Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * analyze.c
4 : * the Postgres statistics generator
5 : *
6 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/commands/analyze.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 : #include "postgres.h"
16 :
17 : #include <math.h>
18 :
19 : #include "access/detoast.h"
20 : #include "access/genam.h"
21 : #include "access/multixact.h"
22 : #include "access/relation.h"
23 : #include "access/table.h"
24 : #include "access/tableam.h"
25 : #include "access/transam.h"
26 : #include "access/tupconvert.h"
27 : #include "access/visibilitymap.h"
28 : #include "access/xact.h"
29 : #include "catalog/index.h"
30 : #include "catalog/indexing.h"
31 : #include "catalog/pg_inherits.h"
32 : #include "commands/progress.h"
33 : #include "commands/tablecmds.h"
34 : #include "commands/vacuum.h"
35 : #include "common/pg_prng.h"
36 : #include "executor/executor.h"
37 : #include "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 8390 : 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 8390 : AcquireSampleRowsFunc acquirefunc = NULL;
115 8390 : BlockNumber relpages = 0;
116 :
117 : /* Select logging level */
118 8390 : if (params.options & VACOPT_VERBOSE)
119 0 : elevel = INFO;
120 : else
121 8390 : elevel = DEBUG2;
122 :
123 : /* Set up static variables */
124 8390 : vac_strategy = bstrategy;
125 :
126 : /*
127 : * Check for user-requested abort.
128 : */
129 8390 : 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 8390 : onerel = vacuum_open_relation(relid, relation, params.options & ~(VACOPT_VACUUM),
141 8390 : params.log_analyze_min_duration >= 0,
142 : ShareUpdateExclusiveLock);
143 :
144 : /* leave if relation could not be opened or locked */
145 8390 : if (!onerel)
146 108 : 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 8386 : if (!vacuum_is_permitted_for_relation(RelationGetRelid(onerel),
156 : onerel->rd_rel,
157 8386 : params.options & ~VACOPT_VACUUM))
158 : {
159 18 : relation_close(onerel, ShareUpdateExclusiveLock);
160 18 : 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 8368 : 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 8368 : if (RelationGetRelid(onerel) == StatisticRelationId)
179 : {
180 86 : relation_close(onerel, ShareUpdateExclusiveLock);
181 86 : return;
182 : }
183 :
184 : /*
185 : * Check that it's of an analyzable relkind, and set up appropriately.
186 : */
187 8282 : if (onerel->rd_rel->relkind == RELKIND_RELATION ||
188 419 : onerel->rd_rel->relkind == RELKIND_MATVIEW)
189 : {
190 : /* Regular table, so we'll use the regular row acquisition function */
191 7865 : acquirefunc = acquire_sample_rows;
192 : /* Also get regular table's size */
193 7865 : relpages = RelationGetNumberOfBlocks(onerel);
194 : }
195 417 : 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 33 : bool ok = false;
203 :
204 33 : fdwroutine = GetFdwRoutineForRelation(onerel, false);
205 :
206 33 : if (fdwroutine->AnalyzeForeignTable != NULL)
207 33 : ok = fdwroutine->AnalyzeForeignTable(onerel,
208 : &acquirefunc,
209 : &relpages);
210 :
211 33 : 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 : }
220 384 : 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 8282 : pgstat_progress_start_command(PROGRESS_COMMAND_ANALYZE,
241 : RelationGetRelid(onerel));
242 8282 : if (AmAutoVacuumWorkerProcess())
243 218 : pgstat_progress_update_param(PROGRESS_ANALYZE_STARTED_BY,
244 : PROGRESS_ANALYZE_STARTED_BY_AUTOVACUUM);
245 : else
246 8064 : 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 8282 : if (onerel->rd_rel->relkind != RELKIND_PARTITIONED_TABLE)
254 7898 : 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 8262 : if (onerel->rd_rel->relhassubclass)
261 448 : 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 8253 : relation_close(onerel, NoLock);
271 :
272 8253 : 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 8346 : 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 8346 : TimestampTz starttime = 0;
307 : MemoryContext caller_context;
308 : Oid save_userid;
309 : int save_sec_context;
310 : int save_nestlevel;
311 8346 : WalUsage startwalusage = pgWalUsage;
312 8346 : BufferUsage startbufferusage = pgBufferUsage;
313 : BufferUsage bufferusage;
314 8346 : PgStat_Counter startreadtime = 0;
315 8346 : PgStat_Counter startwritetime = 0;
316 :
317 8346 : verbose = (params.options & VACOPT_VERBOSE) != 0;
318 8565 : instrument = (verbose || (AmAutoVacuumWorkerProcess() &&
319 219 : params.log_analyze_min_duration >= 0));
320 8346 : if (inh)
321 448 : ereport(elevel,
322 : (errmsg("analyzing \"%s.%s\" inheritance tree",
323 : get_namespace_name(RelationGetNamespace(onerel)),
324 : RelationGetRelationName(onerel))));
325 : else
326 7898 : 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 8346 : anl_context = AllocSetContextCreate(CurrentMemoryContext,
336 : "Analyze",
337 : ALLOCSET_DEFAULT_SIZES);
338 8346 : 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 8346 : GetUserIdAndSecContext(&save_userid, &save_sec_context);
346 8346 : SetUserIdAndSecContext(onerel->rd_rel->relowner,
347 : save_sec_context | SECURITY_RESTRICTED_OPERATION);
348 8346 : save_nestlevel = NewGUCNestLevel();
349 8346 : 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 8346 : if (instrument)
356 : {
357 219 : if (track_io_timing)
358 : {
359 0 : startreadtime = pgStatBlockReadTime;
360 0 : startwritetime = pgStatBlockWriteTime;
361 : }
362 :
363 219 : pg_rusage_init(&ru0);
364 : }
365 :
366 : /* Used for instrumentation and stats report */
367 8346 : 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 8346 : if (va_cols != NIL)
378 : {
379 50 : Bitmapset *unique_cols = NULL;
380 : ListCell *le;
381 :
382 50 : vacattrstats = (VacAttrStats **) palloc(list_length(va_cols) *
383 : sizeof(VacAttrStats *));
384 50 : tcnt = 0;
385 91 : foreach(le, va_cols)
386 : {
387 66 : char *col = strVal(lfirst(le));
388 :
389 66 : i = attnameAttNum(onerel, col, false);
390 66 : if (i == InvalidAttrNumber)
391 19 : ereport(ERROR,
392 : (errcode(ERRCODE_UNDEFINED_COLUMN),
393 : errmsg("column \"%s\" of relation \"%s\" does not exist",
394 : col, RelationGetRelationName(onerel))));
395 47 : if (bms_is_member(i, unique_cols))
396 6 : ereport(ERROR,
397 : (errcode(ERRCODE_DUPLICATE_COLUMN),
398 : errmsg("column \"%s\" of relation \"%s\" appears more than once",
399 : col, RelationGetRelationName(onerel))));
400 41 : unique_cols = bms_add_member(unique_cols, i);
401 :
402 41 : vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
403 41 : if (vacattrstats[tcnt] != NULL)
404 41 : tcnt++;
405 : }
406 25 : attr_cnt = tcnt;
407 : }
408 : else
409 : {
410 8296 : attr_cnt = onerel->rd_att->natts;
411 : vacattrstats = (VacAttrStats **)
412 8296 : palloc(attr_cnt * sizeof(VacAttrStats *));
413 8296 : tcnt = 0;
414 69157 : for (i = 1; i <= attr_cnt; i++)
415 : {
416 60861 : vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
417 60861 : if (vacattrstats[tcnt] != NULL)
418 60825 : tcnt++;
419 : }
420 8296 : 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 8321 : if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
433 : {
434 375 : List *idxs = RelationGetIndexList(onerel);
435 :
436 375 : Irel = NULL;
437 375 : nindexes = 0;
438 375 : hasindex = idxs != NIL;
439 375 : list_free(idxs);
440 : }
441 7946 : else if (!inh)
442 : {
443 7882 : vac_open_indexes(onerel, AccessShareLock, &nindexes, &Irel);
444 7882 : hasindex = nindexes > 0;
445 : }
446 : else
447 : {
448 64 : Irel = NULL;
449 64 : nindexes = 0;
450 64 : hasindex = false;
451 : }
452 8321 : indexdata = NULL;
453 8321 : if (nindexes > 0)
454 : {
455 6184 : indexdata = (AnlIndexData *) palloc0(nindexes * sizeof(AnlIndexData));
456 17827 : for (ind = 0; ind < nindexes; ind++)
457 : {
458 11643 : AnlIndexData *thisdata = &indexdata[ind];
459 : IndexInfo *indexInfo;
460 :
461 11643 : thisdata->indexInfo = indexInfo = BuildIndexInfo(Irel[ind]);
462 11643 : thisdata->tupleFract = 1.0; /* fix later if partial */
463 11643 : if (indexInfo->ii_Expressions != NIL && va_cols == NIL)
464 : {
465 61 : ListCell *indexpr_item = list_head(indexInfo->ii_Expressions);
466 :
467 61 : thisdata->vacattrstats = (VacAttrStats **)
468 61 : palloc(indexInfo->ii_NumIndexAttrs * sizeof(VacAttrStats *));
469 61 : tcnt = 0;
470 125 : for (i = 0; i < indexInfo->ii_NumIndexAttrs; i++)
471 : {
472 64 : int keycol = indexInfo->ii_IndexAttrNumbers[i];
473 :
474 64 : if (keycol == 0)
475 : {
476 : /* Found an index expression */
477 : Node *indexkey;
478 :
479 61 : if (indexpr_item == NULL) /* shouldn't happen */
480 0 : elog(ERROR, "too few entries in indexprs list");
481 61 : indexkey = (Node *) lfirst(indexpr_item);
482 61 : indexpr_item = lnext(indexInfo->ii_Expressions,
483 : indexpr_item);
484 122 : thisdata->vacattrstats[tcnt] =
485 61 : examine_attribute(Irel[ind], i + 1, indexkey);
486 61 : if (thisdata->vacattrstats[tcnt] != NULL)
487 61 : tcnt++;
488 : }
489 : }
490 61 : 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 8321 : targrows = 100;
502 69175 : for (i = 0; i < attr_cnt; i++)
503 : {
504 60854 : if (targrows < vacattrstats[i]->minrows)
505 8285 : targrows = vacattrstats[i]->minrows;
506 : }
507 19964 : for (ind = 0; ind < nindexes; ind++)
508 : {
509 11643 : AnlIndexData *thisdata = &indexdata[ind];
510 :
511 11704 : for (i = 0; i < thisdata->attr_cnt; i++)
512 : {
513 61 : if (targrows < thisdata->vacattrstats[i]->minrows)
514 6 : 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 8321 : minrows = ComputeExtStatisticsRows(onerel, attr_cnt, vacattrstats);
524 :
525 8321 : if (targrows < minrows)
526 0 : targrows = minrows;
527 :
528 : /*
529 : * Acquire the sample rows
530 : */
531 8321 : rows = (HeapTuple *) palloc(targrows * sizeof(HeapTuple));
532 8321 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
533 : inh ? PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS_INH :
534 : PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS);
535 8321 : if (inh)
536 439 : numrows = acquire_inherited_sample_rows(onerel, elevel,
537 : rows, targrows,
538 : &totalrows, &totaldeadrows);
539 : else
540 7882 : 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 8320 : if (numrows > 0)
551 : {
552 : MemoryContext col_context,
553 : old_context;
554 :
555 5627 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
556 : PROGRESS_ANALYZE_PHASE_COMPUTE_STATS);
557 :
558 5627 : col_context = AllocSetContextCreate(anl_context,
559 : "Analyze Column",
560 : ALLOCSET_DEFAULT_SIZES);
561 5627 : old_context = MemoryContextSwitchTo(col_context);
562 :
563 49974 : for (i = 0; i < attr_cnt; i++)
564 : {
565 44347 : VacAttrStats *stats = vacattrstats[i];
566 : AttributeOpts *aopt;
567 :
568 44347 : stats->rows = rows;
569 44347 : stats->tupDesc = onerel->rd_att;
570 44347 : 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 44347 : aopt = get_attribute_options(onerel->rd_id, stats->tupattnum);
