 :
178 : /*
179 : * We can ANALYZE any table except pg_statistic. See update_attstats
180 : */
181 10265 : if (RelationGetRelid(onerel) == StatisticRelationId)
182 : {
183 94 : relation_close(onerel, ShareUpdateExclusiveLock);
184 94 : return;
185 : }
186 :
187 : /*
188 : * Check the given list of columns
189 : */
190 10171 : if (va_cols != NIL)
191 75 : validate_va_cols_list(onerel, va_cols);
192 :
193 : /*
194 : * Initialize progress reporting before setup for regular/foreign tables.
195 : * (For the former, the time spent on it would be negligible, but for the
196 : * latter, if FDWs support statistics import or analysis, they'd do some
197 : * work that needs the remote access, so the time might be
198 : * non-negligible.)
199 : */
200 10136 : pgstat_progress_start_command(PROGRESS_COMMAND_ANALYZE,
201 : RelationGetRelid(onerel));
202 10136 : if (AmAutoVacuumWorkerProcess())
203 321 : pgstat_progress_update_param(PROGRESS_ANALYZE_STARTED_BY,
204 : PROGRESS_ANALYZE_STARTED_BY_AUTOVACUUM);
205 : else
206 9815 : pgstat_progress_update_param(PROGRESS_ANALYZE_STARTED_BY,
207 : PROGRESS_ANALYZE_STARTED_BY_MANUAL);
208 :
209 : /*
210 : * Check that it's of an analyzable relkind, and set up appropriately.
211 : */
212 10136 : if (onerel->rd_rel->relkind == RELKIND_RELATION ||
213 544 : onerel->rd_rel->relkind == RELKIND_MATVIEW)
214 : {
215 : /* Regular table, so we'll use the regular row acquisition function */
216 9594 : acquirefunc = acquire_sample_rows;
217 : /* Also get regular table's size */
218 9594 : relpages = RelationGetNumberOfBlocks(onerel);
219 : }
220 542 : else if (onerel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
221 : {
222 : /*
223 : * For a foreign table, call the FDW's hook functions to see whether
224 : * it supports statistics import or analysis.
225 : */
226 : FdwRoutine *fdwroutine;
227 :
228 44 : fdwroutine = GetFdwRoutineForRelation(onerel, false);
229 :
230 87 : if (fdwroutine->ImportForeignStatistics != NULL &&
231 43 : fdwroutine->ImportForeignStatistics(onerel, va_cols, elevel))
232 6 : stats_imported = true;
233 : else
234 : {
235 38 : bool ok = false;
236 :
237 38 : if (fdwroutine->AnalyzeForeignTable != NULL)
238 38 : ok = fdwroutine->AnalyzeForeignTable(onerel,
239 : &acquirefunc,
240 : &relpages);
241 :
242 38 : if (!ok)
243 : {
244 0 : ereport(WARNING,
245 : errmsg("skipping \"%s\" -- cannot analyze this foreign table.",
246 : RelationGetRelationName(onerel)));
247 0 : relation_close(onerel, ShareUpdateExclusiveLock);
248 0 : goto out;
249 : }
250 : }
251 : }
252 498 : else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
253 : {
254 : /*
255 : * For partitioned tables, we want to do the recursive ANALYZE below.
256 : */
257 : }
258 : else
259 : {
260 : /* No need for a WARNING if we already complained during VACUUM */
261 0 : if (!(params->options & VACOPT_VACUUM))
262 0 : ereport(WARNING,
263 : (errmsg("skipping \"%s\" --- cannot analyze non-tables or special system tables",
264 : RelationGetRelationName(onerel))));
265 0 : relation_close(onerel, ShareUpdateExclusiveLock);
266 0 : goto out;
267 : }
268 :
269 : /*
270 : * Do the normal non-recursive ANALYZE. We can skip this for partitioned
271 : * tables, which don't contain any rows, and foreign tables that
272 : * successfully imported statistics.
273 : */
274 10136 : if ((onerel->rd_rel->relkind != RELKIND_PARTITIONED_TABLE)
275 9638 : && !stats_imported)
276 9632 : do_analyze_rel(onerel, params, va_cols, acquirefunc,
277 : relpages, false, in_outer_xact, elevel);
278 :
279 : /*
280 : * If there are child tables, do recursive ANALYZE.
281 : */
282 10131 : if (onerel->rd_rel->relhassubclass)
283 589 : do_analyze_rel(onerel, params, va_cols, acquirefunc, relpages,
284 : true, in_outer_xact, elevel);
285 :
286 : /*
287 : * Close source relation now, but keep lock so that no one deletes it
288 : * before we commit. (If someone did, they'd fail to clean up the entries
289 : * we made in pg_statistic. Also, releasing the lock before commit would
290 : * expose us to concurrent-update failures in update_attstats.)
291 : */
292 10131 : relation_close(onerel, NoLock);
293 :
294 10131 : out:
295 10131 : pgstat_progress_end_command();
296 : }
297 :
298 : /*
299 : * do_analyze_rel() -- analyze one relation, recursively or not
300 : *
301 : * Note that "acquirefunc" is only relevant for the non-inherited case.
302 : * For the inherited case, acquire_inherited_sample_rows() determines the
303 : * appropriate acquirefunc for each child table.
304 : */
305 : static void
306 10221 : do_analyze_rel(Relation onerel, const VacuumParams *params,
307 : List *va_cols, AcquireSampleRowsFunc acquirefunc,
308 : BlockNumber relpages, bool inh, bool in_outer_xact,
309 : int elevel)
310 : {
311 : int attr_cnt,
312 : tcnt,
313 : i,
314 : ind;
315 : Relation *Irel;
316 : int nindexes;
317 : bool verbose,
318 : instrument,
319 : hasindex;
320 : VacAttrStats **vacattrstats;
321 : AnlIndexData *indexdata;
322 : int targrows,
323 : numrows,
324 : minrows;
325 : double totalrows,
326 : totaldeadrows;
327 : HeapTuple *rows;
328 : PGRUsage ru0;
329 10221 : TimestampTz starttime = 0;
330 : MemoryContext caller_context;
331 : Oid save_userid;
332 : int save_sec_context;
333 : int save_nestlevel;
334 10221 : WalUsage startwalusage = pgWalUsage;
335 10221 : BufferUsage startbufferusage = pgBufferUsage;
336 : BufferUsage bufferusage;
337 10221 : PgStat_Counter startreadtime = 0;
338 10221 : PgStat_Counter startwritetime = 0;
339 :
340 10221 : verbose = (params->options & VACOPT_VERBOSE) != 0;
341 10543 : instrument = (verbose || (AmAutoVacuumWorkerProcess() &&
342 322 : params->log_analyze_min_duration >= 0));
343 10221 : if (inh)
344 589 : ereport(elevel,
345 : (errmsg("analyzing \"%s.%s\" inheritance tree",
346 : get_namespace_name(RelationGetNamespace(onerel)),
347 : RelationGetRelationName(onerel))));
348 : else
349 9632 : ereport(elevel,
350 : (errmsg("analyzing \"%s.%s\"",
351 : get_namespace_name(RelationGetNamespace(onerel)),
352 : RelationGetRelationName(onerel))));
353 :
354 : /*
355 : * Set up a working context so that we can easily free whatever junk gets
356 : * created.
357 : */
358 10221 : anl_context = AllocSetContextCreate(CurrentMemoryContext,
359 : "Analyze",
360 : ALLOCSET_DEFAULT_SIZES);
361 10221 : caller_context = MemoryContextSwitchTo(anl_context);
362 :
363 : /*
364 : * Switch to the table owner's userid, so that any index functions are run
365 : * as that user. Also lock down security-restricted operations and
366 : * arrange to make GUC variable changes local to this command.
367 : */
368 10221 : GetUserIdAndSecContext(&save_userid, &save_sec_context);
369 10221 : SetUserIdAndSecContext(onerel->rd_rel->relowner,
370 : save_sec_context | SECURITY_RESTRICTED_OPERATION);
371 10221 : save_nestlevel = NewGUCNestLevel();
372 10221 : RestrictSearchPath();
373 :
374 : /*
375 : * When verbose or autovacuum logging is used, initialize a resource usage
376 : * snapshot and optionally track I/O timing.
377 : */
378 10221 : if (instrument)
379 : {
380 322 : if (track_io_timing)
381 : {
382 0 : startreadtime = pgStatBlockReadTime;
383 0 : startwritetime = pgStatBlockWriteTime;
384 : }
385 :
386 322 : pg_rusage_init(&ru0);
387 : }
388 :
389 : /* Used for instrumentation and stats report */
390 10221 : starttime = GetCurrentTimestamp();
391 :
392 : /*
393 : * Determine which columns to analyze.
394 : */
395 10221 : if (va_cols != NIL)
396 : {
397 : ListCell *le;
398 :
399 40 : vacattrstats = (VacAttrStats **) palloc(list_length(va_cols) *
400 : sizeof(VacAttrStats *));
401 40 : tcnt = 0;
402 85 : foreach(le, va_cols)
403 : {
404 45 : char *col = strVal(lfirst(le));
405 :
406 45 : i = attnameAttNum(onerel, col, false);
407 : Assert(i != InvalidAttrNumber);
408 45 : vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
409 45 : if (vacattrstats[tcnt] != NULL)
410 45 : tcnt++;
411 : }
412 40 : attr_cnt = tcnt;
413 : }
414 : else
415 : {
416 10181 : attr_cnt = onerel->rd_att->natts;
417 : vacattrstats = (VacAttrStats **)
418 10181 : palloc(attr_cnt * sizeof(VacAttrStats *));
419 10181 : tcnt = 0;
420 81808 : for (i = 1; i <= attr_cnt; i++)
421 : {
422 71627 : vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
423 71627 : if (vacattrstats[tcnt] != NULL)
424 71548 : tcnt++;
425 : }
426 10181 : attr_cnt = tcnt;
427 : }
428 :
429 : /*
430 : * Open all indexes of the relation, and see if there are any analyzable
431 : * columns in the indexes. We do not analyze index columns if there was
432 : * an explicit column list in the ANALYZE command, however.
433 : *
434 : * If we are doing a recursive scan, we don't want to touch the parent's
435 : * indexes at all. If we're processing a partitioned table, we need to
436 : * know if there are any indexes, but we don't want to process them.
437 : */
438 10221 : if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
439 : {
440 498 : List *idxs = RelationGetIndexList(onerel);
441 :
442 498 : Irel = NULL;
443 498 : nindexes = 0;
444 498 : hasindex = idxs != NIL;
445 498 : list_free(idxs);
446 : }
447 9723 : else if (!inh)
448 : {
449 9632 : vac_open_indexes(onerel, AccessShareLock, &nindexes, &Irel);
450 9632 : hasindex = nindexes > 0;
451 : }
452 : else
453 : {
454 91 : Irel = NULL;
455 91 : nindexes = 0;
456 91 : hasindex = false;
457 : }
458 10221 : indexdata = NULL;
459 10221 : if (nindexes > 0)
460 : {
461 7398 : indexdata = (AnlIndexData *) palloc0(nindexes * sizeof(AnlIndexData));
462 21419 : for (ind = 0; ind < nindexes; ind++)
463 : {
464 14021 : AnlIndexData *thisdata = &indexdata[ind];
465 : IndexInfo *indexInfo;
466 :
467 14021 : thisdata->indexInfo = indexInfo = BuildIndexInfo(Irel[ind]);
468 14021 : thisdata->tupleFract = 1.0; /* fix later if partial */
469 14021 : if (indexInfo->ii_Expressions != NIL && va_cols == NIL)
470 : {
471 77 : ListCell *indexpr_item = list_head(indexInfo->ii_Expressions);
472 :
473 77 : thisdata->vacattrstats = (VacAttrStats **)
474 77 : palloc(indexInfo->ii_NumIndexAttrs * sizeof(VacAttrStats *));
475 77 : tcnt = 0;
476 157 : for (i = 0; i < indexInfo->ii_NumIndexAttrs; i++)
477 : {
478 80 : int keycol = indexInfo->ii_IndexAttrNumbers[i];
479 :
480 80 : if (keycol == 0)
481 : {
482 : /* Found an index expression */
483 : Node *indexkey;
484 :
485 77 : if (indexpr_item == NULL) /* shouldn't happen */
486 0 : elog(ERROR, "too few entries in indexprs list");
487 77 : indexkey = (Node *) lfirst(indexpr_item);
488 77 : indexpr_item = lnext(indexInfo->ii_Expressions,
489 : indexpr_item);
490 154 : thisdata->vacattrstats[tcnt] =
491 77 : examine_attribute(Irel[ind], i + 1, indexkey);
492 77 : if (thisdata->vacattrstats[tcnt] != NULL)
493 77 : tcnt++;
494 : }
495 : }
496 77 : thisdata->attr_cnt = tcnt;
497 : }
498 : }
499 : }
500 :
501 : /*
502 : * Determine how many rows we need to sample, using the worst case from
503 : * all analyzable columns. We use a lower bound of 100 rows to avoid
504 : * possible overflow in Vitter's algorithm. (Note: that will also be the
505 : * target in the corner case where there are no analyzable columns.)
506 : */
507 10221 : targrows = 100;
508 81814 : for (i = 0; i < attr_cnt; i++)
509 : {
510 71593 : if (targrows < vacattrstats[i]->minrows)
511 10172 : targrows = vacattrstats[i]->minrows;
512 : }
513 24242 : for (ind = 0; ind < nindexes; ind++)
514 : {
515 14021 : AnlIndexData *thisdata = &indexdata[ind];
516 :
517 14098 : for (i = 0; i < thisdata->attr_cnt; i++)
518 : {
519 77 : if (targrows < thisdata->vacattrstats[i]->minrows)
520 8 : targrows = thisdata->vacattrstats[i]->minrows;
521 : }
522 : }
523 :
524 : /*
525 : * Look at extended statistics objects too, as those may define custom
526 : * statistics target. So we may need to sample more rows and then build
527 : * the statistics with enough detail.
528 : */
529 10221 : minrows = ComputeExtStatisticsRows(onerel, attr_cnt, vacattrstats);
530 :
531 10221 : if (targrows < minrows)
532 0 : targrows = minrows;
533 :
534 : /*
535 : * Acquire the sample rows
536 : */
537 10221 : rows = (HeapTuple *) palloc(targrows * sizeof(HeapTuple));
538 10221 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
539 : inh ? PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS_INH :
540 : PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS);
541 10221 : if (inh)
542 589 : numrows = acquire_inherited_sample_rows(onerel, elevel,
543 : rows, targrows,
544 : &totalrows, &totaldeadrows);
545 : else
546 9632 : numrows = (*acquirefunc) (onerel, elevel,
547 : rows, targrows,
548 : &totalrows, &totaldeadrows);
549 :
550 : /*
551 : * Compute the statistics. Temporary results during the calculations for
552 : * each column are stored in a child context. The calc routines are
553 : * responsible to make sure that whatever they store into the VacAttrStats
554 : * structure is allocated in anl_context.
