Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * analyze.c
4 : * the Postgres statistics generator
5 : *
6 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/commands/analyze.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 : #include "postgres.h"
16 :
17 : #include <math.h>
18 :
19 : #include "access/detoast.h"
20 : #include "access/genam.h"
21 : #include "access/multixact.h"
22 : #include "access/relation.h"
23 : #include "access/table.h"
24 : #include "access/tableam.h"
25 : #include "access/transam.h"
26 : #include "access/tupconvert.h"
27 : #include "access/visibilitymap.h"
28 : #include "access/xact.h"
29 : #include "catalog/index.h"
30 : #include "catalog/indexing.h"
31 : #include "catalog/pg_inherits.h"
32 : #include "commands/dbcommands.h"
33 : #include "commands/progress.h"
34 : #include "commands/tablecmds.h"
35 : #include "commands/vacuum.h"
36 : #include "common/pg_prng.h"
37 : #include "executor/executor.h"
38 : #include "foreign/fdwapi.h"
39 : #include "miscadmin.h"
40 : #include "nodes/nodeFuncs.h"
41 : #include "parser/parse_oper.h"
42 : #include "parser/parse_relation.h"
43 : #include "pgstat.h"
44 : #include "statistics/extended_stats_internal.h"
45 : #include "statistics/statistics.h"
46 : #include "storage/bufmgr.h"
47 : #include "storage/procarray.h"
48 : #include "utils/attoptcache.h"
49 : #include "utils/datum.h"
50 : #include "utils/guc.h"
51 : #include "utils/lsyscache.h"
52 : #include "utils/memutils.h"
53 : #include "utils/pg_rusage.h"
54 : #include "utils/sampling.h"
55 : #include "utils/sortsupport.h"
56 : #include "utils/syscache.h"
57 : #include "utils/timestamp.h"
58 :
59 :
60 : /* Per-index data for ANALYZE */
61 : typedef struct AnlIndexData
62 : {
63 : IndexInfo *indexInfo; /* BuildIndexInfo result */
64 : double tupleFract; /* fraction of rows for partial index */
65 : VacAttrStats **vacattrstats; /* index attrs to analyze */
66 : int attr_cnt;
67 : } AnlIndexData;
68 :
69 :
70 : /* Default statistics target (GUC parameter) */
71 : int default_statistics_target = 100;
72 :
73 : /* A few variables that don't seem worth passing around as parameters */
74 : static MemoryContext anl_context = NULL;
75 : static BufferAccessStrategy vac_strategy;
76 :
77 :
78 : static void do_analyze_rel(Relation onerel,
79 : VacuumParams *params, List *va_cols,
80 : AcquireSampleRowsFunc acquirefunc, BlockNumber relpages,
81 : bool inh, bool in_outer_xact, int elevel);
82 : static void compute_index_stats(Relation onerel, double totalrows,
83 : AnlIndexData *indexdata, int nindexes,
84 : HeapTuple *rows, int numrows,
85 : MemoryContext col_context);
86 : static VacAttrStats *examine_attribute(Relation onerel, int attnum,
87 : Node *index_expr);
88 : static int acquire_sample_rows(Relation onerel, int elevel,
89 : HeapTuple *rows, int targrows,
90 : double *totalrows, double *totaldeadrows);
91 : static int compare_rows(const void *a, const void *b, void *arg);
92 : static int acquire_inherited_sample_rows(Relation onerel, int elevel,
93 : HeapTuple *rows, int targrows,
94 : double *totalrows, double *totaldeadrows);
95 : static void update_attstats(Oid relid, bool inh,
96 : int natts, VacAttrStats **vacattrstats);
97 : static Datum std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
98 : static Datum ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
99 :
100 :
101 : /*
102 : * analyze_rel() -- analyze one relation
103 : *
104 : * relid identifies the relation to analyze. If relation is supplied, use
105 : * the name therein for reporting any failure to open/lock the rel; do not
106 : * use it once we've successfully opened the rel, since it might be stale.
107 : */
108 : void
109 15280 : analyze_rel(Oid relid, RangeVar *relation,
110 : VacuumParams *params, List *va_cols, bool in_outer_xact,
111 : BufferAccessStrategy bstrategy)
112 : {
113 : Relation onerel;
114 : int elevel;
115 15280 : AcquireSampleRowsFunc acquirefunc = NULL;
116 15280 : BlockNumber relpages = 0;
117 :
118 : /* Select logging level */
119 15280 : if (params->options & VACOPT_VERBOSE)
120 0 : elevel = INFO;
121 : else
122 15280 : elevel = DEBUG2;
123 :
124 : /* Set up static variables */
125 15280 : vac_strategy = bstrategy;
126 :
127 : /*
128 : * Check for user-requested abort.
129 : */
130 15280 : CHECK_FOR_INTERRUPTS();
131 :
132 : /*
133 : * Open the relation, getting ShareUpdateExclusiveLock to ensure that two
134 : * ANALYZEs don't run on it concurrently. (This also locks out a
135 : * concurrent VACUUM, which doesn't matter much at the moment but might
136 : * matter if we ever try to accumulate stats on dead tuples.) If the rel
137 : * has been dropped since we last saw it, we don't need to process it.
138 : *
139 : * Make sure to generate only logs for ANALYZE in this case.
140 : */
141 15280 : onerel = vacuum_open_relation(relid, relation, params->options & ~(VACOPT_VACUUM),
142 15280 : params->log_min_duration >= 0,
143 : ShareUpdateExclusiveLock);
144 :
145 : /* leave if relation could not be opened or locked */
146 15280 : if (!onerel)
147 200 : return;
148 :
149 : /*
150 : * Check if relation needs to be skipped based on privileges. This check
151 : * happens also when building the relation list to analyze for a manual
152 : * operation, and needs to be done additionally here as ANALYZE could
153 : * happen across multiple transactions where privileges could have changed
154 : * in-between. Make sure to generate only logs for ANALYZE in this case.
155 : */
156 15272 : if (!vacuum_is_permitted_for_relation(RelationGetRelid(onerel),
157 : onerel->rd_rel,
158 15272 : params->options & ~VACOPT_VACUUM))
159 : {
160 36 : relation_close(onerel, ShareUpdateExclusiveLock);
161 36 : return;
162 : }
163 :
164 : /*
165 : * Silently ignore tables that are temp tables of other backends ---
166 : * trying to analyze these is rather pointless, since their contents are
167 : * probably not up-to-date on disk. (We don't throw a warning here; it
168 : * would just lead to chatter during a database-wide ANALYZE.)
169 : */
170 15236 : if (RELATION_IS_OTHER_TEMP(onerel))
171 : {
172 0 : relation_close(onerel, ShareUpdateExclusiveLock);
173 0 : return;
174 : }
175 :
176 : /*
177 : * We can ANALYZE any table except pg_statistic. See update_attstats
178 : */
179 15236 : if (RelationGetRelid(onerel) == StatisticRelationId)
180 : {
181 156 : relation_close(onerel, ShareUpdateExclusiveLock);
182 156 : return;
183 : }
184 :
185 : /*
186 : * Check that it's of an analyzable relkind, and set up appropriately.
187 : */
188 15080 : if (onerel->rd_rel->relkind == RELKIND_RELATION ||
189 776 : onerel->rd_rel->relkind == RELKIND_MATVIEW)
190 : {
191 : /* Regular table, so we'll use the regular row acquisition function */
192 14308 : acquirefunc = acquire_sample_rows;
193 : /* Also get regular table's size */
194 14308 : relpages = RelationGetNumberOfBlocks(onerel);
195 : }
196 772 : else if (onerel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
197 : {
198 : /*
199 : * For a foreign table, call the FDW's hook function to see whether it
200 : * supports analysis.
201 : */
202 : FdwRoutine *fdwroutine;
203 56 : bool ok = false;
204 :
205 56 : fdwroutine = GetFdwRoutineForRelation(onerel, false);
206 :
207 56 : if (fdwroutine->AnalyzeForeignTable != NULL)
208 56 : ok = fdwroutine->AnalyzeForeignTable(onerel,
209 : &acquirefunc,
210 : &relpages);
211 :
212 56 : if (!ok)
213 : {
214 0 : ereport(WARNING,
215 : (errmsg("skipping \"%s\" --- cannot analyze this foreign table",
216 : RelationGetRelationName(onerel))));
217 0 : relation_close(onerel, ShareUpdateExclusiveLock);
218 0 : return;
219 : }
220 : }
221 716 : else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
222 : {
223 : /*
224 : * For partitioned tables, we want to do the recursive ANALYZE below.
225 : */
226 : }
227 : else
228 : {
229 : /* No need for a WARNING if we already complained during VACUUM */
230 0 : if (!(params->options & VACOPT_VACUUM))
231 0 : ereport(WARNING,
232 : (errmsg("skipping \"%s\" --- cannot analyze non-tables or special system tables",
233 : RelationGetRelationName(onerel))));
234 0 : relation_close(onerel, ShareUpdateExclusiveLock);
235 0 : return;
236 : }
237 :
238 : /*
239 : * OK, let's do it. First, initialize progress reporting.
240 : */
241 15080 : pgstat_progress_start_command(PROGRESS_COMMAND_ANALYZE,
242 : RelationGetRelid(onerel));
243 :
244 : /*
245 : * Do the normal non-recursive ANALYZE. We can skip this for partitioned
246 : * tables, which don't contain any rows.
247 : */
248 15080 : if (onerel->rd_rel->relkind != RELKIND_PARTITIONED_TABLE)
249 14364 : do_analyze_rel(onerel, params, va_cols, acquirefunc,
250 : relpages, false, in_outer_xact, elevel);
251 :
252 : /*
253 : * If there are child tables, do recursive ANALYZE.
254 : */
255 15040 : if (onerel->rd_rel->relhassubclass)
256 830 : do_analyze_rel(onerel, params, va_cols, acquirefunc, relpages,
257 : true, in_outer_xact, elevel);
258 :
259 : /*
260 : * Close source relation now, but keep lock so that no one deletes it
261 : * before we commit. (If someone did, they'd fail to clean up the entries
262 : * we made in pg_statistic. Also, releasing the lock before commit would
263 : * expose us to concurrent-update failures in update_attstats.)
264 : */
265 15022 : relation_close(onerel, NoLock);
266 :
267 15022 : pgstat_progress_end_command();
268 : }
269 :
270 : /*
271 : * do_analyze_rel() -- analyze one relation, recursively or not
272 : *
273 : * Note that "acquirefunc" is only relevant for the non-inherited case.
274 : * For the inherited case, acquire_inherited_sample_rows() determines the
275 : * appropriate acquirefunc for each child table.
276 : */
277 : static void
278 15194 : do_analyze_rel(Relation onerel, VacuumParams *params,
279 : List *va_cols, AcquireSampleRowsFunc acquirefunc,
280 : BlockNumber relpages, bool inh, bool in_outer_xact,
281 : int elevel)
282 : {
283 : int attr_cnt,
284 : tcnt,
285 : i,
286 : ind;
287 : Relation *Irel;
288 : int nindexes;
289 : bool verbose,
290 : instrument,
291 : hasindex;
292 : VacAttrStats **vacattrstats;
293 : AnlIndexData *indexdata;
294 : int targrows,
295 : numrows,
296 : minrows;
297 : double totalrows,
298 : totaldeadrows;
299 : HeapTuple *rows;
300 : PGRUsage ru0;
301 15194 : TimestampTz starttime = 0;
302 : MemoryContext caller_context;
303 : Oid save_userid;
304 : int save_sec_context;
305 : int save_nestlevel;
306 15194 : WalUsage startwalusage = pgWalUsage;
307 15194 : BufferUsage startbufferusage = pgBufferUsage;
308 : BufferUsage bufferusage;
309 15194 : PgStat_Counter startreadtime = 0;
310 15194 : PgStat_Counter startwritetime = 0;
311 :
312 15194 : verbose = (params->options & VACOPT_VERBOSE) != 0;
313 15598 : instrument = (verbose || (AmAutoVacuumWorkerProcess() &&
314 404 : params->log_min_duration >= 0));
315 15194 : if (inh)
316 830 : ereport(elevel,
317 : (errmsg("analyzing \"%s.%s\" inheritance tree",
318 : get_namespace_name(RelationGetNamespace(onerel)),
319 : RelationGetRelationName(onerel))));
320 : else
321 14364 : ereport(elevel,
322 : (errmsg("analyzing \"%s.%s\"",
323 : get_namespace_name(RelationGetNamespace(onerel)),
324 : RelationGetRelationName(onerel))));
325 :
326 : /*
327 : * Set up a working context so that we can easily free whatever junk gets
328 : * created.
329 : */
330 15194 : anl_context = AllocSetContextCreate(CurrentMemoryContext,
331 : "Analyze",
332 : ALLOCSET_DEFAULT_SIZES);
333 15194 : caller_context = MemoryContextSwitchTo(anl_context);
334 :
335 : /*
336 : * Switch to the table owner's userid, so that any index functions are run
337 : * as that user. Also lock down security-restricted operations and
338 : * arrange to make GUC variable changes local to this command.
339 : */
340 15194 : GetUserIdAndSecContext(&save_userid, &save_sec_context);
341 15194 : SetUserIdAndSecContext(onerel->rd_rel->relowner,
342 : save_sec_context | SECURITY_RESTRICTED_OPERATION);
343 15194 : save_nestlevel = NewGUCNestLevel();
344 15194 : RestrictSearchPath();
345 :
346 : /*
347 : * When verbose or autovacuum logging is used, initialize a resource usage
348 : * snapshot and optionally track I/O timing.