580 44347 : if (aopt != NULL)
581 : {
582 : float8 n_distinct;
583 :
584 3 : n_distinct = inh ? aopt->n_distinct_inherited : aopt->n_distinct;
585 3 : if (n_distinct != 0.0)
586 3 : stats->stadistinct = n_distinct;
587 : }
588 :
589 44347 : MemoryContextReset(col_context);
590 : }
591 :
592 5627 : if (nindexes > 0)
593 3641 : compute_index_stats(onerel, totalrows,
594 : indexdata, nindexes,
595 : rows, numrows,
596 : col_context);
597 :
598 5624 : MemoryContextSwitchTo(old_context);
599 5624 : 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 5624 : update_attstats(RelationGetRelid(onerel), inh,
607 : attr_cnt, vacattrstats);
608 :
609 12727 : for (ind = 0; ind < nindexes; ind++)
610 : {
611 7103 : AnlIndexData *thisdata = &indexdata[ind];
612 :
613 7103 : update_attstats(RelationGetRelid(Irel[ind]), false,
614 : thisdata->attr_cnt, thisdata->vacattrstats);
615 : }
616 :
617 : /* Build extended statistics (if there are any). */
618 5624 : BuildRelationExtStatistics(onerel, inh, totalrows, numrows, rows,
619 : attr_cnt, vacattrstats);
620 : }
621 :
622 8317 : 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 8317 : if (!inh)
637 : {
638 7878 : BlockNumber relallvisible = 0;
639 7878 : BlockNumber relallfrozen = 0;
640 :
641 7878 : if (RELKIND_HAS_STORAGE(onerel->rd_rel->relkind))
642 7846 : 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 7878 : CommandCounterIncrement();
649 7878 : 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 19515 : for (ind = 0; ind < nindexes; ind++)
662 : {
663 11637 : AnlIndexData *thisdata = &indexdata[ind];
664 : double totalindexrows;
665 :
666 11637 : totalindexrows = ceil(thisdata->tupleFract * totalrows);
667 11637 : vac_update_relstats(Irel[ind],
668 11637 : RelationGetNumberOfBlocks(Irel[ind]),
669 : totalindexrows,
670 : 0, 0,
671 : false,
672 : InvalidTransactionId,
673 : InvalidMultiXactId,
674 : NULL, NULL,
675 : in_outer_xact);
676 : }
677 : }
678 439 : 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 : */
684 375 : CommandCounterIncrement();
685 375 : 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 8317 : if (!inh)
702 7878 : pgstat_report_analyze(onerel, totalrows, totaldeadrows,
703 : (va_cols == NIL), starttime);
704 439 : else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
705 375 : 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 8317 : if (!(params.options & VACOPT_VACUUM))
715 : {
716 15679 : for (ind = 0; ind < nindexes; ind++)
717 : {
718 : IndexBulkDeleteResult *stats;
719 : IndexVacuumInfo ivinfo;
720 :
721 8981 : ivinfo.index = Irel[ind];
722 8981 : ivinfo.heaprel = onerel;
723 8981 : ivinfo.analyze_only = true;
724 8981 : ivinfo.estimated_count = true;
725 8981 : ivinfo.message_level = elevel;
726 8981 : ivinfo.num_heap_tuples = onerel->rd_rel->reltuples;
727 8981 : ivinfo.strategy = vac_strategy;
728 :
729 8981 : stats = index_vacuum_cleanup(&ivinfo, NULL);
730 :
731 8981 : if (stats)
732 0 : pfree(stats);
733 : }
734 : }
735 :
736 : /* Done with indexes */
737 8317 : vac_close_indexes(nindexes, Irel, NoLock);
738 :
739 : /* Log the action if appropriate */
740 8317 : if (instrument)
741 : {
742 219 : TimestampTz endtime = GetCurrentTimestamp();
743 :
744 419 : if (verbose || params.log_analyze_min_duration == 0 ||
745 200 : TimestampDifferenceExceeds(starttime, endtime,
746 200 : params.log_analyze_min_duration))
747 : {
748 : long delay_in_ms;
749 : WalUsage walusage;
750 19 : double read_rate = 0;
751 19 : 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 19 : memset(&bufferusage, 0, sizeof(BufferUsage));
759 19 : BufferUsageAccumDiff(&bufferusage, &pgBufferUsage, &startbufferusage);
760 19 : memset(&walusage, 0, sizeof(WalUsage));
761 19 : WalUsageAccumDiff(&walusage, &pgWalUsage, &startwalusage);
762 :
763 19 : total_blks_hit = bufferusage.shared_blks_hit +
764 19 : bufferusage.local_blks_hit;
765 19 : total_blks_read = bufferusage.shared_blks_read +
766 19 : bufferusage.local_blks_read;
767 19 : total_blks_dirtied = bufferusage.shared_blks_dirtied +
768 19 : 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 19 : 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 19 : if (delay_in_ms > 0)
793 : {
794 19 : read_rate = (double) BLCKSZ * total_blks_read /
795 19 : (1024 * 1024) / (delay_in_ms / 1000.0);
796 19 : write_rate = (double) BLCKSZ * total_blks_dirtied /
797 19 : (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 19 : initStringInfo(&buf);
806 :
807 19 : if (AmAutoVacuumWorkerProcess())
808 19 : msgfmt = _("automatic analyze of table \"%s.%s.%s\"\n");
809 : else
810 0 : msgfmt = _("finished analyzing table \"%s.%s.%s\"\n");
811 :
812 19 : appendStringInfo(&buf, msgfmt,
813 : get_database_name(MyDatabaseId),
814 19 : get_namespace_name(RelationGetNamespace(onerel)),
815 19 : RelationGetRelationName(onerel));
816 19 : 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 19 : 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 19 : appendStringInfo(&buf, _("avg read rate: %.3f MB/s, avg write rate: %.3f MB/s\n"),
834 : read_rate, write_rate);
835 19 : appendStringInfo(&buf, _("buffer usage: %" PRId64 " hits, %" PRId64 " reads, %" PRId64 " dirtied\n"),
836 : total_blks_hit,
837 : total_blks_read,
838 : total_blks_dirtied);
839 19 : appendStringInfo(&buf,
840 19 : _("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 19 : appendStringInfo(&buf, _("system usage: %s"), pg_rusage_show(&ru0));
847 :
848 19 : ereport(verbose ? INFO : LOG,
849 : (errmsg_internal("%s", buf.data)));
850 :
851 19 : pfree(buf.data);
852 : }
853 : }
854 :
855 : /* Roll back any GUC changes executed by index functions */
856 8317 : AtEOXact_GUC(false, save_nestlevel);
857 :
858 : /* Restore userid and security context */
859 8317 : SetUserIdAndSecContext(save_userid, save_sec_context);
860 :
861 : /* Restore current context and release memory */
862 8317 : MemoryContextSwitchTo(caller_context);
863 8317 : MemoryContextDelete(anl_context);
864 8317 : anl_context = NULL;
865 8317 : }
866 :
867 : /*
868 : * Compute statistics about indexes of a relation
869 : */
870 : static void
871 3641 : 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 3641 : ind_context = AllocSetContextCreate(anl_context,
884 : "Analyze Index",
885 : ALLOCSET_DEFAULT_SIZES);
886 3641 : old_context = MemoryContextSwitchTo(ind_context);
887 :
888 10747 : for (ind = 0; ind < nindexes; ind++)
889 : {
890 7109 : AnlIndexData *thisdata = &indexdata[ind];
891 7109 : IndexInfo *indexInfo = thisdata->indexInfo;
892 7109 : 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 7109 : if (attr_cnt == 0 && indexInfo->ii_Predicate == NIL)
906 7030 : 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 79 : estate = CreateExecutorState();
914 79 : econtext = GetPerTupleExprContext(estate);
915 : /* Need a slot to hold the current heap tuple, too */
916 79 : slot = MakeSingleTupleTableSlot(RelationGetDescr(onerel),
917 : &TTSOpsHeapTuple);
918 :
919 : /* Arrange for econtext's scan tuple to be the tuple under test */
920 79 : econtext->ecxt_scantuple = slot;
921 :
922 : /* Set up execution state for predicate. */
923 79 : predicate = ExecPrepareQual(indexInfo->ii_Predicate, estate);
924 :
925 : /* Compute and save index expression values */
926 79 : exprvals = (Datum *) palloc(numrows * attr_cnt * sizeof(Datum));
927 79 : exprnulls = (bool *) palloc(numrows * attr_cnt * sizeof(bool));
928 79 : numindexrows = 0;
929 79 : tcnt = 0;
930 93735 : for (rowno = 0; rowno < numrows; rowno++)
931 : {
932 93659 : HeapTuple heapTuple = rows[rowno];
933 :
934 93659 : 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 93659 : ResetExprContext(econtext);
941 :
942 : /* Set up for predicate or expression evaluation */
943 93659 : ExecStoreHeapTuple(heapTuple, slot, false);
944 :
945 : /* If index is partial, check predicate */
946 93659 : if (predicate != NULL)
947 : {
948 33033 : if (!ExecQual(predicate, econtext))
949 20664 : continue;
950 : }
951 72995 : numindexrows++;
952 :
953 72995 : 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 60626 : 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 121246 : for (i = 0; i < attr_cnt; i++)
970 : {
971 60623 : VacAttrStats *stats = thisdata->vacattrstats[i];
972 60623 : int attnum = stats->tupattnum;
973 :
974 60623 : if (isnull[attnum - 1])
975 : {
976 6 : exprvals[tcnt] = (Datum) 0;
977 6 : exprnulls[tcnt] = true;
978 : }
979 : else
980 : {
981 121234 : exprvals[tcnt] = datumCopy(values[attnum - 1],
982 60617 : stats->attrtype->typbyval,
983 60617 : stats->attrtype->typlen);
984 60617 : exprnulls[tcnt] = false;
985 : }
986 60623 : 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 76 : thisdata->tupleFract = (double) numindexrows / (double) numrows;
996 76 : totalindexrows = ceil(thisdata->tupleFract * totalrows);
997 :
998 : /*
999 : * Now we can compute the statistics for the expression columns.
1000 : */
1001 76 : if (numindexrows > 0)
1002 : {
1003 72 : MemoryContextSwitchTo(col_context);
1004 121 : for (i = 0; i < attr_cnt; i++)
1005 : {
1006 49 : VacAttrStats *stats = thisdata->vacattrstats[i];
1007 :
1008 49 : stats->exprvals = exprvals + i;
1009 49 : stats->exprnulls = exprnulls + i;
1010 49 : stats->rowstride = attr_cnt;
1011 49 : stats->compute_stats(stats,
1012 : ind_fetch_func,
1013 : numindexrows,
1014 : totalindexrows);
1015 :
1016 49 : MemoryContextReset(col_context);
1017 : }
1018 : }
1019 :
1020 : /* And clean up */
1021 76 : MemoryContextSwitchTo(ind_context);
1022 :
1023 76 : ExecDropSingleTupleTableSlot(slot);
1024 76 : FreeExecutorState(estate);
1025 76 : MemoryContextReset(ind_context);
1026 : }
1027 :
1028 3638 : MemoryContextSwitchTo(old_context);
1029 3638 : MemoryContextDelete(ind_context);
1030 3638 : }
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 60963 : examine_attribute(Relation onerel, int attnum, Node *index_expr)
1043 : {
1044 60963 : 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 60963 : if (attr->attisdropped)
1056 3 : return NULL;
1057 :
1058 : /* Don't analyze virtual generated columns */
1059 60960 : if (attr->attgenerated == ATTRIBUTE_GENERATED_VIRTUAL)
1060 28 : 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 60932 : atttuple = SearchSysCache2(ATTNUM, ObjectIdGetDatum(RelationGetRelid(onerel)), Int16GetDatum(attnum));
1068 60932 : if (!HeapTupleIsValid(atttuple))
1069 0 : elog(ERROR, "cache lookup failed for attribute %d of relation %u",
1070 : attnum, RelationGetRelid(onerel));
1071 60932 : dat = SysCacheGetAttr(ATTNUM, atttuple, Anum_pg_attribute_attstattarget, &isnull);
1072 60932 : attstattarget = isnull ? -1 : DatumGetInt16(dat);
1073 60932 : ReleaseSysCache(atttuple);
1074 :
1075 : /* Don't analyze column if user has specified not to */
1076 60932 : if (attstattarget == 0)
1077 3 : return NULL;
1078 :
1079 : /*
1080 : * Create the VacAttrStats struct.
