Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * analyze.c
4 : * the Postgres statistics generator
5 : *
6 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/commands/analyze.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 : #include "postgres.h"
16 :
17 : #include <math.h>
18 :
19 : #include "access/detoast.h"
20 : #include "access/genam.h"
21 : #include "access/multixact.h"
22 : #include "access/relation.h"
23 : #include "access/table.h"
24 : #include "access/tableam.h"
25 : #include "access/transam.h"
26 : #include "access/tupconvert.h"
27 : #include "access/visibilitymap.h"
28 : #include "access/xact.h"
29 : #include "catalog/index.h"
30 : #include "catalog/indexing.h"
31 : #include "catalog/pg_inherits.h"
32 : #include "commands/progress.h"
33 : #include "commands/tablecmds.h"
34 : #include "commands/vacuum.h"
35 : #include "common/pg_prng.h"
36 : #include "executor/executor.h"
37 : #include "foreign/fdwapi.h"
38 : #include "miscadmin.h"
39 : #include "nodes/nodeFuncs.h"
40 : #include "parser/parse_oper.h"
41 : #include "parser/parse_relation.h"
42 : #include "pgstat.h"
43 : #include "statistics/extended_stats_internal.h"
44 : #include "statistics/statistics.h"
45 : #include "storage/bufmgr.h"
46 : #include "storage/procarray.h"
47 : #include "utils/attoptcache.h"
48 : #include "utils/datum.h"
49 : #include "utils/guc.h"
50 : #include "utils/lsyscache.h"
51 : #include "utils/memutils.h"
52 : #include "utils/pg_rusage.h"
53 : #include "utils/sampling.h"
54 : #include "utils/sortsupport.h"
55 : #include "utils/syscache.h"
56 : #include "utils/timestamp.h"
57 :
58 :
59 : /* Per-index data for ANALYZE */
60 : typedef struct AnlIndexData
61 : {
62 : IndexInfo *indexInfo; /* BuildIndexInfo result */
63 : double tupleFract; /* fraction of rows for partial index */
64 : VacAttrStats **vacattrstats; /* index attrs to analyze */
65 : int attr_cnt;
66 : } AnlIndexData;
67 :
68 :
69 : /* Default statistics target (GUC parameter) */
70 : int default_statistics_target = 100;
71 :
72 : /* A few variables that don't seem worth passing around as parameters */
73 : static MemoryContext anl_context = NULL;
74 : static BufferAccessStrategy vac_strategy;
75 :
76 :
77 : static void do_analyze_rel(Relation onerel,
78 : const VacuumParams params, List *va_cols,
79 : AcquireSampleRowsFunc acquirefunc, BlockNumber relpages,
80 : bool inh, bool in_outer_xact, int elevel);
81 : static void compute_index_stats(Relation onerel, double totalrows,
82 : AnlIndexData *indexdata, int nindexes,
83 : HeapTuple *rows, int numrows,
84 : MemoryContext col_context);
85 : static VacAttrStats *examine_attribute(Relation onerel, int attnum,
86 : Node *index_expr);
87 : static int acquire_sample_rows(Relation onerel, int elevel,
88 : HeapTuple *rows, int targrows,
89 : double *totalrows, double *totaldeadrows);
90 : static int compare_rows(const void *a, const void *b, void *arg);
91 : static int acquire_inherited_sample_rows(Relation onerel, int elevel,
92 : HeapTuple *rows, int targrows,
93 : double *totalrows, double *totaldeadrows);
94 : static void update_attstats(Oid relid, bool inh,
95 : int natts, VacAttrStats **vacattrstats);
96 : static Datum std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
97 : static Datum ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
98 :
99 :
100 : /*
101 : * analyze_rel() -- analyze one relation
102 : *
103 : * relid identifies the relation to analyze. If relation is supplied, use
104 : * the name therein for reporting any failure to open/lock the rel; do not
105 : * use it once we've successfully opened the rel, since it might be stale.
106 : */
107 : void
108 16884 : analyze_rel(Oid relid, RangeVar *relation,
109 : const VacuumParams params, List *va_cols, bool in_outer_xact,
110 : BufferAccessStrategy bstrategy)
111 : {
112 : Relation onerel;
113 : int elevel;
114 16884 : AcquireSampleRowsFunc acquirefunc = NULL;
115 16884 : BlockNumber relpages = 0;
116 :
117 : /* Select logging level */
118 16884 : if (params.options & VACOPT_VERBOSE)
119 0 : elevel = INFO;
120 : else
121 16884 : elevel = DEBUG2;
122 :
123 : /* Set up static variables */
124 16884 : vac_strategy = bstrategy;
125 :
126 : /*
127 : * Check for user-requested abort.
128 : */
129 16884 : CHECK_FOR_INTERRUPTS();
130 :
131 : /*
132 : * Open the relation, getting ShareUpdateExclusiveLock to ensure that two
133 : * ANALYZEs don't run on it concurrently. (This also locks out a
134 : * concurrent VACUUM, which doesn't matter much at the moment but might
135 : * matter if we ever try to accumulate stats on dead tuples.) If the rel
136 : * has been dropped since we last saw it, we don't need to process it.
137 : *
138 : * Make sure to generate only logs for ANALYZE in this case.
139 : */
140 16884 : onerel = vacuum_open_relation(relid, relation, params.options & ~(VACOPT_VACUUM),
141 16884 : params.log_analyze_min_duration >= 0,
142 : ShareUpdateExclusiveLock);
143 :
144 : /* leave if relation could not be opened or locked */
145 16884 : if (!onerel)
146 218 : return;
147 :
148 : /*
149 : * Check if relation needs to be skipped based on privileges. This check
150 : * happens also when building the relation list to analyze for a manual
151 : * operation, and needs to be done additionally here as ANALYZE could
152 : * happen across multiple transactions where privileges could have changed
153 : * in-between. Make sure to generate only logs for ANALYZE in this case.
154 : */
155 16876 : if (!vacuum_is_permitted_for_relation(RelationGetRelid(onerel),
156 : onerel->rd_rel,
157 16876 : params.options & ~VACOPT_VACUUM))
158 : {
159 36 : relation_close(onerel, ShareUpdateExclusiveLock);
160 36 : return;
161 : }
162 :
163 : /*
164 : * Silently ignore tables that are temp tables of other backends ---
165 : * trying to analyze these is rather pointless, since their contents are
166 : * probably not up-to-date on disk. (We don't throw a warning here; it
167 : * would just lead to chatter during a database-wide ANALYZE.)
168 : */
169 16840 : if (RELATION_IS_OTHER_TEMP(onerel))
170 : {
171 0 : relation_close(onerel, ShareUpdateExclusiveLock);
172 0 : return;
173 : }
174 :
175 : /*
176 : * We can ANALYZE any table except pg_statistic. See update_attstats
177 : */
178 16840 : if (RelationGetRelid(onerel) == StatisticRelationId)
179 : {
180 174 : relation_close(onerel, ShareUpdateExclusiveLock);
181 174 : return;
182 : }
183 :
184 : /*
185 : * Check that it's of an analyzable relkind, and set up appropriately.
186 : */
187 16666 : if (onerel->rd_rel->relkind == RELKIND_RELATION ||
188 838 : onerel->rd_rel->relkind == RELKIND_MATVIEW)
189 : {
190 : /* Regular table, so we'll use the regular row acquisition function */
191 15832 : acquirefunc = acquire_sample_rows;
192 : /* Also get regular table's size */
193 15832 : relpages = RelationGetNumberOfBlocks(onerel);
194 : }
195 834 : else if (onerel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
196 : {
197 : /*
198 : * For a foreign table, call the FDW's hook function to see whether it
199 : * supports analysis.
200 : */
201 : FdwRoutine *fdwroutine;
202 66 : bool ok = false;
203 :
204 66 : fdwroutine = GetFdwRoutineForRelation(onerel, false);
205 :
206 66 : if (fdwroutine->AnalyzeForeignTable != NULL)
207 66 : ok = fdwroutine->AnalyzeForeignTable(onerel,
208 : &acquirefunc,
209 : &relpages);
210 :
211 66 : if (!ok)
212 : {
213 0 : ereport(WARNING,
214 : (errmsg("skipping \"%s\" --- cannot analyze this foreign table",
215 : RelationGetRelationName(onerel))));
216 0 : relation_close(onerel, ShareUpdateExclusiveLock);
217 0 : return;
218 : }
219 : }
220 768 : else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
221 : {
222 : /*
223 : * For partitioned tables, we want to do the recursive ANALYZE below.
224 : */
225 : }
226 : else
227 : {
228 : /* No need for a WARNING if we already complained during VACUUM */
229 0 : if (!(params.options & VACOPT_VACUUM))
230 0 : ereport(WARNING,
231 : (errmsg("skipping \"%s\" --- cannot analyze non-tables or special system tables",
232 : RelationGetRelationName(onerel))));
233 0 : relation_close(onerel, ShareUpdateExclusiveLock);
234 0 : return;
235 : }
236 :
237 : /*
238 : * OK, let's do it. First, initialize progress reporting.
239 : */
240 16666 : pgstat_progress_start_command(PROGRESS_COMMAND_ANALYZE,
241 : RelationGetRelid(onerel));
242 16666 : if (AmAutoVacuumWorkerProcess())
243 468 : pgstat_progress_update_param(PROGRESS_ANALYZE_STARTED_BY,
244 : PROGRESS_ANALYZE_STARTED_BY_AUTOVACUUM);
245 : else
246 16198 : pgstat_progress_update_param(PROGRESS_ANALYZE_STARTED_BY,
247 : PROGRESS_ANALYZE_STARTED_BY_MANUAL);
248 :
249 : /*
250 : * Do the normal non-recursive ANALYZE. We can skip this for partitioned
251 : * tables, which don't contain any rows.
252 : */
253 16666 : if (onerel->rd_rel->relkind != RELKIND_PARTITIONED_TABLE)
254 15898 : do_analyze_rel(onerel, params, va_cols, acquirefunc,
255 : relpages, false, in_outer_xact, elevel);
256 :
257 : /*
258 : * If there are child tables, do recursive ANALYZE.
259 : */
260 16626 : if (onerel->rd_rel->relhassubclass)
261 896 : do_analyze_rel(onerel, params, va_cols, acquirefunc, relpages,
262 : true, in_outer_xact, elevel);
263 :
264 : /*
265 : * Close source relation now, but keep lock so that no one deletes it
266 : * before we commit. (If someone did, they'd fail to clean up the entries
267 : * we made in pg_statistic. Also, releasing the lock before commit would
268 : * expose us to concurrent-update failures in update_attstats.)
269 : */
270 16608 : relation_close(onerel, NoLock);
271 :
272 16608 : pgstat_progress_end_command();
273 : }
274 :
275 : /*
276 : * do_analyze_rel() -- analyze one relation, recursively or not
277 : *
278 : * Note that "acquirefunc" is only relevant for the non-inherited case.
279 : * For the inherited case, acquire_inherited_sample_rows() determines the
280 : * appropriate acquirefunc for each child table.
281 : */
282 : static void
283 16794 : do_analyze_rel(Relation onerel, const VacuumParams params,
284 : List *va_cols, AcquireSampleRowsFunc acquirefunc,
285 : BlockNumber relpages, bool inh, bool in_outer_xact,
286 : int elevel)
287 : {
288 : int attr_cnt,
289 : tcnt,
290 : i,
291 : ind;
292 : Relation *Irel;
293 : int nindexes;
294 : bool verbose,
295 : instrument,
296 : hasindex;
297 : VacAttrStats **vacattrstats;
298 : AnlIndexData *indexdata;
299 : int targrows,
300 : numrows,
301 : minrows;
302 : double totalrows,
303 : totaldeadrows;
304 : HeapTuple *rows;
305 : PGRUsage ru0;
306 16794 : TimestampTz starttime = 0;
307 : MemoryContext caller_context;
308 : Oid save_userid;
309 : int save_sec_context;
310 : int save_nestlevel;
311 16794 : WalUsage startwalusage = pgWalUsage;
312 16794 : BufferUsage startbufferusage = pgBufferUsage;
313 : BufferUsage bufferusage;
314 16794 : PgStat_Counter startreadtime = 0;
315 16794 : PgStat_Counter startwritetime = 0;
316 :
317 16794 : verbose = (params.options & VACOPT_VERBOSE) != 0;
318 17264 : instrument = (verbose || (AmAutoVacuumWorkerProcess() &&
319 470 : params.log_analyze_min_duration >= 0));
320 16794 : if (inh)
321 896 : ereport(elevel,
322 : (errmsg("analyzing \"%s.%s\" inheritance tree",
323 : get_namespace_name(RelationGetNamespace(onerel)),
324 : RelationGetRelationName(onerel))));
325 : else
326 15898 : ereport(elevel,
327 : (errmsg("analyzing \"%s.%s\"",
328 : get_namespace_name(RelationGetNamespace(onerel)),
329 : RelationGetRelationName(onerel))));
330 :
331 : /*
332 : * Set up a working context so that we can easily free whatever junk gets
333 : * created.
334 : */
335 16794 : anl_context = AllocSetContextCreate(CurrentMemoryContext,
336 : "Analyze",
337 : ALLOCSET_DEFAULT_SIZES);
338 16794 : caller_context = MemoryContextSwitchTo(anl_context);
339 :
340 : /*
341 : * Switch to the table owner's userid, so that any index functions are run
342 : * as that user. Also lock down security-restricted operations and
343 : * arrange to make GUC variable changes local to this command.
344 : */
345 16794 : GetUserIdAndSecContext(&save_userid, &save_sec_context);
346 16794 : SetUserIdAndSecContext(onerel->rd_rel->relowner,
347 : save_sec_context | SECURITY_RESTRICTED_OPERATION);
348 16794 : save_nestlevel = NewGUCNestLevel();
349 16794 : RestrictSearchPath();
350 :
351 : /*
352 : * When verbose or autovacuum logging is used, initialize a resource usage
353 : * snapshot and optionally track I/O timing.
354 : */
355 16794 : if (instrument)
356 : {
357 470 : if (track_io_timing)
358 : {
359 0 : startreadtime = pgStatBlockReadTime;
360 0 : startwritetime = pgStatBlockWriteTime;
361 : }
362 :
363 470 : pg_rusage_init(&ru0);
364 : }
365 :
366 : /* Used for instrumentation and stats report */
367 16794 : starttime = GetCurrentTimestamp();
368 :
369 : /*
370 : * Determine which columns to analyze
371 : *
372 : * Note that system attributes are never analyzed, so we just reject them
373 : * at the lookup stage. We also reject duplicate column mentions. (We
374 : * could alternatively ignore duplicates, but analyzing a column twice
375 : * won't work; we'd end up making a conflicting update in pg_statistic.)
376 : */
377 16794 : if (va_cols != NIL)
378 : {
379 100 : Bitmapset *unique_cols = NULL;
380 : ListCell *le;
381 :
382 100 : vacattrstats = (VacAttrStats **) palloc(list_length(va_cols) *
383 : sizeof(VacAttrStats *));
384 100 : tcnt = 0;
385 182 : foreach(le, va_cols)
386 : {
387 132 : char *col = strVal(lfirst(le));
388 :
389 132 : i = attnameAttNum(onerel, col, false);
390 132 : if (i == InvalidAttrNumber)
391 38 : ereport(ERROR,
392 : (errcode(ERRCODE_UNDEFINED_COLUMN),
393 : errmsg("column \"%s\" of relation \"%s\" does not exist",
394 : col, RelationGetRelationName(onerel))));
395 94 : if (bms_is_member(i, unique_cols))
396 12 : ereport(ERROR,
397 : (errcode(ERRCODE_DUPLICATE_COLUMN),
398 : errmsg("column \"%s\" of relation \"%s\" appears more than once",
399 : col, RelationGetRelationName(onerel))));
400 82 : unique_cols = bms_add_member(unique_cols, i);
401 :
402 82 : vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
403 82 : if (vacattrstats[tcnt] != NULL)
404 82 : tcnt++;
405 : }
406 50 : attr_cnt = tcnt;
407 : }
408 : else
409 : {
410 16694 : attr_cnt = onerel->rd_att->natts;
411 : vacattrstats = (VacAttrStats **)
412 16694 : palloc(attr_cnt * sizeof(VacAttrStats *));
413 16694 : tcnt = 0;
414 138036 : for (i = 1; i <= attr_cnt; i++)
415 : {
416 121342 : vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
417 121342 : if (vacattrstats[tcnt] != NULL)
418 121274 : tcnt++;
419 : }
420 16694 : attr_cnt = tcnt;
421 : }
422 :
423 : /*
424 : * Open all indexes of the relation, and see if there are any analyzable
425 : * columns in the indexes. We do not analyze index columns if there was
426 : * an explicit column list in the ANALYZE command, however.