349 : */
350 15194 : if (instrument)
351 : {
352 404 : if (track_io_timing)
353 : {
354 0 : startreadtime = pgStatBlockReadTime;
355 0 : startwritetime = pgStatBlockWriteTime;
356 : }
357 :
358 404 : pg_rusage_init(&ru0);
359 : }
360 :
361 : /* Used for instrumentation and stats report */
362 15194 : starttime = GetCurrentTimestamp();
363 :
364 : /*
365 : * Determine which columns to analyze
366 : *
367 : * Note that system attributes are never analyzed, so we just reject them
368 : * at the lookup stage. We also reject duplicate column mentions. (We
369 : * could alternatively ignore duplicates, but analyzing a column twice
370 : * won't work; we'd end up making a conflicting update in pg_statistic.)
371 : */
372 15194 : if (va_cols != NIL)
373 : {
374 100 : Bitmapset *unique_cols = NULL;
375 : ListCell *le;
376 :
377 100 : vacattrstats = (VacAttrStats **) palloc(list_length(va_cols) *
378 : sizeof(VacAttrStats *));
379 100 : tcnt = 0;
380 182 : foreach(le, va_cols)
381 : {
382 132 : char *col = strVal(lfirst(le));
383 :
384 132 : i = attnameAttNum(onerel, col, false);
385 132 : if (i == InvalidAttrNumber)
386 38 : ereport(ERROR,
387 : (errcode(ERRCODE_UNDEFINED_COLUMN),
388 : errmsg("column \"%s\" of relation \"%s\" does not exist",
389 : col, RelationGetRelationName(onerel))));
390 94 : if (bms_is_member(i, unique_cols))
391 12 : ereport(ERROR,
392 : (errcode(ERRCODE_DUPLICATE_COLUMN),
393 : errmsg("column \"%s\" of relation \"%s\" appears more than once",
394 : col, RelationGetRelationName(onerel))));
395 82 : unique_cols = bms_add_member(unique_cols, i);
396 :
397 82 : vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
398 82 : if (vacattrstats[tcnt] != NULL)
399 82 : tcnt++;
400 : }
401 50 : attr_cnt = tcnt;
402 : }
403 : else
404 : {
405 15094 : attr_cnt = onerel->rd_att->natts;
406 : vacattrstats = (VacAttrStats **)
407 15094 : palloc(attr_cnt * sizeof(VacAttrStats *));
408 15094 : tcnt = 0;
409 123968 : for (i = 1; i <= attr_cnt; i++)
410 : {
411 108874 : vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
412 108874 : if (vacattrstats[tcnt] != NULL)
413 108844 : tcnt++;
414 : }
415 15094 : attr_cnt = tcnt;
416 : }
417 :
418 : /*
419 : * Open all indexes of the relation, and see if there are any analyzable
420 : * columns in the indexes. We do not analyze index columns if there was
421 : * an explicit column list in the ANALYZE command, however.
422 : *
423 : * If we are doing a recursive scan, we don't want to touch the parent's
424 : * indexes at all. If we're processing a partitioned table, we need to
425 : * know if there are any indexes, but we don't want to process them.
426 : */
427 15144 : if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
428 : {
429 698 : List *idxs = RelationGetIndexList(onerel);
430 :
431 698 : Irel = NULL;
432 698 : nindexes = 0;
433 698 : hasindex = idxs != NIL;
434 698 : list_free(idxs);
435 : }
436 14446 : else if (!inh)
437 : {
438 14332 : vac_open_indexes(onerel, AccessShareLock, &nindexes, &Irel);
439 14332 : hasindex = nindexes > 0;
440 : }
441 : else
442 : {
443 114 : Irel = NULL;
444 114 : nindexes = 0;
445 114 : hasindex = false;
446 : }
447 15144 : indexdata = NULL;
448 15144 : if (nindexes > 0)
449 : {
450 11224 : indexdata = (AnlIndexData *) palloc0(nindexes * sizeof(AnlIndexData));
451 32342 : for (ind = 0; ind < nindexes; ind++)
452 : {
453 21118 : AnlIndexData *thisdata = &indexdata[ind];
454 : IndexInfo *indexInfo;
455 :
456 21118 : thisdata->indexInfo = indexInfo = BuildIndexInfo(Irel[ind]);
457 21118 : thisdata->tupleFract = 1.0; /* fix later if partial */
458 21118 : if (indexInfo->ii_Expressions != NIL && va_cols == NIL)
459 : {
460 94 : ListCell *indexpr_item = list_head(indexInfo->ii_Expressions);
461 :
462 94 : thisdata->vacattrstats = (VacAttrStats **)
463 94 : palloc(indexInfo->ii_NumIndexAttrs * sizeof(VacAttrStats *));
464 94 : tcnt = 0;
465 194 : for (i = 0; i < indexInfo->ii_NumIndexAttrs; i++)
466 : {
467 100 : int keycol = indexInfo->ii_IndexAttrNumbers[i];
468 :
469 100 : if (keycol == 0)
470 : {
471 : /* Found an index expression */
472 : Node *indexkey;
473 :
474 94 : if (indexpr_item == NULL) /* shouldn't happen */
475 0 : elog(ERROR, "too few entries in indexprs list");
476 94 : indexkey = (Node *) lfirst(indexpr_item);
477 94 : indexpr_item = lnext(indexInfo->ii_Expressions,
478 : indexpr_item);
479 188 : thisdata->vacattrstats[tcnt] =
480 94 : examine_attribute(Irel[ind], i + 1, indexkey);
481 94 : if (thisdata->vacattrstats[tcnt] != NULL)
482 94 : tcnt++;
483 : }
484 : }
485 94 : thisdata->attr_cnt = tcnt;
486 : }
487 : }
488 : }
489 :
490 : /*
491 : * Determine how many rows we need to sample, using the worst case from
492 : * all analyzable columns. We use a lower bound of 100 rows to avoid
493 : * possible overflow in Vitter's algorithm. (Note: that will also be the
494 : * target in the corner case where there are no analyzable columns.)
495 : */
496 15144 : targrows = 100;
497 124046 : for (i = 0; i < attr_cnt; i++)
498 : {
499 108902 : if (targrows < vacattrstats[i]->minrows)
500 15108 : targrows = vacattrstats[i]->minrows;
501 : }
502 36262 : for (ind = 0; ind < nindexes; ind++)
503 : {
504 21118 : AnlIndexData *thisdata = &indexdata[ind];
505 :
506 21212 : for (i = 0; i < thisdata->attr_cnt; i++)
507 : {
508 94 : if (targrows < thisdata->vacattrstats[i]->minrows)
509 0 : targrows = thisdata->vacattrstats[i]->minrows;
510 : }
511 : }
512 :
513 : /*
514 : * Look at extended statistics objects too, as those may define custom
515 : * statistics target. So we may need to sample more rows and then build
516 : * the statistics with enough detail.
517 : */
518 15144 : minrows = ComputeExtStatisticsRows(onerel, attr_cnt, vacattrstats);
519 :
520 15144 : if (targrows < minrows)
521 0 : targrows = minrows;
522 :
523 : /*
524 : * Acquire the sample rows
525 : */
526 15144 : rows = (HeapTuple *) palloc(targrows * sizeof(HeapTuple));
527 15144 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
528 : inh ? PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS_INH :
529 : PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS);
530 15144 : if (inh)
531 812 : numrows = acquire_inherited_sample_rows(onerel, elevel,
532 : rows, targrows,
533 : &totalrows, &totaldeadrows);
534 : else
535 14332 : numrows = (*acquirefunc) (onerel, elevel,
536 : rows, targrows,
537 : &totalrows, &totaldeadrows);
538 :
539 : /*
540 : * Compute the statistics. Temporary results during the calculations for
541 : * each column are stored in a child context. The calc routines are
542 : * responsible to make sure that whatever they store into the VacAttrStats
543 : * structure is allocated in anl_context.
544 : */
545 15142 : if (numrows > 0)
546 : {
547 : MemoryContext col_context,
548 : old_context;
549 :
550 10266 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
551 : PROGRESS_ANALYZE_PHASE_COMPUTE_STATS);
552 :
553 10266 : col_context = AllocSetContextCreate(anl_context,
554 : "Analyze Column",
555 : ALLOCSET_DEFAULT_SIZES);
556 10266 : old_context = MemoryContextSwitchTo(col_context);
557 :
558 90460 : for (i = 0; i < attr_cnt; i++)
559 : {
560 80194 : VacAttrStats *stats = vacattrstats[i];
561 : AttributeOpts *aopt;
562 :
563 80194 : stats->rows = rows;
564 80194 : stats->tupDesc = onerel->rd_att;
565 80194 : stats->compute_stats(stats,
566 : std_fetch_func,
567 : numrows,
568 : totalrows);
569 :
570 : /*
571 : * If the appropriate flavor of the n_distinct option is
572 : * specified, override with the corresponding value.
573 : */
574 80194 : aopt = get_attribute_options(onerel->rd_id, stats->tupattnum);
575 80194 : if (aopt != NULL)
576 : {
577 : float8 n_distinct;
578 :
579 6 : n_distinct = inh ? aopt->n_distinct_inherited : aopt->n_distinct;
580 6 : if (n_distinct != 0.0)
581 6 : stats->stadistinct = n_distinct;
582 : }
583 :
584 80194 : MemoryContextReset(col_context);
585 : }
586 :
587 10266 : if (nindexes > 0)
588 6610 : compute_index_stats(onerel, totalrows,
589 : indexdata, nindexes,
590 : rows, numrows,
591 : col_context);
592 :
593 10260 : MemoryContextSwitchTo(old_context);
594 10260 : MemoryContextDelete(col_context);
595 :
596 : /*
597 : * Emit the completed stats rows into pg_statistic, replacing any
598 : * previous statistics for the target columns. (If there are stats in
599 : * pg_statistic for columns we didn't process, we leave them alone.)
600 : */
601 10260 : update_attstats(RelationGetRelid(onerel), inh,
602 : attr_cnt, vacattrstats);
603 :
604 23142 : for (ind = 0; ind < nindexes; ind++)
605 : {
606 12882 : AnlIndexData *thisdata = &indexdata[ind];
607 :
608 12882 : update_attstats(RelationGetRelid(Irel[ind]), false,
609 : thisdata->attr_cnt, thisdata->vacattrstats);
610 : }
611 :
612 : /* Build extended statistics (if there are any). */
613 10260 : BuildRelationExtStatistics(onerel, inh, totalrows, numrows, rows,
614 : attr_cnt, vacattrstats);
615 : }
616 :
617 15136 : pgstat_progress_update_param(PROGRESS_ANALYZE_PHASE,
618 : PROGRESS_ANALYZE_PHASE_FINALIZE_ANALYZE);
619 :
620 : /*
621 : * Update pages/tuples stats in pg_class ... but not if we're doing
622 : * inherited stats.
623 : *
624 : * We assume that VACUUM hasn't set pg_class.reltuples already, even
625 : * during a VACUUM ANALYZE. Although VACUUM often updates pg_class,
626 : * exceptions exist. A "VACUUM (ANALYZE, INDEX_CLEANUP OFF)" command will
627 : * never update pg_class entries for index relations. It's also possible
628 : * that an individual index's pg_class entry won't be updated during
629 : * VACUUM if the index AM returns NULL from its amvacuumcleanup() routine.
630 : */
631 15136 : if (!inh)
632 : {
633 14324 : BlockNumber relallvisible = 0;
634 14324 : BlockNumber relallfrozen = 0;
635 :
636 14324 : if (RELKIND_HAS_STORAGE(onerel->rd_rel->relkind))
637 14270 : visibilitymap_count(onerel, &relallvisible, &relallfrozen);
638 :
639 : /*
640 : * Update pg_class for table relation. CCI first, in case acquirefunc
641 : * updated pg_class.
642 : */
643 14324 : CommandCounterIncrement();
644 14324 : vac_update_relstats(onerel,
645 : relpages,
646 : totalrows,
647 : relallvisible,
648 : relallfrozen,
649 : hasindex,
650 : InvalidTransactionId,
651 : InvalidMultiXactId,
652 : NULL, NULL,
653 : in_outer_xact);
654 :
655 : /* Same for indexes */
656 35430 : for (ind = 0; ind < nindexes; ind++)
657 : {
658 21106 : AnlIndexData *thisdata = &indexdata[ind];
659 : double totalindexrows;
660 :
661 21106 : totalindexrows = ceil(thisdata->tupleFract * totalrows);
662 21106 : vac_update_relstats(Irel[ind],
663 21106 : RelationGetNumberOfBlocks(Irel[ind]),
664 : totalindexrows,
665 : 0, 0,
666 : false,
667 : InvalidTransactionId,
668 : InvalidMultiXactId,
669 : NULL, NULL,
670 : in_outer_xact);
671 : }
672 : }
673 812 : else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
674 : {
675 : /*
676 : * Partitioned tables don't have storage, so we don't set any fields
677 : * in their pg_class entries except for reltuples and relhasindex.
678 : */
679 698 : CommandCounterIncrement();
680 698 : vac_update_relstats(onerel, -1, totalrows,
681 : 0, 0, hasindex, InvalidTransactionId,
682 : InvalidMultiXactId,
683 : NULL, NULL,
684 : in_outer_xact);
685 : }
686 :
687 : /*
688 : * Now report ANALYZE to the cumulative stats system. For regular tables,
689 : * we do it only if not doing inherited stats. For partitioned tables, we
690 : * only do it for inherited stats. (We're never called for not-inherited
691 : * stats on partitioned tables anyway.)
692 : *
693 : * Reset the changes_since_analyze counter only if we analyzed all
694 : * columns; otherwise, there is still work for auto-analyze to do.