555 : */
556 10220 : if (numrows > 0)
557 : {
558 : MemoryContext col_context,
559 : old_context;
560 :
561 6774 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
562 : PROGRESS_ANALYZE_PHASE_COMPUTE_STATS);
563 :
564 6774 : col_context = AllocSetContextCreate(anl_context,
565 : "Analyze Column",
566 : ALLOCSET_DEFAULT_SIZES);
567 6774 : old_context = MemoryContextSwitchTo(col_context);
568 :
569 57455 : for (i = 0; i < attr_cnt; i++)
570 : {
571 50681 : VacAttrStats *stats = vacattrstats[i];
572 : AttributeOpts *aopt;
573 :
574 50681 : stats->rows = rows;
575 50681 : stats->tupDesc = onerel->rd_att;
576 50681 : stats->compute_stats(stats,
577 : std_fetch_func,
578 : numrows,
579 : totalrows);
580 :
581 : /*
582 : * If the appropriate flavor of the n_distinct option is
583 : * specified, override with the corresponding value.
584 : */
585 50681 : aopt = get_attribute_options(onerel->rd_id, stats->tupattnum);
586 50681 : if (aopt != NULL)
587 : {
588 : float8 n_distinct;
589 :
590 4 : n_distinct = inh ? aopt->n_distinct_inherited : aopt->n_distinct;
591 4 : if (n_distinct != 0.0)
592 4 : stats->stadistinct = n_distinct;
593 : }
594 :
595 50681 : MemoryContextReset(col_context);
596 : }
597 :
598 6774 : if (nindexes > 0)
599 4142 : compute_index_stats(onerel, totalrows,
600 : indexdata, nindexes,
601 : rows, numrows,
602 : col_context);
603 :
604 6770 : MemoryContextSwitchTo(old_context);
605 6770 : MemoryContextDelete(col_context);
606 :
607 : /*
608 : * Emit the completed stats rows into pg_statistic, replacing any
609 : * previous statistics for the target columns. (If there are stats in
610 : * pg_statistic for columns we didn't process, we leave them alone.)
611 : */
612 6770 : update_attstats(RelationGetRelid(onerel), inh,
613 : attr_cnt, vacattrstats);
614 :
615 14788 : for (ind = 0; ind < nindexes; ind++)
616 : {
617 8018 : AnlIndexData *thisdata = &indexdata[ind];
618 :
619 8018 : update_attstats(RelationGetRelid(Irel[ind]), false,
620 : thisdata->attr_cnt, thisdata->vacattrstats);
621 : }
622 :
623 : /* Build extended statistics (if there are any). */
624 6770 : BuildRelationExtStatistics(onerel, inh, totalrows, numrows, rows,
625 : attr_cnt, vacattrstats);
626 : }
627 :
628 10216 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
629 : PROGRESS_ANALYZE_PHASE_FINALIZE_ANALYZE);
630 :
631 : /*
632 : * Update pages/tuples stats in pg_class ... but not if we're doing
633 : * inherited stats.
634 : *
635 : * We assume that VACUUM hasn't set pg_class.reltuples already, even
636 : * during a VACUUM ANALYZE. Although VACUUM often updates pg_class,
637 : * exceptions exist. A "VACUUM (ANALYZE, INDEX_CLEANUP OFF)" command will
638 : * never update pg_class entries for index relations. It's also possible
639 : * that an individual index's pg_class entry won't be updated during
640 : * VACUUM if the index AM returns NULL from its amvacuumcleanup() routine.
641 : */
642 10216 : if (!inh)
643 : {
644 9627 : BlockNumber relallvisible = 0;
645 9627 : BlockNumber relallfrozen = 0;
646 :
647 9627 : if (RELKIND_HAS_STORAGE(onerel->rd_rel->relkind))
648 9590 : visibilitymap_count(onerel, &relallvisible, &relallfrozen);
649 :
650 : /*
651 : * Update pg_class for table relation. CCI first, in case acquirefunc
652 : * updated pg_class.
653 : */
654 9627 : CommandCounterIncrement();
655 9627 : vac_update_relstats(onerel,
656 : relpages,
657 : totalrows,
658 : relallvisible,
659 : relallfrozen,
660 : hasindex,
661 : InvalidTransactionId,
662 : InvalidMultiXactId,
663 : NULL, NULL,
664 : in_outer_xact);
665 :
666 : /* Same for indexes */
667 23640 : for (ind = 0; ind < nindexes; ind++)
668 : {
669 14013 : AnlIndexData *thisdata = &indexdata[ind];
670 : double totalindexrows;
671 :
672 14013 : totalindexrows = ceil(thisdata->tupleFract * totalrows);
673 14013 : vac_update_relstats(Irel[ind],
674 14013 : RelationGetNumberOfBlocks(Irel[ind]),
675 : totalindexrows,
676 : 0, 0,
677 : false,
678 : InvalidTransactionId,
679 : InvalidMultiXactId,
680 : NULL, NULL,
681 : in_outer_xact);
682 : }
683 : }
684 589 : else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
685 : {
686 : /*
687 : * Partitioned tables don't have storage, so we don't set any fields
688 : * in their pg_class entries except for reltuples and relhasindex.
689 : */
690 498 : CommandCounterIncrement();
691 498 : vac_update_relstats(onerel, -1, totalrows,
692 : 0, 0, hasindex, InvalidTransactionId,
693 : InvalidMultiXactId,
694 : NULL, NULL,
695 : in_outer_xact);
696 : }
697 :
698 : /*
699 : * Now report ANALYZE to the cumulative stats system. For regular tables,
700 : * we do it only if not doing inherited stats. For partitioned tables, we
701 : * only do it for inherited stats. (We're never called for not-inherited
702 : * stats on partitioned tables anyway.)
703 : *
704 : * Reset the mod_since_analyze counter only if we analyzed all columns;
705 : * otherwise, there is still work for auto-analyze to do.
706 : */
707 10216 : if (!inh)
708 9627 : pgstat_report_analyze(onerel, totalrows, totaldeadrows,
709 : (va_cols == NIL), starttime);
710 589 : else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
711 498 : pgstat_report_analyze(onerel, 0, 0, (va_cols == NIL), starttime);
712 :
713 : /*
714 : * If this isn't part of VACUUM ANALYZE, let index AMs do cleanup.
715 : *
716 : * Note that most index AMs perform a no-op as a matter of policy for
717 : * amvacuumcleanup() when called in ANALYZE-only mode. The only exception
718 : * among core index AMs is GIN/ginvacuumcleanup().
719 : */
720 10216 : if (!(params->options & VACOPT_VACUUM))
721 : {
722 19059 : for (ind = 0; ind < nindexes; ind++)
723 : {
724 : IndexBulkDeleteResult *stats;
725 : IndexVacuumInfo ivinfo;
726 :
727 10804 : ivinfo.index = Irel[ind];
728 10804 : ivinfo.heaprel = onerel;
729 10804 : ivinfo.analyze_only = true;
730 10804 : ivinfo.estimated_count = true;
731 10804 : ivinfo.message_level = elevel;
732 10804 : ivinfo.num_heap_tuples = onerel->rd_rel->reltuples;
733 10804 : ivinfo.strategy = vac_strategy;
734 :
735 10804 : stats = index_vacuum_cleanup(&ivinfo, NULL);
736 :
737 10804 : if (stats)
738 0 : pfree(stats);
739 : }
740 : }
741 :
742 : /* Done with indexes */
743 10216 : vac_close_indexes(nindexes, Irel, NoLock);
744 :
745 : /* Log the action if appropriate */
746 10216 : if (instrument)
747 : {
748 322 : TimestampTz endtime = GetCurrentTimestamp();
749 :
750 612 : if (verbose || params->log_analyze_min_duration == 0 ||
751 290 : TimestampDifferenceExceeds(starttime, endtime,
752 290 : params->log_analyze_min_duration))
753 : {
754 : long delay_in_ms;
755 : WalUsage walusage;
756 32 : double read_rate = 0;
757 32 : double write_rate = 0;
758 : char *msgfmt;
759 : StringInfoData buf;
760 : int64 total_blks_hit;
761 : int64 total_blks_read;
762 : int64 total_blks_dirtied;
763 :
764 32 : memset(&bufferusage, 0, sizeof(BufferUsage));
765 32 : BufferUsageAccumDiff(&bufferusage, &pgBufferUsage, &startbufferusage);
766 32 : memset(&walusage, 0, sizeof(WalUsage));
767 32 : WalUsageAccumDiff(&walusage, &pgWalUsage, &startwalusage);
768 :
769 32 : total_blks_hit = bufferusage.shared_blks_hit +
770 32 : bufferusage.local_blks_hit;
771 32 : total_blks_read = bufferusage.shared_blks_read +
772 32 : bufferusage.local_blks_read;
773 32 : total_blks_dirtied = bufferusage.shared_blks_dirtied +
774 32 : bufferusage.local_blks_dirtied;
775 :
776 : /*
777 : * We do not expect an analyze to take > 25 days and it simplifies
778 : * things a bit to use TimestampDifferenceMilliseconds.
779 : */
780 32 : delay_in_ms = TimestampDifferenceMilliseconds(starttime, endtime);
781 :
782 : /*
783 : * Note that we are reporting these read/write rates in the same
784 : * manner as VACUUM does, which means that while the 'average read
785 : * rate' here actually corresponds to page misses and resulting
786 : * reads which are also picked up by track_io_timing, if enabled,
787 : * the 'average write rate' is actually talking about the rate of
788 : * pages being dirtied, not being written out, so it's typical to
789 : * have a non-zero 'avg write rate' while I/O timings only reports
790 : * reads.
791 : *
792 : * It's not clear that an ANALYZE will ever result in
793 : * FlushBuffer() being called, but we track and support reporting
794 : * on I/O write time in case that changes as it's practically free
795 : * to do so anyway.
796 : */
797 :
798 32 : if (delay_in_ms > 0)
799 : {
800 32 : read_rate = (double) BLCKSZ * total_blks_read /
801 32 : (1024 * 1024) / (delay_in_ms / 1000.0);
802 32 : write_rate = (double) BLCKSZ * total_blks_dirtied /
803 32 : (1024 * 1024) / (delay_in_ms / 1000.0);
804 : }
805 :
806 : /*
807 : * We split this up so we don't emit empty I/O timing values when
808 : * track_io_timing isn't enabled.
809 : */
810 :
811 32 : initStringInfo(&buf);
812 :
813 32 : if (AmAutoVacuumWorkerProcess())
814 32 : msgfmt = _("automatic analyze of table \"%s.%s.%s\"\n");
815 : else
816 0 : msgfmt = _("finished analyzing table \"%s.%s.%s\"\n");
817 :
818 32 : appendStringInfo(&buf, msgfmt,
819 : get_database_name(MyDatabaseId),
820 32 : get_namespace_name(RelationGetNamespace(onerel)),
821 32 : RelationGetRelationName(onerel));
822 32 : if (track_cost_delay_timing)
823 : {
824 : /*
825 : * We bypass the changecount mechanism because this value is
826 : * only updated by the calling process.
827 : */
828 0 : appendStringInfo(&buf, _("delay time: %.3f ms\n"),
829 0 : (double) MyBEEntry->st_progress_param[PROGRESS_ANALYZE_DELAY_TIME] / 1000000.0);
830 : }
831 32 : if (track_io_timing)
832 : {
833 0 : double read_ms = (double) (pgStatBlockReadTime - startreadtime) / 1000;
834 0 : double write_ms = (double) (pgStatBlockWriteTime - startwritetime) / 1000;
835 :
836 0 : appendStringInfo(&buf, _("I/O timings: read: %.3f ms, write: %.3f ms\n"),
837 : read_ms, write_ms);
838 : }
839 32 : appendStringInfo(&buf, _("avg read rate: %.3f MB/s, avg write rate: %.3f MB/s\n"),
840 : read_rate, write_rate);
841 32 : appendStringInfo(&buf, _("buffer usage: %" PRId64 " hits, %" PRId64 " reads, %" PRId64 " dirtied\n"),
842 : total_blks_hit,
843 : total_blks_read,
844 : total_blks_dirtied);
845 32 : appendStringInfo(&buf,
846 32 : _("WAL usage: %" PRId64 " records, %" PRId64 " full page images, %" PRIu64 " bytes, %" PRIu64 " full page image bytes, %" PRId64 " buffers full\n"),
847 : walusage.wal_records,
848 : walusage.wal_fpi,
849 : walusage.wal_bytes,
850 : walusage.wal_fpi_bytes,
851 : walusage.wal_buffers_full);
852 32 : appendStringInfo(&buf, _("system usage: %s"), pg_rusage_show(&ru0));
853 :
854 32 : ereport(verbose ? INFO : LOG,
855 : (errmsg_internal("%s", buf.data)));
856 :
857 32 : pfree(buf.data);
858 : }
859 : }
860 :
861 : /* Roll back any GUC changes executed by index functions */
862 10216 : AtEOXact_GUC(false, save_nestlevel);
863 :
864 : /* Restore userid and security context */
865 10216 : SetUserIdAndSecContext(save_userid, save_sec_context);
866 :
867 : /* Restore current context and release memory */
868 10216 : MemoryContextSwitchTo(caller_context);
869 10216 : MemoryContextDelete(anl_context);
870 10216 : anl_context = NULL;
871 10216 : }
872 :
873 : /*
874 : * Compute statistics about indexes of a relation
875 : */
876 : static void
877 4142 : compute_index_stats(Relation onerel, double totalrows,
878 : AnlIndexData *indexdata, int nindexes,
879 : HeapTuple *rows, int numrows,
880 : MemoryContext col_context)
881 : {
882 : MemoryContext ind_context,
883 : old_context;
884 : Datum values[INDEX_MAX_KEYS];
885 : bool isnull[INDEX_MAX_KEYS];
886 : int ind,
887 : i;
888 :
889 4142 : ind_context = AllocSetContextCreate(anl_context,
890 : "Analyze Index",
891 : ALLOCSET_DEFAULT_SIZES);
892 4142 : old_context = MemoryContextSwitchTo(ind_context);
893 :
894 12164 : for (ind = 0; ind < nindexes; ind++)
895 : {
896 8026 : AnlIndexData *thisdata = &indexdata[ind];
897 8026 : IndexInfo *indexInfo = thisdata->indexInfo;
898 8026 : int attr_cnt = thisdata->attr_cnt;
899 : TupleTableSlot *slot;
900 : EState *estate;
901 : ExprContext *econtext;
902 : ExprState *predicate;
903 : Datum *exprvals;
904 : bool *exprnulls;
905 : int numindexrows,
906 : tcnt,
907 : rowno;
908 : double totalindexrows;
909 :
910 : /* Ignore index if no columns to analyze and not partial */
911 8026 : if (attr_cnt == 0 && indexInfo->ii_Predicate == NIL)
912 7930 : continue;
913 :
914 : /*
915 : * Need an EState for evaluation of index expressions and
916 : * partial-index predicates. Create it in the per-index context to be
917 : * sure it gets cleaned up at the bottom of the loop.