1081 : */
1082 60929 : stats = palloc0_object(VacAttrStats);
1083 60929 : 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 60929 : if (index_expr)
1095 : {
1096 61 : stats->attrtypid = exprType(index_expr);
1097 61 : 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 61 : if (OidIsValid(onerel->rd_indcollation[attnum - 1]))
1105 7 : stats->attrcollid = onerel->rd_indcollation[attnum - 1];
1106 : else
1107 54 : stats->attrcollid = exprCollation(index_expr);
1108 : }
1109 : else
1110 : {
1111 60868 : stats->attrtypid = attr->atttypid;
1112 60868 : stats->attrtypmod = attr->atttypmod;
1113 60868 : stats->attrcollid = attr->attcollation;
1114 : }
1115 :
1116 60929 : typtuple = SearchSysCacheCopy1(TYPEOID,
1117 : ObjectIdGetDatum(stats->attrtypid));
1118 60929 : if (!HeapTupleIsValid(typtuple))
1119 0 : elog(ERROR, "cache lookup failed for type %u", stats->attrtypid);
1120 60929 : stats->attrtype = (Form_pg_type) GETSTRUCT(typtuple);
1121 60929 : stats->anl_context = anl_context;
1122 60929 : 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 365574 : for (i = 0; i < STATISTIC_NUM_SLOTS; i++)
1130 : {
1131 304645 : stats->statypid[i] = stats->attrtypid;
1132 304645 : stats->statyplen[i] = stats->attrtype->typlen;
1133 304645 : stats->statypbyval[i] = stats->attrtype->typbyval;
1134 304645 : 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 60929 : if (OidIsValid(stats->attrtype->typanalyze))
1142 3891 : ok = DatumGetBool(OidFunctionCall1(stats->attrtype->typanalyze,
1143 : PointerGetDatum(stats)));
1144 : else
1145 57038 : ok = std_typanalyze(stats);
1146 :
1147 60929 : if (!ok || stats->compute_stats == NULL || stats->minrows <= 0)
1148 : {
1149 2 : heap_freetuple(typtuple);
1150 2 : pfree(stats);
1151 2 : return NULL;
1152 : }
1153 :
1154 60927 : 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 78748 : block_sampling_read_stream_next(ReadStream *stream,
1163 : void *callback_private_data,
1164 : void *per_buffer_data)
1165 : {
1166 78748 : BlockSamplerData *bs = callback_private_data;
1167 :
1168 78748 : 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 8877 : acquire_sample_rows(Relation onerel, int elevel,
1206 : HeapTuple *rows, int targrows,
1207 : double *totalrows, double *totaldeadrows)
1208 : {
1209 8877 : int numrows = 0; /* # rows now in reservoir */
1210 8877 : double samplerows = 0; /* total # rows collected */
1211 8877 : double liverows = 0; /* # live rows seen */
1212 8877 : double deadrows = 0; /* # dead rows seen */
1213 8877 : double rowstoskip = -1; /* -1 means not set yet */
1214 : uint32 randseed; /* Seed for block sampler(s) */
1215 : BlockNumber totalblocks;
1216 : BlockSamplerData bs;
1217 : ReservoirStateData rstate;
1218 : TupleTableSlot *slot;
1219 : TableScanDesc scan;
1220 : BlockNumber nblocks;
1221 8877 : BlockNumber blksdone = 0;
1222 : ReadStream *stream;
1223 :
1224 : Assert(targrows > 0);
1225 :
1226 8877 : totalblocks = RelationGetNumberOfBlocks(onerel);
1227 :
1228 : /* Prepare for sampling block numbers */
1229 8877 : randseed = pg_prng_uint32(&pg_global_prng_state);
1230 8877 : nblocks = BlockSampler_Init(&bs, totalblocks, targrows, randseed);
1231 :
1232 : /* Report sampling block numbers */
1233 8877 : pgstat_progress_update_param(PROGRESS_ANALYZE_BLOCKS_TOTAL,
1234 : nblocks);
1235 :
1236 : /* Prepare for sampling rows */
1237 8877 : reservoir_init_selection_state(&rstate, targrows);
1238 :
1239 8877 : scan = table_beginscan_analyze(onerel);
1240 8877 : slot = table_slot_create(onerel, NULL);
1241 :
1242 : /*
1243 : * It is safe to use batching, as block_sampling_read_stream_next never
1244 : * blocks.
1245 : */
1246 8877 : stream = read_stream_begin_relation(READ_STREAM_MAINTENANCE |
1247 : READ_STREAM_USE_BATCHING,
1248 : vac_strategy,
1249 : scan->rs_rd,
1250 : MAIN_FORKNUM,
1251 : block_sampling_read_stream_next,
1252 : &bs,
1253 : 0);
1254 :
1255 : /* Outer loop over blocks to sample */
1256 78748 : while (table_scan_analyze_next_block(scan, stream))
1257 : {
1258 69871 : vacuum_delay_point(true);
1259 :
1260 5473528 : while (table_scan_analyze_next_tuple(scan, &liverows, &deadrows, slot))
1261 : {
1262 : /*
1263 : * The first targrows sample rows are simply copied into the
1264 : * reservoir. Then we start replacing tuples in the sample until
1265 : * we reach the end of the relation. This algorithm is from Jeff
1266 : * Vitter's paper (see full citation in utils/misc/sampling.c). It
1267 : * works by repeatedly computing the number of tuples to skip
1268 : * before selecting a tuple, which replaces a randomly chosen
1269 : * element of the reservoir (current set of tuples). At all times
1270 : * the reservoir is a true random sample of the tuples we've
1271 : * passed over so far, so when we fall off the end of the relation
1272 : * we're done.
1273 : */
1274 5403657 : if (numrows < targrows)
1275 5275868 : rows[numrows++] = ExecCopySlotHeapTuple(slot);
1276 : else
1277 : {
1278 : /*
1279 : * t in Vitter's paper is the number of records already
1280 : * processed. If we need to compute a new S value, we must
1281 : * use the not-yet-incremented value of samplerows as t.
1282 : */
1283 127789 : if (rowstoskip < 0)
1284 58214 : rowstoskip = reservoir_get_next_S(&rstate, samplerows, targrows);
1285 :
1286 127789 : if (rowstoskip <= 0)
1287 : {
1288 : /*
1289 : * Found a suitable tuple, so save it, replacing one old
1290 : * tuple at random
1291 : */
1292 58185 : int k = (int) (targrows * sampler_random_fract(&rstate.randstate));
1293 :
1294 : Assert(k >= 0 && k < targrows);
1295 58185 : heap_freetuple(rows[k]);
1296 58185 : rows[k] = ExecCopySlotHeapTuple(slot);
1297 : }
1298 :
1299 127789 : rowstoskip -= 1;
1300 : }
1301 :
1302 5403657 : samplerows += 1;
1303 : }
1304 :
1305 69871 : pgstat_progress_update_param(PROGRESS_ANALYZE_BLOCKS_DONE,
1306 : ++blksdone);
1307 : }
1308 :
1309 8877 : read_stream_end(stream);
1310 :
1311 8877 : ExecDropSingleTupleTableSlot(slot);
1312 8877 : table_endscan(scan);
1313 :
1314 : /*
1315 : * If we didn't find as many tuples as we wanted then we're done. No sort
1316 : * is needed, since they're already in order.
1317 : *
1318 : * Otherwise we need to sort the collected tuples by position
1319 : * (itempointer). It's not worth worrying about corner cases where the
1320 : * tuples are already sorted.
1321 : */
1322 8877 : if (numrows == targrows)
1323 81 : qsort_interruptible(rows, numrows, sizeof(HeapTuple),
1324 : compare_rows, NULL);
1325 :
1326 : /*
1327 : * Estimate total numbers of live and dead rows in relation, extrapolating
1328 : * on the assumption that the average tuple density in pages we didn't
1329 : * scan is the same as in the pages we did scan. Since what we scanned is
1330 : * a random sample of the pages in the relation, this should be a good
1331 : * assumption.
1332 : */
1333 8877 : if (bs.m > 0)
1334 : {
1335 6229 : *totalrows = floor((liverows / bs.m) * totalblocks + 0.5);
1336 6229 : *totaldeadrows = floor((deadrows / bs.m) * totalblocks + 0.5);
1337 : }
1338 : else
1339 : {
1340 2648 : *totalrows = 0.0;
1341 2648 : *totaldeadrows = 0.0;
1342 : }
1343 :
1344 : /*
1345 : * Emit some interesting relation info
1346 : */
1347 8877 : ereport(elevel,
1348 : (errmsg("\"%s\": scanned %d of %u pages, "
1349 : "containing %.0f live rows and %.0f dead rows; "
1350 : "%d rows in sample, %.0f estimated total rows",
1351 : RelationGetRelationName(onerel),
1352 : bs.m, totalblocks,
1353 : liverows, deadrows,
1354 : numrows, *totalrows)));
1355 :
1356 8877 : return numrows;
1357 : }
1358 :
1359 : /*
1360 : * Comparator for sorting rows[] array
1361 : */
1362 : static int
1363 2019484 : compare_rows(const void *a, const void *b, void *arg)
1364 : {
1365 2019484 : HeapTuple ha = *(const HeapTuple *) a;
1366 2019484 : HeapTuple hb = *(const HeapTuple *) b;
1367 2019484 : BlockNumber ba = ItemPointerGetBlockNumber(&ha->t_self);
1368 2019484 : OffsetNumber oa = ItemPointerGetOffsetNumber(&ha->t_self);
1369 2019484 : BlockNumber bb = ItemPointerGetBlockNumber(&hb->t_self);
1370 2019484 : OffsetNumber ob = ItemPointerGetOffsetNumber(&hb->t_self);
1371 :
1372 2019484 : if (ba < bb)
1373 440864 : return -1;
1374 1578620 : if (ba > bb)
1375 437705 : return 1;
1376 1140915 : if (oa < ob)
1377 762655 : return -1;
1378 378260 : if (oa > ob)
1379 378260 : return 1;
1380 0 : return 0;
1381 : }
1382 :
1383 :
1384 : /*
1385 : * acquire_inherited_sample_rows -- acquire sample rows from inheritance tree
1386 : *
1387 : * This has the same API as acquire_sample_rows, except that rows are
1388 : * collected from all inheritance children as well as the specified table.
1389 : * We fail and return zero if there are no inheritance children, or if all
1390 : * children are foreign tables that don't support ANALYZE.
1391 : */
1392 : static int
1393 439 : acquire_inherited_sample_rows(Relation onerel, int elevel,
1394 : HeapTuple *rows, int targrows,
1395 : double *totalrows, double *totaldeadrows)
1396 : {
1397 : List *tableOIDs;
1398 : Relation *rels;
1399 : AcquireSampleRowsFunc *acquirefuncs;
1400 : double *relblocks;
1401 : double totalblocks;
1402 : int numrows,
1403 : nrels,
1404 : i;
1405 : ListCell *lc;
1406 : bool has_child;
1407 :
1408 : /* Initialize output parameters to zero now, in case we exit early */
1409 439 : *totalrows = 0;
1410 439 : *totaldeadrows = 0;
1411 :
1412 : /*
1413 : * Find all members of inheritance set. We only need AccessShareLock on
1414 : * the children.
1415 : */
1416 : tableOIDs =
1417 439 : find_all_inheritors(RelationGetRelid(onerel), AccessShareLock, NULL);
1418 :
1419 : /*
1420 : * Check that there's at least one descendant, else fail. This could
1421 : * happen despite analyze_rel's relhassubclass check, if table once had a
1422 : * child but no longer does. In that case, we can clear the
1423 : * relhassubclass field so as not to make the same mistake again later.
1424 : * (This is safe because we hold ShareUpdateExclusiveLock.)
1425 : */
1426 439 : if (list_length(tableOIDs) < 2)
1427 : {
1428 : /* CCI because we already updated the pg_class row in this command */
1429 10 : CommandCounterIncrement();
1430 10 : SetRelationHasSubclass(RelationGetRelid(onerel), false);
1431 10 : ereport(elevel,
1432 : (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no child tables",
1433 : get_namespace_name(RelationGetNamespace(onerel)),
1434 : RelationGetRelationName(onerel))));
1435 10 : return 0;
1436 : }
1437 :
1438 : /*
1439 : * Identify acquirefuncs to use, and count blocks in all the relations.
1440 : * The result could overflow BlockNumber, so we use double arithmetic.
1441 : */
1442 429 : rels = (Relation *) palloc(list_length(tableOIDs) * sizeof(Relation));
1443 : acquirefuncs = (AcquireSampleRowsFunc *)
1444 429 : palloc(list_length(tableOIDs) * sizeof(AcquireSampleRowsFunc));
1445 429 : relblocks = (double *) palloc(list_length(tableOIDs) * sizeof(double));
1446 429 : totalblocks = 0;
1447 429 : nrels = 0;
1448 429 : has_child = false;
1449 1963 : foreach(lc, tableOIDs)
1450 : {
1451 1534 : Oid childOID = lfirst_oid(lc);
1452 : Relation childrel;
1453 1534 : AcquireSampleRowsFunc acquirefunc = NULL;
1454 1534 : BlockNumber relpages = 0;
1455 :
1456 : /* We already got the needed lock */
1457 1534 : childrel = table_open(childOID, NoLock);
1458 :
1459 : /* Ignore if temp table of another backend */
1460 1534 : if (RELATION_IS_OTHER_TEMP(childrel))
1461 : {
1462 : /* ... but release the lock on it */
1463 : Assert(childrel != onerel);
1464 0 : table_close(childrel, AccessShareLock);
1465 411 : continue;
1466 : }
1467 :
1468 : /* Check table type (MATVIEW can't happen, but might as well allow) */
1469 1534 : if (childrel->rd_rel->relkind == RELKIND_RELATION ||
1470 426 : childrel->rd_rel->relkind == RELKIND_MATVIEW)
1471 : {
1472 : /* Regular table, so use the regular row acquisition function */
1473 1108 : acquirefunc = acquire_sample_rows;
1474 1108 : relpages = RelationGetNumberOfBlocks(childrel);
1475 : }
1476 426 : else if (childrel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
1477 : {
1478 : /*
1479 : * For a foreign table, call the FDW's hook function to see
1480 : * whether it supports analysis.