695 : */
696 15136 : if (!inh)
697 14324 : pgstat_report_analyze(onerel, totalrows, totaldeadrows,
698 : (va_cols == NIL), starttime);
699 812 : else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
700 698 : pgstat_report_analyze(onerel, 0, 0, (va_cols == NIL), starttime);
701 :
702 : /*
703 : * If this isn't part of VACUUM ANALYZE, let index AMs do cleanup.
704 : *
705 : * Note that most index AMs perform a no-op as a matter of policy for
706 : * amvacuumcleanup() when called in ANALYZE-only mode. The only exception
707 : * among core index AMs is GIN/ginvacuumcleanup().
708 : */
709 15136 : if (!(params->options & VACOPT_VACUUM))
710 : {
711 28528 : for (ind = 0; ind < nindexes; ind++)
712 : {
713 : IndexBulkDeleteResult *stats;
714 : IndexVacuumInfo ivinfo;
715 :
716 16338 : ivinfo.index = Irel[ind];
717 16338 : ivinfo.heaprel = onerel;
718 16338 : ivinfo.analyze_only = true;
719 16338 : ivinfo.estimated_count = true;
720 16338 : ivinfo.message_level = elevel;
721 16338 : ivinfo.num_heap_tuples = onerel->rd_rel->reltuples;
722 16338 : ivinfo.strategy = vac_strategy;
723 :
724 16338 : stats = index_vacuum_cleanup(&ivinfo, NULL);
725 :
726 16338 : if (stats)
727 0 : pfree(stats);
728 : }
729 : }
730 :
731 : /* Done with indexes */
732 15136 : vac_close_indexes(nindexes, Irel, NoLock);
733 :
734 : /* Log the action if appropriate */
735 15136 : if (instrument)
736 : {
737 404 : TimestampTz endtime = GetCurrentTimestamp();
738 :
739 518 : if (verbose || params->log_min_duration == 0 ||
740 114 : TimestampDifferenceExceeds(starttime, endtime,
741 : params->log_min_duration))
742 : {
743 : long delay_in_ms;
744 : WalUsage walusage;
745 290 : double read_rate = 0;
746 290 : double write_rate = 0;
747 : char *msgfmt;
748 : StringInfoData buf;
749 : int64 total_blks_hit;
750 : int64 total_blks_read;
751 : int64 total_blks_dirtied;
752 :
753 290 : memset(&bufferusage, 0, sizeof(BufferUsage));
754 290 : BufferUsageAccumDiff(&bufferusage, &pgBufferUsage, &startbufferusage);
755 290 : memset(&walusage, 0, sizeof(WalUsage));
756 290 : WalUsageAccumDiff(&walusage, &pgWalUsage, &startwalusage);
757 :
758 290 : total_blks_hit = bufferusage.shared_blks_hit +
759 290 : bufferusage.local_blks_hit;
760 290 : total_blks_read = bufferusage.shared_blks_read +
761 290 : bufferusage.local_blks_read;
762 290 : total_blks_dirtied = bufferusage.shared_blks_dirtied +
763 290 : bufferusage.local_blks_dirtied;
764 :
765 : /*
766 : * We do not expect an analyze to take > 25 days and it simplifies
767 : * things a bit to use TimestampDifferenceMilliseconds.
768 : */
769 290 : delay_in_ms = TimestampDifferenceMilliseconds(starttime, endtime);
770 :
771 : /*
772 : * Note that we are reporting these read/write rates in the same
773 : * manner as VACUUM does, which means that while the 'average read
774 : * rate' here actually corresponds to page misses and resulting
775 : * reads which are also picked up by track_io_timing, if enabled,
776 : * the 'average write rate' is actually talking about the rate of
777 : * pages being dirtied, not being written out, so it's typical to
778 : * have a non-zero 'avg write rate' while I/O timings only reports
779 : * reads.
780 : *
781 : * It's not clear that an ANALYZE will ever result in
782 : * FlushBuffer() being called, but we track and support reporting
783 : * on I/O write time in case that changes as it's practically free
784 : * to do so anyway.
785 : */
786 :
787 290 : if (delay_in_ms > 0)
788 : {
789 290 : read_rate = (double) BLCKSZ * total_blks_read /
790 290 : (1024 * 1024) / (delay_in_ms / 1000.0);
791 290 : write_rate = (double) BLCKSZ * total_blks_dirtied /
792 290 : (1024 * 1024) / (delay_in_ms / 1000.0);
793 : }
794 :
795 : /*
796 : * We split this up so we don't emit empty I/O timing values when
797 : * track_io_timing isn't enabled.
798 : */
799 :
800 290 : initStringInfo(&buf);
801 :
802 290 : if (AmAutoVacuumWorkerProcess())
803 290 : msgfmt = _("automatic analyze of table \"%s.%s.%s\"\n");
804 : else
805 0 : msgfmt = _("finished analyzing table \"%s.%s.%s\"\n");
806 :
807 290 : appendStringInfo(&buf, msgfmt,
808 : get_database_name(MyDatabaseId),
809 290 : get_namespace_name(RelationGetNamespace(onerel)),
810 290 : RelationGetRelationName(onerel));
811 290 : if (track_cost_delay_timing)
812 : {
813 : /*
814 : * We bypass the changecount mechanism because this value is
815 : * only updated by the calling process.
816 : */
817 0 : appendStringInfo(&buf, _("delay time: %.3f ms\n"),
818 0 : (double) MyBEEntry->st_progress_param[PROGRESS_ANALYZE_DELAY_TIME] / 1000000.0);
819 : }
820 290 : if (track_io_timing)
821 : {
822 0 : double read_ms = (double) (pgStatBlockReadTime - startreadtime) / 1000;
823 0 : double write_ms = (double) (pgStatBlockWriteTime - startwritetime) / 1000;
824 :
825 0 : appendStringInfo(&buf, _("I/O timings: read: %.3f ms, write: %.3f ms\n"),
826 : read_ms, write_ms);
827 : }
828 290 : appendStringInfo(&buf, _("avg read rate: %.3f MB/s, avg write rate: %.3f MB/s\n"),
829 : read_rate, write_rate);
830 290 : appendStringInfo(&buf, _("buffer usage: %" PRId64 " hits, %" PRId64 " reads, %" PRId64 " dirtied\n"),
831 : total_blks_hit,
832 : total_blks_read,
833 : total_blks_dirtied);
834 290 : appendStringInfo(&buf,
835 290 : _("WAL usage: %" PRId64 " records, %" PRId64 " full page images, %" PRIu64 " bytes, %" PRId64 " buffers full\n"),
836 : walusage.wal_records,
837 : walusage.wal_fpi,
838 : walusage.wal_bytes,
839 : walusage.wal_buffers_full);
840 290 : appendStringInfo(&buf, _("system usage: %s"), pg_rusage_show(&ru0));
841 :
842 290 : ereport(verbose ? INFO : LOG,
843 : (errmsg_internal("%s", buf.data)));
844 :
845 290 : pfree(buf.data);
846 : }
847 : }
848 :
849 : /* Roll back any GUC changes executed by index functions */
850 15136 : AtEOXact_GUC(false, save_nestlevel);
851 :
852 : /* Restore userid and security context */
853 15136 : SetUserIdAndSecContext(save_userid, save_sec_context);
854 :
855 : /* Restore current context and release memory */
856 15136 : MemoryContextSwitchTo(caller_context);
857 15136 : MemoryContextDelete(anl_context);
858 15136 : anl_context = NULL;
859 15136 : }
860 :
861 : /*
862 : * Compute statistics about indexes of a relation
863 : */
864 : static void
865 6610 : compute_index_stats(Relation onerel, double totalrows,
866 : AnlIndexData *indexdata, int nindexes,
867 : HeapTuple *rows, int numrows,
868 : MemoryContext col_context)
869 : {
870 : MemoryContext ind_context,
871 : old_context;
872 : Datum values[INDEX_MAX_KEYS];
873 : bool isnull[INDEX_MAX_KEYS];
874 : int ind,
875 : i;
876 :
877 6610 : ind_context = AllocSetContextCreate(anl_context,
878 : "Analyze Index",
879 : ALLOCSET_DEFAULT_SIZES);
880 6610 : old_context = MemoryContextSwitchTo(ind_context);
881 :
882 19498 : for (ind = 0; ind < nindexes; ind++)
883 : {
884 12894 : AnlIndexData *thisdata = &indexdata[ind];
885 12894 : IndexInfo *indexInfo = thisdata->indexInfo;
886 12894 : int attr_cnt = thisdata->attr_cnt;
887 : TupleTableSlot *slot;
888 : EState *estate;
889 : ExprContext *econtext;
890 : ExprState *predicate;
891 : Datum *exprvals;
892 : bool *exprnulls;
893 : int numindexrows,
894 : tcnt,
895 : rowno;
896 : double totalindexrows;
897 :
898 : /* Ignore index if no columns to analyze and not partial */
899 12894 : if (attr_cnt == 0 && indexInfo->ii_Predicate == NIL)
900 12760 : continue;
901 :
902 : /*
903 : * Need an EState for evaluation of index expressions and
904 : * partial-index predicates. Create it in the per-index context to be
905 : * sure it gets cleaned up at the bottom of the loop.
906 : */
907 134 : estate = CreateExecutorState();
908 134 : econtext = GetPerTupleExprContext(estate);
909 : /* Need a slot to hold the current heap tuple, too */
910 134 : slot = MakeSingleTupleTableSlot(RelationGetDescr(onerel),
911 : &TTSOpsHeapTuple);
912 :
913 : /* Arrange for econtext's scan tuple to be the tuple under test */
914 134 : econtext->ecxt_scantuple = slot;
915 :
916 : /* Set up execution state for predicate. */
917 134 : predicate = ExecPrepareQual(indexInfo->ii_Predicate, estate);
918 :
919 : /* Compute and save index expression values */
920 134 : exprvals = (Datum *) palloc(numrows * attr_cnt * sizeof(Datum));
921 134 : exprnulls = (bool *) palloc(numrows * attr_cnt * sizeof(bool));
922 134 : numindexrows = 0;
923 134 : tcnt = 0;
924 155224 : for (rowno = 0; rowno < numrows; rowno++)
925 : {
926 155096 : HeapTuple heapTuple = rows[rowno];
927 :
928 155096 : vacuum_delay_point(true);
929 :
930 : /*
931 : * Reset the per-tuple context each time, to reclaim any cruft
932 : * left behind by evaluating the predicate or index expressions.
933 : */
934 155096 : ResetExprContext(econtext);
935 :
936 : /* Set up for predicate or expression evaluation */
937 155096 : ExecStoreHeapTuple(heapTuple, slot, false);
938 :
939 : /* If index is partial, check predicate */
940 155096 : if (predicate != NULL)
941 : {
942 40066 : if (!ExecQual(predicate, econtext))
943 21330 : continue;
944 : }
945 133766 : numindexrows++;
946 :
947 133766 : if (attr_cnt > 0)
948 : {
949 : /*
950 : * Evaluate the index row to compute expression values. We
951 : * could do this by hand, but FormIndexDatum is convenient.
952 : */
953 115030 : FormIndexDatum(indexInfo,
954 : slot,
955 : estate,
956 : values,
957 : isnull);
958 :
959 : /*
960 : * Save just the columns we care about. We copy the values
961 : * into ind_context from the estate's per-tuple context.
962 : */
963 230048 : for (i = 0; i < attr_cnt; i++)
964 : {
965 115024 : VacAttrStats *stats = thisdata->vacattrstats[i];
966 115024 : int attnum = stats->tupattnum;
967 :
968 115024 : if (isnull[attnum - 1])
969 : {
970 6 : exprvals[tcnt] = (Datum) 0;
971 6 : exprnulls[tcnt] = true;
972 : }
973 : else
974 : {
975 230036 : exprvals[tcnt] = datumCopy(values[attnum - 1],
976 115018 : stats->attrtype->typbyval,
977 115018 : stats->attrtype->typlen);
978 115018 : exprnulls[tcnt] = false;
979 : }
980 115024 : tcnt++;
981 : }
982 : }
983 : }
984 :
985 : /*
986 : * Having counted the number of rows that pass the predicate in the
987 : * sample, we can estimate the total number of rows in the index.
988 : */
989 128 : thisdata->tupleFract = (double) numindexrows / (double) numrows;
990 128 : totalindexrows = ceil(thisdata->tupleFract * totalrows);
991 :
992 : /*
993 : * Now we can compute the statistics for the expression columns.
994 : */
995 128 : if (numindexrows > 0)
996 : {
997 120 : MemoryContextSwitchTo(col_context);
998 202 : for (i = 0; i < attr_cnt; i++)
999 : {
1000 82 : VacAttrStats *stats = thisdata->vacattrstats[i];
1001 :
1002 82 : stats->exprvals = exprvals + i;
1003 82 : stats->exprnulls = exprnulls + i;
1004 82 : stats->rowstride = attr_cnt;
1005 82 : stats->compute_stats(stats,
1006 : ind_fetch_func,
1007 : numindexrows,
1008 : totalindexrows);
1009 :
1010 82 : MemoryContextReset(col_context);
1011 : }
1012 : }
1013 :
1014 : /* And clean up */
1015 128 : MemoryContextSwitchTo(ind_context);
1016 :
1017 128 : ExecDropSingleTupleTableSlot(slot);
1018 128 : FreeExecutorState(estate);
1019 128 : MemoryContextReset(ind_context);
1020 : }
1021 :
1022 6604 : MemoryContextSwitchTo(old_context);
1023 6604 : MemoryContextDelete(ind_context);
1024 6604 : }
1025 :
1026 : /*
1027 : * examine_attribute -- pre-analysis of a single column
1028 : *
1029 : * Determine whether the column is analyzable; if so, create and initialize
1030 : * a VacAttrStats struct for it. If not, return NULL.
1031 : *
1032 : * If index_expr isn't NULL, then we're trying to analyze an expression index,
1033 : * and index_expr is the expression tree representing the column's data.