918 : */
919 96 : estate = CreateExecutorState();
920 96 : econtext = GetPerTupleExprContext(estate);
921 : /* Need a slot to hold the current heap tuple, too */
922 96 : slot = MakeSingleTupleTableSlot(RelationGetDescr(onerel),
923 : &TTSOpsHeapTuple);
924 :
925 : /* Arrange for econtext's scan tuple to be the tuple under test */
926 96 : econtext->ecxt_scantuple = slot;
927 :
928 : /* Set up execution state for predicate. */
929 96 : predicate = ExecPrepareQual(indexInfo->ii_Predicate, estate);
930 :
931 : /* Compute and save index expression values */
932 96 : exprvals = (Datum *) palloc(numrows * attr_cnt * sizeof(Datum));
933 96 : exprnulls = (bool *) palloc(numrows * attr_cnt * sizeof(bool));
934 96 : numindexrows = 0;
935 96 : tcnt = 0;
936 111173 : for (rowno = 0; rowno < numrows; rowno++)
937 : {
938 111081 : HeapTuple heapTuple = rows[rowno];
939 :
940 111081 : vacuum_delay_point(true);
941 :
942 : /*
943 : * Reset the per-tuple context each time, to reclaim any cruft
944 : * left behind by evaluating the predicate or index expressions.
945 : */
946 111081 : ResetExprContext(econtext);
947 :
948 : /* Set up for predicate or expression evaluation */
949 111081 : ExecStoreHeapTuple(heapTuple, slot, false);
950 :
951 : /* If index is partial, check predicate */
952 111081 : if (predicate != NULL)
953 : {
954 37036 : if (!ExecQual(predicate, econtext))
955 23552 : continue;
956 : }
957 87529 : numindexrows++;
958 :
959 87529 : if (attr_cnt > 0)
960 : {
961 : /*
962 : * Evaluate the index row to compute expression values. We
963 : * could do this by hand, but FormIndexDatum is convenient.
964 : */
965 74045 : FormIndexDatum(indexInfo,
966 : slot,
967 : estate,
968 : values,
969 : isnull);
970 :
971 : /*
972 : * Save just the columns we care about. We copy the values
973 : * into ind_context from the estate's per-tuple context.
974 : */
975 148082 : for (i = 0; i < attr_cnt; i++)
976 : {
977 74041 : VacAttrStats *stats = thisdata->vacattrstats[i];
978 74041 : int attnum = stats->tupattnum;
979 :
980 74041 : if (isnull[attnum - 1])
981 : {
982 8 : exprvals[tcnt] = (Datum) 0;
983 8 : exprnulls[tcnt] = true;
984 : }
985 : else
986 : {
987 148066 : exprvals[tcnt] = datumCopy(values[attnum - 1],
988 74033 : stats->attrtype->typbyval,
989 74033 : stats->attrtype->typlen);
990 74033 : exprnulls[tcnt] = false;
991 : }
992 74041 : tcnt++;
993 : }
994 : }
995 : }
996 :
997 : /*
998 : * Having counted the number of rows that pass the predicate in the
999 : * sample, we can estimate the total number of rows in the index.
1000 : */
1001 92 : thisdata->tupleFract = (double) numindexrows / (double) numrows;
1002 92 : totalindexrows = ceil(thisdata->tupleFract * totalrows);
1003 :
1004 : /*
1005 : * Now we can compute the statistics for the expression columns.
1006 : */
1007 92 : if (numindexrows > 0)
1008 : {
1009 87 : MemoryContextSwitchTo(col_context);
1010 148 : for (i = 0; i < attr_cnt; i++)
1011 : {
1012 61 : VacAttrStats *stats = thisdata->vacattrstats[i];
1013 :
1014 61 : stats->exprvals = exprvals + i;
1015 61 : stats->exprnulls = exprnulls + i;
1016 61 : stats->rowstride = attr_cnt;
1017 61 : stats->compute_stats(stats,
1018 : ind_fetch_func,
1019 : numindexrows,
1020 : totalindexrows);
1021 :
1022 61 : MemoryContextReset(col_context);
1023 : }
1024 : }
1025 :
1026 : /* And clean up */
1027 92 : MemoryContextSwitchTo(ind_context);
1028 :
1029 92 : ExecDropSingleTupleTableSlot(slot);
1030 92 : FreeExecutorState(estate);
1031 92 : MemoryContextReset(ind_context);
1032 : }
1033 :
1034 4138 : MemoryContextSwitchTo(old_context);
1035 4138 : MemoryContextDelete(ind_context);
1036 4138 : }
1037 :
1038 : /*
1039 : * validate_va_cols_list -- validate the columns list given to analyze_rel
1040 : *
1041 : * Note that system attributes are never analyzed, so we just reject them at
1042 : * the lookup stage. We also reject duplicate column mentions. (We could
1043 : * alternatively ignore duplicates, but analyzing a column twice won't work;
1044 : * we'd end up making a conflicting update in pg_statistic.)
1045 : */
1046 : static void
1047 75 : validate_va_cols_list(Relation onerel, List *va_cols)
1048 : {
1049 75 : Bitmapset *unique_cols = NULL;
1050 : ListCell *le;
1051 :
1052 : Assert(va_cols != NIL);
1053 139 : foreach(le, va_cols)
1054 : {
1055 99 : char *col = strVal(lfirst(le));
1056 99 : int i = attnameAttNum(onerel, col, false);
1057 :
1058 99 : if (i == InvalidAttrNumber)
1059 26 : ereport(ERROR,
1060 : (errcode(ERRCODE_UNDEFINED_COLUMN),
1061 : errmsg("column \"%s\" of relation \"%s\" does not exist",
1062 : col, RelationGetRelationName(onerel))));
1063 73 : if (bms_is_member(i, unique_cols))
1064 9 : ereport(ERROR,
1065 : (errcode(ERRCODE_DUPLICATE_COLUMN),
1066 : errmsg("column \"%s\" of relation \"%s\" appears more than once",
1067 : col, RelationGetRelationName(onerel))));
1068 64 : unique_cols = bms_add_member(unique_cols, i);
1069 : }
1070 40 : }
1071 :
1072 : /*
1073 : * examine_attribute -- pre-analysis of a single column
1074 : *
1075 : * Determine whether the column is analyzable; if so, create and initialize
1076 : * a VacAttrStats struct for it. If not, return NULL.
1077 : *
1078 : * If index_expr isn't NULL, then we're trying to analyze an expression index,
1079 : * and index_expr is the expression tree representing the column's data.
1080 : */
1081 : static VacAttrStats *
1082 71749 : examine_attribute(Relation onerel, int attnum, Node *index_expr)
1083 : {
1084 71749 : Form_pg_attribute attr = TupleDescAttr(onerel->rd_att, attnum - 1);
1085 : int attstattarget;
1086 : HeapTuple typtuple;
1087 : VacAttrStats *stats;
1088 : int i;
1089 : bool ok;
1090 :
1091 : /*
1092 : * Check if the column is analyzable.
1093 : */
1094 71749 : if (!attribute_is_analyzable(onerel, attnum, attr, &attstattarget))
1095 77 : return NULL;
1096 :
1097 : /*
1098 : * Create the VacAttrStats struct.
1099 : */
1100 71672 : stats = palloc0_object(VacAttrStats);
1101 71672 : stats->attstattarget = attstattarget;
1102 :
1103 : /*
1104 : * When analyzing an expression index, believe the expression tree's type
1105 : * not the column datatype --- the latter might be the opckeytype storage
1106 : * type of the opclass, which is not interesting for our purposes. (Note:
1107 : * if we did anything with non-expression index columns, we'd need to
1108 : * figure out where to get the correct type info from, but for now that's
1109 : * not a problem.) It's not clear whether anyone will care about the
1110 : * typmod, but we store that too just in case.
1111 : */
1112 71672 : if (index_expr)
1113 : {
1114 77 : stats->attrtypid = exprType(index_expr);
1115 77 : stats->attrtypmod = exprTypmod(index_expr);
1116 :
1117 : /*
1118 : * If a collation has been specified for the index column, use that in
1119 : * preference to anything else; but if not, fall back to whatever we
1120 : * can get from the expression.
1121 : */
1122 77 : if (OidIsValid(onerel->rd_indcollation[attnum - 1]))
1123 9 : stats->attrcollid = onerel->rd_indcollation[attnum - 1];
1124 : else
1125 68 : stats->attrcollid = exprCollation(index_expr);
1126 : }
1127 : else
1128 : {
1129 71595 : stats->attrtypid = attr->atttypid;
1130 71595 : stats->attrtypmod = attr->atttypmod;
1131 71595 : stats->attrcollid = attr->attcollation;
1132 : }
1133 :
1134 71672 : typtuple = SearchSysCacheCopy1(TYPEOID,
1135 : ObjectIdGetDatum(stats->attrtypid));
1136 71672 : if (!HeapTupleIsValid(typtuple))
1137 0 : elog(ERROR, "cache lookup failed for type %u", stats->attrtypid);
1138 71672 : stats->attrtype = (Form_pg_type) GETSTRUCT(typtuple);
1139 71672 : stats->anl_context = anl_context;
1140 71672 : stats->tupattnum = attnum;
1141 :
1142 : /*
1143 : * The fields describing the stats->stavalues[n] element types default to
1144 : * the type of the data being analyzed, but the type-specific typanalyze
1145 : * function can change them if it wants to store something else.
1146 : */
1147 430032 : for (i = 0; i < STATISTIC_NUM_SLOTS; i++)
1148 : {
1149 358360 : stats->statypid[i] = stats->attrtypid;
1150 358360 : stats->statyplen[i] = stats->attrtype->typlen;
1151 358360 : stats->statypbyval[i] = stats->attrtype->typbyval;
1152 358360 : stats->statypalign[i] = stats->attrtype->typalign;
1153 : }
1154 :
1155 : /*
1156 : * Call the type-specific typanalyze function. If none is specified, use
1157 : * std_typanalyze().
1158 : */
1159 71672 : if (OidIsValid(stats->attrtype->typanalyze))
1160 4963 : ok = DatumGetBool(OidFunctionCall1(stats->attrtype->typanalyze,
1161 : PointerGetDatum(stats)));
1162 : else
1163 66709 : ok = std_typanalyze(stats);
1164 :
1165 71672 : if (!ok || stats->compute_stats == NULL || stats->minrows <= 0)
1166 : {
1167 2 : heap_freetuple(typtuple);
1168 2 : pfree(stats);
1169 2 : return NULL;
1170 : }
1171 :
1172 71670 : return stats;
1173 : }
1174 :
1175 : bool
1176 71764 : attribute_is_analyzable(Relation onerel, int attnum, Form_pg_attribute attr,
1177 : int *p_attstattarget)
1178 : {
1179 : int attstattarget;
1180 : HeapTuple atttuple;
1181 : Datum dat;
1182 : bool isnull;
1183 :
1184 : /* Never analyze dropped columns */
1185 71764 : if (attr->attisdropped)
1186 3 : return false;
1187 :
1188 : /* Don't analyze virtual generated columns */
1189 71761 : if (attr->attgenerated == ATTRIBUTE_GENERATED_VIRTUAL)
1190 67 : return false;
1191 :
1192 : /*
1193 : * Get attstattarget value. Set to -1 if null. (Analyze functions expect
1194 : * -1 to mean use default_statistics_target; see for example
1195 : * std_typanalyze.)
1196 : */
1197 71694 : atttuple = SearchSysCache2(ATTNUM, ObjectIdGetDatum(RelationGetRelid(onerel)), Int16GetDatum(attnum));
1198 71694 : if (!HeapTupleIsValid(atttuple))
1199 0 : elog(ERROR, "cache lookup failed for attribute %d of relation %u",
1200 : attnum, RelationGetRelid(onerel));
1201 71694 : dat = SysCacheGetAttr(ATTNUM, atttuple, Anum_pg_attribute_attstattarget, &isnull);
1202 71694 : attstattarget = isnull ? -1 : DatumGetInt16(dat);
1203 71694 : ReleaseSysCache(atttuple);
1204 :
1205 : /* Don't analyze column if user has specified not to */
1206 71694 : if (attstattarget == 0)
1207 7 : return false;
1208 :
1209 71687 : if (p_attstattarget)
1210 71672 : *p_attstattarget = attstattarget;
1211 71687 : return true;
1212 : }
1213 :
1214 : /*
1215 : * Read stream callback returning the next BlockNumber as chosen by the
1216 : * BlockSampling algorithm.
1217 : */
1218 : static BlockNumber
1219 91940 : block_sampling_read_stream_next(ReadStream *stream,
1220 : void *callback_private_data,
1221 : void *per_buffer_data)
1222 : {
1223 91940 : BlockSamplerData *bs = callback_private_data;
1224 :
1225 91940 : return BlockSampler_HasMore(bs) ? BlockSampler_Next(bs) : InvalidBlockNumber;
1226 : }
1227 :
1228 : /*
1229 : * acquire_sample_rows -- acquire a random sample of rows from the table
1230 : *
1231 : * Selected rows are returned in the caller-allocated array rows[], which
1232 : * must have at least targrows entries.
1233 : * The actual number of rows selected is returned as the function result.
1234 : * We also estimate the total numbers of live and dead rows in the table,
1235 : * and return them into *totalrows and *totaldeadrows, respectively.
1236 : *
1237 : * The returned list of tuples is in order by physical position in the table.
1238 : * (We will rely on this later to derive correlation estimates.)
1239 : *
1240 : * As of May 2004 we use a new two-stage method: Stage one selects up
1241 : * to targrows random blocks (or all blocks, if there aren't so many).
1242 : * Stage two scans these blocks and uses the Vitter algorithm to create
1243 : * a random sample of targrows rows (or less, if there are less in the
1244 : * sample of blocks). The two stages are executed simultaneously: each
1245 : * block is processed as soon as stage one returns its number and while
1246 : * the rows are read stage two controls which ones are to be inserted
1247 : * into the sample.
1248 : *
1249 : * Although every row has an equal chance of ending up in the final
1250 : * sample, this sampling method is not perfect: not every possible
1251 : * sample has an equal chance of being selected. For large relations
1252 : * the number of different blocks represented by the sample tends to be
1253 : * too small. We can live with that for now. Improvements are welcome.
1254 : *
1255 : * An important property of this sampling method is that because we do
1256 : * look at a statistically unbiased set of blocks, we should get
1257 : * unbiased estimates of the average numbers of live and dead rows per
1258 : * block. The previous sampling method put too much credence in the row
1259 : * density near the start of the table.
1260 : */
1261 : static int
1262 10985 : acquire_sample_rows(Relation onerel, int elevel,
1263 : HeapTuple *rows, int targrows,
1264 : double *totalrows, double *totaldeadrows)
1265 : {
1266 10985 : int numrows = 0; /* # rows now in reservoir */
1267 10985 : double samplerows = 0; /* total # rows collected */
1268 10985 : double liverows = 0; /* # live rows seen */
1269 10985 : double deadrows = 0; /* # dead rows seen */
1270 10985 : double rowstoskip = -1; /* -1 means not set yet */
1271 : uint32 randseed; /* Seed for block sampler(s) */
1272 : BlockNumber totalblocks;
1273 : BlockSamplerData bs;
1274 : ReservoirStateData rstate;
1275 : TupleTableSlot *slot;
1276 : TableScanDesc scan;
1277 : BlockNumber nblocks;
1278 10985 : BlockNumber blksdone = 0;
1279 : ReadStream *stream;
1280 :
1281 : Assert(targrows > 0);
1282 :
1283 10985 : totalblocks = RelationGetNumberOfBlocks(onerel);
1284 :
1285 : /* Prepare for sampling block numbers */
1286 10985 : randseed = pg_prng_uint32(&pg_global_prng_state);
1287 10985 : nblocks = BlockSampler_Init(&bs, totalblocks, targrows, randseed);
1288 :
1289 : /* Report sampling block numbers */
1290 10985 : pgstat_progress_update_param(PROGRESS_ANALYZE_BLOCKS_TOTAL,
1291 : nblocks);
1292 :
1293 : /* Prepare for sampling rows */
1294 10985 : reservoir_init_selection_state(&rstate, targrows);
1295 :
1296 10985 : scan = table_beginscan_analyze(onerel);
1297 10985 : slot = table_slot_create(onerel, NULL);
1298 :
1299 : /*
1300 : * It is safe to use batching, as block_sampling_read_stream_next never
1301 : * blocks.