427 : *
428 : * If we are doing a recursive scan, we don't want to touch the parent's
429 : * indexes at all. If we're processing a partitioned table, we need to
430 : * know if there are any indexes, but we don't want to process them.
431 : */
432 16744 : if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
433 : {
434 750 : List *idxs = RelationGetIndexList(onerel);
435 :
436 750 : Irel = NULL;
437 750 : nindexes = 0;
438 750 : hasindex = idxs != NIL;
439 750 : list_free(idxs);
440 : }
441 15994 : else if (!inh)
442 : {
443 15866 : vac_open_indexes(onerel, AccessShareLock, &nindexes, &Irel);
444 15866 : hasindex = nindexes > 0;
445 : }
446 : else
447 : {
448 128 : Irel = NULL;
449 128 : nindexes = 0;
450 128 : hasindex = false;
451 : }
452 16744 : indexdata = NULL;
453 16744 : if (nindexes > 0)
454 : {
455 12434 : indexdata = (AnlIndexData *) palloc0(nindexes * sizeof(AnlIndexData));
456 35884 : for (ind = 0; ind < nindexes; ind++)
457 : {
458 23450 : AnlIndexData *thisdata = &indexdata[ind];
459 : IndexInfo *indexInfo;
460 :
461 23450 : thisdata->indexInfo = indexInfo = BuildIndexInfo(Irel[ind]);
462 23450 : thisdata->tupleFract = 1.0; /* fix later if partial */
463 23450 : if (indexInfo->ii_Expressions != NIL && va_cols == NIL)
464 : {
465 112 : ListCell *indexpr_item = list_head(indexInfo->ii_Expressions);
466 :
467 112 : thisdata->vacattrstats = (VacAttrStats **)
468 112 : palloc(indexInfo->ii_NumIndexAttrs * sizeof(VacAttrStats *));
469 112 : tcnt = 0;
470 228 : for (i = 0; i < indexInfo->ii_NumIndexAttrs; i++)
471 : {
472 116 : int keycol = indexInfo->ii_IndexAttrNumbers[i];
473 :
474 116 : if (keycol == 0)
475 : {
476 : /* Found an index expression */
477 : Node *indexkey;
478 :
479 112 : if (indexpr_item == NULL) /* shouldn't happen */
480 0 : elog(ERROR, "too few entries in indexprs list");
481 112 : indexkey = (Node *) lfirst(indexpr_item);
482 112 : indexpr_item = lnext(indexInfo->ii_Expressions,
483 : indexpr_item);
484 224 : thisdata->vacattrstats[tcnt] =
485 112 : examine_attribute(Irel[ind], i + 1, indexkey);
486 112 : if (thisdata->vacattrstats[tcnt] != NULL)
487 112 : tcnt++;
488 : }
489 : }
490 112 : thisdata->attr_cnt = tcnt;
491 : }
492 : }
493 : }
494 :
495 : /*
496 : * Determine how many rows we need to sample, using the worst case from
497 : * all analyzable columns. We use a lower bound of 100 rows to avoid
498 : * possible overflow in Vitter's algorithm. (Note: that will also be the
499 : * target in the corner case where there are no analyzable columns.)
500 : */
501 16744 : targrows = 100;
502 138076 : for (i = 0; i < attr_cnt; i++)
503 : {
504 121332 : if (targrows < vacattrstats[i]->minrows)
505 16672 : targrows = vacattrstats[i]->minrows;
506 : }
507 40194 : for (ind = 0; ind < nindexes; ind++)
508 : {
509 23450 : AnlIndexData *thisdata = &indexdata[ind];
510 :
511 23562 : for (i = 0; i < thisdata->attr_cnt; i++)
512 : {
513 112 : if (targrows < thisdata->vacattrstats[i]->minrows)
514 12 : targrows = thisdata->vacattrstats[i]->minrows;
515 : }
516 : }
517 :
518 : /*
519 : * Look at extended statistics objects too, as those may define custom
520 : * statistics target. So we may need to sample more rows and then build
521 : * the statistics with enough detail.
522 : */
523 16744 : minrows = ComputeExtStatisticsRows(onerel, attr_cnt, vacattrstats);
524 :
525 16744 : if (targrows < minrows)
526 0 : targrows = minrows;
527 :
528 : /*
529 : * Acquire the sample rows
530 : */
531 16744 : rows = (HeapTuple *) palloc(targrows * sizeof(HeapTuple));
532 16744 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
533 : inh ? PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS_INH :
534 : PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS);
535 16744 : if (inh)
536 878 : numrows = acquire_inherited_sample_rows(onerel, elevel,
537 : rows, targrows,
538 : &totalrows, &totaldeadrows);
539 : else
540 15866 : numrows = (*acquirefunc) (onerel, elevel,
541 : rows, targrows,
542 : &totalrows, &totaldeadrows);
543 :
544 : /*
545 : * Compute the statistics. Temporary results during the calculations for
546 : * each column are stored in a child context. The calc routines are
547 : * responsible to make sure that whatever they store into the VacAttrStats
548 : * structure is allocated in anl_context.
549 : */
550 16742 : if (numrows > 0)
551 : {
552 : MemoryContext col_context,
553 : old_context;
554 :
555 11292 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
556 : PROGRESS_ANALYZE_PHASE_COMPUTE_STATS);
557 :
558 11292 : col_context = AllocSetContextCreate(anl_context,
559 : "Analyze Column",
560 : ALLOCSET_DEFAULT_SIZES);
561 11292 : old_context = MemoryContextSwitchTo(col_context);
562 :
563 99986 : for (i = 0; i < attr_cnt; i++)
564 : {
565 88694 : VacAttrStats *stats = vacattrstats[i];
566 : AttributeOpts *aopt;
567 :
568 88694 : stats->rows = rows;
569 88694 : stats->tupDesc = onerel->rd_att;
570 88694 : stats->compute_stats(stats,
571 : std_fetch_func,
572 : numrows,
573 : totalrows);
574 :
575 : /*
576 : * If the appropriate flavor of the n_distinct option is
577 : * specified, override with the corresponding value.
578 : */
579 88694 : aopt = get_attribute_options(onerel->rd_id, stats->tupattnum);
580 88694 : if (aopt != NULL)
581 : {
582 : float8 n_distinct;
583 :
584 6 : n_distinct = inh ? aopt->n_distinct_inherited : aopt->n_distinct;
585 6 : if (n_distinct != 0.0)
586 6 : stats->stadistinct = n_distinct;
587 : }
588 :
589 88694 : MemoryContextReset(col_context);
590 : }
591 :
592 11292 : if (nindexes > 0)
593 7290 : compute_index_stats(onerel, totalrows,
594 : indexdata, nindexes,
595 : rows, numrows,
596 : col_context);
597 :
598 11286 : MemoryContextSwitchTo(old_context);
599 11286 : MemoryContextDelete(col_context);
600 :
601 : /*
602 : * Emit the completed stats rows into pg_statistic, replacing any
603 : * previous statistics for the target columns. (If there are stats in
604 : * pg_statistic for columns we didn't process, we leave them alone.)
605 : */
606 11286 : update_attstats(RelationGetRelid(onerel), inh,
607 : attr_cnt, vacattrstats);
608 :
609 25558 : for (ind = 0; ind < nindexes; ind++)
610 : {
611 14272 : AnlIndexData *thisdata = &indexdata[ind];
612 :
613 14272 : update_attstats(RelationGetRelid(Irel[ind]), false,
614 : thisdata->attr_cnt, thisdata->vacattrstats);
615 : }
616 :
617 : /* Build extended statistics (if there are any). */
618 11286 : BuildRelationExtStatistics(onerel, inh, totalrows, numrows, rows,
619 : attr_cnt, vacattrstats);
620 : }
621 :
622 16736 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
623 : PROGRESS_ANALYZE_PHASE_FINALIZE_ANALYZE);
624 :
625 : /*
626 : * Update pages/tuples stats in pg_class ... but not if we're doing
627 : * inherited stats.
628 : *
629 : * We assume that VACUUM hasn't set pg_class.reltuples already, even
630 : * during a VACUUM ANALYZE. Although VACUUM often updates pg_class,
631 : * exceptions exist. A "VACUUM (ANALYZE, INDEX_CLEANUP OFF)" command will
632 : * never update pg_class entries for index relations. It's also possible
633 : * that an individual index's pg_class entry won't be updated during
634 : * VACUUM if the index AM returns NULL from its amvacuumcleanup() routine.
635 : */
636 16736 : if (!inh)
637 : {
638 15858 : BlockNumber relallvisible = 0;
639 15858 : BlockNumber relallfrozen = 0;
640 :
641 15858 : if (RELKIND_HAS_STORAGE(onerel->rd_rel->relkind))
642 15794 : visibilitymap_count(onerel, &relallvisible, &relallfrozen);
643 :
644 : /*
645 : * Update pg_class for table relation. CCI first, in case acquirefunc
646 : * updated pg_class.
647 : */
648 15858 : CommandCounterIncrement();
649 15858 : vac_update_relstats(onerel,
650 : relpages,
651 : totalrows,
652 : relallvisible,
653 : relallfrozen,
654 : hasindex,
655 : InvalidTransactionId,
656 : InvalidMultiXactId,
657 : NULL, NULL,
658 : in_outer_xact);
659 :
660 : /* Same for indexes */
661 39296 : for (ind = 0; ind < nindexes; ind++)
662 : {
663 23438 : AnlIndexData *thisdata = &indexdata[ind];
664 : double totalindexrows;
665 :
666 23438 : totalindexrows = ceil(thisdata->tupleFract * totalrows);
667 23438 : vac_update_relstats(Irel[ind],
668 23438 : RelationGetNumberOfBlocks(Irel[ind]),
669 : totalindexrows,
670 : 0, 0,
671 : false,
672 : InvalidTransactionId,
673 : InvalidMultiXactId,
674 : NULL, NULL,
675 : in_outer_xact);
676 : }
677 : }
678 878 : else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
679 : {
680 : /*
681 : * Partitioned tables don't have storage, so we don't set any fields
682 : * in their pg_class entries except for reltuples and relhasindex.
683 : */
684 750 : CommandCounterIncrement();
685 750 : vac_update_relstats(onerel, -1, totalrows,
686 : 0, 0, hasindex, InvalidTransactionId,
687 : InvalidMultiXactId,
688 : NULL, NULL,
689 : in_outer_xact);
690 : }
691 :
692 : /*
693 : * Now report ANALYZE to the cumulative stats system. For regular tables,
694 : * we do it only if not doing inherited stats. For partitioned tables, we
695 : * only do it for inherited stats. (We're never called for not-inherited
696 : * stats on partitioned tables anyway.)
697 : *
698 : * Reset the mod_since_analyze counter only if we analyzed all columns;
699 : * otherwise, there is still work for auto-analyze to do.
700 : */
701 16736 : if (!inh)
702 15858 : pgstat_report_analyze(onerel, totalrows, totaldeadrows,
703 : (va_cols == NIL), starttime);
704 878 : else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
705 750 : pgstat_report_analyze(onerel, 0, 0, (va_cols == NIL), starttime);
706 :
707 : /*
708 : * If this isn't part of VACUUM ANALYZE, let index AMs do cleanup.
709 : *
710 : * Note that most index AMs perform a no-op as a matter of policy for
711 : * amvacuumcleanup() when called in ANALYZE-only mode. The only exception
712 : * among core index AMs is GIN/ginvacuumcleanup().
713 : */
714 16736 : if (!(params.options & VACOPT_VACUUM))
715 : {
716 31728 : for (ind = 0; ind < nindexes; ind++)
717 : {
718 : IndexBulkDeleteResult *stats;
719 : IndexVacuumInfo ivinfo;
720 :
721 18196 : ivinfo.index = Irel[ind];
722 18196 : ivinfo.heaprel = onerel;
723 18196 : ivinfo.analyze_only = true;
724 18196 : ivinfo.estimated_count = true;
725 18196 : ivinfo.message_level = elevel;
726 18196 : ivinfo.num_heap_tuples = onerel->rd_rel->reltuples;
727 18196 : ivinfo.strategy = vac_strategy;
728 :
729 18196 : stats = index_vacuum_cleanup(&ivinfo, NULL);
730 :
731 18196 : if (stats)
732 0 : pfree(stats);
733 : }
734 : }
735 :
736 : /* Done with indexes */
737 16736 : vac_close_indexes(nindexes, Irel, NoLock);
738 :
739 : /* Log the action if appropriate */
740 16736 : if (instrument)
741 : {
742 470 : TimestampTz endtime = GetCurrentTimestamp();
743 :
744 906 : if (verbose || params.log_analyze_min_duration == 0 ||
745 436 : TimestampDifferenceExceeds(starttime, endtime,
746 436 : params.log_analyze_min_duration))
747 : {
748 : long delay_in_ms;
749 : WalUsage walusage;
750 34 : double read_rate = 0;
751 34 : double write_rate = 0;
752 : char *msgfmt;
753 : StringInfoData buf;
754 : int64 total_blks_hit;
755 : int64 total_blks_read;
756 : int64 total_blks_dirtied;
757 :
758 34 : memset(&bufferusage, 0, sizeof(BufferUsage));
759 34 : BufferUsageAccumDiff(&bufferusage, &pgBufferUsage, &startbufferusage);
760 34 : memset(&walusage, 0, sizeof(WalUsage));
761 34 : WalUsageAccumDiff(&walusage, &pgWalUsage, &startwalusage);
762 :
763 34 : total_blks_hit = bufferusage.shared_blks_hit +
764 34 : bufferusage.local_blks_hit;
765 34 : total_blks_read = bufferusage.shared_blks_read +
766 34 : bufferusage.local_blks_read;
767 34 : total_blks_dirtied = bufferusage.shared_blks_dirtied +
768 34 : bufferusage.local_blks_dirtied;
769 :
770 : /*
771 : * We do not expect an analyze to take > 25 days and it simplifies
772 : * things a bit to use TimestampDifferenceMilliseconds.
773 : */
774 34 : delay_in_ms = TimestampDifferenceMilliseconds(starttime, endtime);
775 :
776 : /*
777 : * Note that we are reporting these read/write rates in the same
778 : * manner as VACUUM does, which means that while the 'average read
779 : * rate' here actually corresponds to page misses and resulting
780 : * reads which are also picked up by track_io_timing, if enabled,
781 : * the 'average write rate' is actually talking about the rate of
782 : * pages being dirtied, not being written out, so it's typical to
783 : * have a non-zero 'avg write rate' while I/O timings only reports
784 : * reads.
785 : *
786 : * It's not clear that an ANALYZE will ever result in
787 : * FlushBuffer() being called, but we track and support reporting
788 : * on I/O write time in case that changes as it's practically free
789 : * to do so anyway.