1034 : */
1035 : static VacAttrStats *
1036 109050 : examine_attribute(Relation onerel, int attnum, Node *index_expr)
1037 : {
1038 109050 : Form_pg_attribute attr = TupleDescAttr(onerel->rd_att, attnum - 1);
1039 : int attstattarget;
1040 : HeapTuple atttuple;
1041 : Datum dat;
1042 : bool isnull;
1043 : HeapTuple typtuple;
1044 : VacAttrStats *stats;
1045 : int i;
1046 : bool ok;
1047 :
1048 : /* Never analyze dropped columns */
1049 109050 : if (attr->attisdropped)
1050 6 : return NULL;
1051 :
1052 : /* Don't analyze virtual generated columns */
1053 109044 : if (attr->attgenerated == ATTRIBUTE_GENERATED_VIRTUAL)
1054 18 : return NULL;
1055 :
1056 : /*
1057 : * Get attstattarget value. Set to -1 if null. (Analyze functions expect
1058 : * -1 to mean use default_statistics_target; see for example
1059 : * std_typanalyze.)
1060 : */
1061 109026 : atttuple = SearchSysCache2(ATTNUM, ObjectIdGetDatum(RelationGetRelid(onerel)), Int16GetDatum(attnum));
1062 109026 : if (!HeapTupleIsValid(atttuple))
1063 0 : elog(ERROR, "cache lookup failed for attribute %d of relation %u",
1064 : attnum, RelationGetRelid(onerel));
1065 109026 : dat = SysCacheGetAttr(ATTNUM, atttuple, Anum_pg_attribute_attstattarget, &isnull);
1066 109026 : attstattarget = isnull ? -1 : DatumGetInt16(dat);
1067 109026 : ReleaseSysCache(atttuple);
1068 :
1069 : /* Don't analyze column if user has specified not to */
1070 109026 : if (attstattarget == 0)
1071 6 : return NULL;
1072 :
1073 : /*
1074 : * Create the VacAttrStats struct.
1075 : */
1076 109020 : stats = (VacAttrStats *) palloc0(sizeof(VacAttrStats));
1077 109020 : stats->attstattarget = attstattarget;
1078 :
1079 : /*
1080 : * When analyzing an expression index, believe the expression tree's type
1081 : * not the column datatype --- the latter might be the opckeytype storage
1082 : * type of the opclass, which is not interesting for our purposes. (Note:
1083 : * if we did anything with non-expression index columns, we'd need to
1084 : * figure out where to get the correct type info from, but for now that's
1085 : * not a problem.) It's not clear whether anyone will care about the
1086 : * typmod, but we store that too just in case.
1087 : */
1088 109020 : if (index_expr)
1089 : {
1090 94 : stats->attrtypid = exprType(index_expr);
1091 94 : stats->attrtypmod = exprTypmod(index_expr);
1092 :
1093 : /*
1094 : * If a collation has been specified for the index column, use that in
1095 : * preference to anything else; but if not, fall back to whatever we
1096 : * can get from the expression.
1097 : */
1098 94 : if (OidIsValid(onerel->rd_indcollation[attnum - 1]))
1099 12 : stats->attrcollid = onerel->rd_indcollation[attnum - 1];
1100 : else
1101 82 : stats->attrcollid = exprCollation(index_expr);
1102 : }
1103 : else
1104 : {
1105 108926 : stats->attrtypid = attr->atttypid;
1106 108926 : stats->attrtypmod = attr->atttypmod;
1107 108926 : stats->attrcollid = attr->attcollation;
1108 : }
1109 :
1110 109020 : typtuple = SearchSysCacheCopy1(TYPEOID,
1111 : ObjectIdGetDatum(stats->attrtypid));
1112 109020 : if (!HeapTupleIsValid(typtuple))
1113 0 : elog(ERROR, "cache lookup failed for type %u", stats->attrtypid);
1114 109020 : stats->attrtype = (Form_pg_type) GETSTRUCT(typtuple);
1115 109020 : stats->anl_context = anl_context;
1116 109020 : stats->tupattnum = attnum;
1117 :
1118 : /*
1119 : * The fields describing the stats->stavalues[n] element types default to
1120 : * the type of the data being analyzed, but the type-specific typanalyze
1121 : * function can change them if it wants to store something else.
1122 : */
1123 654120 : for (i = 0; i < STATISTIC_NUM_SLOTS; i++)
1124 : {
1125 545100 : stats->statypid[i] = stats->attrtypid;
1126 545100 : stats->statyplen[i] = stats->attrtype->typlen;
1127 545100 : stats->statypbyval[i] = stats->attrtype->typbyval;
1128 545100 : stats->statypalign[i] = stats->attrtype->typalign;
1129 : }
1130 :
1131 : /*
1132 : * Call the type-specific typanalyze function. If none is specified, use
1133 : * std_typanalyze().
1134 : */
1135 109020 : if (OidIsValid(stats->attrtype->typanalyze))
1136 7042 : ok = DatumGetBool(OidFunctionCall1(stats->attrtype->typanalyze,
1137 : PointerGetDatum(stats)));
1138 : else
1139 101978 : ok = std_typanalyze(stats);
1140 :
1141 109020 : if (!ok || stats->compute_stats == NULL || stats->minrows <= 0)
1142 : {
1143 0 : heap_freetuple(typtuple);
1144 0 : pfree(stats);
1145 0 : return NULL;
1146 : }
1147 :
1148 109020 : return stats;
1149 : }
1150 :
1151 : /*
1152 : * Read stream callback returning the next BlockNumber as chosen by the
1153 : * BlockSampling algorithm.
1154 : */
1155 : static BlockNumber
1156 144752 : block_sampling_read_stream_next(ReadStream *stream,
1157 : void *callback_private_data,
1158 : void *per_buffer_data)
1159 : {
1160 144752 : BlockSamplerData *bs = callback_private_data;
1161 :
1162 144752 : return BlockSampler_HasMore(bs) ? BlockSampler_Next(bs) : InvalidBlockNumber;
1163 : }
1164 :
1165 : /*
1166 : * acquire_sample_rows -- acquire a random sample of rows from the table
1167 : *
1168 : * Selected rows are returned in the caller-allocated array rows[], which
1169 : * must have at least targrows entries.
1170 : * The actual number of rows selected is returned as the function result.
1171 : * We also estimate the total numbers of live and dead rows in the table,
1172 : * and return them into *totalrows and *totaldeadrows, respectively.
1173 : *
1174 : * The returned list of tuples is in order by physical position in the table.
1175 : * (We will rely on this later to derive correlation estimates.)
1176 : *
1177 : * As of May 2004 we use a new two-stage method: Stage one selects up
1178 : * to targrows random blocks (or all blocks, if there aren't so many).
1179 : * Stage two scans these blocks and uses the Vitter algorithm to create
1180 : * a random sample of targrows rows (or less, if there are less in the
1181 : * sample of blocks). The two stages are executed simultaneously: each
1182 : * block is processed as soon as stage one returns its number and while
1183 : * the rows are read stage two controls which ones are to be inserted
1184 : * into the sample.
1185 : *
1186 : * Although every row has an equal chance of ending up in the final
1187 : * sample, this sampling method is not perfect: not every possible
1188 : * sample has an equal chance of being selected. For large relations
1189 : * the number of different blocks represented by the sample tends to be
1190 : * too small. We can live with that for now. Improvements are welcome.
1191 : *
1192 : * An important property of this sampling method is that because we do
1193 : * look at a statistically unbiased set of blocks, we should get
1194 : * unbiased estimates of the average numbers of live and dead rows per
1195 : * block. The previous sampling method put too much credence in the row
1196 : * density near the start of the table.
1197 : */
1198 : static int
1199 16192 : acquire_sample_rows(Relation onerel, int elevel,
1200 : HeapTuple *rows, int targrows,
1201 : double *totalrows, double *totaldeadrows)
1202 : {
1203 16192 : int numrows = 0; /* # rows now in reservoir */
1204 16192 : double samplerows = 0; /* total # rows collected */
1205 16192 : double liverows = 0; /* # live rows seen */
1206 16192 : double deadrows = 0; /* # dead rows seen */
1207 16192 : double rowstoskip = -1; /* -1 means not set yet */
1208 : uint32 randseed; /* Seed for block sampler(s) */
1209 : BlockNumber totalblocks;
1210 : TransactionId OldestXmin;
1211 : BlockSamplerData bs;
1212 : ReservoirStateData rstate;
1213 : TupleTableSlot *slot;
1214 : TableScanDesc scan;
1215 : BlockNumber nblocks;
1216 16192 : BlockNumber blksdone = 0;
1217 : ReadStream *stream;
1218 :
1219 : Assert(targrows > 0);
1220 :
1221 16192 : totalblocks = RelationGetNumberOfBlocks(onerel);
1222 :
1223 : /* Need a cutoff xmin for HeapTupleSatisfiesVacuum */
1224 16192 : OldestXmin = GetOldestNonRemovableTransactionId(onerel);
1225 :
1226 : /* Prepare for sampling block numbers */
1227 16192 : randseed = pg_prng_uint32(&pg_global_prng_state);
1228 16192 : nblocks = BlockSampler_Init(&bs, totalblocks, targrows, randseed);
1229 :
1230 : /* Report sampling block numbers */
1231 16192 : pgstat_progress_update_param(PROGRESS_ANALYZE_BLOCKS_TOTAL,
1232 : nblocks);
1233 :
1234 : /* Prepare for sampling rows */
1235 16192 : reservoir_init_selection_state(&rstate, targrows);
1236 :
1237 16192 : scan = table_beginscan_analyze(onerel);
1238 16192 : slot = table_slot_create(onerel, NULL);
1239 :
1240 : /*
1241 : * It is safe to use batching, as block_sampling_read_stream_next never
1242 : * blocks.
1243 : */
1244 16192 : stream = read_stream_begin_relation(READ_STREAM_MAINTENANCE |
1245 : READ_STREAM_USE_BATCHING,
1246 : vac_strategy,
1247 : scan->rs_rd,
1248 : MAIN_FORKNUM,
1249 : block_sampling_read_stream_next,
1250 : &bs,
1251 : 0);
1252 :
1253 : /* Outer loop over blocks to sample */
1254 144752 : while (table_scan_analyze_next_block(scan, stream))
1255 : {
1256 128560 : vacuum_delay_point(true);
1257 :
1258 10283636 : while (table_scan_analyze_next_tuple(scan, OldestXmin, &liverows, &deadrows, slot))
1259 : {
1260 : /*
1261 : * The first targrows sample rows are simply copied into the
1262 : * reservoir. Then we start replacing tuples in the sample until
1263 : * we reach the end of the relation. This algorithm is from Jeff
1264 : * Vitter's paper (see full citation in utils/misc/sampling.c). It
1265 : * works by repeatedly computing the number of tuples to skip
1266 : * before selecting a tuple, which replaces a randomly chosen
1267 : * element of the reservoir (current set of tuples). At all times
1268 : * the reservoir is a true random sample of the tuples we've
1269 : * passed over so far, so when we fall off the end of the relation
1270 : * we're done.
1271 : */
1272 10155076 : if (numrows < targrows)
1273 9903932 : rows[numrows++] = ExecCopySlotHeapTuple(slot);
1274 : else
1275 : {
1276 : /*
1277 : * t in Vitter's paper is the number of records already
1278 : * processed. If we need to compute a new S value, we must
1279 : * use the not-yet-incremented value of samplerows as t.
1280 : */
1281 251144 : if (rowstoskip < 0)
1282 115376 : rowstoskip = reservoir_get_next_S(&rstate, samplerows, targrows);
1283 :
1284 251144 : if (rowstoskip <= 0)
1285 : {
1286 : /*
1287 : * Found a suitable tuple, so save it, replacing one old
1288 : * tuple at random
1289 : */
1290 115326 : int k = (int) (targrows * sampler_random_fract(&rstate.randstate));
1291 :
1292 : Assert(k >= 0 && k < targrows);
1293 115326 : heap_freetuple(rows[k]);
1294 115326 : rows[k] = ExecCopySlotHeapTuple(slot);
1295 : }
1296 :
1297 251144 : rowstoskip -= 1;
1298 : }
1299 :
1300 10155076 : samplerows += 1;
1301 : }
1302 :
1303 128560 : pgstat_progress_update_param(PROGRESS_ANALYZE_BLOCKS_DONE,
1304 : ++blksdone);
1305 : }
1306 :
1307 16192 : read_stream_end(stream);
1308 :
1309 16192 : ExecDropSingleTupleTableSlot(slot);
1310 16192 : table_endscan(scan);
1311 :
1312 : /*
1313 : * If we didn't find as many tuples as we wanted then we're done. No sort
1314 : * is needed, since they're already in order.
1315 : *
1316 : * Otherwise we need to sort the collected tuples by position
1317 : * (itempointer). It's not worth worrying about corner cases where the
1318 : * tuples are already sorted.
1319 : */
1320 16192 : if (numrows == targrows)
1321 162 : qsort_interruptible(rows, numrows, sizeof(HeapTuple),
1322 : compare_rows, NULL);
1323 :
1324 : /*
1325 : * Estimate total numbers of live and dead rows in relation, extrapolating
1326 : * on the assumption that the average tuple density in pages we didn't
1327 : * scan is the same as in the pages we did scan. Since what we scanned is
1328 : * a random sample of the pages in the relation, this should be a good
1329 : * assumption.