1302 : */
1303 10985 : stream = read_stream_begin_relation(READ_STREAM_MAINTENANCE |
1304 : READ_STREAM_USE_BATCHING,
1305 : vac_strategy,
1306 : scan->rs_rd,
1307 : MAIN_FORKNUM,
1308 : block_sampling_read_stream_next,
1309 : &bs,
1310 : 0);
1311 :
1312 : /* Outer loop over blocks to sample */
1313 91940 : while (table_scan_analyze_next_block(scan, stream))
1314 : {
1315 80955 : vacuum_delay_point(true);
1316 :
1317 7219901 : while (table_scan_analyze_next_tuple(scan, &liverows, &deadrows, slot))
1318 : {
1319 : /*
1320 : * The first targrows sample rows are simply copied into the
1321 : * reservoir. Then we start replacing tuples in the sample until
1322 : * we reach the end of the relation. This algorithm is from Jeff
1323 : * Vitter's paper (see full citation in utils/misc/sampling.c). It
1324 : * works by repeatedly computing the number of tuples to skip
1325 : * before selecting a tuple, which replaces a randomly chosen
1326 : * element of the reservoir (current set of tuples). At all times
1327 : * the reservoir is a true random sample of the tuples we've
1328 : * passed over so far, so when we fall off the end of the relation
1329 : * we're done.
1330 : */
1331 7138946 : if (numrows < targrows)
1332 6588967 : rows[numrows++] = ExecCopySlotHeapTuple(slot);
1333 : else
1334 : {
1335 : /*
1336 : * t in Vitter's paper is the number of records already
1337 : * processed. If we need to compute a new S value, we must
1338 : * use the not-yet-incremented value of samplerows as t.
1339 : */
1340 549979 : if (rowstoskip < 0)
1341 275958 : rowstoskip = reservoir_get_next_S(&rstate, samplerows, targrows);
1342 :
1343 549979 : if (rowstoskip <= 0)
1344 : {
1345 : /*
1346 : * Found a suitable tuple, so save it, replacing one old
1347 : * tuple at random
1348 : */
1349 275929 : int k = (int) (targrows * sampler_random_fract(&rstate.randstate));
1350 :
1351 : Assert(k >= 0 && k < targrows);
1352 275929 : heap_freetuple(rows[k]);
1353 275929 : rows[k] = ExecCopySlotHeapTuple(slot);
1354 : }
1355 :
1356 549979 : rowstoskip -= 1;
1357 : }
1358 :
1359 7138946 : samplerows += 1;
1360 : }
1361 :
1362 80955 : pgstat_progress_update_param(PROGRESS_ANALYZE_BLOCKS_DONE,
1363 : ++blksdone);
1364 : }
1365 :
1366 10985 : read_stream_end(stream);
1367 :
1368 10985 : ExecDropSingleTupleTableSlot(slot);
1369 10985 : table_endscan(scan);
1370 :
1371 : /*
1372 : * If we didn't find as many tuples as we wanted then we're done. No sort
1373 : * is needed, since they're already in order.
1374 : *
1375 : * Otherwise we need to sort the collected tuples by position
1376 : * (itempointer). It's not worth worrying about corner cases where the
1377 : * tuples are already sorted.
1378 : */
1379 10985 : if (numrows == targrows)
1380 96 : qsort_interruptible(rows, numrows, sizeof(HeapTuple),
1381 : compare_rows, NULL);
1382 :
1383 : /*
1384 : * Estimate total numbers of live and dead rows in relation, extrapolating
1385 : * on the assumption that the average tuple density in pages we didn't
1386 : * scan is the same as in the pages we did scan. Since what we scanned is
1387 : * a random sample of the pages in the relation, this should be a good
1388 : * assumption.
1389 : */
1390 10985 : if (bs.m > 0)
1391 : {
1392 7594 : *totalrows = floor((liverows / bs.m) * totalblocks + 0.5);
1393 7594 : *totaldeadrows = floor((deadrows / bs.m) * totalblocks + 0.5);
1394 : }
1395 : else
1396 : {
1397 3391 : *totalrows = 0.0;
1398 3391 : *totaldeadrows = 0.0;
1399 : }
1400 :
1401 : /*
1402 : * Emit some interesting relation info
1403 : */
1404 10985 : ereport(elevel,
1405 : (errmsg("\"%s\": scanned %d of %u pages, "
1406 : "containing %.0f live rows and %.0f dead rows; "
1407 : "%d rows in sample, %.0f estimated total rows",
1408 : RelationGetRelationName(onerel),
1409 : bs.m, totalblocks,
1410 : liverows, deadrows,
1411 : numrows, *totalrows)));
1412 :
1413 10985 : return numrows;
1414 : }
1415 :
1416 : /*
1417 : * Comparator for sorting rows[] array
1418 : */
1419 : static int
1420 5128352 : compare_rows(const void *a, const void *b, void *arg)
1421 : {
1422 5128352 : HeapTuple ha = *(const HeapTuple *) a;
1423 5128352 : HeapTuple hb = *(const HeapTuple *) b;
1424 5128352 : BlockNumber ba = ItemPointerGetBlockNumber(&ha->t_self);
1425 5128352 : OffsetNumber oa = ItemPointerGetOffsetNumber(&ha->t_self);
1426 5128352 : BlockNumber bb = ItemPointerGetBlockNumber(&hb->t_self);
1427 5128352 : OffsetNumber ob = ItemPointerGetOffsetNumber(&hb->t_self);
1428 :
1429 5128352 : if (ba < bb)
1430 1365669 : return -1;
1431 3762683 : if (ba > bb)
1432 1351591 : return 1;
1433 2411092 : if (oa < ob)
1434 1408819 : return -1;
1435 1002273 : if (oa > ob)
1436 1002273 : return 1;
1437 0 : return 0;
1438 : }
1439 :
1440 :
1441 : /*
1442 : * acquire_inherited_sample_rows -- acquire sample rows from inheritance tree
1443 : *
1444 : * This has the same API as acquire_sample_rows, except that rows are
1445 : * collected from all inheritance children as well as the specified table.
1446 : * We fail and return zero if there are no inheritance children, or if all
1447 : * children are foreign tables that don't support ANALYZE.
1448 : */
1449 : static int
1450 589 : acquire_inherited_sample_rows(Relation onerel, int elevel,
1451 : HeapTuple *rows, int targrows,
1452 : double *totalrows, double *totaldeadrows)
1453 : {
1454 : List *tableOIDs;
1455 : Relation *rels;
1456 : AcquireSampleRowsFunc *acquirefuncs;
1457 : double *relblocks;
1458 : double totalblocks;
1459 : int numrows,
1460 : nrels,
1461 : i;
1462 : ListCell *lc;
1463 : bool has_child;
1464 :
1465 : /* Initialize output parameters to zero now, in case we exit early */
1466 589 : *totalrows = 0;
1467 589 : *totaldeadrows = 0;
1468 :
1469 : /*
1470 : * Find all members of inheritance set. We only need AccessShareLock on
1471 : * the children.
1472 : */
1473 : tableOIDs =
1474 589 : find_all_inheritors(RelationGetRelid(onerel), AccessShareLock, NULL);
1475 :
1476 : /*
1477 : * Check that there's at least one descendant, else fail. This could
1478 : * happen despite analyze_rel's relhassubclass check, if table once had a
1479 : * child but no longer does. In that case, we can clear the
1480 : * relhassubclass field so as not to make the same mistake again later.
1481 : * (This is safe because we hold ShareUpdateExclusiveLock.)
1482 : */
1483 589 : if (list_length(tableOIDs) < 2)
1484 : {
1485 : /* CCI because we already updated the pg_class row in this command */
1486 13 : CommandCounterIncrement();
1487 13 : SetRelationHasSubclass(RelationGetRelid(onerel), false);
1488 13 : ereport(elevel,
1489 : (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no child tables",
1490 : get_namespace_name(RelationGetNamespace(onerel)),
1491 : RelationGetRelationName(onerel))));
1492 13 : return 0;
1493 : }
1494 :
1495 : /*
1496 : * Identify acquirefuncs to use, and count blocks in all the relations.
1497 : * The result could overflow BlockNumber, so we use double arithmetic.
1498 : */
1499 576 : rels = (Relation *) palloc(list_length(tableOIDs) * sizeof(Relation));
1500 : acquirefuncs = (AcquireSampleRowsFunc *)
1501 576 : palloc(list_length(tableOIDs) * sizeof(AcquireSampleRowsFunc));
1502 576 : relblocks = (double *) palloc(list_length(tableOIDs) * sizeof(double));
1503 576 : totalblocks = 0;
1504 576 : nrels = 0;
1505 576 : has_child = false;
1506 2630 : foreach(lc, tableOIDs)
1507 : {
1508 2054 : Oid childOID = lfirst_oid(lc);
1509 : Relation childrel;
1510 2054 : AcquireSampleRowsFunc acquirefunc = NULL;
1511 2054 : BlockNumber relpages = 0;
1512 :
1513 : /* We already got the needed lock */
1514 2054 : childrel = table_open(childOID, NoLock);
1515 :
1516 : /* Ignore if temp table of another backend */
1517 2054 : if (RELATION_IS_OTHER_TEMP(childrel))
1518 : {
1519 : /* ... but release the lock on it */
1520 : Assert(childrel != onerel);
1521 0 : table_close(childrel, AccessShareLock);
1522 546 : continue;
1523 : }
1524 :
1525 : /* Check table type (MATVIEW can't happen, but might as well allow) */
1526 2054 : if (childrel->rd_rel->relkind == RELKIND_RELATION ||
1527 561 : childrel->rd_rel->relkind == RELKIND_MATVIEW)
1528 : {
1529 : /* Regular table, so use the regular row acquisition function */
1530 1493 : acquirefunc = acquire_sample_rows;
1531 1493 : relpages = RelationGetNumberOfBlocks(childrel);
1532 : }
1533 561 : else if (childrel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
1534 : {
1535 : /*
1536 : * For a foreign table, call the FDW's hook function to see
1537 : * whether it supports analysis.
1538 : */
1539 : FdwRoutine *fdwroutine;
1540 15 : bool ok = false;
1541 :
1542 15 : fdwroutine = GetFdwRoutineForRelation(childrel, false);
1543 :
1544 15 : if (fdwroutine->AnalyzeForeignTable != NULL)
1545 15 : ok = fdwroutine->AnalyzeForeignTable(childrel,
1546 : &acquirefunc,
1547 : &relpages);
1548 :
1549 15 : if (!ok)
1550 : {
1551 : /* ignore, but release the lock on it */
1552 : Assert(childrel != onerel);
1553 0 : table_close(childrel, AccessShareLock);
1554 0 : continue;
1555 : }
1556 : }
1557 : else
1558 : {
1559 : /*
1560 : * ignore, but release the lock on it. don't try to unlock the
1561 : * passed-in relation
1562 : */
1563 : Assert(childrel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE);
1564 546 : if (childrel != onerel)
1565 52 : table_close(childrel, AccessShareLock);
1566 : else
1567 494 : table_close(childrel, NoLock);
1568 546 : continue;
1569 : }
1570 :
1571 : /* OK, we'll process this child */
1572 1508 : has_child = true;
1573 1508 : rels[nrels] = childrel;
1574 1508 : acquirefuncs[nrels] = acquirefunc;
1575 1508 : relblocks[nrels] = (double) relpages;
1576 1508 : totalblocks += (double) relpages;
1577 1508 : nrels++;
1578 : }
1579 :
1580 : /*
1581 : * If we don't have at least one child table to consider, fail. If the
1582 : * relation is a partitioned table, it's not counted as a child table.
1583 : */
1584 576 : if (!has_child)
1585 : {
1586 0 : ereport(elevel,
1587 : (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no analyzable child tables",
1588 : get_namespace_name(RelationGetNamespace(onerel)),
1589 : RelationGetRelationName(onerel))));
1590 0 : return 0;
1591 : }
1592 :
1593 : /*
1594 : * Now sample rows from each relation, proportionally to its fraction of
1595 : * the total block count. (This might be less than desirable if the child
1596 : * rels have radically different free-space percentages, but it's not
1597 : * clear that it's worth working harder.)
1598 : */
1599 576 : pgstat_progress_update_param(PROGRESS_ANALYZE_CHILD_TABLES_TOTAL,
1600 : nrels);
1601 576 : numrows = 0;
1602 2084 : for (i = 0; i < nrels; i++)
1603 : {
1604 1508 : Relation childrel = rels[i];
1605 1508 : AcquireSampleRowsFunc acquirefunc = acquirefuncs[i];
1606 1508 : double childblocks = relblocks[i];
1607 :
1608 : /*
1609 : * Report progress. The sampling function will normally report blocks
1610 : * done/total, but we need to reset them to 0 here, so that they don't
1611 : * show an old value until that.
1612 : */
1613 : {
1614 1508 : const int progress_index[] = {
1615 : PROGRESS_ANALYZE_CURRENT_CHILD_TABLE_RELID,
1616 : PROGRESS_ANALYZE_BLOCKS_DONE,
1617 : PROGRESS_ANALYZE_BLOCKS_TOTAL
1618 : };
1619 1508 : const int64 progress_vals[] = {
1620 1508 : RelationGetRelid(childrel),
1621 : 0,
1622 : 0,
1623 : };
1624 :
1625 1508 : pgstat_progress_update_multi_param(3, progress_index, progress_vals);
1626 : }
1627 :
1628 1508 : if (childblocks > 0)
1629 : {
1630 : int childtargrows;
1631 :
1632 1406 : childtargrows = (int) rint(targrows * childblocks / totalblocks);
1633 : /* Make sure we don't overrun due to roundoff error */
1634 1406 : childtargrows = Min(childtargrows, targrows - numrows);
1635 1406 : if (childtargrows > 0)
1636 : {
1637 : int childrows;
1638 : double trows,
1639 : tdrows;
1640 :
1641 : /* Fetch a random sample of the child's rows */
1642 1406 : childrows = (*acquirefunc) (childrel, elevel,
1643 1406 : rows + numrows, childtargrows,
1644 : &trows, &tdrows);
1645 :
1646 : /* We may need to convert from child's rowtype to parent's */
1647 1406 : if (childrows > 0 &&
1648 1406 : !equalRowTypes(RelationGetDescr(childrel),
1649 : RelationGetDescr(onerel)))
1650 : {
1651 : TupleConversionMap *map;
1652 :
1653 1340 : map = convert_tuples_by_name(RelationGetDescr(childrel),
1654 : RelationGetDescr(onerel));
1655 1340 : if (map != NULL)
1656 : {
1657 : int j;
1658 :
1659 70681 : for (j = 0; j < childrows; j++)
1660 : {
1661 : HeapTuple newtup;
1662 :
1663 70594 : newtup = execute_attr_map_tuple(rows[numrows + j], map);
1664 70594 : heap_freetuple(rows[numrows + j]);
1665 70594 : rows[numrows + j] = newtup;
1666 : }
1667 87 : free_conversion_map(map);
1668 : }
1669 : }
1670 :
1671 : /* And add to counts */
1672 1406 : numrows += childrows;
1673 1406 : *totalrows += trows;
1674 1406 : *totaldeadrows += tdrows;
1675 : }
1676 : }
1677 :
1678 : /*
1679 : * Note: we cannot release the child-table locks, since we may have
1680 : * pointers to their TOAST tables in the sampled rows.