790 : */
791 :
792 34 : if (delay_in_ms > 0)
793 : {
794 34 : read_rate = (double) BLCKSZ * total_blks_read /
795 34 : (1024 * 1024) / (delay_in_ms / 1000.0);
796 34 : write_rate = (double) BLCKSZ * total_blks_dirtied /
797 34 : (1024 * 1024) / (delay_in_ms / 1000.0);
798 : }
799 :
800 : /*
801 : * We split this up so we don't emit empty I/O timing values when
802 : * track_io_timing isn't enabled.
803 : */
804 :
805 34 : initStringInfo(&buf);
806 :
807 34 : if (AmAutoVacuumWorkerProcess())
808 34 : msgfmt = _("automatic analyze of table \"%s.%s.%s\"\n");
809 : else
810 0 : msgfmt = _("finished analyzing table \"%s.%s.%s\"\n");
811 :
812 34 : appendStringInfo(&buf, msgfmt,
813 : get_database_name(MyDatabaseId),
814 34 : get_namespace_name(RelationGetNamespace(onerel)),
815 34 : RelationGetRelationName(onerel));
816 34 : if (track_cost_delay_timing)
817 : {
818 : /*
819 : * We bypass the changecount mechanism because this value is
820 : * only updated by the calling process.
821 : */
822 0 : appendStringInfo(&buf, _("delay time: %.3f ms\n"),
823 0 : (double) MyBEEntry->st_progress_param[PROGRESS_ANALYZE_DELAY_TIME] / 1000000.0);
824 : }
825 34 : if (track_io_timing)
826 : {
827 0 : double read_ms = (double) (pgStatBlockReadTime - startreadtime) / 1000;
828 0 : double write_ms = (double) (pgStatBlockWriteTime - startwritetime) / 1000;
829 :
830 0 : appendStringInfo(&buf, _("I/O timings: read: %.3f ms, write: %.3f ms\n"),
831 : read_ms, write_ms);
832 : }
833 34 : appendStringInfo(&buf, _("avg read rate: %.3f MB/s, avg write rate: %.3f MB/s\n"),
834 : read_rate, write_rate);
835 34 : appendStringInfo(&buf, _("buffer usage: %" PRId64 " hits, %" PRId64 " reads, %" PRId64 " dirtied\n"),
836 : total_blks_hit,
837 : total_blks_read,
838 : total_blks_dirtied);
839 34 : appendStringInfo(&buf,
840 34 : _("WAL usage: %" PRId64 " records, %" PRId64 " full page images, %" PRIu64 " bytes, %" PRIu64 " full page image bytes, %" PRId64 " buffers full\n"),
841 : walusage.wal_records,
842 : walusage.wal_fpi,
843 : walusage.wal_bytes,
844 : walusage.wal_fpi_bytes,
845 : walusage.wal_buffers_full);
846 34 : appendStringInfo(&buf, _("system usage: %s"), pg_rusage_show(&ru0));
847 :
848 34 : ereport(verbose ? INFO : LOG,
849 : (errmsg_internal("%s", buf.data)));
850 :
851 34 : pfree(buf.data);
852 : }
853 : }
854 :
855 : /* Roll back any GUC changes executed by index functions */
856 16736 : AtEOXact_GUC(false, save_nestlevel);
857 :
858 : /* Restore userid and security context */
859 16736 : SetUserIdAndSecContext(save_userid, save_sec_context);
860 :
861 : /* Restore current context and release memory */
862 16736 : MemoryContextSwitchTo(caller_context);
863 16736 : MemoryContextDelete(anl_context);
864 16736 : anl_context = NULL;
865 16736 : }
866 :
867 : /*
868 : * Compute statistics about indexes of a relation
869 : */
870 : static void
871 7290 : compute_index_stats(Relation onerel, double totalrows,
872 : AnlIndexData *indexdata, int nindexes,
873 : HeapTuple *rows, int numrows,
874 : MemoryContext col_context)
875 : {
876 : MemoryContext ind_context,
877 : old_context;
878 : Datum values[INDEX_MAX_KEYS];
879 : bool isnull[INDEX_MAX_KEYS];
880 : int ind,
881 : i;
882 :
883 7290 : ind_context = AllocSetContextCreate(anl_context,
884 : "Analyze Index",
885 : ALLOCSET_DEFAULT_SIZES);
886 7290 : old_context = MemoryContextSwitchTo(ind_context);
887 :
888 21568 : for (ind = 0; ind < nindexes; ind++)
889 : {
890 14284 : AnlIndexData *thisdata = &indexdata[ind];
891 14284 : IndexInfo *indexInfo = thisdata->indexInfo;
892 14284 : int attr_cnt = thisdata->attr_cnt;
893 : TupleTableSlot *slot;
894 : EState *estate;
895 : ExprContext *econtext;
896 : ExprState *predicate;
897 : Datum *exprvals;
898 : bool *exprnulls;
899 : int numindexrows,
900 : tcnt,
901 : rowno;
902 : double totalindexrows;
903 :
904 : /* Ignore index if no columns to analyze and not partial */
905 14284 : if (attr_cnt == 0 && indexInfo->ii_Predicate == NIL)
906 14138 : continue;
907 :
908 : /*
909 : * Need an EState for evaluation of index expressions and
910 : * partial-index predicates. Create it in the per-index context to be
911 : * sure it gets cleaned up at the bottom of the loop.
912 : */
913 146 : estate = CreateExecutorState();
914 146 : econtext = GetPerTupleExprContext(estate);
915 : /* Need a slot to hold the current heap tuple, too */
916 146 : slot = MakeSingleTupleTableSlot(RelationGetDescr(onerel),
917 : &TTSOpsHeapTuple);
918 :
919 : /* Arrange for econtext's scan tuple to be the tuple under test */
920 146 : econtext->ecxt_scantuple = slot;
921 :
922 : /* Set up execution state for predicate. */
923 146 : predicate = ExecPrepareQual(indexInfo->ii_Predicate, estate);
924 :
925 : /* Compute and save index expression values */
926 146 : exprvals = (Datum *) palloc(numrows * attr_cnt * sizeof(Datum));
927 146 : exprnulls = (bool *) palloc(numrows * attr_cnt * sizeof(bool));
928 146 : numindexrows = 0;
929 146 : tcnt = 0;
930 127434 : for (rowno = 0; rowno < numrows; rowno++)
931 : {
932 127294 : HeapTuple heapTuple = rows[rowno];
933 :
934 127294 : vacuum_delay_point(true);
935 :
936 : /*
937 : * Reset the per-tuple context each time, to reclaim any cruft
938 : * left behind by evaluating the predicate or index expressions.
939 : */
940 127294 : ResetExprContext(econtext);
941 :
942 : /* Set up for predicate or expression evaluation */
943 127294 : ExecStoreHeapTuple(heapTuple, slot, false);
944 :
945 : /* If index is partial, check predicate */
946 127294 : if (predicate != NULL)
947 : {
948 46066 : if (!ExecQual(predicate, econtext))
949 39326 : continue;
950 : }
951 87968 : numindexrows++;
952 :
953 87968 : if (attr_cnt > 0)
954 : {
955 : /*
956 : * Evaluate the index row to compute expression values. We
957 : * could do this by hand, but FormIndexDatum is convenient.
958 : */
959 81228 : FormIndexDatum(indexInfo,
960 : slot,
961 : estate,
962 : values,
963 : isnull);
964 :
965 : /*
966 : * Save just the columns we care about. We copy the values
967 : * into ind_context from the estate's per-tuple context.
968 : */
969 162444 : for (i = 0; i < attr_cnt; i++)
970 : {
971 81222 : VacAttrStats *stats = thisdata->vacattrstats[i];
972 81222 : int attnum = stats->tupattnum;
973 :
974 81222 : if (isnull[attnum - 1])
975 : {
976 6 : exprvals[tcnt] = (Datum) 0;
977 6 : exprnulls[tcnt] = true;
978 : }
979 : else
980 : {
981 162432 : exprvals[tcnt] = datumCopy(values[attnum - 1],
982 81216 : stats->attrtype->typbyval,
983 81216 : stats->attrtype->typlen);
984 81216 : exprnulls[tcnt] = false;
985 : }
986 81222 : tcnt++;
987 : }
988 : }
989 : }
990 :
991 : /*
992 : * Having counted the number of rows that pass the predicate in the
993 : * sample, we can estimate the total number of rows in the index.
994 : */
995 140 : thisdata->tupleFract = (double) numindexrows / (double) numrows;
996 140 : totalindexrows = ceil(thisdata->tupleFract * totalrows);
997 :
998 : /*
999 : * Now we can compute the statistics for the expression columns.
1000 : */
1001 140 : if (numindexrows > 0)
1002 : {
1003 132 : MemoryContextSwitchTo(col_context);
1004 220 : for (i = 0; i < attr_cnt; i++)
1005 : {
1006 88 : VacAttrStats *stats = thisdata->vacattrstats[i];
1007 :
1008 88 : stats->exprvals = exprvals + i;
1009 88 : stats->exprnulls = exprnulls + i;
1010 88 : stats->rowstride = attr_cnt;
1011 88 : stats->compute_stats(stats,
1012 : ind_fetch_func,
1013 : numindexrows,
1014 : totalindexrows);
1015 :
1016 88 : MemoryContextReset(col_context);
1017 : }
1018 : }
1019 :
1020 : /* And clean up */
1021 140 : MemoryContextSwitchTo(ind_context);
1022 :
1023 140 : ExecDropSingleTupleTableSlot(slot);
1024 140 : FreeExecutorState(estate);
1025 140 : MemoryContextReset(ind_context);
1026 : }
1027 :
1028 7284 : MemoryContextSwitchTo(old_context);
1029 7284 : MemoryContextDelete(ind_context);
1030 7284 : }
1031 :
1032 : /*
1033 : * examine_attribute -- pre-analysis of a single column
1034 : *
1035 : * Determine whether the column is analyzable; if so, create and initialize
1036 : * a VacAttrStats struct for it. If not, return NULL.
1037 : *
1038 : * If index_expr isn't NULL, then we're trying to analyze an expression index,
1039 : * and index_expr is the expression tree representing the column's data.
1040 : */
1041 : static VacAttrStats *
1042 121536 : examine_attribute(Relation onerel, int attnum, Node *index_expr)
1043 : {
1044 121536 : Form_pg_attribute attr = TupleDescAttr(onerel->rd_att, attnum - 1);
1045 : int attstattarget;
1046 : HeapTuple atttuple;
1047 : Datum dat;
1048 : bool isnull;
1049 : HeapTuple typtuple;
1050 : VacAttrStats *stats;
1051 : int i;
1052 : bool ok;
1053 :
1054 : /* Never analyze dropped columns */
1055 121536 : if (attr->attisdropped)
1056 6 : return NULL;
1057 :
1058 : /* Don't analyze virtual generated columns */
1059 121530 : if (attr->attgenerated == ATTRIBUTE_GENERATED_VIRTUAL)
1060 56 : return NULL;
1061 :
1062 : /*
1063 : * Get attstattarget value. Set to -1 if null. (Analyze functions expect
1064 : * -1 to mean use default_statistics_target; see for example
1065 : * std_typanalyze.)
1066 : */
1067 121474 : atttuple = SearchSysCache2(ATTNUM, ObjectIdGetDatum(RelationGetRelid(onerel)), Int16GetDatum(attnum));
1068 121474 : if (!HeapTupleIsValid(atttuple))
1069 0 : elog(ERROR, "cache lookup failed for attribute %d of relation %u",
1070 : attnum, RelationGetRelid(onerel));
1071 121474 : dat = SysCacheGetAttr(ATTNUM, atttuple, Anum_pg_attribute_attstattarget, &isnull);
1072 121474 : attstattarget = isnull ? -1 : DatumGetInt16(dat);
1073 121474 : ReleaseSysCache(atttuple);
1074 :
1075 : /* Don't analyze column if user has specified not to */
1076 121474 : if (attstattarget == 0)
1077 6 : return NULL;
1078 :
1079 : /*
1080 : * Create the VacAttrStats struct.
1081 : */
1082 121468 : stats = palloc0_object(VacAttrStats);
1083 121468 : stats->attstattarget = attstattarget;
1084 :
1085 : /*
1086 : * When analyzing an expression index, believe the expression tree's type
1087 : * not the column datatype --- the latter might be the opckeytype storage
1088 : * type of the opclass, which is not interesting for our purposes. (Note:
1089 : * if we did anything with non-expression index columns, we'd need to
1090 : * figure out where to get the correct type info from, but for now that's
1091 : * not a problem.) It's not clear whether anyone will care about the
1092 : * typmod, but we store that too just in case.
1093 : */
1094 121468 : if (index_expr)
1095 : {
1096 112 : stats->attrtypid = exprType(index_expr);
1097 112 : stats->attrtypmod = exprTypmod(index_expr);
1098 :
1099 : /*
1100 : * If a collation has been specified for the index column, use that in
1101 : * preference to anything else; but if not, fall back to whatever we
1102 : * can get from the expression.
1103 : */
1104 112 : if (OidIsValid(onerel->rd_indcollation[attnum - 1]))
1105 14 : stats->attrcollid = onerel->rd_indcollation[attnum - 1];
1106 : else
1107 98 : stats->attrcollid = exprCollation(index_expr);
1108 : }
1109 : else
1110 : {
1111 121356 : stats->attrtypid = attr->atttypid;
1112 121356 : stats->attrtypmod = attr->atttypmod;
1113 121356 : stats->attrcollid = attr->attcollation;
1114 : }
1115 :
1116 121468 : typtuple = SearchSysCacheCopy1(TYPEOID,
1117 : ObjectIdGetDatum(stats->attrtypid));
1118 121468 : if (!HeapTupleIsValid(typtuple))
1119 0 : elog(ERROR, "cache lookup failed for type %u", stats->attrtypid);
1120 121468 : stats->attrtype = (Form_pg_type) GETSTRUCT(typtuple);
1121 121468 : stats->anl_context = anl_context;
1122 121468 : stats->tupattnum = attnum;
1123 :
1124 : /*
1125 : * The fields describing the stats->stavalues[n] element types default to
1126 : * the type of the data being analyzed, but the type-specific typanalyze
1127 : * function can change them if it wants to store something else.
1128 : */
1129 728808 : for (i = 0; i < STATISTIC_NUM_SLOTS; i++)
1130 : {
1131 607340 : stats->statypid[i] = stats->attrtypid;
1132 607340 : stats->statyplen[i] = stats->attrtype->typlen;
1133 607340 : stats->statypbyval[i] = stats->attrtype->typbyval;
1134 607340 : stats->statypalign[i] = stats->attrtype->typalign;
1135 : }
1136 :
1137 : /*
1138 : * Call the type-specific typanalyze function. If none is specified, use
1139 : * std_typanalyze().
1140 : */
1141 121468 : if (OidIsValid(stats->attrtype->typanalyze))
1142 7818 : ok = DatumGetBool(OidFunctionCall1(stats->attrtype->typanalyze,
1143 : PointerGetDatum(stats)));
1144 : else
1145 113650 : ok = std_typanalyze(stats);
1146 :
1147 121468 : if (!ok || stats->compute_stats == NULL || stats->minrows <= 0)
1148 : {
1149 0 : heap_freetuple(typtuple);
1150 0 : pfree(stats);
1151 0 : return NULL;
1152 : }
1153 :
1154 121468 : return stats;
1155 : }
1156 :
1157 : /*
1158 : * Read stream callback returning the next BlockNumber as chosen by the
1159 : * BlockSampling algorithm.
1160 : */
1161 : static BlockNumber
1162 157034 : block_sampling_read_stream_next(ReadStream *stream,
1163 : void *callback_private_data,
1164 : void *per_buffer_data)
1165 : {
1166 157034 : BlockSamplerData *bs = callback_private_data;
1167 :
1168 157034 : return BlockSampler_HasMore(bs) ? BlockSampler_Next(bs) : InvalidBlockNumber;
1169 : }
1170 :
1171 : /*
1172 : * acquire_sample_rows -- acquire a random sample of rows from the table
1173 : *
1174 : * Selected rows are returned in the caller-allocated array rows[], which
1175 : * must have at least targrows entries.