1481 : */
1482 : FdwRoutine *fdwroutine;
1483 15 : bool ok = false;
1484 :
1485 15 : fdwroutine = GetFdwRoutineForRelation(childrel, false);
1486 :
1487 15 : if (fdwroutine->AnalyzeForeignTable != NULL)
1488 15 : ok = fdwroutine->AnalyzeForeignTable(childrel,
1489 : &acquirefunc,
1490 : &relpages);
1491 :
1492 15 : if (!ok)
1493 : {
1494 : /* ignore, but release the lock on it */
1495 : Assert(childrel != onerel);
1496 0 : table_close(childrel, AccessShareLock);
1497 0 : continue;
1498 : }
1499 : }
1500 : else
1501 : {
1502 : /*
1503 : * ignore, but release the lock on it. don't try to unlock the
1504 : * passed-in relation
1505 : */
1506 : Assert(childrel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE);
1507 411 : if (childrel != onerel)
1508 39 : table_close(childrel, AccessShareLock);
1509 : else
1510 372 : table_close(childrel, NoLock);
1511 411 : continue;
1512 : }
1513 :
1514 : /* OK, we'll process this child */
1515 1123 : has_child = true;
1516 1123 : rels[nrels] = childrel;
1517 1123 : acquirefuncs[nrels] = acquirefunc;
1518 1123 : relblocks[nrels] = (double) relpages;
1519 1123 : totalblocks += (double) relpages;
1520 1123 : nrels++;
1521 : }
1522 :
1523 : /*
1524 : * If we don't have at least one child table to consider, fail. If the
1525 : * relation is a partitioned table, it's not counted as a child table.
1526 : */
1527 429 : if (!has_child)
1528 : {
1529 0 : ereport(elevel,
1530 : (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no analyzable child tables",
1531 : get_namespace_name(RelationGetNamespace(onerel)),
1532 : RelationGetRelationName(onerel))));
1533 0 : return 0;
1534 : }
1535 :
1536 : /*
1537 : * Now sample rows from each relation, proportionally to its fraction of
1538 : * the total block count. (This might be less than desirable if the child
1539 : * rels have radically different free-space percentages, but it's not
1540 : * clear that it's worth working harder.)
1541 : */
1542 429 : pgstat_progress_update_param(PROGRESS_ANALYZE_CHILD_TABLES_TOTAL,
1543 : nrels);
1544 429 : numrows = 0;
1545 1552 : for (i = 0; i < nrels; i++)
1546 : {
1547 1123 : Relation childrel = rels[i];
1548 1123 : AcquireSampleRowsFunc acquirefunc = acquirefuncs[i];
1549 1123 : double childblocks = relblocks[i];
1550 :
1551 : /*
1552 : * Report progress. The sampling function will normally report blocks
1553 : * done/total, but we need to reset them to 0 here, so that they don't
1554 : * show an old value until that.
1555 : */
1556 : {
1557 1123 : const int progress_index[] = {
1558 : PROGRESS_ANALYZE_CURRENT_CHILD_TABLE_RELID,
1559 : PROGRESS_ANALYZE_BLOCKS_DONE,
1560 : PROGRESS_ANALYZE_BLOCKS_TOTAL
1561 : };
1562 1123 : const int64 progress_vals[] = {
1563 1123 : RelationGetRelid(childrel),
1564 : 0,
1565 : 0,
1566 : };
1567 :
1568 1123 : pgstat_progress_update_multi_param(3, progress_index, progress_vals);
1569 : }
1570 :
1571 1123 : if (childblocks > 0)
1572 : {
1573 : int childtargrows;
1574 :
1575 1043 : childtargrows = (int) rint(targrows * childblocks / totalblocks);
1576 : /* Make sure we don't overrun due to roundoff error */
1577 1043 : childtargrows = Min(childtargrows, targrows - numrows);
1578 1043 : if (childtargrows > 0)
1579 : {
1580 : int childrows;
1581 : double trows,
1582 : tdrows;
1583 :
1584 : /* Fetch a random sample of the child's rows */
1585 1043 : childrows = (*acquirefunc) (childrel, elevel,
1586 1043 : rows + numrows, childtargrows,
1587 : &trows, &tdrows);
1588 :
1589 : /* We may need to convert from child's rowtype to parent's */
1590 1043 : if (childrows > 0 &&
1591 1043 : !equalRowTypes(RelationGetDescr(childrel),
1592 : RelationGetDescr(onerel)))
1593 : {
1594 : TupleConversionMap *map;
1595 :
1596 998 : map = convert_tuples_by_name(RelationGetDescr(childrel),
1597 : RelationGetDescr(onerel));
1598 998 : if (map != NULL)
1599 : {
1600 : int j;
1601 :
1602 53302 : for (j = 0; j < childrows; j++)
1603 : {
1604 : HeapTuple newtup;
1605 :
1606 53236 : newtup = execute_attr_map_tuple(rows[numrows + j], map);
1607 53236 : heap_freetuple(rows[numrows + j]);
1608 53236 : rows[numrows + j] = newtup;
1609 : }
1610 66 : free_conversion_map(map);
1611 : }
1612 : }
1613 :
1614 : /* And add to counts */
1615 1043 : numrows += childrows;
1616 1043 : *totalrows += trows;
1617 1043 : *totaldeadrows += tdrows;
1618 : }
1619 : }
1620 :
1621 : /*
1622 : * Note: we cannot release the child-table locks, since we may have
1623 : * pointers to their TOAST tables in the sampled rows.
1624 : */
1625 1123 : table_close(childrel, NoLock);
1626 1123 : pgstat_progress_update_param(PROGRESS_ANALYZE_CHILD_TABLES_DONE,
1627 1123 : i + 1);
1628 : }
1629 :
1630 429 : return numrows;
1631 : }
1632 :
1633 :
1634 : /*
1635 : * update_attstats() -- update attribute statistics for one relation
1636 : *
1637 : * Statistics are stored in several places: the pg_class row for the
1638 : * relation has stats about the whole relation, and there is a
1639 : * pg_statistic row for each (non-system) attribute that has ever
1640 : * been analyzed. The pg_class values are updated by VACUUM, not here.
1641 : *
1642 : * pg_statistic rows are just added or updated normally. This means
1643 : * that pg_statistic will probably contain some deleted rows at the
1644 : * completion of a vacuum cycle, unless it happens to get vacuumed last.
1645 : *
1646 : * To keep things simple, we punt for pg_statistic, and don't try
1647 : * to compute or store rows for pg_statistic itself in pg_statistic.
1648 : * This could possibly be made to work, but it's not worth the trouble.
1649 : * Note analyze_rel() has seen to it that we won't come here when
1650 : * vacuuming pg_statistic itself.
1651 : *
1652 : * Note: there would be a race condition here if two backends could
1653 : * ANALYZE the same table concurrently. Presently, we lock that out
1654 : * by taking a self-exclusive lock on the relation in analyze_rel().
1655 : */
1656 : static void
1657 12727 : update_attstats(Oid relid, bool inh, int natts, VacAttrStats **vacattrstats)
1658 : {
1659 : Relation sd;
1660 : int attno;
1661 12727 : CatalogIndexState indstate = NULL;
1662 :
1663 12727 : if (natts <= 0)
1664 7060 : return; /* nothing to do */
1665 :
1666 5667 : sd = table_open(StatisticRelationId, RowExclusiveLock);
1667 :
1668 50063 : for (attno = 0; attno < natts; attno++)
1669 : {
1670 44396 : VacAttrStats *stats = vacattrstats[attno];
1671 : HeapTuple stup,
1672 : oldtup;
1673 : int i,
1674 : k,
1675 : n;
1676 : Datum values[Natts_pg_statistic];
1677 : bool nulls[Natts_pg_statistic];
1678 : bool replaces[Natts_pg_statistic];
1679 :
1680 : /* Ignore attr if we weren't able to collect stats */
1681 44396 : if (!stats->stats_valid)
1682 5 : continue;
1683 :
1684 : /*
1685 : * Construct a new pg_statistic tuple
1686 : */
1687 1420512 : for (i = 0; i < Natts_pg_statistic; ++i)
1688 : {
1689 1376121 : nulls[i] = false;
1690 1376121 : replaces[i] = true;
1691 : }
1692 :
1693 44391 : values[Anum_pg_statistic_starelid - 1] = ObjectIdGetDatum(relid);
1694 44391 : values[Anum_pg_statistic_staattnum - 1] = Int16GetDatum(stats->tupattnum);
1695 44391 : values[Anum_pg_statistic_stainherit - 1] = BoolGetDatum(inh);
1696 44391 : values[Anum_pg_statistic_stanullfrac - 1] = Float4GetDatum(stats->stanullfrac);
1697 44391 : values[Anum_pg_statistic_stawidth - 1] = Int32GetDatum(stats->stawidth);
1698 44391 : values[Anum_pg_statistic_stadistinct - 1] = Float4GetDatum(stats->stadistinct);
1699 44391 : i = Anum_pg_statistic_stakind1 - 1;
1700 266346 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1701 : {
1702 221955 : values[i++] = Int16GetDatum(stats->stakind[k]); /* stakindN */
1703 : }
1704 44391 : i = Anum_pg_statistic_staop1 - 1;
1705 266346 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1706 : {
1707 221955 : values[i++] = ObjectIdGetDatum(stats->staop[k]); /* staopN */
1708 : }
1709 44391 : i = Anum_pg_statistic_stacoll1 - 1;
1710 266346 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1711 : {
1712 221955 : values[i++] = ObjectIdGetDatum(stats->stacoll[k]); /* stacollN */
1713 : }
1714 44391 : i = Anum_pg_statistic_stanumbers1 - 1;
1715 266346 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1716 : {
1717 221955 : if (stats->stanumbers[k] != NULL)
1718 : {
1719 69163 : int nnum = stats->numnumbers[k];
1720 69163 : Datum *numdatums = (Datum *) palloc(nnum * sizeof(Datum));
1721 : ArrayType *arry;
1722 :
1723 574810 : for (n = 0; n < nnum; n++)
1724 505647 : numdatums[n] = Float4GetDatum(stats->stanumbers[k][n]);
1725 69163 : arry = construct_array_builtin(numdatums, nnum, FLOAT4OID);
1726 69163 : values[i++] = PointerGetDatum(arry); /* stanumbersN */
1727 : }
1728 : else
1729 : {
1730 152792 : nulls[i] = true;
1731 152792 : values[i++] = (Datum) 0;
1732 : }
1733 : }
1734 44391 : i = Anum_pg_statistic_stavalues1 - 1;
1735 266346 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1736 : {
1737 221955 : if (stats->stavalues[k] != NULL)
1738 : {
1739 : ArrayType *arry;
1740 :
1741 48864 : arry = construct_array(stats->stavalues[k],
1742 : stats->numvalues[k],
1743 : stats->statypid[k],
1744 48864 : stats->statyplen[k],
1745 48864 : stats->statypbyval[k],
1746 48864 : stats->statypalign[k]);
1747 48864 : values[i++] = PointerGetDatum(arry); /* stavaluesN */
1748 : }
1749 : else
1750 : {
1751 173091 : nulls[i] = true;
1752 173091 : values[i++] = (Datum) 0;
1753 : }
1754 : }
1755 :
1756 : /* Is there already a pg_statistic tuple for this attribute? */
1757 88782 : oldtup = SearchSysCache3(STATRELATTINH,
1758 : ObjectIdGetDatum(relid),
1759 44391 : Int16GetDatum(stats->tupattnum),
1760 : BoolGetDatum(inh));
1761 :
1762 : /* Open index information when we know we need it */
1763 44391 : if (indstate == NULL)
1764 5664 : indstate = CatalogOpenIndexes(sd);
1765 :
1766 44391 : if (HeapTupleIsValid(oldtup))
1767 : {
1768 : /* Yes, replace it */
1769 19789 : stup = heap_modify_tuple(oldtup,
1770 : RelationGetDescr(sd),
1771 : values,
1772 : nulls,
1773 : replaces);
1774 19789 : ReleaseSysCache(oldtup);
1775 19789 : CatalogTupleUpdateWithInfo(sd, &stup->t_self, stup, indstate);
1776 : }
1777 : else
1778 : {
1779 : /* No, insert new tuple */
1780 24602 : stup = heap_form_tuple(RelationGetDescr(sd), values, nulls);
1781 24602 : CatalogTupleInsertWithInfo(sd, stup, indstate);
1782 : }
1783 :
1784 44391 : heap_freetuple(stup);
1785 : }
1786 :
1787 5667 : if (indstate != NULL)
1788 5664 : CatalogCloseIndexes(indstate);
1789 5667 : table_close(sd, RowExclusiveLock);
1790 : }
1791 :
1792 : /*
1793 : * Standard fetch function for use by compute_stats subroutines.
1794 : *
1795 : * This exists to provide some insulation between compute_stats routines
1796 : * and the actual storage of the sample data.
1797 : */
1798 : static Datum
1799 42334449 : std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1800 : {
1801 42334449 : int attnum = stats->tupattnum;
1802 42334449 : HeapTuple tuple = stats->rows[rownum];
1803 42334449 : TupleDesc tupDesc = stats->tupDesc;
1804 :
1805 42334449 : return heap_getattr(tuple, attnum, tupDesc, isNull);
1806 : }
1807 :
1808 : /*
1809 : * Fetch function for analyzing index expressions.