1681 : */
1682 1508 : table_close(childrel, NoLock);
1683 1508 : pgstat_progress_update_param(PROGRESS_ANALYZE_CHILD_TABLES_DONE,
1684 1508 : i + 1);
1685 : }
1686 :
1687 576 : return numrows;
1688 : }
1689 :
1690 :
1691 : /*
1692 : * update_attstats() -- update attribute statistics for one relation
1693 : *
1694 : * Statistics are stored in several places: the pg_class row for the
1695 : * relation has stats about the whole relation, and there is a
1696 : * pg_statistic row for each (non-system) attribute that has ever
1697 : * been analyzed. The pg_class values are updated by VACUUM, not here.
1698 : *
1699 : * pg_statistic rows are just added or updated normally. This means
1700 : * that pg_statistic will probably contain some deleted rows at the
1701 : * completion of a vacuum cycle, unless it happens to get vacuumed last.
1702 : *
1703 : * To keep things simple, we punt for pg_statistic, and don't try
1704 : * to compute or store rows for pg_statistic itself in pg_statistic.
1705 : * This could possibly be made to work, but it's not worth the trouble.
1706 : * Note analyze_rel() has seen to it that we won't come here when
1707 : * vacuuming pg_statistic itself.
1708 : *
1709 : * Note: there would be a race condition here if two backends could
1710 : * ANALYZE the same table concurrently. Presently, we lock that out
1711 : * by taking a self-exclusive lock on the relation in analyze_rel().
1712 : */
1713 : static void
1714 14788 : update_attstats(Oid relid, bool inh, int natts, VacAttrStats **vacattrstats)
1715 : {
1716 : Relation sd;
1717 : int attno;
1718 14788 : CatalogIndexState indstate = NULL;
1719 :
1720 14788 : if (natts <= 0)
1721 7966 : return; /* nothing to do */
1722 :
1723 6822 : sd = table_open(StatisticRelationId, RowExclusiveLock);
1724 :
1725 57564 : for (attno = 0; attno < natts; attno++)
1726 : {
1727 50742 : VacAttrStats *stats = vacattrstats[attno];
1728 : HeapTuple stup,
1729 : oldtup;
1730 : int i,
1731 : k,
1732 : n;
1733 : Datum values[Natts_pg_statistic];
1734 : bool nulls[Natts_pg_statistic];
1735 : bool replaces[Natts_pg_statistic];
1736 :
1737 : /* Ignore attr if we weren't able to collect stats */
1738 50742 : if (!stats->stats_valid)
1739 6 : continue;
1740 :
1741 : /*
1742 : * Construct a new pg_statistic tuple
1743 : */
1744 1623552 : for (i = 0; i < Natts_pg_statistic; ++i)
1745 : {
1746 1572816 : nulls[i] = false;
1747 1572816 : replaces[i] = true;
1748 : }
1749 :
1750 50736 : values[Anum_pg_statistic_starelid - 1] = ObjectIdGetDatum(relid);
1751 50736 : values[Anum_pg_statistic_staattnum - 1] = Int16GetDatum(stats->tupattnum);
1752 50736 : values[Anum_pg_statistic_stainherit - 1] = BoolGetDatum(inh);
1753 50736 : values[Anum_pg_statistic_stanullfrac - 1] = Float4GetDatum(stats->stanullfrac);
1754 50736 : values[Anum_pg_statistic_stawidth - 1] = Int32GetDatum(stats->stawidth);
1755 50736 : values[Anum_pg_statistic_stadistinct - 1] = Float4GetDatum(stats->stadistinct);
1756 50736 : i = Anum_pg_statistic_stakind1 - 1;
1757 304416 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1758 : {
1759 253680 : values[i++] = Int16GetDatum(stats->stakind[k]); /* stakindN */
1760 : }
1761 50736 : i = Anum_pg_statistic_staop1 - 1;
1762 304416 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1763 : {
1764 253680 : values[i++] = ObjectIdGetDatum(stats->staop[k]); /* staopN */
1765 : }
1766 50736 : i = Anum_pg_statistic_stacoll1 - 1;
1767 304416 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1768 : {
1769 253680 : values[i++] = ObjectIdGetDatum(stats->stacoll[k]); /* stacollN */
1770 : }
1771 50736 : i = Anum_pg_statistic_stanumbers1 - 1;
1772 304416 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1773 : {
1774 253680 : if (stats->stanumbers[k] != NULL)
1775 : {
1776 78792 : int nnum = stats->numnumbers[k];
1777 78792 : Datum *numdatums = (Datum *) palloc(nnum * sizeof(Datum));
1778 : ArrayType *arry;
1779 :
1780 657563 : for (n = 0; n < nnum; n++)
1781 578771 : numdatums[n] = Float4GetDatum(stats->stanumbers[k][n]);
1782 78792 : arry = construct_array_builtin(numdatums, nnum, FLOAT4OID);
1783 78792 : values[i++] = PointerGetDatum(arry); /* stanumbersN */
1784 : }
1785 : else
1786 : {
1787 174888 : nulls[i] = true;
1788 174888 : values[i++] = (Datum) 0;
1789 : }
1790 : }
1791 50736 : i = Anum_pg_statistic_stavalues1 - 1;
1792 304416 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1793 : {
1794 253680 : if (stats->stavalues[k] != NULL)
1795 : {
1796 : ArrayType *arry;
1797 :
1798 55605 : arry = construct_array(stats->stavalues[k],
1799 : stats->numvalues[k],
1800 : stats->statypid[k],
1801 55605 : stats->statyplen[k],
1802 55605 : stats->statypbyval[k],
1803 55605 : stats->statypalign[k]);
1804 55605 : values[i++] = PointerGetDatum(arry); /* stavaluesN */
1805 : }
1806 : else
1807 : {
1808 198075 : nulls[i] = true;
1809 198075 : values[i++] = (Datum) 0;
1810 : }
1811 : }
1812 :
1813 : /* Is there already a pg_statistic tuple for this attribute? */
1814 101472 : oldtup = SearchSysCache3(STATRELATTINH,
1815 : ObjectIdGetDatum(relid),
1816 50736 : Int16GetDatum(stats->tupattnum),
1817 : BoolGetDatum(inh));
1818 :
1819 : /* Open index information when we know we need it */
1820 50736 : if (indstate == NULL)
1821 6818 : indstate = CatalogOpenIndexes(sd);
1822 :
1823 50736 : if (HeapTupleIsValid(oldtup))
1824 : {
1825 : /* Yes, replace it */
1826 22261 : stup = heap_modify_tuple(oldtup,
1827 : RelationGetDescr(sd),
1828 : values,
1829 : nulls,
1830 : replaces);
1831 22261 : ReleaseSysCache(oldtup);
1832 22261 : CatalogTupleUpdateWithInfo(sd, &stup->t_self, stup, indstate);
1833 : }
1834 : else
1835 : {
1836 : /* No, insert new tuple */
1837 28475 : stup = heap_form_tuple(RelationGetDescr(sd), values, nulls);
1838 28475 : CatalogTupleInsertWithInfo(sd, stup, indstate);
1839 : }
1840 :
1841 50736 : heap_freetuple(stup);
1842 : }
1843 :
1844 6822 : if (indstate != NULL)
1845 6818 : CatalogCloseIndexes(indstate);
1846 6822 : table_close(sd, RowExclusiveLock);
1847 : }
1848 :
1849 : /*
1850 : * Standard fetch function for use by compute_stats subroutines.
1851 : *
1852 : * This exists to provide some insulation between compute_stats routines
1853 : * and the actual storage of the sample data.
1854 : */
1855 : static Datum
1856 50006284 : std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1857 : {
1858 50006284 : int attnum = stats->tupattnum;
1859 50006284 : HeapTuple tuple = stats->rows[rownum];
1860 50006284 : TupleDesc tupDesc = stats->tupDesc;
1861 :
1862 50006284 : return heap_getattr(tuple, attnum, tupDesc, isNull);
1863 : }
1864 :
1865 : /*
1866 : * Fetch function for analyzing index expressions.
1867 : *
1868 : * We have not bothered to construct index tuples, instead the data is
1869 : * just in Datum arrays.
1870 : */
1871 : static Datum
1872 74041 : ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1873 : {
1874 : int i;
1875 :
1876 : /* exprvals and exprnulls are already offset for proper column */
1877 74041 : i = rownum * stats->rowstride;
1878 74041 : *isNull = stats->exprnulls[i];
1879 74041 : return stats->exprvals[i];
1880 : }
1881 :
1882 :
1883 : /*==========================================================================
1884 : *
1885 : * Code below this point represents the "standard" type-specific statistics
1886 : * analysis algorithms. This code can be replaced on a per-data-type basis
1887 : * by setting a nonzero value in pg_type.typanalyze.
1888 : *
1889 : *==========================================================================
1890 : */
1891 :
1892 :
1893 : /*
1894 : * To avoid consuming too much memory during analysis and/or too much space
1895 : * in the resulting pg_statistic rows, we ignore varlena datums that are wider
1896 : * than WIDTH_THRESHOLD (after detoasting!). This is legitimate for MCV
1897 : * and distinct-value calculations since a wide value is unlikely to be
1898 : * duplicated at all, much less be a most-common value. For the same reason,
1899 : * ignoring wide values will not affect our estimates of histogram bin
1900 : * boundaries very much.
1901 : */
1902 : #define WIDTH_THRESHOLD 1024
1903 :
1904 : #define swapInt(a,b) do {int _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1905 : #define swapDatum(a,b) do {Datum _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1906 :
1907 : /*
1908 : * Extra information used by the default analysis routines
1909 : */
1910 : typedef struct
1911 : {
1912 : int count; /* # of duplicates */
1913 : int first; /* values[] index of first occurrence */
1914 : } ScalarMCVItem;
1915 :
1916 : typedef struct
1917 : {
1918 : SortSupport ssup;
1919 : int *tupnoLink;
1920 : } CompareScalarsContext;
1921 :
1922 :
1923 : static void compute_trivial_stats(VacAttrStatsP stats,
1924 : AnalyzeAttrFetchFunc fetchfunc,
1925 : int samplerows,
1926 : double totalrows);
1927 : static void compute_distinct_stats(VacAttrStatsP stats,
1928 : AnalyzeAttrFetchFunc fetchfunc,
1929 : int samplerows,
1930 : double totalrows);
1931 : static void compute_scalar_stats(VacAttrStatsP stats,
1932 : AnalyzeAttrFetchFunc fetchfunc,
1933 : int samplerows,
1934 : double totalrows);
1935 : static int compare_scalars(const void *a, const void *b, void *arg);
1936 : static int compare_mcvs(const void *a, const void *b, void *arg);
1937 : static int analyze_mcv_list(int *mcv_counts,
1938 : int num_mcv,
1939 : double stadistinct,
1940 : double stanullfrac,
1941 : int samplerows,
1942 : double totalrows);
1943 :
1944 :
1945 : /*
1946 : * std_typanalyze -- the default type-specific typanalyze function
1947 : */
1948 : bool
1949 72848 : std_typanalyze(VacAttrStats *stats)
1950 : {
1951 : Oid ltopr;
1952 : Oid eqopr;
1953 : StdAnalyzeData *mystats;
1954 :
1955 : /* If the attstattarget column is negative, use the default value */
1956 72848 : if (stats->attstattarget < 0)
1957 72232 : stats->attstattarget = default_statistics_target;
1958 :
1959 : /* Look for default "<" and "=" operators for column's type */
1960 72848 : get_sort_group_operators(stats->attrtypid,
1961 : false, false, false,
1962 : <opr, &eqopr, NULL,
1963 : NULL);
1964 :
1965 : /* Save the operator info for compute_stats routines */
1966 72848 : mystats = palloc_object(StdAnalyzeData);
1967 72848 : mystats->eqopr = eqopr;
1968 72848 : mystats->eqfunc = OidIsValid(eqopr) ? get_opcode(eqopr) : InvalidOid;
1969 72848 : mystats->ltopr = ltopr;
1970 72848 : stats->extra_data = mystats;
1971 :
1972 : /*
1973 : * Determine which standard statistics algorithm to use
1974 : */
1975 72848 : if (OidIsValid(eqopr) && OidIsValid(ltopr))
1976 : {
1977 : /* Seems to be a scalar datatype */
1978 70771 : stats->compute_stats = compute_scalar_stats;
1979 : /*--------------------
1980 : * The following choice of minrows is based on the paper
1981 : * "Random sampling for histogram construction: how much is enough?"
1982 : * by Surajit Chaudhuri, Rajeev Motwani and Vivek Narasayya, in
1983 : * Proceedings of ACM SIGMOD International Conference on Management
1984 : * of Data, 1998, Pages 436-447. Their Corollary 1 to Theorem 5
1985 : * says that for table size n, histogram size k, maximum relative
1986 : * error in bin size f, and error probability gamma, the minimum
1987 : * random sample size is
1988 : * r = 4 * k * ln(2*n/gamma) / f^2
1989 : * Taking f = 0.5, gamma = 0.01, n = 10^6 rows, we obtain
1990 : * r = 305.82 * k
1991 : * Note that because of the log function, the dependence on n is
1992 : * quite weak; even at n = 10^12, a 300*k sample gives <= 0.66
1993 : * bin size error with probability 0.99. So there's no real need to
1994 : * scale for n, which is a good thing because we don't necessarily
1995 : * know it at this point.
1996 : *--------------------
1997 : */
1998 70771 : stats->minrows = 300 * stats->attstattarget;
1999 : }
2000 2077 : else if (OidIsValid(eqopr))
2001 : {
2002 : /* We can still recognize distinct values */
2003 1784 : stats->compute_stats = compute_distinct_stats;
2004 : /* Might as well use the same minrows as above */
2005 1784 : stats->minrows = 300 * stats->attstattarget;
2006 : }
2007 : else
2008 : {
2009 : /* Can't do much but the trivial stuff */
2010 293 : stats->compute_stats = compute_trivial_stats;
2011 : /* Might as well use the same minrows as above */
2012 293 : stats->minrows = 300 * stats->attstattarget;
2013 : }
2014 :
2015 72848 : return true;
2016 : }
2017 :
2018 :
2019 : /*
2020 : * compute_trivial_stats() -- compute very basic column statistics
2021 : *
2022 : * We use this when we cannot find a hash "=" operator for the datatype.