1176 : * The actual number of rows selected is returned as the function result.
1177 : * We also estimate the total numbers of live and dead rows in the table,
1178 : * and return them into *totalrows and *totaldeadrows, respectively.
1179 : *
1180 : * The returned list of tuples is in order by physical position in the table.
1181 : * (We will rely on this later to derive correlation estimates.)
1182 : *
1183 : * As of May 2004 we use a new two-stage method: Stage one selects up
1184 : * to targrows random blocks (or all blocks, if there aren't so many).
1185 : * Stage two scans these blocks and uses the Vitter algorithm to create
1186 : * a random sample of targrows rows (or less, if there are less in the
1187 : * sample of blocks). The two stages are executed simultaneously: each
1188 : * block is processed as soon as stage one returns its number and while
1189 : * the rows are read stage two controls which ones are to be inserted
1190 : * into the sample.
1191 : *
1192 : * Although every row has an equal chance of ending up in the final
1193 : * sample, this sampling method is not perfect: not every possible
1194 : * sample has an equal chance of being selected. For large relations
1195 : * the number of different blocks represented by the sample tends to be
1196 : * too small. We can live with that for now. Improvements are welcome.
1197 : *
1198 : * An important property of this sampling method is that because we do
1199 : * look at a statistically unbiased set of blocks, we should get
1200 : * unbiased estimates of the average numbers of live and dead rows per
1201 : * block. The previous sampling method put too much credence in the row
1202 : * density near the start of the table.
1203 : */
1204 : static int
1205 17850 : acquire_sample_rows(Relation onerel, int elevel,
1206 : HeapTuple *rows, int targrows,
1207 : double *totalrows, double *totaldeadrows)
1208 : {
1209 17850 : int numrows = 0; /* # rows now in reservoir */
1210 17850 : double samplerows = 0; /* total # rows collected */
1211 17850 : double liverows = 0; /* # live rows seen */
1212 17850 : double deadrows = 0; /* # dead rows seen */
1213 17850 : double rowstoskip = -1; /* -1 means not set yet */
1214 : uint32 randseed; /* Seed for block sampler(s) */
1215 : BlockNumber totalblocks;
1216 : TransactionId OldestXmin;
1217 : BlockSamplerData bs;
1218 : ReservoirStateData rstate;
1219 : TupleTableSlot *slot;
1220 : TableScanDesc scan;
1221 : BlockNumber nblocks;
1222 17850 : BlockNumber blksdone = 0;
1223 : ReadStream *stream;
1224 :
1225 : Assert(targrows > 0);
1226 :
1227 17850 : totalblocks = RelationGetNumberOfBlocks(onerel);
1228 :
1229 : /* Need a cutoff xmin for HeapTupleSatisfiesVacuum */
1230 17850 : OldestXmin = GetOldestNonRemovableTransactionId(onerel);
1231 :
1232 : /* Prepare for sampling block numbers */
1233 17850 : randseed = pg_prng_uint32(&pg_global_prng_state);
1234 17850 : nblocks = BlockSampler_Init(&bs, totalblocks, targrows, randseed);
1235 :
1236 : /* Report sampling block numbers */
1237 17850 : pgstat_progress_update_param(PROGRESS_ANALYZE_BLOCKS_TOTAL,
1238 : nblocks);
1239 :
1240 : /* Prepare for sampling rows */
1241 17850 : reservoir_init_selection_state(&rstate, targrows);
1242 :
1243 17850 : scan = table_beginscan_analyze(onerel);
1244 17850 : slot = table_slot_create(onerel, NULL);
1245 :
1246 : /*
1247 : * It is safe to use batching, as block_sampling_read_stream_next never
1248 : * blocks.
1249 : */
1250 17850 : stream = read_stream_begin_relation(READ_STREAM_MAINTENANCE |
1251 : READ_STREAM_USE_BATCHING,
1252 : vac_strategy,
1253 : scan->rs_rd,
1254 : MAIN_FORKNUM,
1255 : block_sampling_read_stream_next,
1256 : &bs,
1257 : 0);
1258 :
1259 : /* Outer loop over blocks to sample */
1260 157034 : while (table_scan_analyze_next_block(scan, stream))
1261 : {
1262 139184 : vacuum_delay_point(true);
1263 :
1264 10771818 : while (table_scan_analyze_next_tuple(scan, OldestXmin, &liverows, &deadrows, slot))
1265 : {
1266 : /*
1267 : * The first targrows sample rows are simply copied into the
1268 : * reservoir. Then we start replacing tuples in the sample until
1269 : * we reach the end of the relation. This algorithm is from Jeff
1270 : * Vitter's paper (see full citation in utils/misc/sampling.c). It
1271 : * works by repeatedly computing the number of tuples to skip
1272 : * before selecting a tuple, which replaces a randomly chosen
1273 : * element of the reservoir (current set of tuples). At all times
1274 : * the reservoir is a true random sample of the tuples we've
1275 : * passed over so far, so when we fall off the end of the relation
1276 : * we're done.
1277 : */
1278 10632634 : if (numrows < targrows)
1279 10379552 : rows[numrows++] = ExecCopySlotHeapTuple(slot);
1280 : else
1281 : {
1282 : /*
1283 : * t in Vitter's paper is the number of records already
1284 : * processed. If we need to compute a new S value, we must
1285 : * use the not-yet-incremented value of samplerows as t.
1286 : */
1287 253082 : if (rowstoskip < 0)
1288 115804 : rowstoskip = reservoir_get_next_S(&rstate, samplerows, targrows);
1289 :
1290 253082 : if (rowstoskip <= 0)
1291 : {
1292 : /*
1293 : * Found a suitable tuple, so save it, replacing one old
1294 : * tuple at random
1295 : */
1296 115750 : int k = (int) (targrows * sampler_random_fract(&rstate.randstate));
1297 :
1298 : Assert(k >= 0 && k < targrows);
1299 115750 : heap_freetuple(rows[k]);
1300 115750 : rows[k] = ExecCopySlotHeapTuple(slot);
1301 : }
1302 :
1303 253082 : rowstoskip -= 1;
1304 : }
1305 :
1306 10632634 : samplerows += 1;
1307 : }
1308 :
1309 139184 : pgstat_progress_update_param(PROGRESS_ANALYZE_BLOCKS_DONE,
1310 : ++blksdone);
1311 : }
1312 :
1313 17850 : read_stream_end(stream);
1314 :
1315 17850 : ExecDropSingleTupleTableSlot(slot);
1316 17850 : table_endscan(scan);
1317 :
1318 : /*
1319 : * If we didn't find as many tuples as we wanted then we're done. No sort
1320 : * is needed, since they're already in order.
1321 : *
1322 : * Otherwise we need to sort the collected tuples by position
1323 : * (itempointer). It's not worth worrying about corner cases where the
1324 : * tuples are already sorted.
1325 : */
1326 17850 : if (numrows == targrows)
1327 162 : qsort_interruptible(rows, numrows, sizeof(HeapTuple),
1328 : compare_rows, NULL);
1329 :
1330 : /*
1331 : * Estimate total numbers of live and dead rows in relation, extrapolating
1332 : * on the assumption that the average tuple density in pages we didn't
1333 : * scan is the same as in the pages we did scan. Since what we scanned is
1334 : * a random sample of the pages in the relation, this should be a good
1335 : * assumption.
1336 : */
1337 17850 : if (bs.m > 0)
1338 : {
1339 12492 : *totalrows = floor((liverows / bs.m) * totalblocks + 0.5);
1340 12492 : *totaldeadrows = floor((deadrows / bs.m) * totalblocks + 0.5);
1341 : }
1342 : else
1343 : {
1344 5358 : *totalrows = 0.0;
1345 5358 : *totaldeadrows = 0.0;
1346 : }
1347 :
1348 : /*
1349 : * Emit some interesting relation info
1350 : */
1351 17850 : ereport(elevel,
1352 : (errmsg("\"%s\": scanned %d of %u pages, "
1353 : "containing %.0f live rows and %.0f dead rows; "
1354 : "%d rows in sample, %.0f estimated total rows",
1355 : RelationGetRelationName(onerel),
1356 : bs.m, totalblocks,
1357 : liverows, deadrows,
1358 : numrows, *totalrows)));
1359 :
1360 17850 : return numrows;
1361 : }
1362 :
1363 : /*
1364 : * Comparator for sorting rows[] array
1365 : */
1366 : static int
1367 4016284 : compare_rows(const void *a, const void *b, void *arg)
1368 : {
1369 4016284 : HeapTuple ha = *(const HeapTuple *) a;
1370 4016284 : HeapTuple hb = *(const HeapTuple *) b;
1371 4016284 : BlockNumber ba = ItemPointerGetBlockNumber(&ha->t_self);
1372 4016284 : OffsetNumber oa = ItemPointerGetOffsetNumber(&ha->t_self);
1373 4016284 : BlockNumber bb = ItemPointerGetBlockNumber(&hb->t_self);
1374 4016284 : OffsetNumber ob = ItemPointerGetOffsetNumber(&hb->t_self);
1375 :
1376 4016284 : if (ba < bb)
1377 876586 : return -1;
1378 3139698 : if (ba > bb)
1379 870434 : return 1;
1380 2269264 : if (oa < ob)
1381 1523414 : return -1;
1382 745850 : if (oa > ob)
1383 745850 : return 1;
1384 0 : return 0;
1385 : }
1386 :
1387 :
1388 : /*
1389 : * acquire_inherited_sample_rows -- acquire sample rows from inheritance tree
1390 : *
1391 : * This has the same API as acquire_sample_rows, except that rows are
1392 : * collected from all inheritance children as well as the specified table.
1393 : * We fail and return zero if there are no inheritance children, or if all
1394 : * children are foreign tables that don't support ANALYZE.
1395 : */
1396 : static int
1397 878 : acquire_inherited_sample_rows(Relation onerel, int elevel,
1398 : HeapTuple *rows, int targrows,
1399 : double *totalrows, double *totaldeadrows)
1400 : {
1401 : List *tableOIDs;
1402 : Relation *rels;
1403 : AcquireSampleRowsFunc *acquirefuncs;
1404 : double *relblocks;
1405 : double totalblocks;
1406 : int numrows,
1407 : nrels,
1408 : i;
1409 : ListCell *lc;
1410 : bool has_child;
1411 :
1412 : /* Initialize output parameters to zero now, in case we exit early */
1413 878 : *totalrows = 0;
1414 878 : *totaldeadrows = 0;
1415 :
1416 : /*
1417 : * Find all members of inheritance set. We only need AccessShareLock on
1418 : * the children.
1419 : */
1420 : tableOIDs =
1421 878 : find_all_inheritors(RelationGetRelid(onerel), AccessShareLock, NULL);
1422 :
1423 : /*
1424 : * Check that there's at least one descendant, else fail. This could
1425 : * happen despite analyze_rel's relhassubclass check, if table once had a
1426 : * child but no longer does. In that case, we can clear the
1427 : * relhassubclass field so as not to make the same mistake again later.
1428 : * (This is safe because we hold ShareUpdateExclusiveLock.)
1429 : */
1430 878 : if (list_length(tableOIDs) < 2)
1431 : {
1432 : /* CCI because we already updated the pg_class row in this command */
1433 20 : CommandCounterIncrement();
1434 20 : SetRelationHasSubclass(RelationGetRelid(onerel), false);
1435 20 : ereport(elevel,
1436 : (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no child tables",
1437 : get_namespace_name(RelationGetNamespace(onerel)),
1438 : RelationGetRelationName(onerel))));
1439 20 : return 0;
1440 : }
1441 :
1442 : /*
1443 : * Identify acquirefuncs to use, and count blocks in all the relations.
1444 : * The result could overflow BlockNumber, so we use double arithmetic.
1445 : */
1446 858 : rels = (Relation *) palloc(list_length(tableOIDs) * sizeof(Relation));
1447 : acquirefuncs = (AcquireSampleRowsFunc *)
1448 858 : palloc(list_length(tableOIDs) * sizeof(AcquireSampleRowsFunc));
1449 858 : relblocks = (double *) palloc(list_length(tableOIDs) * sizeof(double));
1450 858 : totalblocks = 0;
1451 858 : nrels = 0;
1452 858 : has_child = false;
1453 3926 : foreach(lc, tableOIDs)
1454 : {
1455 3068 : Oid childOID = lfirst_oid(lc);
1456 : Relation childrel;
1457 3068 : AcquireSampleRowsFunc acquirefunc = NULL;
1458 3068 : BlockNumber relpages = 0;
1459 :
1460 : /* We already got the needed lock */
1461 3068 : childrel = table_open(childOID, NoLock);
1462 :
1463 : /* Ignore if temp table of another backend */
1464 3068 : if (RELATION_IS_OTHER_TEMP(childrel))
1465 : {
1466 : /* ... but release the lock on it */
1467 : Assert(childrel != onerel);
1468 0 : table_close(childrel, AccessShareLock);
1469 822 : continue;
1470 : }
1471 :
1472 : /* Check table type (MATVIEW can't happen, but might as well allow) */
1473 3068 : if (childrel->rd_rel->relkind == RELKIND_RELATION ||
1474 852 : childrel->rd_rel->relkind == RELKIND_MATVIEW)
1475 : {
1476 : /* Regular table, so use the regular row acquisition function */
1477 2216 : acquirefunc = acquire_sample_rows;
1478 2216 : relpages = RelationGetNumberOfBlocks(childrel);
1479 : }
1480 852 : else if (childrel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
1481 : {
1482 : /*
1483 : * For a foreign table, call the FDW's hook function to see
1484 : * whether it supports analysis.
1485 : */
1486 : FdwRoutine *fdwroutine;
1487 30 : bool ok = false;
1488 :
1489 30 : fdwroutine = GetFdwRoutineForRelation(childrel, false);
1490 :
1491 30 : if (fdwroutine->AnalyzeForeignTable != NULL)
1492 30 : ok = fdwroutine->AnalyzeForeignTable(childrel,
1493 : &acquirefunc,
1494 : &relpages);
1495 :
1496 30 : if (!ok)
1497 : {
1498 : /* ignore, but release the lock on it */
1499 : Assert(childrel != onerel);
1500 0 : table_close(childrel, AccessShareLock);
1501 0 : continue;
1502 : }
1503 : }
1504 : else
1505 : {
1506 : /*
1507 : * ignore, but release the lock on it. don't try to unlock the
1508 : * passed-in relation
1509 : */
1510 : Assert(childrel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE);
1511 822 : if (childrel != onerel)
1512 78 : table_close(childrel, AccessShareLock);
1513 : else
1514 744 : table_close(childrel, NoLock);
1515 822 : continue;
1516 : }
1517 :
1518 : /* OK, we'll process this child */
1519 2246 : has_child = true;
1520 2246 : rels[nrels] = childrel;
1521 2246 : acquirefuncs[nrels] = acquirefunc;
1522 2246 : relblocks[nrels] = (double) relpages;
1523 2246 : totalblocks += (double) relpages;
1524 2246 : nrels++;
1525 : }
1526 :
1527 : /*
1528 : * If we don't have at least one child table to consider, fail. If the
1529 : * relation is a partitioned table, it's not counted as a child table.
1530 : */
1531 858 : if (!has_child)
1532 : {
1533 0 : ereport(elevel,
1534 : (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no analyzable child tables",
1535 : get_namespace_name(RelationGetNamespace(onerel)),
1536 : RelationGetRelationName(onerel))));
1537 0 : return 0;
1538 : }
1539 :
1540 : /*
1541 : * Now sample rows from each relation, proportionally to its fraction of
1542 : * the total block count. (This might be less than desirable if the child
1543 : * rels have radically different free-space percentages, but it's not
1544 : * clear that it's worth working harder.)