1810 : *
1811 : * We have not bothered to construct index tuples, instead the data is
1812 : * just in Datum arrays.
1813 : */
1814 : static Datum
1815 60623 : ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1816 : {
1817 : int i;
1818 :
1819 : /* exprvals and exprnulls are already offset for proper column */
1820 60623 : i = rownum * stats->rowstride;
1821 60623 : *isNull = stats->exprnulls[i];
1822 60623 : return stats->exprvals[i];
1823 : }
1824 :
1825 :
1826 : /*==========================================================================
1827 : *
1828 : * Code below this point represents the "standard" type-specific statistics
1829 : * analysis algorithms. This code can be replaced on a per-data-type basis
1830 : * by setting a nonzero value in pg_type.typanalyze.
1831 : *
1832 : *==========================================================================
1833 : */
1834 :
1835 :
1836 : /*
1837 : * To avoid consuming too much memory during analysis and/or too much space
1838 : * in the resulting pg_statistic rows, we ignore varlena datums that are wider
1839 : * than WIDTH_THRESHOLD (after detoasting!). This is legitimate for MCV
1840 : * and distinct-value calculations since a wide value is unlikely to be
1841 : * duplicated at all, much less be a most-common value. For the same reason,
1842 : * ignoring wide values will not affect our estimates of histogram bin
1843 : * boundaries very much.
1844 : */
1845 : #define WIDTH_THRESHOLD 1024
1846 :
1847 : #define swapInt(a,b) do {int _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1848 : #define swapDatum(a,b) do {Datum _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1849 :
1850 : /*
1851 : * Extra information used by the default analysis routines
1852 : */
1853 : typedef struct
1854 : {
1855 : int count; /* # of duplicates */
1856 : int first; /* values[] index of first occurrence */
1857 : } ScalarMCVItem;
1858 :
1859 : typedef struct
1860 : {
1861 : SortSupport ssup;
1862 : int *tupnoLink;
1863 : } CompareScalarsContext;
1864 :
1865 :
1866 : static void compute_trivial_stats(VacAttrStatsP stats,
1867 : AnalyzeAttrFetchFunc fetchfunc,
1868 : int samplerows,
1869 : double totalrows);
1870 : static void compute_distinct_stats(VacAttrStatsP stats,
1871 : AnalyzeAttrFetchFunc fetchfunc,
1872 : int samplerows,
1873 : double totalrows);
1874 : static void compute_scalar_stats(VacAttrStatsP stats,
1875 : AnalyzeAttrFetchFunc fetchfunc,
1876 : int samplerows,
1877 : double totalrows);
1878 : static int compare_scalars(const void *a, const void *b, void *arg);
1879 : static int compare_mcvs(const void *a, const void *b, void *arg);
1880 : static int analyze_mcv_list(int *mcv_counts,
1881 : int num_mcv,
1882 : double stadistinct,
1883 : double stanullfrac,
1884 : int samplerows,
1885 : double totalrows);
1886 :
1887 :
1888 : /*
1889 : * std_typanalyze -- the default type-specific typanalyze function
1890 : */
1891 : bool
1892 61740 : std_typanalyze(VacAttrStats *stats)
1893 : {
1894 : Oid ltopr;
1895 : Oid eqopr;
1896 : StdAnalyzeData *mystats;
1897 :
1898 : /* If the attstattarget column is negative, use the default value */
1899 61740 : if (stats->attstattarget < 0)
1900 61313 : stats->attstattarget = default_statistics_target;
1901 :
1902 : /* Look for default "<" and "=" operators for column's type */
1903 61740 : get_sort_group_operators(stats->attrtypid,
1904 : false, false, false,
1905 : <opr, &eqopr, NULL,
1906 : NULL);
1907 :
1908 : /* Save the operator info for compute_stats routines */
1909 61740 : mystats = palloc_object(StdAnalyzeData);
1910 61740 : mystats->eqopr = eqopr;
1911 61740 : mystats->eqfunc = OidIsValid(eqopr) ? get_opcode(eqopr) : InvalidOid;
1912 61740 : mystats->ltopr = ltopr;
1913 61740 : stats->extra_data = mystats;
1914 :
1915 : /*
1916 : * Determine which standard statistics algorithm to use
1917 : */
1918 61740 : if (OidIsValid(eqopr) && OidIsValid(ltopr))
1919 : {
1920 : /* Seems to be a scalar datatype */
1921 59874 : stats->compute_stats = compute_scalar_stats;
1922 : /*--------------------
1923 : * The following choice of minrows is based on the paper
1924 : * "Random sampling for histogram construction: how much is enough?"
1925 : * by Surajit Chaudhuri, Rajeev Motwani and Vivek Narasayya, in
1926 : * Proceedings of ACM SIGMOD International Conference on Management
1927 : * of Data, 1998, Pages 436-447. Their Corollary 1 to Theorem 5
1928 : * says that for table size n, histogram size k, maximum relative
1929 : * error in bin size f, and error probability gamma, the minimum
1930 : * random sample size is
1931 : * r = 4 * k * ln(2*n/gamma) / f^2
1932 : * Taking f = 0.5, gamma = 0.01, n = 10^6 rows, we obtain
1933 : * r = 305.82 * k
1934 : * Note that because of the log function, the dependence on n is
1935 : * quite weak; even at n = 10^12, a 300*k sample gives <= 0.66
1936 : * bin size error with probability 0.99. So there's no real need to
1937 : * scale for n, which is a good thing because we don't necessarily
1938 : * know it at this point.
1939 : *--------------------
1940 : */
1941 59874 : stats->minrows = 300 * stats->attstattarget;
1942 : }
1943 1866 : else if (OidIsValid(eqopr))
1944 : {
1945 : /* We can still recognize distinct values */
1946 1610 : stats->compute_stats = compute_distinct_stats;
1947 : /* Might as well use the same minrows as above */
1948 1610 : stats->minrows = 300 * stats->attstattarget;
1949 : }
1950 : else
1951 : {
1952 : /* Can't do much but the trivial stuff */
1953 256 : stats->compute_stats = compute_trivial_stats;
1954 : /* Might as well use the same minrows as above */
1955 256 : stats->minrows = 300 * stats->attstattarget;
1956 : }
1957 :
1958 61740 : return true;
1959 : }
1960 :
1961 :
1962 : /*
1963 : * compute_trivial_stats() -- compute very basic column statistics
1964 : *
1965 : * We use this when we cannot find a hash "=" operator for the datatype.
1966 : *
1967 : * We determine the fraction of non-null rows and the average datum width.
1968 : */
1969 : static void
1970 171 : compute_trivial_stats(VacAttrStatsP stats,
1971 : AnalyzeAttrFetchFunc fetchfunc,
1972 : int samplerows,
1973 : double totalrows)
1974 : {
1975 : int i;
1976 171 : int null_cnt = 0;
1977 171 : int nonnull_cnt = 0;
1978 171 : double total_width = 0;
1979 342 : bool is_varlena = (!stats->attrtype->typbyval &&
1980 171 : stats->attrtype->typlen == -1);
1981 342 : bool is_varwidth = (!stats->attrtype->typbyval &&
1982 171 : stats->attrtype->typlen < 0);
1983 :
1984 529953 : for (i = 0; i < samplerows; i++)
1985 : {
1986 : Datum value;
1987 : bool isnull;
1988 :
1989 529782 : vacuum_delay_point(true);
1990 :
1991 529782 : value = fetchfunc(stats, i, &isnull);
1992 :
1993 : /* Check for null/nonnull */
1994 529782 : if (isnull)
1995 : {
1996 325341 : null_cnt++;
1997 325341 : continue;
1998 : }
1999 204441 : nonnull_cnt++;
2000 :
2001 : /*
2002 : * If it's a variable-width field, add up widths for average width
2003 : * calculation. Note that if the value is toasted, we use the toasted
2004 : * width. We don't bother with this calculation if it's a fixed-width
2005 : * type.
2006 : */
2007 204441 : if (is_varlena)
2008 : {
2009 39945 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2010 : }
2011 164496 : else if (is_varwidth)
2012 : {
2013 : /* must be cstring */
2014 0 : total_width += strlen(DatumGetCString(value)) + 1;
2015 : }
2016 : }
2017 :
2018 : /* We can only compute average width if we found some non-null values. */
2019 171 : if (nonnull_cnt > 0)
2020 : {
2021 77 : stats->stats_valid = true;
2022 : /* Do the simple null-frac and width stats */
2023 77 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2024 77 : if (is_varwidth)
2025 37 : stats->stawidth = total_width / (double) nonnull_cnt;
2026 : else
2027 40 : stats->stawidth = stats->attrtype->typlen;
2028 77 : stats->stadistinct = 0.0; /* "unknown" */
2029 : }
2030 94 : else if (null_cnt > 0)
2031 : {
2032 : /* We found only nulls; assume the column is entirely null */
2033 94 : stats->stats_valid = true;
2034 94 : stats->stanullfrac = 1.0;
2035 94 : if (is_varwidth)
2036 94 : stats->stawidth = 0; /* "unknown" */
2037 : else
2038 0 : stats->stawidth = stats->attrtype->typlen;
2039 94 : stats->stadistinct = 0.0; /* "unknown" */
2040 : }
2041 171 : }
2042 :
2043 :
2044 : /*
2045 : * compute_distinct_stats() -- compute column statistics including ndistinct
2046 : *
2047 : * We use this when we can find only an "=" operator for the datatype.
2048 : *
2049 : * We determine the fraction of non-null rows, the average width, the
2050 : * most common values, and the (estimated) number of distinct values.
2051 : *
2052 : * The most common values are determined by brute force: we keep a list
2053 : * of previously seen values, ordered by number of times seen, as we scan
2054 : * the samples. A newly seen value is inserted just after the last
2055 : * multiply-seen value, causing the bottommost (oldest) singly-seen value
2056 : * to drop off the list. The accuracy of this method, and also its cost,
2057 : * depend mainly on the length of the list we are willing to keep.
2058 : */
2059 : static void
2060 1178 : compute_distinct_stats(VacAttrStatsP stats,
2061 : AnalyzeAttrFetchFunc fetchfunc,
2062 : int samplerows,
2063 : double totalrows)
2064 : {
2065 : int i;
2066 1178 : int null_cnt = 0;
2067 1178 : int nonnull_cnt = 0;
2068 1178 : int toowide_cnt = 0;
2069 1178 : double total_width = 0;
2070 1992 : bool is_varlena = (!stats->attrtype->typbyval &&
2071 814 : stats->attrtype->typlen == -1);
2072 1992 : bool is_varwidth = (!stats->attrtype->typbyval &&
2073 814 : stats->attrtype->typlen < 0);
2074 : FmgrInfo f_cmpeq;
2075 : typedef struct
2076 : {
2077 : Datum value;
2078 : int count;
2079 : } TrackItem;
2080 : TrackItem *track;
2081 : int track_cnt,
2082 : track_max;
2083 1178 : int num_mcv = stats->attstattarget;
2084 1178 : StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2085 :
2086 : /*
2087 : * We track up to 2*n values for an n-element MCV list; but at least 10
2088 : */
2089 1178 : track_max = 2 * num_mcv;
2090 1178 : if (track_max < 10)
2091 39 : track_max = 10;
2092 1178 : track = (TrackItem *) palloc(track_max * sizeof(TrackItem));
2093 1178 : track_cnt = 0;
2094 :
2095 1178 : fmgr_info(mystats->eqfunc, &f_cmpeq);
2096 :
2097 833023 : for (i = 0; i < samplerows; i++)
2098 : {
2099 : Datum value;
2100 : bool isnull;
2101 : bool match;
2102 : int firstcount1,
2103 : j;
2104 :
2105 831845 : vacuum_delay_point(true);
2106 :
2107 831845 : value = fetchfunc(stats, i, &isnull);
2108 :
2109 : /* Check for null/nonnull */
2110 831845 : if (isnull)
2111 : {
2112 695689 : null_cnt++;
2113 695689 : continue;
2114 : }
2115 136156 : nonnull_cnt++;
2116 :
2117 : /*
2118 : * If it's a variable-width field, add up widths for average width
2119 : * calculation. Note that if the value is toasted, we use the toasted
2120 : * width. We don't bother with this calculation if it's a fixed-width
2121 : * type.
2122 : */
2123 136156 : if (is_varlena)
2124 : {
2125 49736 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2126 :
2127 : /*
2128 : * If the value is toasted, we want to detoast it just once to
2129 : * avoid repeated detoastings and resultant excess memory usage
2130 : * during the comparisons. Also, check to see if the value is
2131 : * excessively wide, and if so don't detoast at all --- just
2132 : * ignore the value.