2023 : *
2024 : * We determine the fraction of non-null rows and the average datum width.
2025 : */
2026 : static void
2027 200 : compute_trivial_stats(VacAttrStatsP stats,
2028 : AnalyzeAttrFetchFunc fetchfunc,
2029 : int samplerows,
2030 : double totalrows)
2031 : {
2032 : int i;
2033 200 : int null_cnt = 0;
2034 200 : int nonnull_cnt = 0;
2035 200 : double total_width = 0;
2036 400 : bool is_varlena = (!stats->attrtype->typbyval &&
2037 200 : stats->attrtype->typlen == -1);
2038 400 : bool is_varwidth = (!stats->attrtype->typbyval &&
2039 200 : stats->attrtype->typlen < 0);
2040 :
2041 636051 : for (i = 0; i < samplerows; i++)
2042 : {
2043 : Datum value;
2044 : bool isnull;
2045 :
2046 635851 : vacuum_delay_point(true);
2047 :
2048 635851 : value = fetchfunc(stats, i, &isnull);
2049 :
2050 : /* Check for null/nonnull */
2051 635851 : if (isnull)
2052 : {
2053 393123 : null_cnt++;
2054 393123 : continue;
2055 : }
2056 242728 : nonnull_cnt++;
2057 :
2058 : /*
2059 : * If it's a variable-width field, add up widths for average width
2060 : * calculation. Note that if the value is toasted, we use the toasted
2061 : * width. We don't bother with this calculation if it's a fixed-width
2062 : * type.
2063 : */
2064 242728 : if (is_varlena)
2065 : {
2066 45668 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2067 : }
2068 197060 : else if (is_varwidth)
2069 : {
2070 : /* must be cstring */
2071 0 : total_width += strlen(DatumGetCString(value)) + 1;
2072 : }
2073 : }
2074 :
2075 : /* We can only compute average width if we found some non-null values. */
2076 200 : if (nonnull_cnt > 0)
2077 : {
2078 95 : stats->stats_valid = true;
2079 : /* Do the simple null-frac and width stats */
2080 95 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2081 95 : if (is_varwidth)
2082 46 : stats->stawidth = total_width / (double) nonnull_cnt;
2083 : else
2084 49 : stats->stawidth = stats->attrtype->typlen;
2085 95 : stats->stadistinct = 0.0; /* "unknown" */
2086 : }
2087 105 : else if (null_cnt > 0)
2088 : {
2089 : /* We found only nulls; assume the column is entirely null */
2090 105 : stats->stats_valid = true;
2091 105 : stats->stanullfrac = 1.0;
2092 105 : if (is_varwidth)
2093 105 : stats->stawidth = 0; /* "unknown" */
2094 : else
2095 0 : stats->stawidth = stats->attrtype->typlen;
2096 105 : stats->stadistinct = 0.0; /* "unknown" */
2097 : }
2098 200 : }
2099 :
2100 :
2101 : /*
2102 : * compute_distinct_stats() -- compute column statistics including ndistinct
2103 : *
2104 : * We use this when we can find only an "=" operator for the datatype.
2105 : *
2106 : * We determine the fraction of non-null rows, the average width, the
2107 : * most common values, and the (estimated) number of distinct values.
2108 : *
2109 : * The most common values are determined by brute force: we keep a list
2110 : * of previously seen values, ordered by number of times seen, as we scan
2111 : * the samples. A newly seen value is inserted just after the last
2112 : * multiply-seen value, causing the bottommost (oldest) singly-seen value
2113 : * to drop off the list. The accuracy of this method, and also its cost,
2114 : * depend mainly on the length of the list we are willing to keep.
2115 : */
2116 : static void
2117 1302 : compute_distinct_stats(VacAttrStatsP stats,
2118 : AnalyzeAttrFetchFunc fetchfunc,
2119 : int samplerows,
2120 : double totalrows)
2121 : {
2122 : int i;
2123 1302 : int null_cnt = 0;
2124 1302 : int nonnull_cnt = 0;
2125 1302 : int toowide_cnt = 0;
2126 1302 : double total_width = 0;
2127 2198 : bool is_varlena = (!stats->attrtype->typbyval &&
2128 896 : stats->attrtype->typlen == -1);
2129 2198 : bool is_varwidth = (!stats->attrtype->typbyval &&
2130 896 : stats->attrtype->typlen < 0);
2131 : FmgrInfo f_cmpeq;
2132 : typedef struct
2133 : {
2134 : Datum value;
2135 : int count;
2136 : } TrackItem;
2137 : TrackItem *track;
2138 : int track_cnt,
2139 : track_max;
2140 1302 : int num_mcv = stats->attstattarget;
2141 1302 : StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2142 :
2143 : /*
2144 : * We track up to 2*n values for an n-element MCV list; but at least 10
2145 : */
2146 1302 : track_max = 2 * num_mcv;
2147 1302 : if (track_max < 10)
2148 39 : track_max = 10;
2149 1302 : track = (TrackItem *) palloc(track_max * sizeof(TrackItem));
2150 1302 : track_cnt = 0;
2151 :
2152 1302 : fmgr_info(mystats->eqfunc, &f_cmpeq);
2153 :
2154 968771 : for (i = 0; i < samplerows; i++)
2155 : {
2156 : Datum value;
2157 : bool isnull;
2158 : bool match;
2159 : int firstcount1,
2160 : j;
2161 :
2162 967469 : vacuum_delay_point(true);
2163 :
2164 967469 : value = fetchfunc(stats, i, &isnull);
2165 :
2166 : /* Check for null/nonnull */
2167 967469 : if (isnull)
2168 : {
2169 802736 : null_cnt++;
2170 802736 : continue;
2171 : }
2172 164733 : nonnull_cnt++;
2173 :
2174 : /*
2175 : * If it's a variable-width field, add up widths for average width
2176 : * calculation. Note that if the value is toasted, we use the toasted
2177 : * width. We don't bother with this calculation if it's a fixed-width
2178 : * type.
2179 : */
2180 164733 : if (is_varlena)
2181 : {
2182 58689 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2183 :
2184 : /*
2185 : * If the value is toasted, we want to detoast it just once to
2186 : * avoid repeated detoastings and resultant excess memory usage
2187 : * during the comparisons. Also, check to see if the value is
2188 : * excessively wide, and if so don't detoast at all --- just
2189 : * ignore the value.
2190 : */
2191 58689 : if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
2192 : {
2193 0 : toowide_cnt++;
2194 0 : continue;
2195 : }
2196 58689 : value = PointerGetDatum(PG_DETOAST_DATUM(value));
2197 : }
2198 106044 : else if (is_varwidth)
2199 : {
2200 : /* must be cstring */
2201 0 : total_width += strlen(DatumGetCString(value)) + 1;
2202 : }
2203 :
2204 : /*
2205 : * See if the value matches anything we're already tracking.
2206 : */
2207 164733 : match = false;
2208 164733 : firstcount1 = track_cnt;
2209 402529 : for (j = 0; j < track_cnt; j++)
2210 : {
2211 398277 : if (DatumGetBool(FunctionCall2Coll(&f_cmpeq,
2212 : stats->attrcollid,
2213 398277 : value, track[j].value)))
2214 : {
2215 160481 : match = true;
2216 160481 : break;
2217 : }
2218 237796 : if (j < firstcount1 && track[j].count == 1)
2219 2819 : firstcount1 = j;
2220 : }
2221 :
2222 164733 : if (match)
2223 : {
2224 : /* Found a match */
2225 160481 : track[j].count++;
2226 : /* This value may now need to "bubble up" in the track list */
2227 167694 : while (j > 0 && track[j].count > track[j - 1].count)
2228 : {
2229 7213 : swapDatum(track[j].value, track[j - 1].value);
2230 7213 : swapInt(track[j].count, track[j - 1].count);
2231 7213 : j--;
2232 : }
2233 : }
2234 : else
2235 : {
2236 : /* No match. Insert at head of count-1 list */
2237 4252 : if (track_cnt < track_max)
2238 3869 : track_cnt++;
2239 84303 : for (j = track_cnt - 1; j > firstcount1; j--)
2240 : {
2241 80051 : track[j].value = track[j - 1].value;
2242 80051 : track[j].count = track[j - 1].count;
2243 : }
2244 4252 : if (firstcount1 < track_cnt)
2245 : {
2246 4252 : track[firstcount1].value = value;
2247 4252 : track[firstcount1].count = 1;
2248 : }
2249 : }
2250 : }
2251 :
2252 : /* We can only compute real stats if we found some non-null values. */
2253 1302 : if (nonnull_cnt > 0)
2254 : {
2255 : int nmultiple,
2256 : summultiple;
2257 :
2258 952 : stats->stats_valid = true;
2259 : /* Do the simple null-frac and width stats */
2260 952 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2261 952 : if (is_varwidth)
2262 546 : stats->stawidth = total_width / (double) nonnull_cnt;
2263 : else
2264 406 : stats->stawidth = stats->attrtype->typlen;
2265 :
2266 : /* Count the number of values we found multiple times */
2267 952 : summultiple = 0;
2268 3743 : for (nmultiple = 0; nmultiple < track_cnt; nmultiple++)
2269 : {
2270 3277 : if (track[nmultiple].count == 1)
2271 486 : break;
2272 2791 : summultiple += track[nmultiple].count;
2273 : }
2274 :
2275 952 : if (nmultiple == 0)
2276 : {
2277 : /*
2278 : * If we found no repeated non-null values, assume it's a unique
2279 : * column; but be sure to discount for any nulls we found.
2280 : */
2281 112 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2282 : }
2283 840 : else if (track_cnt < track_max && toowide_cnt == 0 &&
2284 : nmultiple == track_cnt)
2285 : {
2286 : /*
2287 : * Our track list includes every value in the sample, and every
2288 : * value appeared more than once. Assume the column has just
2289 : * these values. (This case is meant to address columns with
2290 : * small, fixed sets of possible values, such as boolean or enum
2291 : * columns. If there are any values that appear just once in the
2292 : * sample, including too-wide values, we should assume that that's
2293 : * not what we're dealing with.)
2294 : */
2295 466 : stats->stadistinct = track_cnt;
2296 : }
2297 : else
2298 : {
2299 : /*----------
2300 : * Estimate the number of distinct values using the estimator
2301 : * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2302 : * n*d / (n - f1 + f1*n/N)
2303 : * where f1 is the number of distinct values that occurred
2304 : * exactly once in our sample of n rows (from a total of N),
2305 : * and d is the total number of distinct values in the sample.
2306 : * This is their Duj1 estimator; the other estimators they
2307 : * recommend are considerably more complex, and are numerically
2308 : * very unstable when n is much smaller than N.
2309 : *
2310 : * In this calculation, we consider only non-nulls. We used to
2311 : * include rows with null values in the n and N counts, but that
2312 : * leads to inaccurate answers in columns with many nulls, and
2313 : * it's intuitively bogus anyway considering the desired result is
2314 : * the number of distinct non-null values.
2315 : *
2316 : * We assume (not very reliably!) that all the multiply-occurring
2317 : * values are reflected in the final track[] list, and the other
2318 : * nonnull values all appeared but once. (XXX this usually
2319 : * results in a drastic overestimate of ndistinct. Can we do
2320 : * any better?)
2321 : *----------
2322 : */
2323 374 : int f1 = nonnull_cnt - summultiple;
2324 374 : int d = f1 + nmultiple;
2325 374 : double n = samplerows - null_cnt;
2326 374 : double N = totalrows * (1.0 - stats->stanullfrac);
2327 : double stadistinct;
2328 :
2329 : /* N == 0 shouldn't happen, but just in case ... */
2330 374 : if (N > 0)
2331 374 : stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2332 : else
2333 0 : stadistinct = 0;
2334 :
2335 : /* Clamp to sane range in case of roundoff error */
2336 374 : if (stadistinct < d)
2337 3 : stadistinct = d;
2338 374 : if (stadistinct > N)
2339 0 : stadistinct = N;
2340 : /* And round to integer */
2341 374 : stats->stadistinct = floor(stadistinct + 0.5);
2342 : }
2343 :
2344 : /*
2345 : * If we estimated the number of distinct values at more than 10% of
2346 : * the total row count (a very arbitrary limit), then assume that
2347 : * stadistinct should scale with the row count rather than be a fixed
2348 : * value.
2349 : */
2350 952 : if (stats->stadistinct > 0.1 * totalrows)
2351 213 : stats->stadistinct = -(stats->stadistinct / totalrows);
2352 :
2353 : /*
2354 : * Decide how many values are worth storing as most-common values. If
2355 : * we are able to generate a complete MCV list (all the values in the
2356 : * sample will fit, and we think these are all the ones in the table),
2357 : * then do so. Otherwise, store only those values that are
2358 : * significantly more common than the values not in the list.
2359 : *
2360 : * Note: the first of these cases is meant to address columns with
2361 : * small, fixed sets of possible values, such as boolean or enum
2362 : * columns. If we can *completely* represent the column population by
2363 : * an MCV list that will fit into the stats target, then we should do
2364 : * so and thus provide the planner with complete information. But if
2365 : * the MCV list is not complete, it's generally worth being more
2366 : * selective, and not just filling it all the way up to the stats
2367 : * target.
2368 : */
2369 952 : if (track_cnt < track_max && toowide_cnt == 0 &&
2370 947 : stats->stadistinct > 0 &&
2371 : track_cnt <= num_mcv)
2372 : {
2373 : /* Track list includes all values seen, and all will fit */
2374 613 : num_mcv = track_cnt;
2375 : }
2376 : else
2377 : {
2378 : int *mcv_counts;
2379 :
2380 : /* Incomplete list; decide how many values are worth keeping */
2381 339 : if (num_mcv > track_cnt)
2382 309 : num_mcv = track_cnt;
2383 :
2384 339 : if (num_mcv > 0)
2385 : {
2386 339 : mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2387 1081 : for (i = 0; i < num_mcv; i++)
2388 742 : mcv_counts[i] = track[i].count;
2389 :
2390 339 : num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2391 339 : stats->stadistinct,
2392 339 : stats->stanullfrac,
2393 : samplerows, totalrows);
2394 : }
2395 : }
2396 :
2397 : /* Generate MCV slot entry */
2398 952 : if (num_mcv > 0)
2399 : {
2400 : MemoryContext old_context;
2401 : Datum *mcv_values;
2402 : float4 *mcv_freqs;
2403 :
2404 : /* Must copy the target values into anl_context */
2405 949 : old_context = MemoryContextSwitchTo(stats->anl_context);
2406 949 : mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2407 949 : mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2408 4484 : for (i = 0; i < num_mcv; i++)
2409 : {
2410 7070 : mcv_values[i] = datumCopy(track[i].value,
2411 3535 : stats->attrtype->typbyval,
2412 3535 : stats->attrtype->typlen);
2413 3535 : mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2414 : }
2415 949 : MemoryContextSwitchTo(old_context);
2416 :
2417 949 : stats->stakind[0] = STATISTIC_KIND_MCV;
2418 949 : stats->staop[0] = mystats->eqopr;
2419 949 : stats->stacoll[0] = stats->attrcollid;
2420 949 : stats->stanumbers[0] = mcv_freqs;
2421 949 : stats->numnumbers[0] = num_mcv;
2422 949 : stats->stavalues[0] = mcv_values;
2423 949 : stats->numvalues[0] = num_mcv;
2424 :
2425 : /*
2426 : * Accept the defaults for stats->statypid and others. They have
2427 : * been set before we were called (see vacuum.h)
2428 : */
2429 : }
2430 : }
2431 350 : else if (null_cnt > 0)
2432 : {
2433 : /* We found only nulls; assume the column is entirely null */
2434 350 : stats->stats_valid = true;
2435 350 : stats->stanullfrac = 1.0;
2436 350 : if (is_varwidth)
2437 350 : stats->stawidth = 0; /* "unknown" */
2438 : else
2439 0 : stats->stawidth = stats->attrtype->typlen;
2440 350 : stats->stadistinct = 0.0; /* "unknown" */
2441 : }
2442 :
2443 : /* We don't need to bother cleaning up any of our temporary palloc's */
2444 1302 : }
2445 :
2446 :
2447 : /*
2448 : * compute_scalar_stats() -- compute column statistics
2449 : *
2450 : * We use this when we can find "=" and "<" operators for the datatype.