1330 : */
1331 16192 : if (bs.m > 0)
1332 : {
1333 11400 : *totalrows = floor((liverows / bs.m) * totalblocks + 0.5);
1334 11400 : *totaldeadrows = floor((deadrows / bs.m) * totalblocks + 0.5);
1335 : }
1336 : else
1337 : {
1338 4792 : *totalrows = 0.0;
1339 4792 : *totaldeadrows = 0.0;
1340 : }
1341 :
1342 : /*
1343 : * Emit some interesting relation info
1344 : */
1345 16192 : ereport(elevel,
1346 : (errmsg("\"%s\": scanned %d of %u pages, "
1347 : "containing %.0f live rows and %.0f dead rows; "
1348 : "%d rows in sample, %.0f estimated total rows",
1349 : RelationGetRelationName(onerel),
1350 : bs.m, totalblocks,
1351 : liverows, deadrows,
1352 : numrows, *totalrows)));
1353 :
1354 16192 : return numrows;
1355 : }
1356 :
1357 : /*
1358 : * Comparator for sorting rows[] array
1359 : */
1360 : static int
1361 4040622 : compare_rows(const void *a, const void *b, void *arg)
1362 : {
1363 4040622 : HeapTuple ha = *(const HeapTuple *) a;
1364 4040622 : HeapTuple hb = *(const HeapTuple *) b;
1365 4040622 : BlockNumber ba = ItemPointerGetBlockNumber(&ha->t_self);
1366 4040622 : OffsetNumber oa = ItemPointerGetOffsetNumber(&ha->t_self);
1367 4040622 : BlockNumber bb = ItemPointerGetBlockNumber(&hb->t_self);
1368 4040622 : OffsetNumber ob = ItemPointerGetOffsetNumber(&hb->t_self);
1369 :
1370 4040622 : if (ba < bb)
1371 886372 : return -1;
1372 3154250 : if (ba > bb)
1373 872554 : return 1;
1374 2281696 : if (oa < ob)
1375 1539830 : return -1;
1376 741866 : if (oa > ob)
1377 741866 : return 1;
1378 0 : return 0;
1379 : }
1380 :
1381 :
1382 : /*
1383 : * acquire_inherited_sample_rows -- acquire sample rows from inheritance tree
1384 : *
1385 : * This has the same API as acquire_sample_rows, except that rows are
1386 : * collected from all inheritance children as well as the specified table.
1387 : * We fail and return zero if there are no inheritance children, or if all
1388 : * children are foreign tables that don't support ANALYZE.
1389 : */
1390 : static int
1391 812 : acquire_inherited_sample_rows(Relation onerel, int elevel,
1392 : HeapTuple *rows, int targrows,
1393 : double *totalrows, double *totaldeadrows)
1394 : {
1395 : List *tableOIDs;
1396 : Relation *rels;
1397 : AcquireSampleRowsFunc *acquirefuncs;
1398 : double *relblocks;
1399 : double totalblocks;
1400 : int numrows,
1401 : nrels,
1402 : i;
1403 : ListCell *lc;
1404 : bool has_child;
1405 :
1406 : /* Initialize output parameters to zero now, in case we exit early */
1407 812 : *totalrows = 0;
1408 812 : *totaldeadrows = 0;
1409 :
1410 : /*
1411 : * Find all members of inheritance set. We only need AccessShareLock on
1412 : * the children.
1413 : */
1414 : tableOIDs =
1415 812 : find_all_inheritors(RelationGetRelid(onerel), AccessShareLock, NULL);
1416 :
1417 : /*
1418 : * Check that there's at least one descendant, else fail. This could
1419 : * happen despite analyze_rel's relhassubclass check, if table once had a
1420 : * child but no longer does. In that case, we can clear the
1421 : * relhassubclass field so as not to make the same mistake again later.
1422 : * (This is safe because we hold ShareUpdateExclusiveLock.)
1423 : */
1424 812 : if (list_length(tableOIDs) < 2)
1425 : {
1426 : /* CCI because we already updated the pg_class row in this command */
1427 14 : CommandCounterIncrement();
1428 14 : SetRelationHasSubclass(RelationGetRelid(onerel), false);
1429 14 : ereport(elevel,
1430 : (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no child tables",
1431 : get_namespace_name(RelationGetNamespace(onerel)),
1432 : RelationGetRelationName(onerel))));
1433 14 : return 0;
1434 : }
1435 :
1436 : /*
1437 : * Identify acquirefuncs to use, and count blocks in all the relations.
1438 : * The result could overflow BlockNumber, so we use double arithmetic.
1439 : */
1440 798 : rels = (Relation *) palloc(list_length(tableOIDs) * sizeof(Relation));
1441 : acquirefuncs = (AcquireSampleRowsFunc *)
1442 798 : palloc(list_length(tableOIDs) * sizeof(AcquireSampleRowsFunc));
1443 798 : relblocks = (double *) palloc(list_length(tableOIDs) * sizeof(double));
1444 798 : totalblocks = 0;
1445 798 : nrels = 0;
1446 798 : has_child = false;
1447 3656 : foreach(lc, tableOIDs)
1448 : {
1449 2858 : Oid childOID = lfirst_oid(lc);
1450 : Relation childrel;
1451 2858 : AcquireSampleRowsFunc acquirefunc = NULL;
1452 2858 : BlockNumber relpages = 0;
1453 :
1454 : /* We already got the needed lock */
1455 2858 : childrel = table_open(childOID, NoLock);
1456 :
1457 : /* Ignore if temp table of another backend */
1458 2858 : if (RELATION_IS_OTHER_TEMP(childrel))
1459 : {
1460 : /* ... but release the lock on it */
1461 : Assert(childrel != onerel);
1462 0 : table_close(childrel, AccessShareLock);
1463 758 : continue;
1464 : }
1465 :
1466 : /* Check table type (MATVIEW can't happen, but might as well allow) */
1467 2858 : if (childrel->rd_rel->relkind == RELKIND_RELATION ||
1468 788 : childrel->rd_rel->relkind == RELKIND_MATVIEW)
1469 : {
1470 : /* Regular table, so use the regular row acquisition function */
1471 2070 : acquirefunc = acquire_sample_rows;
1472 2070 : relpages = RelationGetNumberOfBlocks(childrel);
1473 : }
1474 788 : else if (childrel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
1475 : {
1476 : /*
1477 : * For a foreign table, call the FDW's hook function to see
1478 : * whether it supports analysis.
1479 : */
1480 : FdwRoutine *fdwroutine;
1481 30 : bool ok = false;
1482 :
1483 30 : fdwroutine = GetFdwRoutineForRelation(childrel, false);
1484 :
1485 30 : if (fdwroutine->AnalyzeForeignTable != NULL)
1486 30 : ok = fdwroutine->AnalyzeForeignTable(childrel,
1487 : &acquirefunc,
1488 : &relpages);
1489 :
1490 30 : if (!ok)
1491 : {
1492 : /* ignore, but release the lock on it */
1493 : Assert(childrel != onerel);
1494 0 : table_close(childrel, AccessShareLock);
1495 0 : continue;
1496 : }
1497 : }
1498 : else
1499 : {
1500 : /*
1501 : * ignore, but release the lock on it. don't try to unlock the
1502 : * passed-in relation
1503 : */
1504 : Assert(childrel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE);
1505 758 : if (childrel != onerel)
1506 66 : table_close(childrel, AccessShareLock);
1507 : else
1508 692 : table_close(childrel, NoLock);
1509 758 : continue;
1510 : }
1511 :
1512 : /* OK, we'll process this child */
1513 2100 : has_child = true;
1514 2100 : rels[nrels] = childrel;
1515 2100 : acquirefuncs[nrels] = acquirefunc;
1516 2100 : relblocks[nrels] = (double) relpages;
1517 2100 : totalblocks += (double) relpages;
1518 2100 : nrels++;
1519 : }
1520 :
1521 : /*
1522 : * If we don't have at least one child table to consider, fail. If the
1523 : * relation is a partitioned table, it's not counted as a child table.
1524 : */
1525 798 : if (!has_child)
1526 : {
1527 0 : ereport(elevel,
1528 : (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no analyzable child tables",
1529 : get_namespace_name(RelationGetNamespace(onerel)),
1530 : RelationGetRelationName(onerel))));
1531 0 : return 0;
1532 : }
1533 :
1534 : /*
1535 : * Now sample rows from each relation, proportionally to its fraction of
1536 : * the total block count. (This might be less than desirable if the child
1537 : * rels have radically different free-space percentages, but it's not
1538 : * clear that it's worth working harder.)
1539 : */
1540 798 : pgstat_progress_update_param(PROGRESS_ANALYZE_CHILD_TABLES_TOTAL,
1541 : nrels);
1542 798 : numrows = 0;
1543 2898 : for (i = 0; i < nrels; i++)
1544 : {
1545 2100 : Relation childrel = rels[i];
1546 2100 : AcquireSampleRowsFunc acquirefunc = acquirefuncs[i];
1547 2100 : double childblocks = relblocks[i];
1548 :
1549 : /*
1550 : * Report progress. The sampling function will normally report blocks
1551 : * done/total, but we need to reset them to 0 here, so that they don't
1552 : * show an old value until that.
1553 : */
1554 : {
1555 2100 : const int progress_index[] = {
1556 : PROGRESS_ANALYZE_CURRENT_CHILD_TABLE_RELID,
1557 : PROGRESS_ANALYZE_BLOCKS_DONE,
1558 : PROGRESS_ANALYZE_BLOCKS_TOTAL
1559 : };
1560 2100 : const int64 progress_vals[] = {
1561 2100 : RelationGetRelid(childrel),
1562 : 0,
1563 : 0,
1564 : };
1565 :
1566 2100 : pgstat_progress_update_multi_param(3, progress_index, progress_vals);
1567 : }
1568 :
1569 2100 : if (childblocks > 0)
1570 : {
1571 : int childtargrows;
1572 :
1573 1946 : childtargrows = (int) rint(targrows * childblocks / totalblocks);
1574 : /* Make sure we don't overrun due to roundoff error */
1575 1946 : childtargrows = Min(childtargrows, targrows - numrows);
1576 1946 : if (childtargrows > 0)
1577 : {
1578 : int childrows;
1579 : double trows,
1580 : tdrows;
1581 :
1582 : /* Fetch a random sample of the child's rows */
1583 1946 : childrows = (*acquirefunc) (childrel, elevel,
1584 1946 : rows + numrows, childtargrows,
1585 : &trows, &tdrows);
1586 :
1587 : /* We may need to convert from child's rowtype to parent's */
1588 1946 : if (childrows > 0 &&
1589 1946 : !equalRowTypes(RelationGetDescr(childrel),
1590 : RelationGetDescr(onerel)))
1591 : {
1592 : TupleConversionMap *map;
1593 :
1594 1858 : map = convert_tuples_by_name(RelationGetDescr(childrel),
1595 : RelationGetDescr(onerel));
1596 1858 : if (map != NULL)
1597 : {
1598 : int j;
1599 :
1600 106604 : for (j = 0; j < childrows; j++)
1601 : {
1602 : HeapTuple newtup;
1603 :
1604 106472 : newtup = execute_attr_map_tuple(rows[numrows + j], map);
1605 106472 : heap_freetuple(rows[numrows + j]);
1606 106472 : rows[numrows + j] = newtup;
1607 : }
1608 132 : free_conversion_map(map);
1609 : }
1610 : }
1611 :
1612 : /* And add to counts */
1613 1946 : numrows += childrows;
1614 1946 : *totalrows += trows;
1615 1946 : *totaldeadrows += tdrows;
1616 : }
1617 : }
1618 :
1619 : /*
1620 : * Note: we cannot release the child-table locks, since we may have
1621 : * pointers to their TOAST tables in the sampled rows.
1622 : */
1623 2100 : table_close(childrel, NoLock);
1624 2100 : pgstat_progress_update_param(PROGRESS_ANALYZE_CHILD_TABLES_DONE,
1625 2100 : i + 1);
1626 : }
1627 :
1628 798 : return numrows;
1629 : }
1630 :
1631 :
1632 : /*
1633 : * update_attstats() -- update attribute statistics for one relation
1634 : *
1635 : * Statistics are stored in several places: the pg_class row for the
1636 : * relation has stats about the whole relation, and there is a
1637 : * pg_statistic row for each (non-system) attribute that has ever
1638 : * been analyzed. The pg_class values are updated by VACUUM, not here.
1639 : *
1640 : * pg_statistic rows are just added or updated normally. This means
1641 : * that pg_statistic will probably contain some deleted rows at the
1642 : * completion of a vacuum cycle, unless it happens to get vacuumed last.
1643 : *
1644 : * To keep things simple, we punt for pg_statistic, and don't try
1645 : * to compute or store rows for pg_statistic itself in pg_statistic.
1646 : * This could possibly be made to work, but it's not worth the trouble.
1647 : * Note analyze_rel() has seen to it that we won't come here when
1648 : * vacuuming pg_statistic itself.
1649 : *
1650 : * Note: there would be a race condition here if two backends could
1651 : * ANALYZE the same table concurrently. Presently, we lock that out
1652 : * by taking a self-exclusive lock on the relation in analyze_rel().