1545 : */
1546 858 : pgstat_progress_update_param(PROGRESS_ANALYZE_CHILD_TABLES_TOTAL,
1547 : nrels);
1548 858 : numrows = 0;
1549 3104 : for (i = 0; i < nrels; i++)
1550 : {
1551 2246 : Relation childrel = rels[i];
1552 2246 : AcquireSampleRowsFunc acquirefunc = acquirefuncs[i];
1553 2246 : double childblocks = relblocks[i];
1554 :
1555 : /*
1556 : * Report progress. The sampling function will normally report blocks
1557 : * done/total, but we need to reset them to 0 here, so that they don't
1558 : * show an old value until that.
1559 : */
1560 : {
1561 2246 : const int progress_index[] = {
1562 : PROGRESS_ANALYZE_CURRENT_CHILD_TABLE_RELID,
1563 : PROGRESS_ANALYZE_BLOCKS_DONE,
1564 : PROGRESS_ANALYZE_BLOCKS_TOTAL
1565 : };
1566 2246 : const int64 progress_vals[] = {
1567 2246 : RelationGetRelid(childrel),
1568 : 0,
1569 : 0,
1570 : };
1571 :
1572 2246 : pgstat_progress_update_multi_param(3, progress_index, progress_vals);
1573 : }
1574 :
1575 2246 : if (childblocks > 0)
1576 : {
1577 : int childtargrows;
1578 :
1579 2080 : childtargrows = (int) rint(targrows * childblocks / totalblocks);
1580 : /* Make sure we don't overrun due to roundoff error */
1581 2080 : childtargrows = Min(childtargrows, targrows - numrows);
1582 2080 : if (childtargrows > 0)
1583 : {
1584 : int childrows;
1585 : double trows,
1586 : tdrows;
1587 :
1588 : /* Fetch a random sample of the child's rows */
1589 2080 : childrows = (*acquirefunc) (childrel, elevel,
1590 2080 : rows + numrows, childtargrows,
1591 : &trows, &tdrows);
1592 :
1593 : /* We may need to convert from child's rowtype to parent's */
1594 2080 : if (childrows > 0 &&
1595 2080 : !equalRowTypes(RelationGetDescr(childrel),
1596 : RelationGetDescr(onerel)))
1597 : {
1598 : TupleConversionMap *map;
1599 :
1600 1990 : map = convert_tuples_by_name(RelationGetDescr(childrel),
1601 : RelationGetDescr(onerel));
1602 1990 : if (map != NULL)
1603 : {
1604 : int j;
1605 :
1606 106604 : for (j = 0; j < childrows; j++)
1607 : {
1608 : HeapTuple newtup;
1609 :
1610 106472 : newtup = execute_attr_map_tuple(rows[numrows + j], map);
1611 106472 : heap_freetuple(rows[numrows + j]);
1612 106472 : rows[numrows + j] = newtup;
1613 : }
1614 132 : free_conversion_map(map);
1615 : }
1616 : }
1617 :
1618 : /* And add to counts */
1619 2080 : numrows += childrows;
1620 2080 : *totalrows += trows;
1621 2080 : *totaldeadrows += tdrows;
1622 : }
1623 : }
1624 :
1625 : /*
1626 : * Note: we cannot release the child-table locks, since we may have
1627 : * pointers to their TOAST tables in the sampled rows.
1628 : */
1629 2246 : table_close(childrel, NoLock);
1630 2246 : pgstat_progress_update_param(PROGRESS_ANALYZE_CHILD_TABLES_DONE,
1631 2246 : i + 1);
1632 : }
1633 :
1634 858 : return numrows;
1635 : }
1636 :
1637 :
1638 : /*
1639 : * update_attstats() -- update attribute statistics for one relation
1640 : *
1641 : * Statistics are stored in several places: the pg_class row for the
1642 : * relation has stats about the whole relation, and there is a
1643 : * pg_statistic row for each (non-system) attribute that has ever
1644 : * been analyzed. The pg_class values are updated by VACUUM, not here.
1645 : *
1646 : * pg_statistic rows are just added or updated normally. This means
1647 : * that pg_statistic will probably contain some deleted rows at the
1648 : * completion of a vacuum cycle, unless it happens to get vacuumed last.
1649 : *
1650 : * To keep things simple, we punt for pg_statistic, and don't try
1651 : * to compute or store rows for pg_statistic itself in pg_statistic.
1652 : * This could possibly be made to work, but it's not worth the trouble.
1653 : * Note analyze_rel() has seen to it that we won't come here when
1654 : * vacuuming pg_statistic itself.
1655 : *
1656 : * Note: there would be a race condition here if two backends could
1657 : * ANALYZE the same table concurrently. Presently, we lock that out
1658 : * by taking a self-exclusive lock on the relation in analyze_rel().
1659 : */
1660 : static void
1661 25558 : update_attstats(Oid relid, bool inh, int natts, VacAttrStats **vacattrstats)
1662 : {
1663 : Relation sd;
1664 : int attno;
1665 25558 : CatalogIndexState indstate = NULL;
1666 :
1667 25558 : if (natts <= 0)
1668 14196 : return; /* nothing to do */
1669 :
1670 11362 : sd = table_open(StatisticRelationId, RowExclusiveLock);
1671 :
1672 100144 : for (attno = 0; attno < natts; attno++)
1673 : {
1674 88782 : VacAttrStats *stats = vacattrstats[attno];
1675 : HeapTuple stup,
1676 : oldtup;
1677 : int i,
1678 : k,
1679 : n;
1680 : Datum values[Natts_pg_statistic];
1681 : bool nulls[Natts_pg_statistic];
1682 : bool replaces[Natts_pg_statistic];
1683 :
1684 : /* Ignore attr if we weren't able to collect stats */
1685 88782 : if (!stats->stats_valid)
1686 6 : continue;
1687 :
1688 : /*
1689 : * Construct a new pg_statistic tuple
1690 : */
1691 2840832 : for (i = 0; i < Natts_pg_statistic; ++i)
1692 : {
1693 2752056 : nulls[i] = false;
1694 2752056 : replaces[i] = true;
1695 : }
1696 :
1697 88776 : values[Anum_pg_statistic_starelid - 1] = ObjectIdGetDatum(relid);
1698 88776 : values[Anum_pg_statistic_staattnum - 1] = Int16GetDatum(stats->tupattnum);
1699 88776 : values[Anum_pg_statistic_stainherit - 1] = BoolGetDatum(inh);
1700 88776 : values[Anum_pg_statistic_stanullfrac - 1] = Float4GetDatum(stats->stanullfrac);
1701 88776 : values[Anum_pg_statistic_stawidth - 1] = Int32GetDatum(stats->stawidth);
1702 88776 : values[Anum_pg_statistic_stadistinct - 1] = Float4GetDatum(stats->stadistinct);
1703 88776 : i = Anum_pg_statistic_stakind1 - 1;
1704 532656 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1705 : {
1706 443880 : values[i++] = Int16GetDatum(stats->stakind[k]); /* stakindN */
1707 : }
1708 88776 : i = Anum_pg_statistic_staop1 - 1;
1709 532656 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1710 : {
1711 443880 : values[i++] = ObjectIdGetDatum(stats->staop[k]); /* staopN */
1712 : }
1713 88776 : i = Anum_pg_statistic_stacoll1 - 1;
1714 532656 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1715 : {
1716 443880 : values[i++] = ObjectIdGetDatum(stats->stacoll[k]); /* stacollN */
1717 : }
1718 88776 : i = Anum_pg_statistic_stanumbers1 - 1;
1719 532656 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1720 : {
1721 443880 : if (stats->stanumbers[k] != NULL)
1722 : {
1723 138926 : int nnum = stats->numnumbers[k];
1724 138926 : Datum *numdatums = (Datum *) palloc(nnum * sizeof(Datum));
1725 : ArrayType *arry;
1726 :
1727 1146962 : for (n = 0; n < nnum; n++)
1728 1008036 : numdatums[n] = Float4GetDatum(stats->stanumbers[k][n]);
1729 138926 : arry = construct_array_builtin(numdatums, nnum, FLOAT4OID);
1730 138926 : values[i++] = PointerGetDatum(arry); /* stanumbersN */
1731 : }
1732 : else
1733 : {
1734 304954 : nulls[i] = true;
1735 304954 : values[i++] = (Datum) 0;
1736 : }
1737 : }
1738 88776 : i = Anum_pg_statistic_stavalues1 - 1;
1739 532656 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1740 : {
1741 443880 : if (stats->stavalues[k] != NULL)
1742 : {
1743 : ArrayType *arry;
1744 :
1745 97768 : arry = construct_array(stats->stavalues[k],
1746 : stats->numvalues[k],
1747 : stats->statypid[k],
1748 97768 : stats->statyplen[k],
1749 97768 : stats->statypbyval[k],
1750 97768 : stats->statypalign[k]);
1751 97768 : values[i++] = PointerGetDatum(arry); /* stavaluesN */
1752 : }
1753 : else
1754 : {
1755 346112 : nulls[i] = true;
1756 346112 : values[i++] = (Datum) 0;
1757 : }
1758 : }
1759 :
1760 : /* Is there already a pg_statistic tuple for this attribute? */
1761 177552 : oldtup = SearchSysCache3(STATRELATTINH,
1762 : ObjectIdGetDatum(relid),
1763 88776 : Int16GetDatum(stats->tupattnum),
1764 : BoolGetDatum(inh));
1765 :
1766 : /* Open index information when we know we need it */
1767 88776 : if (indstate == NULL)
1768 11356 : indstate = CatalogOpenIndexes(sd);
1769 :
1770 88776 : if (HeapTupleIsValid(oldtup))
1771 : {
1772 : /* Yes, replace it */
1773 40308 : stup = heap_modify_tuple(oldtup,
1774 : RelationGetDescr(sd),
1775 : values,
1776 : nulls,
1777 : replaces);
1778 40308 : ReleaseSysCache(oldtup);
1779 40308 : CatalogTupleUpdateWithInfo(sd, &stup->t_self, stup, indstate);
1780 : }
1781 : else
1782 : {
1783 : /* No, insert new tuple */
1784 48468 : stup = heap_form_tuple(RelationGetDescr(sd), values, nulls);
1785 48468 : CatalogTupleInsertWithInfo(sd, stup, indstate);
1786 : }
1787 :
1788 88776 : heap_freetuple(stup);
1789 : }
1790 :
1791 11362 : if (indstate != NULL)
1792 11356 : CatalogCloseIndexes(indstate);
1793 11362 : table_close(sd, RowExclusiveLock);
1794 : }
1795 :
1796 : /*
1797 : * Standard fetch function for use by compute_stats subroutines.
1798 : *
1799 : * This exists to provide some insulation between compute_stats routines
1800 : * and the actual storage of the sample data.
1801 : */
1802 : static Datum
1803 84441968 : std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1804 : {
1805 84441968 : int attnum = stats->tupattnum;
1806 84441968 : HeapTuple tuple = stats->rows[rownum];
1807 84441968 : TupleDesc tupDesc = stats->tupDesc;
1808 :
1809 84441968 : return heap_getattr(tuple, attnum, tupDesc, isNull);
1810 : }
1811 :
1812 : /*
1813 : * Fetch function for analyzing index expressions.
1814 : *
1815 : * We have not bothered to construct index tuples, instead the data is
1816 : * just in Datum arrays.
1817 : */
1818 : static Datum
1819 81222 : ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1820 : {
1821 : int i;
1822 :
1823 : /* exprvals and exprnulls are already offset for proper column */
1824 81222 : i = rownum * stats->rowstride;
1825 81222 : *isNull = stats->exprnulls[i];
1826 81222 : return stats->exprvals[i];
1827 : }
1828 :
1829 :
1830 : /*==========================================================================
1831 : *
1832 : * Code below this point represents the "standard" type-specific statistics
1833 : * analysis algorithms. This code can be replaced on a per-data-type basis
1834 : * by setting a nonzero value in pg_type.typanalyze.
1835 : *
1836 : *==========================================================================
1837 : */
1838 :
1839 :
1840 : /*
1841 : * To avoid consuming too much memory during analysis and/or too much space
1842 : * in the resulting pg_statistic rows, we ignore varlena datums that are wider
1843 : * than WIDTH_THRESHOLD (after detoasting!). This is legitimate for MCV
1844 : * and distinct-value calculations since a wide value is unlikely to be
1845 : * duplicated at all, much less be a most-common value. For the same reason,
1846 : * ignoring wide values will not affect our estimates of histogram bin
1847 : * boundaries very much.
1848 : */
1849 : #define WIDTH_THRESHOLD 1024
1850 :
1851 : #define swapInt(a,b) do {int _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1852 : #define swapDatum(a,b) do {Datum _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1853 :
1854 : /*
1855 : * Extra information used by the default analysis routines
1856 : */
1857 : typedef struct
1858 : {
1859 : int count; /* # of duplicates */
1860 : int first; /* values[] index of first occurrence */
1861 : } ScalarMCVItem;
1862 :
1863 : typedef struct
1864 : {
1865 : SortSupport ssup;
1866 : int *tupnoLink;
1867 : } CompareScalarsContext;
1868 :
1869 :
1870 : static void compute_trivial_stats(VacAttrStatsP stats,
1871 : AnalyzeAttrFetchFunc fetchfunc,
1872 : int samplerows,
1873 : double totalrows);
1874 : static void compute_distinct_stats(VacAttrStatsP stats,
1875 : AnalyzeAttrFetchFunc fetchfunc,
1876 : int samplerows,
1877 : double totalrows);
1878 : static void compute_scalar_stats(VacAttrStatsP stats,
1879 : AnalyzeAttrFetchFunc fetchfunc,
1880 : int samplerows,
1881 : double totalrows);
1882 : static int compare_scalars(const void *a, const void *b, void *arg);
1883 : static int compare_mcvs(const void *a, const void *b, void *arg);
1884 : static int analyze_mcv_list(int *mcv_counts,
1885 : int num_mcv,
1886 : double stadistinct,
1887 : double stanullfrac,
1888 : int samplerows,
1889 : double totalrows);
1890 :
1891 :
1892 : /*
1893 : * std_typanalyze -- the default type-specific typanalyze function
1894 : */
1895 : bool
1896 122652 : std_typanalyze(VacAttrStats *stats)
1897 : {
1898 : Oid ltopr;
1899 : Oid eqopr;
1900 : StdAnalyzeData *mystats;
1901 :
1902 : /* If the attstattarget column is negative, use the default value */
1903 122652 : if (stats->attstattarget < 0)
1904 122034 : stats->attstattarget = default_statistics_target;
1905 :
1906 : /* Look for default "<" and "=" operators for column's type */
1907 122652 : get_sort_group_operators(stats->attrtypid,
1908 : false, false, false,
1909 : <opr, &eqopr, NULL,
1910 : NULL);
1911 :
1912 : /* Save the operator info for compute_stats routines */
1913 122652 : mystats = palloc_object(StdAnalyzeData);
1914 122652 : mystats->eqopr = eqopr;
1915 122652 : mystats->eqfunc = OidIsValid(eqopr) ? get_opcode(eqopr) : InvalidOid;
1916 122652 : mystats->ltopr = ltopr;
1917 122652 : stats->extra_data = mystats;
1918 :
1919 : /*
1920 : * Determine which standard statistics algorithm to use
1921 : */
1922 122652 : if (OidIsValid(eqopr) && OidIsValid(ltopr))
1923 : {
1924 : /* Seems to be a scalar datatype */
1925 118878 : stats->compute_stats = compute_scalar_stats;
1926 : /*--------------------
1927 : * The following choice of minrows is based on the paper
1928 : * "Random sampling for histogram construction: how much is enough?"