2451 : *
2452 : * We determine the fraction of non-null rows, the average width, the
2453 : * most common values, the (estimated) number of distinct values, the
2454 : * distribution histogram, and the correlation of physical to logical order.
2455 : *
2456 : * The desired stats can be determined fairly easily after sorting the
2457 : * data values into order.
2458 : */
2459 : static void
2460 49500 : compute_scalar_stats(VacAttrStatsP stats,
2461 : AnalyzeAttrFetchFunc fetchfunc,
2462 : int samplerows,
2463 : double totalrows)
2464 : {
2465 : int i;
2466 49500 : int null_cnt = 0;
2467 49500 : int nonnull_cnt = 0;
2468 49500 : int toowide_cnt = 0;
2469 49500 : double total_width = 0;
2470 61569 : bool is_varlena = (!stats->attrtype->typbyval &&
2471 12069 : stats->attrtype->typlen == -1);
2472 61569 : bool is_varwidth = (!stats->attrtype->typbyval &&
2473 12069 : stats->attrtype->typlen < 0);
2474 : double corr_xysum;
2475 : SortSupportData ssup;
2476 : ScalarItem *values;
2477 49500 : int values_cnt = 0;
2478 : int *tupnoLink;
2479 : ScalarMCVItem *track;
2480 49500 : int track_cnt = 0;
2481 49500 : int num_mcv = stats->attstattarget;
2482 49500 : int num_bins = stats->attstattarget;
2483 49500 : StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2484 :
2485 49500 : values = (ScalarItem *) palloc(samplerows * sizeof(ScalarItem));
2486 49500 : tupnoLink = (int *) palloc(samplerows * sizeof(int));
2487 49500 : track = (ScalarMCVItem *) palloc(num_mcv * sizeof(ScalarMCVItem));
2488 :
2489 49500 : memset(&ssup, 0, sizeof(ssup));
2490 49500 : ssup.ssup_cxt = CurrentMemoryContext;
2491 49500 : ssup.ssup_collation = stats->attrcollid;
2492 49500 : ssup.ssup_nulls_first = false;
2493 :
2494 : /*
2495 : * For now, don't perform abbreviated key conversion, because full values
2496 : * are required for MCV slot generation. Supporting that optimization
2497 : * would necessitate teaching compare_scalars() to call a tie-breaker.
2498 : */
2499 49500 : ssup.abbreviate = false;
2500 :
2501 49500 : PrepareSortSupportFromOrderingOp(mystats->ltopr, &ssup);
2502 :
2503 : /* Initial scan to find sortable values */
2504 45766988 : for (i = 0; i < samplerows; i++)
2505 : {
2506 : Datum value;
2507 : bool isnull;
2508 :
2509 45717488 : vacuum_delay_point(true);
2510 :
2511 45717488 : value = fetchfunc(stats, i, &isnull);
2512 :
2513 : /* Check for null/nonnull */
2514 45717488 : if (isnull)
2515 : {
2516 5941964 : null_cnt++;
2517 5964337 : continue;
2518 : }
2519 39775524 : nonnull_cnt++;
2520 :
2521 : /*
2522 : * If it's a variable-width field, add up widths for average width
2523 : * calculation. Note that if the value is toasted, we use the toasted
2524 : * width. We don't bother with this calculation if it's a fixed-width
2525 : * type.
2526 : */
2527 39775524 : if (is_varlena)
2528 : {
2529 4837530 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2530 :
2531 : /*
2532 : * If the value is toasted, we want to detoast it just once to
2533 : * avoid repeated detoastings and resultant excess memory usage
2534 : * during the comparisons. Also, check to see if the value is
2535 : * excessively wide, and if so don't detoast at all --- just
2536 : * ignore the value.
2537 : */
2538 4837530 : if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
2539 : {
2540 22373 : toowide_cnt++;
2541 22373 : continue;
2542 : }
2543 4815157 : value = PointerGetDatum(PG_DETOAST_DATUM(value));
2544 : }
2545 34937994 : else if (is_varwidth)
2546 : {
2547 : /* must be cstring */
2548 0 : total_width += strlen(DatumGetCString(value)) + 1;
2549 : }
2550 :
2551 : /* Add it to the list to be sorted */
2552 39753151 : values[values_cnt].value = value;
2553 39753151 : values[values_cnt].tupno = values_cnt;
2554 39753151 : tupnoLink[values_cnt] = values_cnt;
2555 39753151 : values_cnt++;
2556 : }
2557 :
2558 : /* We can only compute real stats if we found some sortable values. */
2559 49500 : if (values_cnt > 0)
2560 : {
2561 : int ndistinct, /* # distinct values in sample */
2562 : nmultiple, /* # that appear multiple times */
2563 : num_hist,
2564 : dups_cnt;
2565 46274 : int slot_idx = 0;
2566 : CompareScalarsContext cxt;
2567 :
2568 : /* Sort the collected values */
2569 46274 : cxt.ssup = &ssup;
2570 46274 : cxt.tupnoLink = tupnoLink;
2571 46274 : qsort_interruptible(values, values_cnt, sizeof(ScalarItem),
2572 : compare_scalars, &cxt);
2573 :
2574 : /*
2575 : * Now scan the values in order, find the most common ones, and also
2576 : * accumulate ordering-correlation statistics.
2577 : *
2578 : * To determine which are most common, we first have to count the
2579 : * number of duplicates of each value. The duplicates are adjacent in
2580 : * the sorted list, so a brute-force approach is to compare successive
2581 : * datum values until we find two that are not equal. However, that
2582 : * requires N-1 invocations of the datum comparison routine, which are
2583 : * completely redundant with work that was done during the sort. (The
2584 : * sort algorithm must at some point have compared each pair of items
2585 : * that are adjacent in the sorted order; otherwise it could not know
2586 : * that it's ordered the pair correctly.) We exploit this by having
2587 : * compare_scalars remember the highest tupno index that each
2588 : * ScalarItem has been found equal to. At the end of the sort, a
2589 : * ScalarItem's tupnoLink will still point to itself if and only if it
2590 : * is the last item of its group of duplicates (since the group will
2591 : * be ordered by tupno).
2592 : */
2593 46274 : corr_xysum = 0;
2594 46274 : ndistinct = 0;
2595 46274 : nmultiple = 0;
2596 46274 : dups_cnt = 0;
2597 39799425 : for (i = 0; i < values_cnt; i++)
2598 : {
2599 39753151 : int tupno = values[i].tupno;
2600 :
2601 39753151 : corr_xysum += ((double) i) * ((double) tupno);
2602 39753151 : dups_cnt++;
2603 39753151 : if (tupnoLink[tupno] == tupno)
2604 : {
2605 : /* Reached end of duplicates of this value */
2606 8015846 : ndistinct++;
2607 8015846 : if (dups_cnt > 1)
2608 : {
2609 688664 : nmultiple++;
2610 688664 : if (track_cnt < num_mcv ||
2611 283500 : dups_cnt > track[track_cnt - 1].count)
2612 : {
2613 : /*
2614 : * Found a new item for the mcv list; find its
2615 : * position, bubbling down old items if needed. Loop
2616 : * invariant is that j points at an empty/ replaceable
2617 : * slot.
2618 : */
2619 : int j;
2620 :
2621 466961 : if (track_cnt < num_mcv)
2622 405164 : track_cnt++;
2623 6162745 : for (j = track_cnt - 1; j > 0; j--)
2624 : {
2625 6112249 : if (dups_cnt <= track[j - 1].count)
2626 416465 : break;
2627 5695784 : track[j].count = track[j - 1].count;
2628 5695784 : track[j].first = track[j - 1].first;
2629 : }
2630 466961 : track[j].count = dups_cnt;
2631 466961 : track[j].first = i + 1 - dups_cnt;
2632 : }
2633 : }
2634 8015846 : dups_cnt = 0;
2635 : }
2636 : }
2637 :
2638 46274 : stats->stats_valid = true;
2639 : /* Do the simple null-frac and width stats */
2640 46274 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2641 46274 : if (is_varwidth)
2642 6842 : stats->stawidth = total_width / (double) nonnull_cnt;
2643 : else
2644 39432 : stats->stawidth = stats->attrtype->typlen;
2645 :
2646 46274 : if (nmultiple == 0)
2647 : {
2648 : /*
2649 : * If we found no repeated non-null values, assume it's a unique
2650 : * column; but be sure to discount for any nulls we found.
2651 : */
2652 12691 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2653 : }
2654 33583 : else if (toowide_cnt == 0 && nmultiple == ndistinct)
2655 : {
2656 : /*
2657 : * Every value in the sample appeared more than once. Assume the
2658 : * column has just these values. (This case is meant to address
2659 : * columns with small, fixed sets of possible values, such as
2660 : * boolean or enum columns. If there are any values that appear
2661 : * just once in the sample, including too-wide values, we should
2662 : * assume that that's not what we're dealing with.)
2663 : */
2664 20787 : stats->stadistinct = ndistinct;
2665 : }
2666 : else
2667 : {
2668 : /*----------
2669 : * Estimate the number of distinct values using the estimator
2670 : * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2671 : * n*d / (n - f1 + f1*n/N)
2672 : * where f1 is the number of distinct values that occurred
2673 : * exactly once in our sample of n rows (from a total of N),
2674 : * and d is the total number of distinct values in the sample.
2675 : * This is their Duj1 estimator; the other estimators they
2676 : * recommend are considerably more complex, and are numerically
2677 : * very unstable when n is much smaller than N.
2678 : *
2679 : * In this calculation, we consider only non-nulls. We used to
2680 : * include rows with null values in the n and N counts, but that
2681 : * leads to inaccurate answers in columns with many nulls, and
2682 : * it's intuitively bogus anyway considering the desired result is
2683 : * the number of distinct non-null values.
2684 : *
2685 : * Overwidth values are assumed to have been distinct.
2686 : *----------
2687 : */
2688 12796 : int f1 = ndistinct - nmultiple + toowide_cnt;
2689 12796 : int d = f1 + nmultiple;
2690 12796 : double n = samplerows - null_cnt;
2691 12796 : double N = totalrows * (1.0 - stats->stanullfrac);
2692 : double stadistinct;
2693 :
2694 : /* N == 0 shouldn't happen, but just in case ... */
2695 12796 : if (N > 0)
2696 12796 : stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2697 : else
2698 0 : stadistinct = 0;
2699 :
2700 : /* Clamp to sane range in case of roundoff error */
2701 12796 : if (stadistinct < d)
2702 549 : stadistinct = d;
2703 12796 : if (stadistinct > N)
2704 0 : stadistinct = N;
2705 : /* And round to integer */
2706 12796 : stats->stadistinct = floor(stadistinct + 0.5);
2707 : }
2708 :
2709 : /*
2710 : * If we estimated the number of distinct values at more than 10% of
2711 : * the total row count (a very arbitrary limit), then assume that
2712 : * stadistinct should scale with the row count rather than be a fixed
2713 : * value.
2714 : */
2715 46274 : if (stats->stadistinct > 0.1 * totalrows)
2716 9569 : stats->stadistinct = -(stats->stadistinct / totalrows);
2717 :
2718 : /*
2719 : * Decide how many values are worth storing as most-common values. If
2720 : * we are able to generate a complete MCV list (all the values in the
2721 : * sample will fit, and we think these are all the ones in the table),
2722 : * then do so. Otherwise, store only those values that are
2723 : * significantly more common than the values not in the list.
2724 : *
2725 : * Note: the first of these cases is meant to address columns with
2726 : * small, fixed sets of possible values, such as boolean or enum
2727 : * columns. If we can *completely* represent the column population by
2728 : * an MCV list that will fit into the stats target, then we should do
2729 : * so and thus provide the planner with complete information. But if
2730 : * the MCV list is not complete, it's generally worth being more
2731 : * selective, and not just filling it all the way up to the stats
2732 : * target.
2733 : */
2734 46274 : if (track_cnt == ndistinct && toowide_cnt == 0 &&
2735 20406 : stats->stadistinct > 0 &&
2736 : track_cnt <= num_mcv)
2737 : {
2738 : /* Track list includes all values seen, and all will fit */
2739 18257 : num_mcv = track_cnt;
2740 : }
2741 : else
2742 : {
2743 : int *mcv_counts;
2744 :
2745 : /* Incomplete list; decide how many values are worth keeping */
2746 28017 : if (num_mcv > track_cnt)
2747 25836 : num_mcv = track_cnt;
2748 :
2749 28017 : if (num_mcv > 0)
2750 : {
2751 15326 : mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2752 318764 : for (i = 0; i < num_mcv; i++)
2753 303438 : mcv_counts[i] = track[i].count;
2754 :
2755 15326 : num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2756 15326 : stats->stadistinct,
2757 15326 : stats->stanullfrac,
2758 : samplerows, totalrows);
2759 : }
2760 : }
2761 :
2762 : /* Generate MCV slot entry */
2763 46274 : if (num_mcv > 0)
2764 : {
2765 : MemoryContext old_context;
2766 : Datum *mcv_values;
2767 : float4 *mcv_freqs;
2768 :
2769 : /* Must copy the target values into anl_context */
2770 33561 : old_context = MemoryContextSwitchTo(stats->anl_context);
2771 33561 : mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2772 33561 : mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2773 438612 : for (i = 0; i < num_mcv; i++)
2774 : {
2775 810102 : mcv_values[i] = datumCopy(values[track[i].first].value,
2776 405051 : stats->attrtype->typbyval,
2777 405051 : stats->attrtype->typlen);
2778 405051 : mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2779 : }
2780 33561 : MemoryContextSwitchTo(old_context);
2781 :
2782 33561 : stats->stakind[slot_idx] = STATISTIC_KIND_MCV;
2783 33561 : stats->staop[slot_idx] = mystats->eqopr;
2784 33561 : stats->stacoll[slot_idx] = stats->attrcollid;
2785 33561 : stats->stanumbers[slot_idx] = mcv_freqs;
2786 33561 : stats->numnumbers[slot_idx] = num_mcv;
2787 33561 : stats->stavalues[slot_idx] = mcv_values;
2788 33561 : stats->numvalues[slot_idx] = num_mcv;
2789 :
2790 : /*
2791 : * Accept the defaults for stats->statypid and others. They have
2792 : * been set before we were called (see vacuum.h)
2793 : */
2794 33561 : slot_idx++;
2795 : }
2796 :
2797 : /*
2798 : * Generate a histogram slot entry if there are at least two distinct
2799 : * values not accounted for in the MCV list. (This ensures the
2800 : * histogram won't collapse to empty or a singleton.)