1653 : */
1654 : static void
1655 23142 : update_attstats(Oid relid, bool inh, int natts, VacAttrStats **vacattrstats)
1656 : {
1657 : Relation sd;
1658 : int attno;
1659 23142 : CatalogIndexState indstate = NULL;
1660 :
1661 23142 : if (natts <= 0)
1662 12812 : return; /* nothing to do */
1663 :
1664 10330 : sd = table_open(StatisticRelationId, RowExclusiveLock);
1665 :
1666 90606 : for (attno = 0; attno < natts; attno++)
1667 : {
1668 80276 : VacAttrStats *stats = vacattrstats[attno];
1669 : HeapTuple stup,
1670 : oldtup;
1671 : int i,
1672 : k,
1673 : n;
1674 : Datum values[Natts_pg_statistic];
1675 : bool nulls[Natts_pg_statistic];
1676 : bool replaces[Natts_pg_statistic];
1677 :
1678 : /* Ignore attr if we weren't able to collect stats */
1679 80276 : if (!stats->stats_valid)
1680 6 : continue;
1681 :
1682 : /*
1683 : * Construct a new pg_statistic tuple
1684 : */
1685 2568640 : for (i = 0; i < Natts_pg_statistic; ++i)
1686 : {
1687 2488370 : nulls[i] = false;
1688 2488370 : replaces[i] = true;
1689 : }
1690 :
1691 80270 : values[Anum_pg_statistic_starelid - 1] = ObjectIdGetDatum(relid);
1692 80270 : values[Anum_pg_statistic_staattnum - 1] = Int16GetDatum(stats->tupattnum);
1693 80270 : values[Anum_pg_statistic_stainherit - 1] = BoolGetDatum(inh);
1694 80270 : values[Anum_pg_statistic_stanullfrac - 1] = Float4GetDatum(stats->stanullfrac);
1695 80270 : values[Anum_pg_statistic_stawidth - 1] = Int32GetDatum(stats->stawidth);
1696 80270 : values[Anum_pg_statistic_stadistinct - 1] = Float4GetDatum(stats->stadistinct);
1697 80270 : i = Anum_pg_statistic_stakind1 - 1;
1698 481620 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1699 : {
1700 401350 : values[i++] = Int16GetDatum(stats->stakind[k]); /* stakindN */
1701 : }
1702 80270 : i = Anum_pg_statistic_staop1 - 1;
1703 481620 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1704 : {
1705 401350 : values[i++] = ObjectIdGetDatum(stats->staop[k]); /* staopN */
1706 : }
1707 80270 : i = Anum_pg_statistic_stacoll1 - 1;
1708 481620 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1709 : {
1710 401350 : values[i++] = ObjectIdGetDatum(stats->stacoll[k]); /* stacollN */
1711 : }
1712 80270 : i = Anum_pg_statistic_stanumbers1 - 1;
1713 481620 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1714 : {
1715 401350 : int nnum = stats->numnumbers[k];
1716 :
1717 401350 : if (nnum > 0)
1718 : {
1719 125302 : Datum *numdatums = (Datum *) palloc(nnum * sizeof(Datum));
1720 : ArrayType *arry;
1721 :
1722 1028382 : for (n = 0; n < nnum; n++)
1723 903080 : numdatums[n] = Float4GetDatum(stats->stanumbers[k][n]);
1724 125302 : arry = construct_array_builtin(numdatums, nnum, FLOAT4OID);
1725 125302 : values[i++] = PointerGetDatum(arry); /* stanumbersN */
1726 : }
1727 : else
1728 : {
1729 276048 : nulls[i] = true;
1730 276048 : values[i++] = (Datum) 0;
1731 : }
1732 : }
1733 80270 : i = Anum_pg_statistic_stavalues1 - 1;
1734 481620 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1735 : {
1736 401350 : if (stats->numvalues[k] > 0)
1737 : {
1738 : ArrayType *arry;
1739 :
1740 88294 : arry = construct_array(stats->stavalues[k],
1741 : stats->numvalues[k],
1742 : stats->statypid[k],
1743 88294 : stats->statyplen[k],
1744 88294 : stats->statypbyval[k],
1745 88294 : stats->statypalign[k]);
1746 88294 : values[i++] = PointerGetDatum(arry); /* stavaluesN */
1747 : }
1748 : else
1749 : {
1750 313056 : nulls[i] = true;
1751 313056 : values[i++] = (Datum) 0;
1752 : }
1753 : }
1754 :
1755 : /* Is there already a pg_statistic tuple for this attribute? */
1756 160540 : oldtup = SearchSysCache3(STATRELATTINH,
1757 : ObjectIdGetDatum(relid),
1758 80270 : Int16GetDatum(stats->tupattnum),
1759 : BoolGetDatum(inh));
1760 :
1761 : /* Open index information when we know we need it */
1762 80270 : if (indstate == NULL)
1763 10324 : indstate = CatalogOpenIndexes(sd);
1764 :
1765 80270 : if (HeapTupleIsValid(oldtup))
1766 : {
1767 : /* Yes, replace it */
1768 37260 : stup = heap_modify_tuple(oldtup,
1769 : RelationGetDescr(sd),
1770 : values,
1771 : nulls,
1772 : replaces);
1773 37260 : ReleaseSysCache(oldtup);
1774 37260 : CatalogTupleUpdateWithInfo(sd, &stup->t_self, stup, indstate);
1775 : }
1776 : else
1777 : {
1778 : /* No, insert new tuple */
1779 43010 : stup = heap_form_tuple(RelationGetDescr(sd), values, nulls);
1780 43010 : CatalogTupleInsertWithInfo(sd, stup, indstate);
1781 : }
1782 :
1783 80270 : heap_freetuple(stup);
1784 : }
1785 :
1786 10330 : if (indstate != NULL)
1787 10324 : CatalogCloseIndexes(indstate);
1788 10330 : table_close(sd, RowExclusiveLock);
1789 : }
1790 :
1791 : /*
1792 : * Standard fetch function for use by compute_stats subroutines.
1793 : *
1794 : * This exists to provide some insulation between compute_stats routines
1795 : * and the actual storage of the sample data.
1796 : */
1797 : static Datum
1798 75948356 : std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1799 : {
1800 75948356 : int attnum = stats->tupattnum;
1801 75948356 : HeapTuple tuple = stats->rows[rownum];
1802 75948356 : TupleDesc tupDesc = stats->tupDesc;
1803 :
1804 75948356 : return heap_getattr(tuple, attnum, tupDesc, isNull);
1805 : }
1806 :
1807 : /*
1808 : * Fetch function for analyzing index expressions.
1809 : *
1810 : * We have not bothered to construct index tuples, instead the data is
1811 : * just in Datum arrays.
1812 : */
1813 : static Datum
1814 115024 : ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1815 : {
1816 : int i;
1817 :
1818 : /* exprvals and exprnulls are already offset for proper column */
1819 115024 : i = rownum * stats->rowstride;
1820 115024 : *isNull = stats->exprnulls[i];
1821 115024 : return stats->exprvals[i];
1822 : }
1823 :
1824 :
1825 : /*==========================================================================
1826 : *
1827 : * Code below this point represents the "standard" type-specific statistics
1828 : * analysis algorithms. This code can be replaced on a per-data-type basis
1829 : * by setting a nonzero value in pg_type.typanalyze.
1830 : *
1831 : *==========================================================================
1832 : */
1833 :
1834 :
1835 : /*
1836 : * To avoid consuming too much memory during analysis and/or too much space
1837 : * in the resulting pg_statistic rows, we ignore varlena datums that are wider
1838 : * than WIDTH_THRESHOLD (after detoasting!). This is legitimate for MCV
1839 : * and distinct-value calculations since a wide value is unlikely to be
1840 : * duplicated at all, much less be a most-common value. For the same reason,
1841 : * ignoring wide values will not affect our estimates of histogram bin
1842 : * boundaries very much.
1843 : */
1844 : #define WIDTH_THRESHOLD 1024
1845 :
1846 : #define swapInt(a,b) do {int _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1847 : #define swapDatum(a,b) do {Datum _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1848 :
1849 : /*
1850 : * Extra information used by the default analysis routines
1851 : */
1852 : typedef struct
1853 : {
1854 : int count; /* # of duplicates */
1855 : int first; /* values[] index of first occurrence */
1856 : } ScalarMCVItem;
1857 :
1858 : typedef struct
1859 : {
1860 : SortSupport ssup;
1861 : int *tupnoLink;
1862 : } CompareScalarsContext;
1863 :
1864 :
1865 : static void compute_trivial_stats(VacAttrStatsP stats,
1866 : AnalyzeAttrFetchFunc fetchfunc,
1867 : int samplerows,
1868 : double totalrows);
1869 : static void compute_distinct_stats(VacAttrStatsP stats,
1870 : AnalyzeAttrFetchFunc fetchfunc,
1871 : int samplerows,
1872 : double totalrows);
1873 : static void compute_scalar_stats(VacAttrStatsP stats,
1874 : AnalyzeAttrFetchFunc fetchfunc,
1875 : int samplerows,
1876 : double totalrows);
1877 : static int compare_scalars(const void *a, const void *b, void *arg);
1878 : static int compare_mcvs(const void *a, const void *b, void *arg);
1879 : static int analyze_mcv_list(int *mcv_counts,
1880 : int num_mcv,
1881 : double stadistinct,
1882 : double stanullfrac,
1883 : int samplerows,
1884 : double totalrows);
1885 :
1886 :
1887 : /*
1888 : * std_typanalyze -- the default type-specific typanalyze function
1889 : */
1890 : bool
1891 110180 : std_typanalyze(VacAttrStats *stats)
1892 : {
1893 : Oid ltopr;
1894 : Oid eqopr;
1895 : StdAnalyzeData *mystats;
1896 :
1897 : /* If the attstattarget column is negative, use the default value */
1898 110180 : if (stats->attstattarget < 0)
1899 109574 : stats->attstattarget = default_statistics_target;
1900 :
1901 : /* Look for default "<" and "=" operators for column's type */
1902 110180 : get_sort_group_operators(stats->attrtypid,
1903 : false, false, false,
1904 : <opr, &eqopr, NULL,
1905 : NULL);
1906 :
1907 : /* Save the operator info for compute_stats routines */
1908 110180 : mystats = (StdAnalyzeData *) palloc(sizeof(StdAnalyzeData));
1909 110180 : mystats->eqopr = eqopr;
1910 110180 : mystats->eqfunc = OidIsValid(eqopr) ? get_opcode(eqopr) : InvalidOid;
1911 110180 : mystats->ltopr = ltopr;
1912 110180 : stats->extra_data = mystats;
1913 :
1914 : /*
1915 : * Determine which standard statistics algorithm to use
1916 : */
1917 110180 : if (OidIsValid(eqopr) && OidIsValid(ltopr))
1918 : {
1919 : /* Seems to be a scalar datatype */
1920 106764 : stats->compute_stats = compute_scalar_stats;
1921 : /*--------------------
1922 : * The following choice of minrows is based on the paper
1923 : * "Random sampling for histogram construction: how much is enough?"
1924 : * by Surajit Chaudhuri, Rajeev Motwani and Vivek Narasayya, in
1925 : * Proceedings of ACM SIGMOD International Conference on Management
1926 : * of Data, 1998, Pages 436-447. Their Corollary 1 to Theorem 5
1927 : * says that for table size n, histogram size k, maximum relative
1928 : * error in bin size f, and error probability gamma, the minimum
1929 : * random sample size is
1930 : * r = 4 * k * ln(2*n/gamma) / f^2
1931 : * Taking f = 0.5, gamma = 0.01, n = 10^6 rows, we obtain
1932 : * r = 305.82 * k
1933 : * Note that because of the log function, the dependence on n is
1934 : * quite weak; even at n = 10^12, a 300*k sample gives <= 0.66
1935 : * bin size error with probability 0.99. So there's no real need to
1936 : * scale for n, which is a good thing because we don't necessarily
1937 : * know it at this point.
1938 : *--------------------
1939 : */
1940 106764 : stats->minrows = 300 * stats->attstattarget;
1941 : }
1942 3416 : else if (OidIsValid(eqopr))
1943 : {
1944 : /* We can still recognize distinct values */
1945 2934 : stats->compute_stats = compute_distinct_stats;
1946 : /* Might as well use the same minrows as above */
1947 2934 : stats->minrows = 300 * stats->attstattarget;
1948 : }
1949 : else
1950 : {
1951 : /* Can't do much but the trivial stuff */
1952 482 : stats->compute_stats = compute_trivial_stats;
1953 : /* Might as well use the same minrows as above */
1954 482 : stats->minrows = 300 * stats->attstattarget;
1955 : }
1956 :
1957 110180 : return true;
1958 : }
1959 :
1960 :
1961 : /*
1962 : * compute_trivial_stats() -- compute very basic column statistics
1963 : *
1964 : * We use this when we cannot find a hash "=" operator for the datatype.
1965 : *
1966 : * We determine the fraction of non-null rows and the average datum width.
1967 : */
1968 : static void
1969 326 : compute_trivial_stats(VacAttrStatsP stats,
1970 : AnalyzeAttrFetchFunc fetchfunc,
1971 : int samplerows,
1972 : double totalrows)
1973 : {
1974 : int i;
1975 326 : int null_cnt = 0;
1976 326 : int nonnull_cnt = 0;
1977 326 : double total_width = 0;
1978 652 : bool is_varlena = (!stats->attrtype->typbyval &&
1979 326 : stats->attrtype->typlen == -1);
1980 652 : bool is_varwidth = (!stats->attrtype->typbyval &&
1981 326 : stats->attrtype->typlen < 0);
1982 :
1983 970194 : for (i = 0; i < samplerows; i++)
1984 : {
1985 : Datum value;
1986 : bool isnull;
1987 :
1988 969868 : vacuum_delay_point(true);
1989 :
1990 969868 : value = fetchfunc(stats, i, &isnull);
1991 :
1992 : /* Check for null/nonnull */
1993 969868 : if (isnull)
1994 : {
1995 555992 : null_cnt++;
1996 555992 : continue;
1997 : }
1998 413876 : nonnull_cnt++;
1999 :
2000 : /*
2001 : * If it's a variable-width field, add up widths for average width
2002 : * calculation. Note that if the value is toasted, we use the toasted
2003 : * width. We don't bother with this calculation if it's a fixed-width
2004 : * type.