1929 : * by Surajit Chaudhuri, Rajeev Motwani and Vivek Narasayya, in
1930 : * Proceedings of ACM SIGMOD International Conference on Management
1931 : * of Data, 1998, Pages 436-447. Their Corollary 1 to Theorem 5
1932 : * says that for table size n, histogram size k, maximum relative
1933 : * error in bin size f, and error probability gamma, the minimum
1934 : * random sample size is
1935 : * r = 4 * k * ln(2*n/gamma) / f^2
1936 : * Taking f = 0.5, gamma = 0.01, n = 10^6 rows, we obtain
1937 : * r = 305.82 * k
1938 : * Note that because of the log function, the dependence on n is
1939 : * quite weak; even at n = 10^12, a 300*k sample gives <= 0.66
1940 : * bin size error with probability 0.99. So there's no real need to
1941 : * scale for n, which is a good thing because we don't necessarily
1942 : * know it at this point.
1943 : *--------------------
1944 : */
1945 118878 : stats->minrows = 300 * stats->attstattarget;
1946 : }
1947 3774 : else if (OidIsValid(eqopr))
1948 : {
1949 : /* We can still recognize distinct values */
1950 3260 : stats->compute_stats = compute_distinct_stats;
1951 : /* Might as well use the same minrows as above */
1952 3260 : stats->minrows = 300 * stats->attstattarget;
1953 : }
1954 : else
1955 : {
1956 : /* Can't do much but the trivial stuff */
1957 514 : stats->compute_stats = compute_trivial_stats;
1958 : /* Might as well use the same minrows as above */
1959 514 : stats->minrows = 300 * stats->attstattarget;
1960 : }
1961 :
1962 122652 : return true;
1963 : }
1964 :
1965 :
1966 : /*
1967 : * compute_trivial_stats() -- compute very basic column statistics
1968 : *
1969 : * We use this when we cannot find a hash "=" operator for the datatype.
1970 : *
1971 : * We determine the fraction of non-null rows and the average datum width.
1972 : */
1973 : static void
1974 342 : compute_trivial_stats(VacAttrStatsP stats,
1975 : AnalyzeAttrFetchFunc fetchfunc,
1976 : int samplerows,
1977 : double totalrows)
1978 : {
1979 : int i;
1980 342 : int null_cnt = 0;
1981 342 : int nonnull_cnt = 0;
1982 342 : double total_width = 0;
1983 684 : bool is_varlena = (!stats->attrtype->typbyval &&
1984 342 : stats->attrtype->typlen == -1);
1985 684 : bool is_varwidth = (!stats->attrtype->typbyval &&
1986 342 : stats->attrtype->typlen < 0);
1987 :
1988 1070496 : for (i = 0; i < samplerows; i++)
1989 : {
1990 : Datum value;
1991 : bool isnull;
1992 :
1993 1070154 : vacuum_delay_point(true);
1994 :
1995 1070154 : value = fetchfunc(stats, i, &isnull);
1996 :
1997 : /* Check for null/nonnull */
1998 1070154 : if (isnull)
1999 : {
2000 667476 : null_cnt++;
2001 667476 : continue;
2002 : }
2003 402678 : nonnull_cnt++;
2004 :
2005 : /*
2006 : * If it's a variable-width field, add up widths for average width
2007 : * calculation. Note that if the value is toasted, we use the toasted
2008 : * width. We don't bother with this calculation if it's a fixed-width
2009 : * type.
2010 : */
2011 402678 : if (is_varlena)
2012 : {
2013 79886 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2014 : }
2015 322792 : else if (is_varwidth)
2016 : {
2017 : /* must be cstring */
2018 0 : total_width += strlen(DatumGetCString(value)) + 1;
2019 : }
2020 : }
2021 :
2022 : /* We can only compute average width if we found some non-null values. */
2023 342 : if (nonnull_cnt > 0)
2024 : {
2025 148 : stats->stats_valid = true;
2026 : /* Do the simple null-frac and width stats */
2027 148 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2028 148 : if (is_varwidth)
2029 70 : stats->stawidth = total_width / (double) nonnull_cnt;
2030 : else
2031 78 : stats->stawidth = stats->attrtype->typlen;
2032 148 : stats->stadistinct = 0.0; /* "unknown" */
2033 : }
2034 194 : else if (null_cnt > 0)
2035 : {
2036 : /* We found only nulls; assume the column is entirely null */
2037 194 : stats->stats_valid = true;
2038 194 : stats->stanullfrac = 1.0;
2039 194 : if (is_varwidth)
2040 194 : stats->stawidth = 0; /* "unknown" */
2041 : else
2042 0 : stats->stawidth = stats->attrtype->typlen;
2043 194 : stats->stadistinct = 0.0; /* "unknown" */
2044 : }
2045 342 : }
2046 :
2047 :
2048 : /*
2049 : * compute_distinct_stats() -- compute column statistics including ndistinct
2050 : *
2051 : * We use this when we can find only an "=" operator for the datatype.
2052 : *
2053 : * We determine the fraction of non-null rows, the average width, the
2054 : * most common values, and the (estimated) number of distinct values.
2055 : *
2056 : * The most common values are determined by brute force: we keep a list
2057 : * of previously seen values, ordered by number of times seen, as we scan
2058 : * the samples. A newly seen value is inserted just after the last
2059 : * multiply-seen value, causing the bottommost (oldest) singly-seen value
2060 : * to drop off the list. The accuracy of this method, and also its cost,
2061 : * depend mainly on the length of the list we are willing to keep.
2062 : */
2063 : static void
2064 2390 : compute_distinct_stats(VacAttrStatsP stats,
2065 : AnalyzeAttrFetchFunc fetchfunc,
2066 : int samplerows,
2067 : double totalrows)
2068 : {
2069 : int i;
2070 2390 : int null_cnt = 0;
2071 2390 : int nonnull_cnt = 0;
2072 2390 : int toowide_cnt = 0;
2073 2390 : double total_width = 0;
2074 4040 : bool is_varlena = (!stats->attrtype->typbyval &&
2075 1650 : stats->attrtype->typlen == -1);
2076 4040 : bool is_varwidth = (!stats->attrtype->typbyval &&
2077 1650 : stats->attrtype->typlen < 0);
2078 : FmgrInfo f_cmpeq;
2079 : typedef struct
2080 : {
2081 : Datum value;
2082 : int count;
2083 : } TrackItem;
2084 : TrackItem *track;
2085 : int track_cnt,
2086 : track_max;
2087 2390 : int num_mcv = stats->attstattarget;
2088 2390 : StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2089 :
2090 : /*
2091 : * We track up to 2*n values for an n-element MCV list; but at least 10
2092 : */
2093 2390 : track_max = 2 * num_mcv;
2094 2390 : if (track_max < 10)
2095 78 : track_max = 10;
2096 2390 : track = (TrackItem *) palloc(track_max * sizeof(TrackItem));
2097 2390 : track_cnt = 0;
2098 :
2099 2390 : fmgr_info(mystats->eqfunc, &f_cmpeq);
2100 :
2101 1689444 : for (i = 0; i < samplerows; i++)
2102 : {
2103 : Datum value;
2104 : bool isnull;
2105 : bool match;
2106 : int firstcount1,
2107 : j;
2108 :
2109 1687054 : vacuum_delay_point(true);
2110 :
2111 1687054 : value = fetchfunc(stats, i, &isnull);
2112 :
2113 : /* Check for null/nonnull */
2114 1687054 : if (isnull)
2115 : {
2116 1408788 : null_cnt++;
2117 1408788 : continue;
2118 : }
2119 278266 : nonnull_cnt++;
2120 :
2121 : /*
2122 : * If it's a variable-width field, add up widths for average width
2123 : * calculation. Note that if the value is toasted, we use the toasted
2124 : * width. We don't bother with this calculation if it's a fixed-width
2125 : * type.
2126 : */
2127 278266 : if (is_varlena)
2128 : {
2129 100202 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2130 :
2131 : /*
2132 : * If the value is toasted, we want to detoast it just once to
2133 : * avoid repeated detoastings and resultant excess memory usage
2134 : * during the comparisons. Also, check to see if the value is
2135 : * excessively wide, and if so don't detoast at all --- just
2136 : * ignore the value.
2137 : */
2138 100202 : if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
2139 : {
2140 0 : toowide_cnt++;
2141 0 : continue;
2142 : }
2143 100202 : value = PointerGetDatum(PG_DETOAST_DATUM(value));
2144 : }
2145 178064 : else if (is_varwidth)
2146 : {
2147 : /* must be cstring */
2148 0 : total_width += strlen(DatumGetCString(value)) + 1;
2149 : }
2150 :
2151 : /*
2152 : * See if the value matches anything we're already tracking.
2153 : */
2154 278266 : match = false;
2155 278266 : firstcount1 = track_cnt;
2156 614828 : for (j = 0; j < track_cnt; j++)
2157 : {
2158 607212 : if (DatumGetBool(FunctionCall2Coll(&f_cmpeq,
2159 : stats->attrcollid,
2160 607212 : value, track[j].value)))
2161 : {
2162 270650 : match = true;
2163 270650 : break;
2164 : }
2165 336562 : if (j < firstcount1 && track[j].count == 1)
2166 5270 : firstcount1 = j;
2167 : }
2168 :
2169 278266 : if (match)
2170 : {
2171 : /* Found a match */
2172 270650 : track[j].count++;
2173 : /* This value may now need to "bubble up" in the track list */
2174 283036 : while (j > 0 && track[j].count > track[j - 1].count)
2175 : {
2176 12386 : swapDatum(track[j].value, track[j - 1].value);
2177 12386 : swapInt(track[j].count, track[j - 1].count);
2178 12386 : j--;
2179 : }
2180 : }
2181 : else
2182 : {
2183 : /* No match. Insert at head of count-1 list */
2184 7616 : if (track_cnt < track_max)
2185 7282 : track_cnt++;
2186 142892 : for (j = track_cnt - 1; j > firstcount1; j--)
2187 : {
2188 135276 : track[j].value = track[j - 1].value;
2189 135276 : track[j].count = track[j - 1].count;
2190 : }
2191 7616 : if (firstcount1 < track_cnt)
2192 : {
2193 7616 : track[firstcount1].value = value;
2194 7616 : track[firstcount1].count = 1;
2195 : }
2196 : }
2197 : }
2198 :
2199 : /* We can only compute real stats if we found some non-null values. */
2200 2390 : if (nonnull_cnt > 0)
2201 : {
2202 : int nmultiple,
2203 : summultiple;
2204 :
2205 1750 : stats->stats_valid = true;
2206 : /* Do the simple null-frac and width stats */
2207 1750 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2208 1750 : if (is_varwidth)
2209 1010 : stats->stawidth = total_width / (double) nonnull_cnt;
2210 : else
2211 740 : stats->stawidth = stats->attrtype->typlen;
2212 :
2213 : /* Count the number of values we found multiple times */
2214 1750 : summultiple = 0;
2215 6630 : for (nmultiple = 0; nmultiple < track_cnt; nmultiple++)
2216 : {
2217 5764 : if (track[nmultiple].count == 1)
2218 884 : break;
2219 4880 : summultiple += track[nmultiple].count;
2220 : }
2221 :
2222 1750 : if (nmultiple == 0)
2223 : {
2224 : /*
2225 : * If we found no repeated non-null values, assume it's a unique
2226 : * column; but be sure to discount for any nulls we found.
2227 : */
2228 200 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2229 : }
2230 1550 : else if (track_cnt < track_max && toowide_cnt == 0 &&
2231 : nmultiple == track_cnt)
2232 : {
2233 : /*
2234 : * Our track list includes every value in the sample, and every
2235 : * value appeared more than once. Assume the column has just
2236 : * these values. (This case is meant to address columns with
2237 : * small, fixed sets of possible values, such as boolean or enum
2238 : * columns. If there are any values that appear just once in the
2239 : * sample, including too-wide values, we should assume that that's
2240 : * not what we're dealing with.)
2241 : */
2242 866 : stats->stadistinct = track_cnt;
2243 : }
2244 : else
2245 : {
2246 : /*----------
2247 : * Estimate the number of distinct values using the estimator
2248 : * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2249 : * n*d / (n - f1 + f1*n/N)
2250 : * where f1 is the number of distinct values that occurred
2251 : * exactly once in our sample of n rows (from a total of N),
2252 : * and d is the total number of distinct values in the sample.
2253 : * This is their Duj1 estimator; the other estimators they
2254 : * recommend are considerably more complex, and are numerically
2255 : * very unstable when n is much smaller than N.
2256 : *
2257 : * In this calculation, we consider only non-nulls. We used to
2258 : * include rows with null values in the n and N counts, but that
2259 : * leads to inaccurate answers in columns with many nulls, and
2260 : * it's intuitively bogus anyway considering the desired result is
2261 : * the number of distinct non-null values.
2262 : *
2263 : * We assume (not very reliably!) that all the multiply-occurring
2264 : * values are reflected in the final track[] list, and the other
2265 : * nonnull values all appeared but once. (XXX this usually
2266 : * results in a drastic overestimate of ndistinct. Can we do
2267 : * any better?)
2268 : *----------
2269 : */
2270 684 : int f1 = nonnull_cnt - summultiple;
2271 684 : int d = f1 + nmultiple;
2272 684 : double n = samplerows - null_cnt;
2273 684 : double N = totalrows * (1.0 - stats->stanullfrac);
2274 : double stadistinct;
2275 :
2276 : /* N == 0 shouldn't happen, but just in case ... */
2277 684 : if (N > 0)
2278 684 : stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2279 : else
2280 0 : stadistinct = 0;
2281 :
2282 : /* Clamp to sane range in case of roundoff error */
2283 684 : if (stadistinct < d)
2284 194 : stadistinct = d;
2285 684 : if (stadistinct > N)
2286 0 : stadistinct = N;
2287 : /* And round to integer */
2288 684 : stats->stadistinct = floor(stadistinct + 0.5);
2289 : }
2290 :
2291 : /*
2292 : * If we estimated the number of distinct values at more than 10% of
2293 : * the total row count (a very arbitrary limit), then assume that
2294 : * stadistinct should scale with the row count rather than be a fixed
2295 : * value.
2296 : */
2297 1750 : if (stats->stadistinct > 0.1 * totalrows)
2298 408 : stats->stadistinct = -(stats->stadistinct / totalrows);
2299 :
2300 : /*
2301 : * Decide how many values are worth storing as most-common values. If
2302 : * we are able to generate a complete MCV list (all the values in the
2303 : * sample will fit, and we think these are all the ones in the table),
2304 : * then do so. Otherwise, store only those values that are
2305 : * significantly more common than the values not in the list.
2306 : *
2307 : * Note: the first of these cases is meant to address columns with
2308 : * small, fixed sets of possible values, such as boolean or enum
2309 : * columns. If we can *completely* represent the column population by
2310 : * an MCV list that will fit into the stats target, then we should do
2311 : * so and thus provide the planner with complete information. But if
2312 : * the MCV list is not complete, it's generally worth being more
2313 : * selective, and not just filling it all the way up to the stats
2314 : * target.