2801 : */
2802 46274 : num_hist = ndistinct - num_mcv;
2803 46274 : if (num_hist > num_bins)
2804 7266 : num_hist = num_bins + 1;
2805 46274 : if (num_hist >= 2)
2806 : {
2807 : MemoryContext old_context;
2808 : Datum *hist_values;
2809 : int nvals;
2810 : int pos,
2811 : posfrac,
2812 : delta,
2813 : deltafrac;
2814 :
2815 : /* Sort the MCV items into position order to speed next loop */
2816 20746 : qsort_interruptible(track, num_mcv, sizeof(ScalarMCVItem),
2817 : compare_mcvs, NULL);
2818 :
2819 : /*
2820 : * Collapse out the MCV items from the values[] array.
2821 : *
2822 : * Note we destroy the values[] array here... but we don't need it
2823 : * for anything more. We do, however, still need values_cnt.
2824 : * nvals will be the number of remaining entries in values[].
2825 : */
2826 20746 : if (num_mcv > 0)
2827 : {
2828 : int src,
2829 : dest;
2830 : int j;
2831 :
2832 10707 : src = dest = 0;
2833 10707 : j = 0; /* index of next interesting MCV item */
2834 391525 : while (src < values_cnt)
2835 : {
2836 : int ncopy;
2837 :
2838 380818 : if (j < num_mcv)
2839 : {
2840 372859 : int first = track[j].first;
2841 :
2842 372859 : if (src >= first)
2843 : {
2844 : /* advance past this MCV item */
2845 271290 : src = first + track[j].count;
2846 271290 : j++;
2847 271290 : continue;
2848 : }
2849 101569 : ncopy = first - src;
2850 : }
2851 : else
2852 7959 : ncopy = values_cnt - src;
2853 109528 : memmove(&values[dest], &values[src],
2854 : ncopy * sizeof(ScalarItem));
2855 109528 : src += ncopy;
2856 109528 : dest += ncopy;
2857 : }
2858 10707 : nvals = dest;
2859 : }
2860 : else
2861 10039 : nvals = values_cnt;
2862 : Assert(nvals >= num_hist);
2863 :
2864 : /* Must copy the target values into anl_context */
2865 20746 : old_context = MemoryContextSwitchTo(stats->anl_context);
2866 20746 : hist_values = (Datum *) palloc(num_hist * sizeof(Datum));
2867 :
2868 : /*
2869 : * The object of this loop is to copy the first and last values[]
2870 : * entries along with evenly-spaced values in between. So the
2871 : * i'th value is values[(i * (nvals - 1)) / (num_hist - 1)]. But
2872 : * computing that subscript directly risks integer overflow when
2873 : * the stats target is more than a couple thousand. Instead we
2874 : * add (nvals - 1) / (num_hist - 1) to pos at each step, tracking
2875 : * the integral and fractional parts of the sum separately.
2876 : */
2877 20746 : delta = (nvals - 1) / (num_hist - 1);
2878 20746 : deltafrac = (nvals - 1) % (num_hist - 1);
2879 20746 : pos = posfrac = 0;
2880 :
2881 1092515 : for (i = 0; i < num_hist; i++)
2882 : {
2883 2143538 : hist_values[i] = datumCopy(values[pos].value,
2884 1071769 : stats->attrtype->typbyval,
2885 1071769 : stats->attrtype->typlen);
2886 1071769 : pos += delta;
2887 1071769 : posfrac += deltafrac;
2888 1071769 : if (posfrac >= (num_hist - 1))
2889 : {
2890 : /* fractional part exceeds 1, carry to integer part */
2891 353522 : pos++;
2892 353522 : posfrac -= (num_hist - 1);
2893 : }
2894 : }
2895 :
2896 20746 : MemoryContextSwitchTo(old_context);
2897 :
2898 20746 : stats->stakind[slot_idx] = STATISTIC_KIND_HISTOGRAM;
2899 20746 : stats->staop[slot_idx] = mystats->ltopr;
2900 20746 : stats->stacoll[slot_idx] = stats->attrcollid;
2901 20746 : stats->stavalues[slot_idx] = hist_values;
2902 20746 : stats->numvalues[slot_idx] = num_hist;
2903 :
2904 : /*
2905 : * Accept the defaults for stats->statypid and others. They have
2906 : * been set before we were called (see vacuum.h)
2907 : */
2908 20746 : slot_idx++;
2909 : }
2910 :
2911 : /* Generate a correlation entry if there are multiple values */
2912 46274 : if (values_cnt > 1)
2913 : {
2914 : MemoryContext old_context;
2915 : float4 *corrs;
2916 : double corr_xsum,
2917 : corr_x2sum;
2918 :
2919 : /* Must copy the target values into anl_context */
2920 43600 : old_context = MemoryContextSwitchTo(stats->anl_context);
2921 43600 : corrs = palloc_object(float4);
2922 43600 : MemoryContextSwitchTo(old_context);
2923 :
2924 : /*----------
2925 : * Since we know the x and y value sets are both
2926 : * 0, 1, ..., values_cnt-1
2927 : * we have sum(x) = sum(y) =
2928 : * (values_cnt-1)*values_cnt / 2
2929 : * and sum(x^2) = sum(y^2) =
2930 : * (values_cnt-1)*values_cnt*(2*values_cnt-1) / 6.
2931 : *----------
2932 : */
2933 43600 : corr_xsum = ((double) (values_cnt - 1)) *
2934 43600 : ((double) values_cnt) / 2.0;
2935 43600 : corr_x2sum = ((double) (values_cnt - 1)) *
2936 43600 : ((double) values_cnt) * (double) (2 * values_cnt - 1) / 6.0;
2937 :
2938 : /* And the correlation coefficient reduces to */
2939 43600 : corrs[0] = (values_cnt * corr_xysum - corr_xsum * corr_xsum) /
2940 43600 : (values_cnt * corr_x2sum - corr_xsum * corr_xsum);
2941 :
2942 43600 : stats->stakind[slot_idx] = STATISTIC_KIND_CORRELATION;
2943 43600 : stats->staop[slot_idx] = mystats->ltopr;
2944 43600 : stats->stacoll[slot_idx] = stats->attrcollid;
2945 43600 : stats->stanumbers[slot_idx] = corrs;
2946 43600 : stats->numnumbers[slot_idx] = 1;
2947 43600 : slot_idx++;
2948 : }
2949 : }
2950 3226 : else if (nonnull_cnt > 0)
2951 : {
2952 : /* We found some non-null values, but they were all too wide */
2953 : Assert(nonnull_cnt == toowide_cnt);
2954 206 : stats->stats_valid = true;
2955 : /* Do the simple null-frac and width stats */
2956 206 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2957 206 : if (is_varwidth)
2958 206 : stats->stawidth = total_width / (double) nonnull_cnt;
2959 : else
2960 0 : stats->stawidth = stats->attrtype->typlen;
2961 : /* Assume all too-wide values are distinct, so it's a unique column */
2962 206 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2963 : }
2964 3020 : else if (null_cnt > 0)
2965 : {
2966 : /* We found only nulls; assume the column is entirely null */
2967 3020 : stats->stats_valid = true;
2968 3020 : stats->stanullfrac = 1.0;
2969 3020 : if (is_varwidth)
2970 2614 : stats->stawidth = 0; /* "unknown" */
2971 : else
2972 406 : stats->stawidth = stats->attrtype->typlen;
2973 3020 : stats->stadistinct = 0.0; /* "unknown" */
2974 : }
2975 :
2976 : /* We don't need to bother cleaning up any of our temporary palloc's */
2977 49500 : }
2978 :
2979 : /*
2980 : * Comparator for sorting ScalarItems
2981 : *
2982 : * Aside from sorting the items, we update the tupnoLink[] array
2983 : * whenever two ScalarItems are found to contain equal datums. The array
2984 : * is indexed by tupno; for each ScalarItem, it contains the highest
2985 : * tupno that that item's datum has been found to be equal to. This allows
2986 : * us to avoid additional comparisons in compute_scalar_stats().
2987 : */
2988 : static int
2989 369477202 : compare_scalars(const void *a, const void *b, void *arg)
2990 : {
2991 369477202 : Datum da = ((const ScalarItem *) a)->value;
2992 369477202 : int ta = ((const ScalarItem *) a)->tupno;
2993 369477202 : Datum db = ((const ScalarItem *) b)->value;
2994 369477202 : int tb = ((const ScalarItem *) b)->tupno;
2995 369477202 : CompareScalarsContext *cxt = (CompareScalarsContext *) arg;
2996 : int compare;
2997 :
2998 369477202 : compare = ApplySortComparator(da, false, db, false, cxt->ssup);
2999 369477202 : if (compare != 0)
3000 135926693 : return compare;
3001 :
3002 : /*
3003 : * The two datums are equal, so update cxt->tupnoLink[].
3004 : */
3005 233550509 : if (cxt->tupnoLink[ta] < tb)
3006 32970854 : cxt->tupnoLink[ta] = tb;
3007 233550509 : if (cxt->tupnoLink[tb] < ta)
3008 2302436 : cxt->tupnoLink[tb] = ta;
3009 :
3010 : /*
3011 : * For equal datums, sort by tupno
3012 : */
3013 233550509 : return ta - tb;
3014 : }
3015 :
3016 : /*
3017 : * Comparator for sorting ScalarMCVItems by position
3018 : */
3019 : static int
3020 1424386 : compare_mcvs(const void *a, const void *b, void *arg)
3021 : {
3022 1424386 : int da = ((const ScalarMCVItem *) a)->first;
3023 1424386 : int db = ((const ScalarMCVItem *) b)->first;
3024 :
3025 1424386 : return da - db;
3026 : }
3027 :
3028 : /*
3029 : * Analyze the list of common values in the sample and decide how many are
3030 : * worth storing in the table's MCV list.
3031 : *
3032 : * mcv_counts is assumed to be a list of the counts of the most common values
3033 : * seen in the sample, starting with the most common. The return value is the
3034 : * number that are significantly more common than the values not in the list,
3035 : * and which are therefore deemed worth storing in the table's MCV list.
3036 : */
3037 : static int
3038 15665 : analyze_mcv_list(int *mcv_counts,
3039 : int num_mcv,
3040 : double stadistinct,
3041 : double stanullfrac,
3042 : int samplerows,
3043 : double totalrows)
3044 : {
3045 : double ndistinct_table;
3046 : double sumcount;
3047 : int i;
3048 :
3049 : /*
3050 : * If the entire table was sampled, keep the whole list. This also
3051 : * protects us against division by zero in the code below.
3052 : */
3053 15665 : if (samplerows == totalrows || totalrows <= 1.0)
3054 15248 : return num_mcv;
3055 :
3056 : /* Re-extract the estimated number of distinct nonnull values in table */
3057 417 : ndistinct_table = stadistinct;
3058 417 : if (ndistinct_table < 0)
3059 74 : ndistinct_table = -ndistinct_table * totalrows;
3060 :
3061 : /*
3062 : * Exclude the least common values from the MCV list, if they are not
3063 : * significantly more common than the estimated selectivity they would
3064 : * have if they weren't in the list. All non-MCV values are assumed to be
3065 : * equally common, after taking into account the frequencies of all the
3066 : * values in the MCV list and the number of nulls (c.f. eqsel()).
3067 : *
3068 : * Here sumcount tracks the total count of all but the last (least common)
3069 : * value in the MCV list, allowing us to determine the effect of excluding
3070 : * that value from the list.
3071 : *
3072 : * Note that we deliberately do this by removing values from the full
3073 : * list, rather than starting with an empty list and adding values,
3074 : * because the latter approach can fail to add any values if all the most
3075 : * common values have around the same frequency and make up the majority
3076 : * of the table, so that the overall average frequency of all values is
3077 : * roughly the same as that of the common values. This would lead to any
3078 : * uncommon values being significantly overestimated.
3079 : */
3080 417 : sumcount = 0.0;
3081 874 : for (i = 0; i < num_mcv - 1; i++)
3082 457 : sumcount += mcv_counts[i];
3083 :
3084 508 : while (num_mcv > 0)
3085 : {
3086 : double selec,
3087 : otherdistinct,
3088 : N,
3089 : n,
3090 : K,
3091 : variance,
3092 : stddev;
3093 :
3094 : /*
3095 : * Estimated selectivity the least common value would have if it
3096 : * wasn't in the MCV list (c.f. eqsel()).
3097 : */
3098 508 : selec = 1.0 - sumcount / samplerows - stanullfrac;
3099 508 : if (selec < 0.0)
3100 0 : selec = 0.0;
3101 508 : if (selec > 1.0)
3102 0 : selec = 1.0;
3103 508 : otherdistinct = ndistinct_table - (num_mcv - 1);
3104 508 : if (otherdistinct > 1)
3105 508 : selec /= otherdistinct;
3106 :
3107 : /*
3108 : * If the value is kept in the MCV list, its population frequency is
3109 : * assumed to equal its sample frequency. We use the lower end of a
3110 : * textbook continuity-corrected Wald-type confidence interval to
3111 : * determine if that is significantly more common than the non-MCV
3112 : * frequency --- specifically we assume the population frequency is
3113 : * highly likely to be within around 2 standard errors of the sample
3114 : * frequency, which equates to an interval of 2 standard deviations
3115 : * either side of the sample count, plus an additional 0.5 for the
3116 : * continuity correction. Since we are sampling without replacement,
3117 : * this is a hypergeometric distribution.
3118 : *
3119 : * XXX: Empirically, this approach seems to work quite well, but it
3120 : * may be worth considering more advanced techniques for estimating
3121 : * the confidence interval of the hypergeometric distribution.
3122 : */
3123 508 : N = totalrows;
3124 508 : n = samplerows;
3125 508 : K = N * mcv_counts[num_mcv - 1] / n;
3126 508 : variance = n * K * (N - K) * (N - n) / (N * N * (N - 1));
3127 508 : stddev = sqrt(variance);
3128 :
3129 508 : if (mcv_counts[num_mcv - 1] > selec * samplerows + 2 * stddev + 0.5)
3130 : {
3131 : /*
3132 : * The value is significantly more common than the non-MCV
3133 : * selectivity would suggest. Keep it, and all the other more
3134 : * common values in the list.
3135 : */
3136 392 : break;
3137 : }
3138 : else
3139 : {
3140 : /* Discard this value and consider the next least common value */
3141 116 : num_mcv--;
3142 116 : if (num_mcv == 0)
3143 25 : break;
3144 91 : sumcount -= mcv_counts[num_mcv - 1];
3145 : }
3146 : }
3147 417 : return num_mcv;
3148 : }
|