2133 : */
2134 49736 : if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
2135 : {
2136 0 : toowide_cnt++;
2137 0 : continue;
2138 : }
2139 49736 : value = PointerGetDatum(PG_DETOAST_DATUM(value));
2140 : }
2141 86420 : else if (is_varwidth)
2142 : {
2143 : /* must be cstring */
2144 0 : total_width += strlen(DatumGetCString(value)) + 1;
2145 : }
2146 :
2147 : /*
2148 : * See if the value matches anything we're already tracking.
2149 : */
2150 136156 : match = false;
2151 136156 : firstcount1 = track_cnt;
2152 302450 : for (j = 0; j < track_cnt; j++)
2153 : {
2154 298698 : if (DatumGetBool(FunctionCall2Coll(&f_cmpeq,
2155 : stats->attrcollid,
2156 298698 : value, track[j].value)))
2157 : {
2158 132404 : match = true;
2159 132404 : break;
2160 : }
2161 166294 : if (j < firstcount1 && track[j].count == 1)
2162 2446 : firstcount1 = j;
2163 : }
2164 :
2165 136156 : if (match)
2166 : {
2167 : /* Found a match */
2168 132404 : track[j].count++;
2169 : /* This value may now need to "bubble up" in the track list */
2170 136018 : while (j > 0 && track[j].count > track[j - 1].count)
2171 : {
2172 3614 : swapDatum(track[j].value, track[j - 1].value);
2173 3614 : swapInt(track[j].count, track[j - 1].count);
2174 3614 : j--;
2175 : }
2176 : }
2177 : else
2178 : {
2179 : /* No match. Insert at head of count-1 list */
2180 3752 : if (track_cnt < track_max)
2181 3576 : track_cnt++;
2182 81288 : for (j = track_cnt - 1; j > firstcount1; j--)
2183 : {
2184 77536 : track[j].value = track[j - 1].value;
2185 77536 : track[j].count = track[j - 1].count;
2186 : }
2187 3752 : if (firstcount1 < track_cnt)
2188 : {
2189 3752 : track[firstcount1].value = value;
2190 3752 : track[firstcount1].count = 1;
2191 : }
2192 : }
2193 : }
2194 :
2195 : /* We can only compute real stats if we found some non-null values. */
2196 1178 : if (nonnull_cnt > 0)
2197 : {
2198 : int nmultiple,
2199 : summultiple;
2200 :
2201 862 : stats->stats_valid = true;
2202 : /* Do the simple null-frac and width stats */
2203 862 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2204 862 : if (is_varwidth)
2205 498 : stats->stawidth = total_width / (double) nonnull_cnt;
2206 : else
2207 364 : stats->stawidth = stats->attrtype->typlen;
2208 :
2209 : /* Count the number of values we found multiple times */
2210 862 : summultiple = 0;
2211 3251 : for (nmultiple = 0; nmultiple < track_cnt; nmultiple++)
2212 : {
2213 2824 : if (track[nmultiple].count == 1)
2214 435 : break;
2215 2389 : summultiple += track[nmultiple].count;
2216 : }
2217 :
2218 862 : if (nmultiple == 0)
2219 : {
2220 : /*
2221 : * If we found no repeated non-null values, assume it's a unique
2222 : * column; but be sure to discount for any nulls we found.
2223 : */
2224 101 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2225 : }
2226 761 : else if (track_cnt < track_max && toowide_cnt == 0 &&
2227 : nmultiple == track_cnt)
2228 : {
2229 : /*
2230 : * Our track list includes every value in the sample, and every
2231 : * value appeared more than once. Assume the column has just
2232 : * these values. (This case is meant to address columns with
2233 : * small, fixed sets of possible values, such as boolean or enum
2234 : * columns. If there are any values that appear just once in the
2235 : * sample, including too-wide values, we should assume that that's
2236 : * not what we're dealing with.)
2237 : */
2238 427 : stats->stadistinct = track_cnt;
2239 : }
2240 : else
2241 : {
2242 : /*----------
2243 : * Estimate the number of distinct values using the estimator
2244 : * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2245 : * n*d / (n - f1 + f1*n/N)
2246 : * where f1 is the number of distinct values that occurred
2247 : * exactly once in our sample of n rows (from a total of N),
2248 : * and d is the total number of distinct values in the sample.
2249 : * This is their Duj1 estimator; the other estimators they
2250 : * recommend are considerably more complex, and are numerically
2251 : * very unstable when n is much smaller than N.
2252 : *
2253 : * In this calculation, we consider only non-nulls. We used to
2254 : * include rows with null values in the n and N counts, but that
2255 : * leads to inaccurate answers in columns with many nulls, and
2256 : * it's intuitively bogus anyway considering the desired result is
2257 : * the number of distinct non-null values.
2258 : *
2259 : * We assume (not very reliably!) that all the multiply-occurring
2260 : * values are reflected in the final track[] list, and the other
2261 : * nonnull values all appeared but once. (XXX this usually
2262 : * results in a drastic overestimate of ndistinct. Can we do
2263 : * any better?)
2264 : *----------
2265 : */
2266 334 : int f1 = nonnull_cnt - summultiple;
2267 334 : int d = f1 + nmultiple;
2268 334 : double n = samplerows - null_cnt;
2269 334 : double N = totalrows * (1.0 - stats->stanullfrac);
2270 : double stadistinct;
2271 :
2272 : /* N == 0 shouldn't happen, but just in case ... */
2273 334 : if (N > 0)
2274 334 : stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2275 : else
2276 0 : stadistinct = 0;
2277 :
2278 : /* Clamp to sane range in case of roundoff error */
2279 334 : if (stadistinct < d)
2280 4 : stadistinct = d;
2281 334 : if (stadistinct > N)
2282 0 : stadistinct = N;
2283 : /* And round to integer */
2284 334 : stats->stadistinct = floor(stadistinct + 0.5);
2285 : }
2286 :
2287 : /*
2288 : * If we estimated the number of distinct values at more than 10% of
2289 : * the total row count (a very arbitrary limit), then assume that
2290 : * stadistinct should scale with the row count rather than be a fixed
2291 : * value.
2292 : */
2293 862 : if (stats->stadistinct > 0.1 * totalrows)
2294 200 : stats->stadistinct = -(stats->stadistinct / totalrows);
2295 :
2296 : /*
2297 : * Decide how many values are worth storing as most-common values. If
2298 : * we are able to generate a complete MCV list (all the values in the
2299 : * sample will fit, and we think these are all the ones in the table),
2300 : * then do so. Otherwise, store only those values that are
2301 : * significantly more common than the values not in the list.
2302 : *
2303 : * Note: the first of these cases is meant to address columns with
2304 : * small, fixed sets of possible values, such as boolean or enum
2305 : * columns. If we can *completely* represent the column population by
2306 : * an MCV list that will fit into the stats target, then we should do
2307 : * so and thus provide the planner with complete information. But if
2308 : * the MCV list is not complete, it's generally worth being more
2309 : * selective, and not just filling it all the way up to the stats
2310 : * target.
2311 : */
2312 862 : if (track_cnt < track_max && toowide_cnt == 0 &&
2313 856 : stats->stadistinct > 0 &&
2314 : track_cnt <= num_mcv)
2315 : {
2316 : /* Track list includes all values seen, and all will fit */
2317 547 : num_mcv = track_cnt;
2318 : }
2319 : else
2320 : {
2321 : int *mcv_counts;
2322 :
2323 : /* Incomplete list; decide how many values are worth keeping */
2324 315 : if (num_mcv > track_cnt)
2325 284 : num_mcv = track_cnt;
2326 :
2327 315 : if (num_mcv > 0)
2328 : {
2329 315 : mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2330 1075 : for (i = 0; i < num_mcv; i++)
2331 760 : mcv_counts[i] = track[i].count;
2332 :
2333 315 : num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2334 315 : stats->stadistinct,
2335 315 : stats->stanullfrac,
2336 : samplerows, totalrows);
2337 : }
2338 : }
2339 :
2340 : /* Generate MCV slot entry */
2341 862 : if (num_mcv > 0)
2342 : {
2343 : MemoryContext old_context;
2344 : Datum *mcv_values;
2345 : float4 *mcv_freqs;
2346 :
2347 : /* Must copy the target values into anl_context */
2348 858 : old_context = MemoryContextSwitchTo(stats->anl_context);
2349 858 : mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2350 858 : mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2351 4078 : for (i = 0; i < num_mcv; i++)
2352 : {
2353 6440 : mcv_values[i] = datumCopy(track[i].value,
2354 3220 : stats->attrtype->typbyval,
2355 3220 : stats->attrtype->typlen);
2356 3220 : mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2357 : }
2358 858 : MemoryContextSwitchTo(old_context);
2359 :
2360 858 : stats->stakind[0] = STATISTIC_KIND_MCV;
2361 858 : stats->staop[0] = mystats->eqopr;
2362 858 : stats->stacoll[0] = stats->attrcollid;
2363 858 : stats->stanumbers[0] = mcv_freqs;
2364 858 : stats->numnumbers[0] = num_mcv;
2365 858 : stats->stavalues[0] = mcv_values;
2366 858 : stats->numvalues[0] = num_mcv;
2367 :
2368 : /*
2369 : * Accept the defaults for stats->statypid and others. They have
2370 : * been set before we were called (see vacuum.h)
2371 : */
2372 : }
2373 : }
2374 316 : else if (null_cnt > 0)
2375 : {
2376 : /* We found only nulls; assume the column is entirely null */
2377 316 : stats->stats_valid = true;
2378 316 : stats->stanullfrac = 1.0;
2379 316 : if (is_varwidth)
2380 316 : stats->stawidth = 0; /* "unknown" */
2381 : else
2382 0 : stats->stawidth = stats->attrtype->typlen;
2383 316 : stats->stadistinct = 0.0; /* "unknown" */
2384 : }
2385 :
2386 : /* We don't need to bother cleaning up any of our temporary palloc's */
2387 1178 : }
2388 :
2389 :
2390 : /*
2391 : * compute_scalar_stats() -- compute column statistics
2392 : *
2393 : * We use this when we can find "=" and "<" operators for the datatype.
2394 : *
2395 : * We determine the fraction of non-null rows, the average width, the
2396 : * most common values, the (estimated) number of distinct values, the
2397 : * distribution histogram, and the correlation of physical to logical order.
2398 : *
2399 : * The desired stats can be determined fairly easily after sorting the
2400 : * data values into order.
2401 : */
2402 : static void
2403 43224 : compute_scalar_stats(VacAttrStatsP stats,
2404 : AnalyzeAttrFetchFunc fetchfunc,
2405 : int samplerows,
2406 : double totalrows)
2407 : {
2408 : int i;
2409 43224 : int null_cnt = 0;
2410 43224 : int nonnull_cnt = 0;
2411 43224 : int toowide_cnt = 0;
2412 43224 : double total_width = 0;
2413 53666 : bool is_varlena = (!stats->attrtype->typbyval &&
2414 10442 : stats->attrtype->typlen == -1);
2415 53666 : bool is_varwidth = (!stats->attrtype->typbyval &&
2416 10442 : stats->attrtype->typlen < 0);
2417 : double corr_xysum;
2418 : SortSupportData ssup;
2419 : ScalarItem *values;
2420 43224 : int values_cnt = 0;
2421 : int *tupnoLink;
2422 : ScalarMCVItem *track;
2423 43224 : int track_cnt = 0;
2424 43224 : int num_mcv = stats->attstattarget;
2425 43224 : int num_bins = stats->attstattarget;
2426 43224 : StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2427 :
2428 43224 : values = (ScalarItem *) palloc(samplerows * sizeof(ScalarItem));
2429 43224 : tupnoLink = (int *) palloc(samplerows * sizeof(int));
2430 43224 : track = (ScalarMCVItem *) palloc(num_mcv * sizeof(ScalarMCVItem));
2431 :
2432 43224 : memset(&ssup, 0, sizeof(ssup));
2433 43224 : ssup.ssup_cxt = CurrentMemoryContext;
2434 43224 : ssup.ssup_collation = stats->attrcollid;
2435 43224 : ssup.ssup_nulls_first = false;
2436 :
2437 : /*
2438 : * For now, don't perform abbreviated key conversion, because full values
2439 : * are required for MCV slot generation. Supporting that optimization
2440 : * would necessitate teaching compare_scalars() to call a tie-breaker.
2441 : */
2442 43224 : ssup.abbreviate = false;
2443 :
2444 43224 : PrepareSortSupportFromOrderingOp(mystats->ltopr, &ssup);
2445 :
2446 : /* Initial scan to find sortable values */
2447 38641351 : for (i = 0; i < samplerows; i++)
2448 : {
2449 : Datum value;
2450 : bool isnull;
2451 :
2452 38598127 : vacuum_delay_point(true);
2453 :
2454 38598127 : value = fetchfunc(stats, i, &isnull);
2455 :
2456 : /* Check for null/nonnull */
2457 38598127 : if (isnull)
2458 : {
2459 5063720 : null_cnt++;
2460 5082906 : continue;
2461 : }
2462 33534407 : nonnull_cnt++;
2463 :
2464 : /*
2465 : * If it's a variable-width field, add up widths for average width
2466 : * calculation. Note that if the value is toasted, we use the toasted
2467 : * width. We don't bother with this calculation if it's a fixed-width
2468 : * type.