2005 : */
2006 413876 : if (is_varlena)
2007 : {
2008 79884 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2009 : }
2010 333992 : else if (is_varwidth)
2011 : {
2012 : /* must be cstring */
2013 0 : total_width += strlen(DatumGetCString(value)) + 1;
2014 : }
2015 : }
2016 :
2017 : /* We can only compute average width if we found some non-null values. */
2018 326 : if (nonnull_cnt > 0)
2019 : {
2020 150 : stats->stats_valid = true;
2021 : /* Do the simple null-frac and width stats */
2022 150 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2023 150 : if (is_varwidth)
2024 68 : stats->stawidth = total_width / (double) nonnull_cnt;
2025 : else
2026 82 : stats->stawidth = stats->attrtype->typlen;
2027 150 : stats->stadistinct = 0.0; /* "unknown" */
2028 : }
2029 176 : else if (null_cnt > 0)
2030 : {
2031 : /* We found only nulls; assume the column is entirely null */
2032 176 : stats->stats_valid = true;
2033 176 : stats->stanullfrac = 1.0;
2034 176 : if (is_varwidth)
2035 176 : stats->stawidth = 0; /* "unknown" */
2036 : else
2037 0 : stats->stawidth = stats->attrtype->typlen;
2038 176 : stats->stadistinct = 0.0; /* "unknown" */
2039 : }
2040 326 : }
2041 :
2042 :
2043 : /*
2044 : * compute_distinct_stats() -- compute column statistics including ndistinct
2045 : *
2046 : * We use this when we can find only an "=" operator for the datatype.
2047 : *
2048 : * We determine the fraction of non-null rows, the average width, the
2049 : * most common values, and the (estimated) number of distinct values.
2050 : *
2051 : * The most common values are determined by brute force: we keep a list
2052 : * of previously seen values, ordered by number of times seen, as we scan
2053 : * the samples. A newly seen value is inserted just after the last
2054 : * multiply-seen value, causing the bottommost (oldest) singly-seen value
2055 : * to drop off the list. The accuracy of this method, and also its cost,
2056 : * depend mainly on the length of the list we are willing to keep.
2057 : */
2058 : static void
2059 2154 : compute_distinct_stats(VacAttrStatsP stats,
2060 : AnalyzeAttrFetchFunc fetchfunc,
2061 : int samplerows,
2062 : double totalrows)
2063 : {
2064 : int i;
2065 2154 : int null_cnt = 0;
2066 2154 : int nonnull_cnt = 0;
2067 2154 : int toowide_cnt = 0;
2068 2154 : double total_width = 0;
2069 3640 : bool is_varlena = (!stats->attrtype->typbyval &&
2070 1486 : stats->attrtype->typlen == -1);
2071 3640 : bool is_varwidth = (!stats->attrtype->typbyval &&
2072 1486 : stats->attrtype->typlen < 0);
2073 : FmgrInfo f_cmpeq;
2074 : typedef struct
2075 : {
2076 : Datum value;
2077 : int count;
2078 : } TrackItem;
2079 : TrackItem *track;
2080 : int track_cnt,
2081 : track_max;
2082 2154 : int num_mcv = stats->attstattarget;
2083 2154 : StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2084 :
2085 : /*
2086 : * We track up to 2*n values for an n-element MCV list; but at least 10
2087 : */
2088 2154 : track_max = 2 * num_mcv;
2089 2154 : if (track_max < 10)
2090 78 : track_max = 10;
2091 2154 : track = (TrackItem *) palloc(track_max * sizeof(TrackItem));
2092 2154 : track_cnt = 0;
2093 :
2094 2154 : fmgr_info(mystats->eqfunc, &f_cmpeq);
2095 :
2096 1445060 : for (i = 0; i < samplerows; i++)
2097 : {
2098 : Datum value;
2099 : bool isnull;
2100 : bool match;
2101 : int firstcount1,
2102 : j;
2103 :
2104 1442906 : vacuum_delay_point(true);
2105 :
2106 1442906 : value = fetchfunc(stats, i, &isnull);
2107 :
2108 : /* Check for null/nonnull */
2109 1442906 : if (isnull)
2110 : {
2111 1203964 : null_cnt++;
2112 1203964 : continue;
2113 : }
2114 238942 : nonnull_cnt++;
2115 :
2116 : /*
2117 : * If it's a variable-width field, add up widths for average width
2118 : * calculation. Note that if the value is toasted, we use the toasted
2119 : * width. We don't bother with this calculation if it's a fixed-width
2120 : * type.
2121 : */
2122 238942 : if (is_varlena)
2123 : {
2124 87982 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2125 :
2126 : /*
2127 : * If the value is toasted, we want to detoast it just once to
2128 : * avoid repeated detoastings and resultant excess memory usage
2129 : * during the comparisons. Also, check to see if the value is
2130 : * excessively wide, and if so don't detoast at all --- just
2131 : * ignore the value.
2132 : */
2133 87982 : if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
2134 : {
2135 0 : toowide_cnt++;
2136 0 : continue;
2137 : }
2138 87982 : value = PointerGetDatum(PG_DETOAST_DATUM(value));
2139 : }
2140 150960 : else if (is_varwidth)
2141 : {
2142 : /* must be cstring */
2143 0 : total_width += strlen(DatumGetCString(value)) + 1;
2144 : }
2145 :
2146 : /*
2147 : * See if the value matches anything we're already tracking.
2148 : */
2149 238942 : match = false;
2150 238942 : firstcount1 = track_cnt;
2151 385882 : for (j = 0; j < track_cnt; j++)
2152 : {
2153 380088 : if (DatumGetBool(FunctionCall2Coll(&f_cmpeq,
2154 : stats->attrcollid,
2155 380088 : value, track[j].value)))
2156 : {
2157 233148 : match = true;
2158 233148 : break;
2159 : }
2160 146940 : if (j < firstcount1 && track[j].count == 1)
2161 3382 : firstcount1 = j;
2162 : }
2163 :
2164 238942 : if (match)
2165 : {
2166 : /* Found a match */
2167 233148 : track[j].count++;
2168 : /* This value may now need to "bubble up" in the track list */
2169 237110 : while (j > 0 && track[j].count > track[j - 1].count)
2170 : {
2171 3962 : swapDatum(track[j].value, track[j - 1].value);
2172 3962 : swapInt(track[j].count, track[j - 1].count);
2173 3962 : j--;
2174 : }
2175 : }
2176 : else
2177 : {
2178 : /* No match. Insert at head of count-1 list */
2179 5794 : if (track_cnt < track_max)
2180 5748 : track_cnt++;
2181 49240 : for (j = track_cnt - 1; j > firstcount1; j--)
2182 : {
2183 43446 : track[j].value = track[j - 1].value;
2184 43446 : track[j].count = track[j - 1].count;
2185 : }
2186 5794 : if (firstcount1 < track_cnt)
2187 : {
2188 5794 : track[firstcount1].value = value;
2189 5794 : track[firstcount1].count = 1;
2190 : }
2191 : }
2192 : }
2193 :
2194 : /* We can only compute real stats if we found some non-null values. */
2195 2154 : if (nonnull_cnt > 0)
2196 : {
2197 : int nmultiple,
2198 : summultiple;
2199 :
2200 1582 : stats->stats_valid = true;
2201 : /* Do the simple null-frac and width stats */
2202 1582 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2203 1582 : if (is_varwidth)
2204 914 : stats->stawidth = total_width / (double) nonnull_cnt;
2205 : else
2206 668 : stats->stawidth = stats->attrtype->typlen;
2207 :
2208 : /* Count the number of values we found multiple times */
2209 1582 : summultiple = 0;
2210 5664 : for (nmultiple = 0; nmultiple < track_cnt; nmultiple++)
2211 : {
2212 4884 : if (track[nmultiple].count == 1)
2213 802 : break;
2214 4082 : summultiple += track[nmultiple].count;
2215 : }
2216 :
2217 1582 : if (nmultiple == 0)
2218 : {
2219 : /*
2220 : * If we found no repeated non-null values, assume it's a unique
2221 : * column; but be sure to discount for any nulls we found.
2222 : */
2223 178 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2224 : }
2225 1404 : else if (track_cnt < track_max && toowide_cnt == 0 &&
2226 : nmultiple == track_cnt)
2227 : {
2228 : /*
2229 : * Our track list includes every value in the sample, and every
2230 : * value appeared more than once. Assume the column has just
2231 : * these values. (This case is meant to address columns with
2232 : * small, fixed sets of possible values, such as boolean or enum
2233 : * columns. If there are any values that appear just once in the
2234 : * sample, including too-wide values, we should assume that that's
2235 : * not what we're dealing with.)
2236 : */
2237 780 : stats->stadistinct = track_cnt;
2238 : }
2239 : else
2240 : {
2241 : /*----------
2242 : * Estimate the number of distinct values using the estimator
2243 : * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2244 : * n*d / (n - f1 + f1*n/N)
2245 : * where f1 is the number of distinct values that occurred
2246 : * exactly once in our sample of n rows (from a total of N),
2247 : * and d is the total number of distinct values in the sample.
2248 : * This is their Duj1 estimator; the other estimators they
2249 : * recommend are considerably more complex, and are numerically
2250 : * very unstable when n is much smaller than N.
2251 : *
2252 : * In this calculation, we consider only non-nulls. We used to
2253 : * include rows with null values in the n and N counts, but that
2254 : * leads to inaccurate answers in columns with many nulls, and
2255 : * it's intuitively bogus anyway considering the desired result is
2256 : * the number of distinct non-null values.
2257 : *
2258 : * We assume (not very reliably!) that all the multiply-occurring
2259 : * values are reflected in the final track[] list, and the other
2260 : * nonnull values all appeared but once. (XXX this usually
2261 : * results in a drastic overestimate of ndistinct. Can we do
2262 : * any better?)
2263 : *----------
2264 : */
2265 624 : int f1 = nonnull_cnt - summultiple;
2266 624 : int d = f1 + nmultiple;
2267 624 : double n = samplerows - null_cnt;
2268 624 : double N = totalrows * (1.0 - stats->stanullfrac);
2269 : double stadistinct;
2270 :
2271 : /* N == 0 shouldn't happen, but just in case ... */
2272 624 : if (N > 0)
2273 624 : stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2274 : else
2275 0 : stadistinct = 0;
2276 :
2277 : /* Clamp to sane range in case of roundoff error */
2278 624 : if (stadistinct < d)
2279 170 : stadistinct = d;
2280 624 : if (stadistinct > N)
2281 0 : stadistinct = N;
2282 : /* And round to integer */
2283 624 : stats->stadistinct = floor(stadistinct + 0.5);
2284 : }
2285 :
2286 : /*
2287 : * If we estimated the number of distinct values at more than 10% of
2288 : * the total row count (a very arbitrary limit), then assume that
2289 : * stadistinct should scale with the row count rather than be a fixed
2290 : * value.
2291 : */
2292 1582 : if (stats->stadistinct > 0.1 * totalrows)
2293 368 : stats->stadistinct = -(stats->stadistinct / totalrows);
2294 :
2295 : /*
2296 : * Decide how many values are worth storing as most-common values. If
2297 : * we are able to generate a complete MCV list (all the values in the
2298 : * sample will fit, and we think these are all the ones in the table),
2299 : * then do so. Otherwise, store only those values that are
2300 : * significantly more common than the values not in the list.
2301 : *
2302 : * Note: the first of these cases is meant to address columns with
2303 : * small, fixed sets of possible values, such as boolean or enum
2304 : * columns. If we can *completely* represent the column population by
2305 : * an MCV list that will fit into the stats target, then we should do
2306 : * so and thus provide the planner with complete information. But if
2307 : * the MCV list is not complete, it's generally worth being more
2308 : * selective, and not just filling it all the way up to the stats
2309 : * target.
2310 : */
2311 1582 : if (track_cnt < track_max && toowide_cnt == 0 &&
2312 1574 : stats->stadistinct > 0 &&
2313 : track_cnt <= num_mcv)
2314 : {
2315 : /* Track list includes all values seen, and all will fit */
2316 1012 : num_mcv = track_cnt;
2317 : }
2318 : else
2319 : {
2320 : int *mcv_counts;
2321 :
2322 : /* Incomplete list; decide how many values are worth keeping */
2323 570 : if (num_mcv > track_cnt)
2324 514 : num_mcv = track_cnt;
2325 :
2326 570 : if (num_mcv > 0)
2327 : {
2328 570 : mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2329 1510 : for (i = 0; i < num_mcv; i++)
2330 940 : mcv_counts[i] = track[i].count;
2331 :
2332 570 : num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2333 570 : stats->stadistinct,
2334 570 : stats->stanullfrac,
2335 : samplerows, totalrows);
2336 : }
2337 : }
2338 :
2339 : /* Generate MCV slot entry */
2340 1582 : if (num_mcv > 0)
2341 : {
2342 : MemoryContext old_context;
2343 : Datum *mcv_values;
2344 : float4 *mcv_freqs;
2345 :
2346 : /* Must copy the target values into anl_context */
2347 1576 : old_context = MemoryContextSwitchTo(stats->anl_context);
2348 1576 : mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2349 1576 : mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2350 7014 : for (i = 0; i < num_mcv; i++)
2351 : {
2352 10876 : mcv_values[i] = datumCopy(track[i].value,
2353 5438 : stats->attrtype->typbyval,
2354 5438 : stats->attrtype->typlen);
2355 5438 : mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2356 : }
2357 1576 : MemoryContextSwitchTo(old_context);
2358 :
2359 1576 : stats->stakind[0] = STATISTIC_KIND_MCV;
2360 1576 : stats->staop[0] = mystats->eqopr;
2361 1576 : stats->stacoll[0] = stats->attrcollid;
2362 1576 : stats->stanumbers[0] = mcv_freqs;
2363 1576 : stats->numnumbers[0] = num_mcv;
2364 1576 : stats->stavalues[0] = mcv_values;
2365 1576 : stats->numvalues[0] = num_mcv;
2366 :
2367 : /*
2368 : * Accept the defaults for stats->statypid and others. They have
2369 : * been set before we were called (see vacuum.h)
2370 : */
2371 : }
2372 : }
2373 572 : else if (null_cnt > 0)
2374 : {
2375 : /* We found only nulls; assume the column is entirely null */
2376 572 : stats->stats_valid = true;
2377 572 : stats->stanullfrac = 1.0;
2378 572 : if (is_varwidth)
2379 572 : stats->stawidth = 0; /* "unknown" */
2380 : else
2381 0 : stats->stawidth = stats->attrtype->typlen;
2382 572 : stats->stadistinct = 0.0; /* "unknown" */
2383 : }
2384 :
2385 : /* We don't need to bother cleaning up any of our temporary palloc's */
2386 2154 : }
2387 :
2388 :
2389 : /*
2390 : * compute_scalar_stats() -- compute column statistics
2391 : *
2392 : * We use this when we can find "=" and "<" operators for the datatype.