2315 : */
2316 1750 : if (track_cnt < track_max && toowide_cnt == 0 &&
2317 1740 : stats->stadistinct > 0 &&
2318 : track_cnt <= num_mcv)
2319 : {
2320 : /* Track list includes all values seen, and all will fit */
2321 1116 : num_mcv = track_cnt;
2322 : }
2323 : else
2324 : {
2325 : int *mcv_counts;
2326 :
2327 : /* Incomplete list; decide how many values are worth keeping */
2328 634 : if (num_mcv > track_cnt)
2329 574 : num_mcv = track_cnt;
2330 :
2331 634 : if (num_mcv > 0)
2332 : {
2333 634 : mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2334 2256 : for (i = 0; i < num_mcv; i++)
2335 1622 : mcv_counts[i] = track[i].count;
2336 :
2337 634 : num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2338 634 : stats->stadistinct,
2339 634 : stats->stanullfrac,
2340 : samplerows, totalrows);
2341 : }
2342 : }
2343 :
2344 : /* Generate MCV slot entry */
2345 1750 : if (num_mcv > 0)
2346 : {
2347 : MemoryContext old_context;
2348 : Datum *mcv_values;
2349 : float4 *mcv_freqs;
2350 :
2351 : /* Must copy the target values into anl_context */
2352 1746 : old_context = MemoryContextSwitchTo(stats->anl_context);
2353 1746 : mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2354 1746 : mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2355 8414 : for (i = 0; i < num_mcv; i++)
2356 : {
2357 13336 : mcv_values[i] = datumCopy(track[i].value,
2358 6668 : stats->attrtype->typbyval,
2359 6668 : stats->attrtype->typlen);
2360 6668 : mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2361 : }
2362 1746 : MemoryContextSwitchTo(old_context);
2363 :
2364 1746 : stats->stakind[0] = STATISTIC_KIND_MCV;
2365 1746 : stats->staop[0] = mystats->eqopr;
2366 1746 : stats->stacoll[0] = stats->attrcollid;
2367 1746 : stats->stanumbers[0] = mcv_freqs;
2368 1746 : stats->numnumbers[0] = num_mcv;
2369 1746 : stats->stavalues[0] = mcv_values;
2370 1746 : stats->numvalues[0] = num_mcv;
2371 :
2372 : /*
2373 : * Accept the defaults for stats->statypid and others. They have
2374 : * been set before we were called (see vacuum.h)
2375 : */
2376 : }
2377 : }
2378 640 : else if (null_cnt > 0)
2379 : {
2380 : /* We found only nulls; assume the column is entirely null */
2381 640 : stats->stats_valid = true;
2382 640 : stats->stanullfrac = 1.0;
2383 640 : if (is_varwidth)
2384 640 : stats->stawidth = 0; /* "unknown" */
2385 : else
2386 0 : stats->stawidth = stats->attrtype->typlen;
2387 640 : stats->stadistinct = 0.0; /* "unknown" */
2388 : }
2389 :
2390 : /* We don't need to bother cleaning up any of our temporary palloc's */
2391 2390 : }
2392 :
2393 :
2394 : /*
2395 : * compute_scalar_stats() -- compute column statistics
2396 : *
2397 : * We use this when we can find "=" and "<" operators for the datatype.
2398 : *
2399 : * We determine the fraction of non-null rows, the average width, the
2400 : * most common values, the (estimated) number of distinct values, the
2401 : * distribution histogram, and the correlation of physical to logical order.
2402 : *
2403 : * The desired stats can be determined fairly easily after sorting the
2404 : * data values into order.
2405 : */
2406 : static void
2407 86316 : compute_scalar_stats(VacAttrStatsP stats,
2408 : AnalyzeAttrFetchFunc fetchfunc,
2409 : int samplerows,
2410 : double totalrows)
2411 : {
2412 : int i;
2413 86316 : int null_cnt = 0;
2414 86316 : int nonnull_cnt = 0;
2415 86316 : int toowide_cnt = 0;
2416 86316 : double total_width = 0;
2417 107260 : bool is_varlena = (!stats->attrtype->typbyval &&
2418 20944 : stats->attrtype->typlen == -1);
2419 107260 : bool is_varwidth = (!stats->attrtype->typbyval &&
2420 20944 : stats->attrtype->typlen < 0);
2421 : double corr_xysum;
2422 : SortSupportData ssup;
2423 : ScalarItem *values;
2424 86316 : int values_cnt = 0;
2425 : int *tupnoLink;
2426 : ScalarMCVItem *track;
2427 86316 : int track_cnt = 0;
2428 86316 : int num_mcv = stats->attstattarget;
2429 86316 : int num_bins = stats->attstattarget;
2430 86316 : StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2431 :
2432 86316 : values = (ScalarItem *) palloc(samplerows * sizeof(ScalarItem));
2433 86316 : tupnoLink = (int *) palloc(samplerows * sizeof(int));
2434 86316 : track = (ScalarMCVItem *) palloc(num_mcv * sizeof(ScalarMCVItem));
2435 :
2436 86316 : memset(&ssup, 0, sizeof(ssup));
2437 86316 : ssup.ssup_cxt = CurrentMemoryContext;
2438 86316 : ssup.ssup_collation = stats->attrcollid;
2439 86316 : ssup.ssup_nulls_first = false;
2440 :
2441 : /*
2442 : * For now, don't perform abbreviated key conversion, because full values
2443 : * are required for MCV slot generation. Supporting that optimization
2444 : * would necessitate teaching compare_scalars() to call a tie-breaker.
2445 : */
2446 86316 : ssup.abbreviate = false;
2447 :
2448 86316 : PrepareSortSupportFromOrderingOp(mystats->ltopr, &ssup);
2449 :
2450 : /* Initial scan to find sortable values */
2451 76966110 : for (i = 0; i < samplerows; i++)
2452 : {
2453 : Datum value;
2454 : bool isnull;
2455 :
2456 76879794 : vacuum_delay_point(true);
2457 :
2458 76879794 : value = fetchfunc(stats, i, &isnull);
2459 :
2460 : /* Check for null/nonnull */
2461 76879794 : if (isnull)
2462 : {
2463 10122458 : null_cnt++;
2464 10161124 : continue;
2465 : }
2466 66757336 : nonnull_cnt++;
2467 :
2468 : /*
2469 : * If it's a variable-width field, add up widths for average width
2470 : * calculation. Note that if the value is toasted, we use the toasted
2471 : * width. We don't bother with this calculation if it's a fixed-width
2472 : * type.
2473 : */
2474 66757336 : if (is_varlena)
2475 : {
2476 8129132 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2477 :
2478 : /*
2479 : * If the value is toasted, we want to detoast it just once to
2480 : * avoid repeated detoastings and resultant excess memory usage
2481 : * during the comparisons. Also, check to see if the value is
2482 : * excessively wide, and if so don't detoast at all --- just
2483 : * ignore the value.
2484 : */
2485 8129132 : if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
2486 : {
2487 38666 : toowide_cnt++;
2488 38666 : continue;
2489 : }
2490 8090466 : value = PointerGetDatum(PG_DETOAST_DATUM(value));
2491 : }
2492 58628204 : else if (is_varwidth)
2493 : {
2494 : /* must be cstring */
2495 0 : total_width += strlen(DatumGetCString(value)) + 1;
2496 : }
2497 :
2498 : /* Add it to the list to be sorted */
2499 66718670 : values[values_cnt].value = value;
2500 66718670 : values[values_cnt].tupno = values_cnt;
2501 66718670 : tupnoLink[values_cnt] = values_cnt;
2502 66718670 : values_cnt++;
2503 : }
2504 :
2505 : /* We can only compute real stats if we found some sortable values. */
2506 86316 : if (values_cnt > 0)
2507 : {
2508 : int ndistinct, /* # distinct values in sample */
2509 : nmultiple, /* # that appear multiple times */
2510 : num_hist,
2511 : dups_cnt;
2512 80634 : int slot_idx = 0;
2513 : CompareScalarsContext cxt;
2514 :
2515 : /* Sort the collected values */
2516 80634 : cxt.ssup = &ssup;
2517 80634 : cxt.tupnoLink = tupnoLink;
2518 80634 : qsort_interruptible(values, values_cnt, sizeof(ScalarItem),
2519 : compare_scalars, &cxt);
2520 :
2521 : /*
2522 : * Now scan the values in order, find the most common ones, and also
2523 : * accumulate ordering-correlation statistics.
2524 : *
2525 : * To determine which are most common, we first have to count the
2526 : * number of duplicates of each value. The duplicates are adjacent in
2527 : * the sorted list, so a brute-force approach is to compare successive
2528 : * datum values until we find two that are not equal. However, that
2529 : * requires N-1 invocations of the datum comparison routine, which are
2530 : * completely redundant with work that was done during the sort. (The
2531 : * sort algorithm must at some point have compared each pair of items
2532 : * that are adjacent in the sorted order; otherwise it could not know
2533 : * that it's ordered the pair correctly.) We exploit this by having
2534 : * compare_scalars remember the highest tupno index that each
2535 : * ScalarItem has been found equal to. At the end of the sort, a
2536 : * ScalarItem's tupnoLink will still point to itself if and only if it
2537 : * is the last item of its group of duplicates (since the group will
2538 : * be ordered by tupno).
2539 : */
2540 80634 : corr_xysum = 0;
2541 80634 : ndistinct = 0;
2542 80634 : nmultiple = 0;
2543 80634 : dups_cnt = 0;
2544 66799304 : for (i = 0; i < values_cnt; i++)
2545 : {
2546 66718670 : int tupno = values[i].tupno;
2547 :
2548 66718670 : corr_xysum += ((double) i) * ((double) tupno);
2549 66718670 : dups_cnt++;
2550 66718670 : if (tupnoLink[tupno] == tupno)
2551 : {
2552 : /* Reached end of duplicates of this value */
2553 13536058 : ndistinct++;
2554 13536058 : if (dups_cnt > 1)
2555 : {
2556 1187124 : nmultiple++;
2557 1187124 : if (track_cnt < num_mcv ||
2558 480992 : dups_cnt > track[track_cnt - 1].count)
2559 : {
2560 : /*
2561 : * Found a new item for the mcv list; find its
2562 : * position, bubbling down old items if needed. Loop
2563 : * invariant is that j points at an empty/ replaceable
2564 : * slot.
2565 : */
2566 : int j;
2567 :
2568 813824 : if (track_cnt < num_mcv)
2569 706132 : track_cnt++;
2570 10936308 : for (j = track_cnt - 1; j > 0; j--)
2571 : {
2572 10845796 : if (dups_cnt <= track[j - 1].count)
2573 723312 : break;
2574 10122484 : track[j].count = track[j - 1].count;
2575 10122484 : track[j].first = track[j - 1].first;
2576 : }
2577 813824 : track[j].count = dups_cnt;
2578 813824 : track[j].first = i + 1 - dups_cnt;
2579 : }
2580 : }
2581 13536058 : dups_cnt = 0;
2582 : }
2583 : }
2584 :
2585 80634 : stats->stats_valid = true;
2586 : /* Do the simple null-frac and width stats */
2587 80634 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2588 80634 : if (is_varwidth)
2589 11578 : stats->stawidth = total_width / (double) nonnull_cnt;
2590 : else
2591 69056 : stats->stawidth = stats->attrtype->typlen;
2592 :
2593 80634 : if (nmultiple == 0)
2594 : {
2595 : /*
2596 : * If we found no repeated non-null values, assume it's a unique
2597 : * column; but be sure to discount for any nulls we found.
2598 : */
2599 20812 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2600 : }
2601 59822 : else if (toowide_cnt == 0 && nmultiple == ndistinct)
2602 : {
2603 : /*
2604 : * Every value in the sample appeared more than once. Assume the
2605 : * column has just these values. (This case is meant to address
2606 : * columns with small, fixed sets of possible values, such as
2607 : * boolean or enum columns. If there are any values that appear
2608 : * just once in the sample, including too-wide values, we should
2609 : * assume that that's not what we're dealing with.)
2610 : */
2611 36678 : stats->stadistinct = ndistinct;
2612 : }
2613 : else
2614 : {
2615 : /*----------
2616 : * Estimate the number of distinct values using the estimator
2617 : * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2618 : * n*d / (n - f1 + f1*n/N)
2619 : * where f1 is the number of distinct values that occurred
2620 : * exactly once in our sample of n rows (from a total of N),
2621 : * and d is the total number of distinct values in the sample.
2622 : * This is their Duj1 estimator; the other estimators they
2623 : * recommend are considerably more complex, and are numerically
2624 : * very unstable when n is much smaller than N.
2625 : *
2626 : * In this calculation, we consider only non-nulls. We used to
2627 : * include rows with null values in the n and N counts, but that
2628 : * leads to inaccurate answers in columns with many nulls, and
2629 : * it's intuitively bogus anyway considering the desired result is
2630 : * the number of distinct non-null values.
2631 : *
2632 : * Overwidth values are assumed to have been distinct.
2633 : *----------
2634 : */
2635 23144 : int f1 = ndistinct - nmultiple + toowide_cnt;
2636 23144 : int d = f1 + nmultiple;
2637 23144 : double n = samplerows - null_cnt;
2638 23144 : double N = totalrows * (1.0 - stats->stanullfrac);
2639 : double stadistinct;
2640 :
2641 : /* N == 0 shouldn't happen, but just in case ... */
2642 23144 : if (N > 0)
2643 23144 : stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2644 : else
2645 0 : stadistinct = 0;
2646 :
2647 : /* Clamp to sane range in case of roundoff error */
2648 23144 : if (stadistinct < d)
2649 934 : stadistinct = d;
2650 23144 : if (stadistinct > N)
2651 0 : stadistinct = N;
2652 : /* And round to integer */
2653 23144 : stats->stadistinct = floor(stadistinct + 0.5);
2654 : }
2655 :
2656 : /*
2657 : * If we estimated the number of distinct values at more than 10% of
2658 : * the total row count (a very arbitrary limit), then assume that
2659 : * stadistinct should scale with the row count rather than be a fixed
2660 : * value.
2661 : */
2662 80634 : if (stats->stadistinct > 0.1 * totalrows)
2663 17152 : stats->stadistinct = -(stats->stadistinct / totalrows);
2664 :
2665 : /*
2666 : * Decide how many values are worth storing as most-common values. If
2667 : * we are able to generate a complete MCV list (all the values in the
2668 : * sample will fit, and we think these are all the ones in the table),
2669 : * then do so. Otherwise, store only those values that are
2670 : * significantly more common than the values not in the list.
2671 : *
2672 : * Note: the first of these cases is meant to address columns with
2673 : * small, fixed sets of possible values, such as boolean or enum
2674 : * columns. If we can *completely* represent the column population by
2675 : * an MCV list that will fit into the stats target, then we should do
2676 : * so and thus provide the planner with complete information. But if
2677 : * the MCV list is not complete, it's generally worth being more
2678 : * selective, and not just filling it all the way up to the stats
2679 : * target.
2680 : */
2681 80634 : if (track_cnt == ndistinct && toowide_cnt == 0 &&
2682 35984 : stats->stadistinct > 0 &&
2683 : track_cnt <= num_mcv)
2684 : {
2685 : /* Track list includes all values seen, and all will fit */
2686 32230 : num_mcv = track_cnt;
2687 : }
2688 : else
2689 : {
2690 : int *mcv_counts;
2691 :
2692 : /* Incomplete list; decide how many values are worth keeping */
2693 48404 : if (num_mcv > track_cnt)
2694 44356 : num_mcv = track_cnt;
2695 :
2696 48404 : if (num_mcv > 0)
2697 : {
2698 27592 : mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2699 564658 : for (i = 0; i < num_mcv; i++)
2700 537066 : mcv_counts[i] = track[i].count;
2701 :
2702 27592 : num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2703 27592 : stats->stadistinct,
2704 27592 : stats->stanullfrac,
2705 : samplerows, totalrows);
2706 : }
2707 : }
2708 :
2709 : /* Generate MCV slot entry */
2710 80634 : if (num_mcv > 0)
2711 : {
2712 : MemoryContext old_context;
2713 : Datum *mcv_values;
2714 : float4 *mcv_freqs;
2715 :
2716 : /* Must copy the target values into anl_context */
2717 59778 : old_context = MemoryContextSwitchTo(stats->anl_context);
2718 59778 : mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2719 59778 : mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2720 765686 : for (i = 0; i < num_mcv; i++)
2721 : {
2722 1411816 : mcv_values[i] = datumCopy(values[track[i].first].value,
2723 705908 : stats->attrtype->typbyval,
2724 705908 : stats->attrtype->typlen);
2725 705908 : mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2726 : }
2727 59778 : MemoryContextSwitchTo(old_context);
2728 :
2729 59778 : stats->stakind[slot_idx] = STATISTIC_KIND_MCV;
2730 59778 : stats->staop[slot_idx] = mystats->eqopr;
2731 59778 : stats->stacoll[slot_idx] = stats->attrcollid;
2732 59778 : stats->stanumbers[slot_idx] = mcv_freqs;
2733 59778 : stats->numnumbers[slot_idx] = num_mcv;
2734 59778 : stats->stavalues[slot_idx] = mcv_values;
2735 59778 : stats->numvalues[slot_idx] = num_mcv;
2736 :
2737 : /*
2738 : * Accept the defaults for stats->statypid and others. They have
2739 : * been set before we were called (see vacuum.h)
2740 : */
2741 59778 : slot_idx++;
2742 : }
2743 :
2744 : /*
2745 : * Generate a histogram slot entry if there are at least two distinct
2746 : * values not accounted for in the MCV list. (This ensures the
2747 : * histogram won't collapse to empty or a singleton.)