2469 : */
2470 33534407 : if (is_varlena)
2471 : {
2472 4114792 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2473 :
2474 : /*
2475 : * If the value is toasted, we want to detoast it just once to
2476 : * avoid repeated detoastings and resultant excess memory usage
2477 : * during the comparisons. Also, check to see if the value is
2478 : * excessively wide, and if so don't detoast at all --- just
2479 : * ignore the value.
2480 : */
2481 4114792 : if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
2482 : {
2483 19186 : toowide_cnt++;
2484 19186 : continue;
2485 : }
2486 4095606 : value = PointerGetDatum(PG_DETOAST_DATUM(value));
2487 : }
2488 29419615 : else if (is_varwidth)
2489 : {
2490 : /* must be cstring */
2491 0 : total_width += strlen(DatumGetCString(value)) + 1;
2492 : }
2493 :
2494 : /* Add it to the list to be sorted */
2495 33515221 : values[values_cnt].value = value;
2496 33515221 : values[values_cnt].tupno = values_cnt;
2497 33515221 : tupnoLink[values_cnt] = values_cnt;
2498 33515221 : values_cnt++;
2499 : }
2500 :
2501 : /* We can only compute real stats if we found some sortable values. */
2502 43224 : if (values_cnt > 0)
2503 : {
2504 : int ndistinct, /* # distinct values in sample */
2505 : nmultiple, /* # that appear multiple times */
2506 : num_hist,
2507 : dups_cnt;
2508 40417 : int slot_idx = 0;
2509 : CompareScalarsContext cxt;
2510 :
2511 : /* Sort the collected values */
2512 40417 : cxt.ssup = &ssup;
2513 40417 : cxt.tupnoLink = tupnoLink;
2514 40417 : qsort_interruptible(values, values_cnt, sizeof(ScalarItem),
2515 : compare_scalars, &cxt);
2516 :
2517 : /*
2518 : * Now scan the values in order, find the most common ones, and also
2519 : * accumulate ordering-correlation statistics.
2520 : *
2521 : * To determine which are most common, we first have to count the
2522 : * number of duplicates of each value. The duplicates are adjacent in
2523 : * the sorted list, so a brute-force approach is to compare successive
2524 : * datum values until we find two that are not equal. However, that
2525 : * requires N-1 invocations of the datum comparison routine, which are
2526 : * completely redundant with work that was done during the sort. (The
2527 : * sort algorithm must at some point have compared each pair of items
2528 : * that are adjacent in the sorted order; otherwise it could not know
2529 : * that it's ordered the pair correctly.) We exploit this by having
2530 : * compare_scalars remember the highest tupno index that each
2531 : * ScalarItem has been found equal to. At the end of the sort, a
2532 : * ScalarItem's tupnoLink will still point to itself if and only if it
2533 : * is the last item of its group of duplicates (since the group will
2534 : * be ordered by tupno).
2535 : */
2536 40417 : corr_xysum = 0;
2537 40417 : ndistinct = 0;
2538 40417 : nmultiple = 0;
2539 40417 : dups_cnt = 0;
2540 33555638 : for (i = 0; i < values_cnt; i++)
2541 : {
2542 33515221 : int tupno = values[i].tupno;
2543 :
2544 33515221 : corr_xysum += ((double) i) * ((double) tupno);
2545 33515221 : dups_cnt++;
2546 33515221 : if (tupnoLink[tupno] == tupno)
2547 : {
2548 : /* Reached end of duplicates of this value */
2549 6925982 : ndistinct++;
2550 6925982 : if (dups_cnt > 1)
2551 : {
2552 598684 : nmultiple++;
2553 598684 : if (track_cnt < num_mcv ||
2554 246646 : dups_cnt > track[track_cnt - 1].count)
2555 : {
2556 : /*
2557 : * Found a new item for the mcv list; find its
2558 : * position, bubbling down old items if needed. Loop
2559 : * invariant is that j points at an empty/ replaceable
2560 : * slot.
2561 : */
2562 : int j;
2563 :
2564 405683 : if (track_cnt < num_mcv)
2565 352038 : track_cnt++;
2566 5438432 : for (j = track_cnt - 1; j > 0; j--)
2567 : {
2568 5393583 : if (dups_cnt <= track[j - 1].count)
2569 360834 : break;
2570 5032749 : track[j].count = track[j - 1].count;
2571 5032749 : track[j].first = track[j - 1].first;
2572 : }
2573 405683 : track[j].count = dups_cnt;
2574 405683 : track[j].first = i + 1 - dups_cnt;
2575 : }
2576 : }
2577 6925982 : dups_cnt = 0;
2578 : }
2579 : }
2580 :
2581 40417 : stats->stats_valid = true;
2582 : /* Do the simple null-frac and width stats */
2583 40417 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2584 40417 : if (is_varwidth)
2585 5810 : stats->stawidth = total_width / (double) nonnull_cnt;
2586 : else
2587 34607 : stats->stawidth = stats->attrtype->typlen;
2588 :
2589 40417 : if (nmultiple == 0)
2590 : {
2591 : /*
2592 : * If we found no repeated non-null values, assume it's a unique
2593 : * column; but be sure to discount for any nulls we found.
2594 : */
2595 10837 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2596 : }
2597 29580 : else if (toowide_cnt == 0 && nmultiple == ndistinct)
2598 : {
2599 : /*
2600 : * Every value in the sample appeared more than once. Assume the
2601 : * column has just these values. (This case is meant to address
2602 : * columns with small, fixed sets of possible values, such as
2603 : * boolean or enum columns. If there are any values that appear
2604 : * just once in the sample, including too-wide values, we should
2605 : * assume that that's not what we're dealing with.)
2606 : */
2607 18057 : stats->stadistinct = ndistinct;
2608 : }
2609 : else
2610 : {
2611 : /*----------
2612 : * Estimate the number of distinct values using the estimator
2613 : * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2614 : * n*d / (n - f1 + f1*n/N)
2615 : * where f1 is the number of distinct values that occurred
2616 : * exactly once in our sample of n rows (from a total of N),
2617 : * and d is the total number of distinct values in the sample.
2618 : * This is their Duj1 estimator; the other estimators they
2619 : * recommend are considerably more complex, and are numerically
2620 : * very unstable when n is much smaller than N.
2621 : *
2622 : * In this calculation, we consider only non-nulls. We used to
2623 : * include rows with null values in the n and N counts, but that
2624 : * leads to inaccurate answers in columns with many nulls, and
2625 : * it's intuitively bogus anyway considering the desired result is
2626 : * the number of distinct non-null values.
2627 : *
2628 : * Overwidth values are assumed to have been distinct.
2629 : *----------
2630 : */
2631 11523 : int f1 = ndistinct - nmultiple + toowide_cnt;
2632 11523 : int d = f1 + nmultiple;
2633 11523 : double n = samplerows - null_cnt;
2634 11523 : double N = totalrows * (1.0 - stats->stanullfrac);
2635 : double stadistinct;
2636 :
2637 : /* N == 0 shouldn't happen, but just in case ... */
2638 11523 : if (N > 0)
2639 11523 : stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2640 : else
2641 0 : stadistinct = 0;
2642 :
2643 : /* Clamp to sane range in case of roundoff error */
2644 11523 : if (stadistinct < d)
2645 643 : stadistinct = d;
2646 11523 : if (stadistinct > N)
2647 0 : stadistinct = N;
2648 : /* And round to integer */
2649 11523 : stats->stadistinct = floor(stadistinct + 0.5);
2650 : }
2651 :
2652 : /*
2653 : * If we estimated the number of distinct values at more than 10% of
2654 : * the total row count (a very arbitrary limit), then assume that
2655 : * stadistinct should scale with the row count rather than be a fixed
2656 : * value.
2657 : */
2658 40417 : if (stats->stadistinct > 0.1 * totalrows)
2659 8500 : stats->stadistinct = -(stats->stadistinct / totalrows);
2660 :
2661 : /*
2662 : * Decide how many values are worth storing as most-common values. If
2663 : * we are able to generate a complete MCV list (all the values in the
2664 : * sample will fit, and we think these are all the ones in the table),
2665 : * then do so. Otherwise, store only those values that are
2666 : * significantly more common than the values not in the list.
2667 : *
2668 : * Note: the first of these cases is meant to address columns with
2669 : * small, fixed sets of possible values, such as boolean or enum
2670 : * columns. If we can *completely* represent the column population by
2671 : * an MCV list that will fit into the stats target, then we should do
2672 : * so and thus provide the planner with complete information. But if
2673 : * the MCV list is not complete, it's generally worth being more
2674 : * selective, and not just filling it all the way up to the stats
2675 : * target.
2676 : */
2677 40417 : if (track_cnt == ndistinct && toowide_cnt == 0 &&
2678 17711 : stats->stadistinct > 0 &&
2679 : track_cnt <= num_mcv)
2680 : {
2681 : /* Track list includes all values seen, and all will fit */
2682 15849 : num_mcv = track_cnt;
2683 : }
2684 : else
2685 : {
2686 : int *mcv_counts;
2687 :
2688 : /* Incomplete list; decide how many values are worth keeping */
2689 24568 : if (num_mcv > track_cnt)
2690 22553 : num_mcv = track_cnt;
2691 :
2692 24568 : if (num_mcv > 0)
2693 : {
2694 13731 : mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2695 282406 : for (i = 0; i < num_mcv; i++)
2696 268675 : mcv_counts[i] = track[i].count;
2697 :
2698 13731 : num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2699 13731 : stats->stadistinct,
2700 13731 : stats->stanullfrac,
2701 : samplerows, totalrows);
2702 : }
2703 : }
2704 :
2705 : /* Generate MCV slot entry */
2706 40417 : if (num_mcv > 0)
2707 : {
2708 : MemoryContext old_context;
2709 : Datum *mcv_values;
2710 : float4 *mcv_freqs;
2711 :
2712 : /* Must copy the target values into anl_context */
2713 29556 : old_context = MemoryContextSwitchTo(stats->anl_context);
2714 29556 : mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2715 29556 : mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2716 381487 : for (i = 0; i < num_mcv; i++)
2717 : {
2718 703862 : mcv_values[i] = datumCopy(values[track[i].first].value,
2719 351931 : stats->attrtype->typbyval,
2720 351931 : stats->attrtype->typlen);
2721 351931 : mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2722 : }
2723 29556 : MemoryContextSwitchTo(old_context);
2724 :
2725 29556 : stats->stakind[slot_idx] = STATISTIC_KIND_MCV;
2726 29556 : stats->staop[slot_idx] = mystats->eqopr;
2727 29556 : stats->stacoll[slot_idx] = stats->attrcollid;
2728 29556 : stats->stanumbers[slot_idx] = mcv_freqs;
2729 29556 : stats->numnumbers[slot_idx] = num_mcv;
2730 29556 : stats->stavalues[slot_idx] = mcv_values;
2731 29556 : stats->numvalues[slot_idx] = num_mcv;
2732 :
2733 : /*
2734 : * Accept the defaults for stats->statypid and others. They have
2735 : * been set before we were called (see vacuum.h)
2736 : */
2737 29556 : slot_idx++;
2738 : }
2739 :
2740 : /*
2741 : * Generate a histogram slot entry if there are at least two distinct
2742 : * values not accounted for in the MCV list. (This ensures the
2743 : * histogram won't collapse to empty or a singleton.)
2744 : */
2745 40417 : num_hist = ndistinct - num_mcv;
2746 40417 : if (num_hist > num_bins)
2747 6477 : num_hist = num_bins + 1;
2748 40417 : if (num_hist >= 2)
2749 : {
2750 : MemoryContext old_context;
2751 : Datum *hist_values;
2752 : int nvals;
2753 : int pos,
2754 : posfrac,
2755 : delta,
2756 : deltafrac;
2757 :
2758 : /* Sort the MCV items into position order to speed next loop */
2759 18086 : qsort_interruptible(track, num_mcv, sizeof(ScalarMCVItem),
2760 : compare_mcvs, NULL);
2761 :
2762 : /*
2763 : * Collapse out the MCV items from the values[] array.
2764 : *
2765 : * Note we destroy the values[] array here... but we don't need it
2766 : * for anything more. We do, however, still need values_cnt.
2767 : * nvals will be the number of remaining entries in values[].