2748 : */
2749 80634 : num_hist = ndistinct - num_mcv;
2750 80634 : if (num_hist > num_bins)
2751 13002 : num_hist = num_bins + 1;
2752 80634 : if (num_hist >= 2)
2753 : {
2754 : MemoryContext old_context;
2755 : Datum *hist_values;
2756 : int nvals;
2757 : int pos,
2758 : posfrac,
2759 : delta,
2760 : deltafrac;
2761 :
2762 : /* Sort the MCV items into position order to speed next loop */
2763 35442 : qsort_interruptible(track, num_mcv, sizeof(ScalarMCVItem),
2764 : compare_mcvs, NULL);
2765 :
2766 : /*
2767 : * Collapse out the MCV items from the values[] array.
2768 : *
2769 : * Note we destroy the values[] array here... but we don't need it
2770 : * for anything more. We do, however, still need values_cnt.
2771 : * nvals will be the number of remaining entries in values[].
2772 : */
2773 35442 : if (num_mcv > 0)
2774 : {
2775 : int src,
2776 : dest;
2777 : int j;
2778 :
2779 19368 : src = dest = 0;
2780 19368 : j = 0; /* index of next interesting MCV item */
2781 700454 : while (src < values_cnt)
2782 : {
2783 : int ncopy;
2784 :
2785 681086 : if (j < num_mcv)
2786 : {
2787 666368 : int first = track[j].first;
2788 :
2789 666368 : if (src >= first)
2790 : {
2791 : /* advance past this MCV item */
2792 485914 : src = first + track[j].count;
2793 485914 : j++;
2794 485914 : continue;
2795 : }
2796 180454 : ncopy = first - src;
2797 : }
2798 : else
2799 14718 : ncopy = values_cnt - src;
2800 195172 : memmove(&values[dest], &values[src],
2801 : ncopy * sizeof(ScalarItem));
2802 195172 : src += ncopy;
2803 195172 : dest += ncopy;
2804 : }
2805 19368 : nvals = dest;
2806 : }
2807 : else
2808 16074 : nvals = values_cnt;
2809 : Assert(nvals >= num_hist);
2810 :
2811 : /* Must copy the target values into anl_context */
2812 35442 : old_context = MemoryContextSwitchTo(stats->anl_context);
2813 35442 : hist_values = (Datum *) palloc(num_hist * sizeof(Datum));
2814 :
2815 : /*
2816 : * The object of this loop is to copy the first and last values[]
2817 : * entries along with evenly-spaced values in between. So the
2818 : * i'th value is values[(i * (nvals - 1)) / (num_hist - 1)]. But
2819 : * computing that subscript directly risks integer overflow when
2820 : * the stats target is more than a couple thousand. Instead we
2821 : * add (nvals - 1) / (num_hist - 1) to pos at each step, tracking
2822 : * the integral and fractional parts of the sum separately.
2823 : */
2824 35442 : delta = (nvals - 1) / (num_hist - 1);
2825 35442 : deltafrac = (nvals - 1) % (num_hist - 1);
2826 35442 : pos = posfrac = 0;
2827 :
2828 1844158 : for (i = 0; i < num_hist; i++)
2829 : {
2830 3617432 : hist_values[i] = datumCopy(values[pos].value,
2831 1808716 : stats->attrtype->typbyval,
2832 1808716 : stats->attrtype->typlen);
2833 1808716 : pos += delta;
2834 1808716 : posfrac += deltafrac;
2835 1808716 : if (posfrac >= (num_hist - 1))
2836 : {
2837 : /* fractional part exceeds 1, carry to integer part */
2838 588872 : pos++;
2839 588872 : posfrac -= (num_hist - 1);
2840 : }
2841 : }
2842 :
2843 35442 : MemoryContextSwitchTo(old_context);
2844 :
2845 35442 : stats->stakind[slot_idx] = STATISTIC_KIND_HISTOGRAM;
2846 35442 : stats->staop[slot_idx] = mystats->ltopr;
2847 35442 : stats->stacoll[slot_idx] = stats->attrcollid;
2848 35442 : stats->stavalues[slot_idx] = hist_values;
2849 35442 : stats->numvalues[slot_idx] = num_hist;
2850 :
2851 : /*
2852 : * Accept the defaults for stats->statypid and others. They have
2853 : * been set before we were called (see vacuum.h)
2854 : */
2855 35442 : slot_idx++;
2856 : }
2857 :
2858 : /* Generate a correlation entry if there are multiple values */
2859 80634 : if (values_cnt > 1)
2860 : {
2861 : MemoryContext old_context;
2862 : float4 *corrs;
2863 : double corr_xsum,
2864 : corr_x2sum;
2865 :
2866 : /* Must copy the target values into anl_context */
2867 75852 : old_context = MemoryContextSwitchTo(stats->anl_context);
2868 75852 : corrs = palloc_object(float4);
2869 75852 : MemoryContextSwitchTo(old_context);
2870 :
2871 : /*----------
2872 : * Since we know the x and y value sets are both
2873 : * 0, 1, ..., values_cnt-1
2874 : * we have sum(x) = sum(y) =
2875 : * (values_cnt-1)*values_cnt / 2
2876 : * and sum(x^2) = sum(y^2) =
2877 : * (values_cnt-1)*values_cnt*(2*values_cnt-1) / 6.
2878 : *----------
2879 : */
2880 75852 : corr_xsum = ((double) (values_cnt - 1)) *
2881 75852 : ((double) values_cnt) / 2.0;
2882 75852 : corr_x2sum = ((double) (values_cnt - 1)) *
2883 75852 : ((double) values_cnt) * (double) (2 * values_cnt - 1) / 6.0;
2884 :
2885 : /* And the correlation coefficient reduces to */
2886 75852 : corrs[0] = (values_cnt * corr_xysum - corr_xsum * corr_xsum) /
2887 75852 : (values_cnt * corr_x2sum - corr_xsum * corr_xsum);
2888 :
2889 75852 : stats->stakind[slot_idx] = STATISTIC_KIND_CORRELATION;
2890 75852 : stats->staop[slot_idx] = mystats->ltopr;
2891 75852 : stats->stacoll[slot_idx] = stats->attrcollid;
2892 75852 : stats->stanumbers[slot_idx] = corrs;
2893 75852 : stats->numnumbers[slot_idx] = 1;
2894 75852 : slot_idx++;
2895 : }
2896 : }
2897 5682 : else if (nonnull_cnt > 0)
2898 : {
2899 : /* We found some non-null values, but they were all too wide */
2900 : Assert(nonnull_cnt == toowide_cnt);
2901 364 : stats->stats_valid = true;
2902 : /* Do the simple null-frac and width stats */
2903 364 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2904 364 : if (is_varwidth)
2905 364 : stats->stawidth = total_width / (double) nonnull_cnt;
2906 : else
2907 0 : stats->stawidth = stats->attrtype->typlen;
2908 : /* Assume all too-wide values are distinct, so it's a unique column */
2909 364 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2910 : }
2911 5318 : else if (null_cnt > 0)
2912 : {
2913 : /* We found only nulls; assume the column is entirely null */
2914 5318 : stats->stats_valid = true;
2915 5318 : stats->stanullfrac = 1.0;
2916 5318 : if (is_varwidth)
2917 4632 : stats->stawidth = 0; /* "unknown" */
2918 : else
2919 686 : stats->stawidth = stats->attrtype->typlen;
2920 5318 : stats->stadistinct = 0.0; /* "unknown" */
2921 : }
2922 :
2923 : /* We don't need to bother cleaning up any of our temporary palloc's */
2924 86316 : }
2925 :
2926 : /*
2927 : * Comparator for sorting ScalarItems
2928 : *
2929 : * Aside from sorting the items, we update the tupnoLink[] array
2930 : * whenever two ScalarItems are found to contain equal datums. The array
2931 : * is indexed by tupno; for each ScalarItem, it contains the highest
2932 : * tupno that that item's datum has been found to be equal to. This allows
2933 : * us to avoid additional comparisons in compute_scalar_stats().
2934 : */
2935 : static int
2936 618404850 : compare_scalars(const void *a, const void *b, void *arg)
2937 : {
2938 618404850 : Datum da = ((const ScalarItem *) a)->value;
2939 618404850 : int ta = ((const ScalarItem *) a)->tupno;
2940 618404850 : Datum db = ((const ScalarItem *) b)->value;
2941 618404850 : int tb = ((const ScalarItem *) b)->tupno;
2942 618404850 : CompareScalarsContext *cxt = (CompareScalarsContext *) arg;
2943 : int compare;
2944 :
2945 618404850 : compare = ApplySortComparator(da, false, db, false, cxt->ssup);
2946 618404850 : if (compare != 0)
2947 231178546 : return compare;
2948 :
2949 : /*
2950 : * The two datums are equal, so update cxt->tupnoLink[].
2951 : */
2952 387226304 : if (cxt->tupnoLink[ta] < tb)
2953 55691846 : cxt->tupnoLink[ta] = tb;
2954 387226304 : if (cxt->tupnoLink[tb] < ta)
2955 3894134 : cxt->tupnoLink[tb] = ta;
2956 :
2957 : /*
2958 : * For equal datums, sort by tupno
2959 : */
2960 387226304 : return ta - tb;
2961 : }
2962 :
2963 : /*
2964 : * Comparator for sorting ScalarMCVItems by position
2965 : */
2966 : static int
2967 2555806 : compare_mcvs(const void *a, const void *b, void *arg)
2968 : {
2969 2555806 : int da = ((const ScalarMCVItem *) a)->first;
2970 2555806 : int db = ((const ScalarMCVItem *) b)->first;
2971 :
2972 2555806 : return da - db;
2973 : }
2974 :
2975 : /*
2976 : * Analyze the list of common values in the sample and decide how many are
2977 : * worth storing in the table's MCV list.
2978 : *
2979 : * mcv_counts is assumed to be a list of the counts of the most common values
2980 : * seen in the sample, starting with the most common. The return value is the
2981 : * number that are significantly more common than the values not in the list,
2982 : * and which are therefore deemed worth storing in the table's MCV list.
2983 : */
2984 : static int
2985 28226 : analyze_mcv_list(int *mcv_counts,
2986 : int num_mcv,
2987 : double stadistinct,
2988 : double stanullfrac,
2989 : int samplerows,
2990 : double totalrows)
2991 : {
2992 : double ndistinct_table;
2993 : double sumcount;
2994 : int i;
2995 :
2996 : /*
2997 : * If the entire table was sampled, keep the whole list. This also
2998 : * protects us against division by zero in the code below.
2999 : */
3000 28226 : if (samplerows == totalrows || totalrows <= 1.0)
3001 27392 : return num_mcv;
3002 :
3003 : /* Re-extract the estimated number of distinct nonnull values in table */
3004 834 : ndistinct_table = stadistinct;
3005 834 : if (ndistinct_table < 0)
3006 152 : ndistinct_table = -ndistinct_table * totalrows;
3007 :
3008 : /*
3009 : * Exclude the least common values from the MCV list, if they are not
3010 : * significantly more common than the estimated selectivity they would
3011 : * have if they weren't in the list. All non-MCV values are assumed to be
3012 : * equally common, after taking into account the frequencies of all the
3013 : * values in the MCV list and the number of nulls (c.f. eqsel()).
3014 : *
3015 : * Here sumcount tracks the total count of all but the last (least common)
3016 : * value in the MCV list, allowing us to determine the effect of excluding
3017 : * that value from the list.
3018 : *
3019 : * Note that we deliberately do this by removing values from the full
3020 : * list, rather than starting with an empty list and adding values,
3021 : * because the latter approach can fail to add any values if all the most
3022 : * common values have around the same frequency and make up the majority
3023 : * of the table, so that the overall average frequency of all values is
3024 : * roughly the same as that of the common values. This would lead to any
3025 : * uncommon values being significantly overestimated.
3026 : */
3027 834 : sumcount = 0.0;
3028 1750 : for (i = 0; i < num_mcv - 1; i++)
3029 916 : sumcount += mcv_counts[i];
3030 :
3031 1014 : while (num_mcv > 0)
3032 : {
3033 : double selec,
3034 : otherdistinct,
3035 : N,
3036 : n,
3037 : K,
3038 : variance,
3039 : stddev;
3040 :
3041 : /*
3042 : * Estimated selectivity the least common value would have if it
3043 : * wasn't in the MCV list (c.f. eqsel()).
3044 : */
3045 1014 : selec = 1.0 - sumcount / samplerows - stanullfrac;
3046 1014 : if (selec < 0.0)
3047 0 : selec = 0.0;
3048 1014 : if (selec > 1.0)
3049 0 : selec = 1.0;
3050 1014 : otherdistinct = ndistinct_table - (num_mcv - 1);
3051 1014 : if (otherdistinct > 1)
3052 1014 : selec /= otherdistinct;
3053 :
3054 : /*
3055 : * If the value is kept in the MCV list, its population frequency is
3056 : * assumed to equal its sample frequency. We use the lower end of a
3057 : * textbook continuity-corrected Wald-type confidence interval to
3058 : * determine if that is significantly more common than the non-MCV
3059 : * frequency --- specifically we assume the population frequency is
3060 : * highly likely to be within around 2 standard errors of the sample
3061 : * frequency, which equates to an interval of 2 standard deviations
3062 : * either side of the sample count, plus an additional 0.5 for the
3063 : * continuity correction. Since we are sampling without replacement,
3064 : * this is a hypergeometric distribution.
3065 : *
3066 : * XXX: Empirically, this approach seems to work quite well, but it
3067 : * may be worth considering more advanced techniques for estimating
3068 : * the confidence interval of the hypergeometric distribution.
3069 : */
3070 1014 : N = totalrows;
3071 1014 : n = samplerows;
3072 1014 : K = N * mcv_counts[num_mcv - 1] / n;
3073 1014 : variance = n * K * (N - K) * (N - n) / (N * N * (N - 1));
3074 1014 : stddev = sqrt(variance);
3075 :
3076 1014 : if (mcv_counts[num_mcv - 1] > selec * samplerows + 2 * stddev + 0.5)
3077 : {
3078 : /*
3079 : * The value is significantly more common than the non-MCV
3080 : * selectivity would suggest. Keep it, and all the other more
3081 : * common values in the list.
3082 : */
3083 786 : break;
3084 : }
3085 : else
3086 : {
3087 : /* Discard this value and consider the next least common value */
3088 228 : num_mcv--;
3089 228 : if (num_mcv == 0)
3090 48 : break;
3091 180 : sumcount -= mcv_counts[num_mcv - 1];
3092 : }
3093 : }
3094 834 : return num_mcv;
3095 : }
|