2768 : */
2769 18086 : if (num_mcv > 0)
2770 : {
2771 : int src,
2772 : dest;
2773 : int j;
2774 :
2775 9608 : src = dest = 0;
2776 9608 : j = 0; /* index of next interesting MCV item */
2777 348528 : while (src < values_cnt)
2778 : {
2779 : int ncopy;
2780 :
2781 338920 : if (j < num_mcv)
2782 : {
2783 331799 : int first = track[j].first;
2784 :
2785 331799 : if (src >= first)
2786 : {
2787 : /* advance past this MCV item */
2788 241593 : src = first + track[j].count;
2789 241593 : j++;
2790 241593 : continue;
2791 : }
2792 90206 : ncopy = first - src;
2793 : }
2794 : else
2795 7121 : ncopy = values_cnt - src;
2796 97327 : memmove(&values[dest], &values[src],
2797 : ncopy * sizeof(ScalarItem));
2798 97327 : src += ncopy;
2799 97327 : dest += ncopy;
2800 : }
2801 9608 : nvals = dest;
2802 : }
2803 : else
2804 8478 : nvals = values_cnt;
2805 : Assert(nvals >= num_hist);
2806 :
2807 : /* Must copy the target values into anl_context */
2808 18086 : old_context = MemoryContextSwitchTo(stats->anl_context);
2809 18086 : hist_values = (Datum *) palloc(num_hist * sizeof(Datum));
2810 :
2811 : /*
2812 : * The object of this loop is to copy the first and last values[]
2813 : * entries along with evenly-spaced values in between. So the
2814 : * i'th value is values[(i * (nvals - 1)) / (num_hist - 1)]. But
2815 : * computing that subscript directly risks integer overflow when
2816 : * the stats target is more than a couple thousand. Instead we
2817 : * add (nvals - 1) / (num_hist - 1) to pos at each step, tracking
2818 : * the integral and fractional parts of the sum separately.
2819 : */
2820 18086 : delta = (nvals - 1) / (num_hist - 1);
2821 18086 : deltafrac = (nvals - 1) % (num_hist - 1);
2822 18086 : pos = posfrac = 0;
2823 :
2824 919347 : for (i = 0; i < num_hist; i++)
2825 : {
2826 1802522 : hist_values[i] = datumCopy(values[pos].value,
2827 901261 : stats->attrtype->typbyval,
2828 901261 : stats->attrtype->typlen);
2829 901261 : pos += delta;
2830 901261 : posfrac += deltafrac;
2831 901261 : if (posfrac >= (num_hist - 1))
2832 : {
2833 : /* fractional part exceeds 1, carry to integer part */
2834 283226 : pos++;
2835 283226 : posfrac -= (num_hist - 1);
2836 : }
2837 : }
2838 :
2839 18086 : MemoryContextSwitchTo(old_context);
2840 :
2841 18086 : stats->stakind[slot_idx] = STATISTIC_KIND_HISTOGRAM;
2842 18086 : stats->staop[slot_idx] = mystats->ltopr;
2843 18086 : stats->stacoll[slot_idx] = stats->attrcollid;
2844 18086 : stats->stavalues[slot_idx] = hist_values;
2845 18086 : stats->numvalues[slot_idx] = num_hist;
2846 :
2847 : /*
2848 : * Accept the defaults for stats->statypid and others. They have
2849 : * been set before we were called (see vacuum.h)
2850 : */
2851 18086 : slot_idx++;
2852 : }
2853 :
2854 : /* Generate a correlation entry if there are multiple values */
2855 40417 : if (values_cnt > 1)
2856 : {
2857 : MemoryContext old_context;
2858 : float4 *corrs;
2859 : double corr_xsum,
2860 : corr_x2sum;
2861 :
2862 : /* Must copy the target values into anl_context */
2863 38034 : old_context = MemoryContextSwitchTo(stats->anl_context);
2864 38034 : corrs = palloc_object(float4);
2865 38034 : MemoryContextSwitchTo(old_context);
2866 :
2867 : /*----------
2868 : * Since we know the x and y value sets are both
2869 : * 0, 1, ..., values_cnt-1
2870 : * we have sum(x) = sum(y) =
2871 : * (values_cnt-1)*values_cnt / 2
2872 : * and sum(x^2) = sum(y^2) =
2873 : * (values_cnt-1)*values_cnt*(2*values_cnt-1) / 6.
2874 : *----------
2875 : */
2876 38034 : corr_xsum = ((double) (values_cnt - 1)) *
2877 38034 : ((double) values_cnt) / 2.0;
2878 38034 : corr_x2sum = ((double) (values_cnt - 1)) *
2879 38034 : ((double) values_cnt) * (double) (2 * values_cnt - 1) / 6.0;
2880 :
2881 : /* And the correlation coefficient reduces to */
2882 38034 : corrs[0] = (values_cnt * corr_xysum - corr_xsum * corr_xsum) /
2883 38034 : (values_cnt * corr_x2sum - corr_xsum * corr_xsum);
2884 :
2885 38034 : stats->stakind[slot_idx] = STATISTIC_KIND_CORRELATION;
2886 38034 : stats->staop[slot_idx] = mystats->ltopr;
2887 38034 : stats->stacoll[slot_idx] = stats->attrcollid;
2888 38034 : stats->stanumbers[slot_idx] = corrs;
2889 38034 : stats->numnumbers[slot_idx] = 1;
2890 38034 : slot_idx++;
2891 : }
2892 : }
2893 2807 : else if (nonnull_cnt > 0)
2894 : {
2895 : /* We found some non-null values, but they were all too wide */
2896 : Assert(nonnull_cnt == toowide_cnt);
2897 179 : stats->stats_valid = true;
2898 : /* Do the simple null-frac and width stats */
2899 179 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2900 179 : if (is_varwidth)
2901 179 : stats->stawidth = total_width / (double) nonnull_cnt;
2902 : else
2903 0 : stats->stawidth = stats->attrtype->typlen;
2904 : /* Assume all too-wide values are distinct, so it's a unique column */
2905 179 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2906 : }
2907 2628 : else if (null_cnt > 0)
2908 : {
2909 : /* We found only nulls; assume the column is entirely null */
2910 2628 : stats->stats_valid = true;
2911 2628 : stats->stanullfrac = 1.0;
2912 2628 : if (is_varwidth)
2913 2290 : stats->stawidth = 0; /* "unknown" */
2914 : else
2915 338 : stats->stawidth = stats->attrtype->typlen;
2916 2628 : stats->stadistinct = 0.0; /* "unknown" */
2917 : }
2918 :
2919 : /* We don't need to bother cleaning up any of our temporary palloc's */
2920 43224 : }
2921 :
2922 : /*
2923 : * Comparator for sorting ScalarItems
2924 : *
2925 : * Aside from sorting the items, we update the tupnoLink[] array
2926 : * whenever two ScalarItems are found to contain equal datums. The array
2927 : * is indexed by tupno; for each ScalarItem, it contains the highest
2928 : * tupno that that item's datum has been found to be equal to. This allows
2929 : * us to avoid additional comparisons in compute_scalar_stats().
2930 : */
2931 : static int
2932 310860754 : compare_scalars(const void *a, const void *b, void *arg)
2933 : {
2934 310860754 : Datum da = ((const ScalarItem *) a)->value;
2935 310860754 : int ta = ((const ScalarItem *) a)->tupno;
2936 310860754 : Datum db = ((const ScalarItem *) b)->value;
2937 310860754 : int tb = ((const ScalarItem *) b)->tupno;
2938 310860754 : CompareScalarsContext *cxt = (CompareScalarsContext *) arg;
2939 : int compare;
2940 :
2941 310860754 : compare = ApplySortComparator(da, false, db, false, cxt->ssup);
2942 310860754 : if (compare != 0)
2943 117269400 : return compare;
2944 :
2945 : /*
2946 : * The two datums are equal, so update cxt->tupnoLink[].
2947 : */
2948 193591354 : if (cxt->tupnoLink[ta] < tb)
2949 27711870 : cxt->tupnoLink[ta] = tb;
2950 193591354 : if (cxt->tupnoLink[tb] < ta)
2951 1968519 : cxt->tupnoLink[tb] = ta;
2952 :
2953 : /*
2954 : * For equal datums, sort by tupno
2955 : */
2956 193591354 : return ta - tb;
2957 : }
2958 :
2959 : /*
2960 : * Comparator for sorting ScalarMCVItems by position
2961 : */
2962 : static int
2963 1279138 : compare_mcvs(const void *a, const void *b, void *arg)
2964 : {
2965 1279138 : int da = ((const ScalarMCVItem *) a)->first;
2966 1279138 : int db = ((const ScalarMCVItem *) b)->first;
2967 :
2968 1279138 : return da - db;
2969 : }
2970 :
2971 : /*
2972 : * Analyze the list of common values in the sample and decide how many are
2973 : * worth storing in the table's MCV list.
2974 : *
2975 : * mcv_counts is assumed to be a list of the counts of the most common values
2976 : * seen in the sample, starting with the most common. The return value is the
2977 : * number that are significantly more common than the values not in the list,
2978 : * and which are therefore deemed worth storing in the table's MCV list.
2979 : */
2980 : static int
2981 14046 : analyze_mcv_list(int *mcv_counts,
2982 : int num_mcv,
2983 : double stadistinct,
2984 : double stanullfrac,
2985 : int samplerows,
2986 : double totalrows)
2987 : {
2988 : double ndistinct_table;
2989 : double sumcount;
2990 : int i;
2991 :
2992 : /*
2993 : * If the entire table was sampled, keep the whole list. This also
2994 : * protects us against division by zero in the code below.
2995 : */
2996 14046 : if (samplerows == totalrows || totalrows <= 1.0)
2997 13627 : return num_mcv;
2998 :
2999 : /* Re-extract the estimated number of distinct nonnull values in table */
3000 419 : ndistinct_table = stadistinct;
3001 419 : if (ndistinct_table < 0)
3002 79 : ndistinct_table = -ndistinct_table * totalrows;
3003 :
3004 : /*
3005 : * Exclude the least common values from the MCV list, if they are not
3006 : * significantly more common than the estimated selectivity they would
3007 : * have if they weren't in the list. All non-MCV values are assumed to be
3008 : * equally common, after taking into account the frequencies of all the
3009 : * values in the MCV list and the number of nulls (c.f. eqsel()).
3010 : *
3011 : * Here sumcount tracks the total count of all but the last (least common)
3012 : * value in the MCV list, allowing us to determine the effect of excluding
3013 : * that value from the list.
3014 : *
3015 : * Note that we deliberately do this by removing values from the full
3016 : * list, rather than starting with an empty list and adding values,
3017 : * because the latter approach can fail to add any values if all the most
3018 : * common values have around the same frequency and make up the majority
3019 : * of the table, so that the overall average frequency of all values is
3020 : * roughly the same as that of the common values. This would lead to any
3021 : * uncommon values being significantly overestimated.
3022 : */
3023 419 : sumcount = 0.0;
3024 875 : for (i = 0; i < num_mcv - 1; i++)
3025 456 : sumcount += mcv_counts[i];
3026 :
3027 502 : while (num_mcv > 0)
3028 : {
3029 : double selec,
3030 : otherdistinct,
3031 : N,
3032 : n,
3033 : K,
3034 : variance,
3035 : stddev;
3036 :
3037 : /*
3038 : * Estimated selectivity the least common value would have if it
3039 : * wasn't in the MCV list (c.f. eqsel()).
3040 : */
3041 502 : selec = 1.0 - sumcount / samplerows - stanullfrac;
3042 502 : if (selec < 0.0)
3043 0 : selec = 0.0;
3044 502 : if (selec > 1.0)
3045 0 : selec = 1.0;
3046 502 : otherdistinct = ndistinct_table - (num_mcv - 1);
3047 502 : if (otherdistinct > 1)
3048 502 : selec /= otherdistinct;
3049 :
3050 : /*
3051 : * If the value is kept in the MCV list, its population frequency is
3052 : * assumed to equal its sample frequency. We use the lower end of a
3053 : * textbook continuity-corrected Wald-type confidence interval to
3054 : * determine if that is significantly more common than the non-MCV
3055 : * frequency --- specifically we assume the population frequency is
3056 : * highly likely to be within around 2 standard errors of the sample
3057 : * frequency, which equates to an interval of 2 standard deviations
3058 : * either side of the sample count, plus an additional 0.5 for the
3059 : * continuity correction. Since we are sampling without replacement,
3060 : * this is a hypergeometric distribution.
3061 : *
3062 : * XXX: Empirically, this approach seems to work quite well, but it
3063 : * may be worth considering more advanced techniques for estimating
3064 : * the confidence interval of the hypergeometric distribution.
3065 : */
3066 502 : N = totalrows;
3067 502 : n = samplerows;
3068 502 : K = N * mcv_counts[num_mcv - 1] / n;
3069 502 : variance = n * K * (N - K) * (N - n) / (N * N * (N - 1));
3070 502 : stddev = sqrt(variance);
3071 :
3072 502 : if (mcv_counts[num_mcv - 1] > selec * samplerows + 2 * stddev + 0.5)
3073 : {
3074 : /*
3075 : * The value is significantly more common than the non-MCV
3076 : * selectivity would suggest. Keep it, and all the other more
3077 : * common values in the list.
3078 : */
3079 391 : break;
3080 : }
3081 : else
3082 : {
3083 : /* Discard this value and consider the next least common value */
3084 111 : num_mcv--;
3085 111 : if (num_mcv == 0)
3086 28 : break;
3087 83 : sumcount -= mcv_counts[num_mcv - 1];
3088 : }
3089 : }
3090 419 : return num_mcv;
3091 : }
|