2393 : *
2394 : * We determine the fraction of non-null rows, the average width, the
2395 : * most common values, the (estimated) number of distinct values, the
2396 : * distribution histogram, and the correlation of physical to logical order.
2397 : *
2398 : * The desired stats can be determined fairly easily after sorting the
2399 : * data values into order.
2400 : */
2401 : static void
2402 78056 : compute_scalar_stats(VacAttrStatsP stats,
2403 : AnalyzeAttrFetchFunc fetchfunc,
2404 : int samplerows,
2405 : double totalrows)
2406 : {
2407 : int i;
2408 78056 : int null_cnt = 0;
2409 78056 : int nonnull_cnt = 0;
2410 78056 : int toowide_cnt = 0;
2411 78056 : double total_width = 0;
2412 97182 : bool is_varlena = (!stats->attrtype->typbyval &&
2413 19126 : stats->attrtype->typlen == -1);
2414 97182 : bool is_varwidth = (!stats->attrtype->typbyval &&
2415 19126 : stats->attrtype->typlen < 0);
2416 : double corr_xysum;
2417 : SortSupportData ssup;
2418 : ScalarItem *values;
2419 78056 : int values_cnt = 0;
2420 : int *tupnoLink;
2421 : ScalarMCVItem *track;
2422 78056 : int track_cnt = 0;
2423 78056 : int num_mcv = stats->attstattarget;
2424 78056 : int num_bins = stats->attstattarget;
2425 78056 : StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2426 :
2427 78056 : values = (ScalarItem *) palloc(samplerows * sizeof(ScalarItem));
2428 78056 : tupnoLink = (int *) palloc(samplerows * sizeof(int));
2429 78056 : track = (ScalarMCVItem *) palloc(num_mcv * sizeof(ScalarMCVItem));
2430 :
2431 78056 : memset(&ssup, 0, sizeof(ssup));
2432 78056 : ssup.ssup_cxt = CurrentMemoryContext;
2433 78056 : ssup.ssup_collation = stats->attrcollid;
2434 78056 : ssup.ssup_nulls_first = false;
2435 :
2436 : /*
2437 : * For now, don't perform abbreviated key conversion, because full values
2438 : * are required for MCV slot generation. Supporting that optimization
2439 : * would necessitate teaching compare_scalars() to call a tie-breaker.
2440 : */
2441 78056 : ssup.abbreviate = false;
2442 :
2443 78056 : PrepareSortSupportFromOrderingOp(mystats->ltopr, &ssup);
2444 :
2445 : /* Initial scan to find sortable values */
2446 69390846 : for (i = 0; i < samplerows; i++)
2447 : {
2448 : Datum value;
2449 : bool isnull;
2450 :
2451 69312790 : vacuum_delay_point(true);
2452 :
2453 69312790 : value = fetchfunc(stats, i, &isnull);
2454 :
2455 : /* Check for null/nonnull */
2456 69312790 : if (isnull)
2457 : {
2458 9092308 : null_cnt++;
2459 9126444 : continue;
2460 : }
2461 60220482 : nonnull_cnt++;
2462 :
2463 : /*
2464 : * If it's a variable-width field, add up widths for average width
2465 : * calculation. Note that if the value is toasted, we use the toasted
2466 : * width. We don't bother with this calculation if it's a fixed-width
2467 : * type.
2468 : */
2469 60220482 : if (is_varlena)
2470 : {
2471 7523856 : total_width += VARSIZE_ANY(DatumGetPointer(value));
2472 :
2473 : /*
2474 : * If the value is toasted, we want to detoast it just once to
2475 : * avoid repeated detoastings and resultant excess memory usage
2476 : * during the comparisons. Also, check to see if the value is
2477 : * excessively wide, and if so don't detoast at all --- just
2478 : * ignore the value.
2479 : */
2480 7523856 : if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
2481 : {
2482 34136 : toowide_cnt++;
2483 34136 : continue;
2484 : }
2485 7489720 : value = PointerGetDatum(PG_DETOAST_DATUM(value));
2486 : }
2487 52696626 : else if (is_varwidth)
2488 : {
2489 : /* must be cstring */
2490 0 : total_width += strlen(DatumGetCString(value)) + 1;
2491 : }
2492 :
2493 : /* Add it to the list to be sorted */
2494 60186346 : values[values_cnt].value = value;
2495 60186346 : values[values_cnt].tupno = values_cnt;
2496 60186346 : tupnoLink[values_cnt] = values_cnt;
2497 60186346 : values_cnt++;
2498 : }
2499 :
2500 : /* We can only compute real stats if we found some sortable values. */
2501 78056 : if (values_cnt > 0)
2502 : {
2503 : int ndistinct, /* # distinct values in sample */
2504 : nmultiple, /* # that appear multiple times */
2505 : num_hist,
2506 : dups_cnt;
2507 72878 : int slot_idx = 0;
2508 : CompareScalarsContext cxt;
2509 :
2510 : /* Sort the collected values */
2511 72878 : cxt.ssup = &ssup;
2512 72878 : cxt.tupnoLink = tupnoLink;
2513 72878 : qsort_interruptible(values, values_cnt, sizeof(ScalarItem),
2514 : compare_scalars, &cxt);
2515 :
2516 : /*
2517 : * Now scan the values in order, find the most common ones, and also
2518 : * accumulate ordering-correlation statistics.
2519 : *
2520 : * To determine which are most common, we first have to count the
2521 : * number of duplicates of each value. The duplicates are adjacent in
2522 : * the sorted list, so a brute-force approach is to compare successive
2523 : * datum values until we find two that are not equal. However, that
2524 : * requires N-1 invocations of the datum comparison routine, which are
2525 : * completely redundant with work that was done during the sort. (The
2526 : * sort algorithm must at some point have compared each pair of items
2527 : * that are adjacent in the sorted order; otherwise it could not know
2528 : * that it's ordered the pair correctly.) We exploit this by having
2529 : * compare_scalars remember the highest tupno index that each
2530 : * ScalarItem has been found equal to. At the end of the sort, a
2531 : * ScalarItem's tupnoLink will still point to itself if and only if it
2532 : * is the last item of its group of duplicates (since the group will
2533 : * be ordered by tupno).
2534 : */
2535 72878 : corr_xysum = 0;
2536 72878 : ndistinct = 0;
2537 72878 : nmultiple = 0;
2538 72878 : dups_cnt = 0;
2539 60259224 : for (i = 0; i < values_cnt; i++)
2540 : {
2541 60186346 : int tupno = values[i].tupno;
2542 :
2543 60186346 : corr_xysum += ((double) i) * ((double) tupno);
2544 60186346 : dups_cnt++;
2545 60186346 : if (tupnoLink[tupno] == tupno)
2546 : {
2547 : /* Reached end of duplicates of this value */
2548 12825890 : ndistinct++;
2549 12825890 : if (dups_cnt > 1)
2550 : {
2551 1085206 : nmultiple++;
2552 1085206 : if (track_cnt < num_mcv ||
2553 449702 : dups_cnt > track[track_cnt - 1].count)
2554 : {
2555 : /*
2556 : * Found a new item for the mcv list; find its
2557 : * position, bubbling down old items if needed. Loop
2558 : * invariant is that j points at an empty/ replaceable
2559 : * slot.
2560 : */
2561 : int j;
2562 :
2563 729194 : if (track_cnt < num_mcv)
2564 635504 : track_cnt++;
2565 9649756 : for (j = track_cnt - 1; j > 0; j--)
2566 : {
2567 9568782 : if (dups_cnt <= track[j - 1].count)
2568 648220 : break;
2569 8920562 : track[j].count = track[j - 1].count;
2570 8920562 : track[j].first = track[j - 1].first;
2571 : }
2572 729194 : track[j].count = dups_cnt;
2573 729194 : track[j].first = i + 1 - dups_cnt;
2574 : }
2575 : }
2576 12825890 : dups_cnt = 0;
2577 : }
2578 : }
2579 :
2580 72878 : stats->stats_valid = true;
2581 : /* Do the simple null-frac and width stats */
2582 72878 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2583 72878 : if (is_varwidth)
2584 10624 : stats->stawidth = total_width / (double) nonnull_cnt;
2585 : else
2586 62254 : stats->stawidth = stats->attrtype->typlen;
2587 :
2588 72878 : if (nmultiple == 0)
2589 : {
2590 : /*
2591 : * If we found no repeated non-null values, assume it's a unique
2592 : * column; but be sure to discount for any nulls we found.
2593 : */
2594 19036 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2595 : }
2596 53842 : else if (toowide_cnt == 0 && nmultiple == ndistinct)
2597 : {
2598 : /*
2599 : * Every value in the sample appeared more than once. Assume the
2600 : * column has just these values. (This case is meant to address
2601 : * columns with small, fixed sets of possible values, such as
2602 : * boolean or enum columns. If there are any values that appear
2603 : * just once in the sample, including too-wide values, we should
2604 : * assume that that's not what we're dealing with.)
2605 : */
2606 33094 : stats->stadistinct = ndistinct;
2607 : }
2608 : else
2609 : {
2610 : /*----------
2611 : * Estimate the number of distinct values using the estimator
2612 : * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2613 : * n*d / (n - f1 + f1*n/N)
2614 : * where f1 is the number of distinct values that occurred
2615 : * exactly once in our sample of n rows (from a total of N),
2616 : * and d is the total number of distinct values in the sample.
2617 : * This is their Duj1 estimator; the other estimators they
2618 : * recommend are considerably more complex, and are numerically
2619 : * very unstable when n is much smaller than N.
2620 : *
2621 : * In this calculation, we consider only non-nulls. We used to
2622 : * include rows with null values in the n and N counts, but that
2623 : * leads to inaccurate answers in columns with many nulls, and
2624 : * it's intuitively bogus anyway considering the desired result is
2625 : * the number of distinct non-null values.
2626 : *
2627 : * Overwidth values are assumed to have been distinct.
2628 : *----------
2629 : */
2630 20748 : int f1 = ndistinct - nmultiple + toowide_cnt;
2631 20748 : int d = f1 + nmultiple;
2632 20748 : double n = samplerows - null_cnt;
2633 20748 : double N = totalrows * (1.0 - stats->stanullfrac);
2634 : double stadistinct;
2635 :
2636 : /* N == 0 shouldn't happen, but just in case ... */
2637 20748 : if (N > 0)
2638 20748 : stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2639 : else
2640 0 : stadistinct = 0;
2641 :
2642 : /* Clamp to sane range in case of roundoff error */
2643 20748 : if (stadistinct < d)
2644 868 : stadistinct = d;
2645 20748 : if (stadistinct > N)
2646 0 : stadistinct = N;
2647 : /* And round to integer */
2648 20748 : stats->stadistinct = floor(stadistinct + 0.5);
2649 : }
2650 :
2651 : /*
2652 : * If we estimated the number of distinct values at more than 10% of
2653 : * the total row count (a very arbitrary limit), then assume that
2654 : * stadistinct should scale with the row count rather than be a fixed
2655 : * value.
2656 : */
2657 72878 : if (stats->stadistinct > 0.1 * totalrows)
2658 15440 : stats->stadistinct = -(stats->stadistinct / totalrows);
2659 :
2660 : /*
2661 : * Decide how many values are worth storing as most-common values. If
2662 : * we are able to generate a complete MCV list (all the values in the
2663 : * sample will fit, and we think these are all the ones in the table),
2664 : * then do so. Otherwise, store only those values that are
2665 : * significantly more common than the values not in the list.
2666 : *
2667 : * Note: the first of these cases is meant to address columns with
2668 : * small, fixed sets of possible values, such as boolean or enum
2669 : * columns. If we can *completely* represent the column population by
2670 : * an MCV list that will fit into the stats target, then we should do
2671 : * so and thus provide the planner with complete information. But if
2672 : * the MCV list is not complete, it's generally worth being more
2673 : * selective, and not just filling it all the way up to the stats
2674 : * target.
2675 : */
2676 72878 : if (track_cnt == ndistinct && toowide_cnt == 0 &&
2677 32412 : stats->stadistinct > 0 &&
2678 : track_cnt <= num_mcv)
2679 : {
2680 : /* Track list includes all values seen, and all will fit */
2681 28992 : num_mcv = track_cnt;
2682 : }
2683 : else
2684 : {
2685 : int *mcv_counts;
2686 :
2687 : /* Incomplete list; decide how many values are worth keeping */
2688 43886 : if (num_mcv > track_cnt)
2689 40080 : num_mcv = track_cnt;
2690 :
2691 43886 : if (num_mcv > 0)
2692 : {
2693 24850 : mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2694 504988 : for (i = 0; i < num_mcv; i++)
2695 480138 : mcv_counts[i] = track[i].count;
2696 :
2697 24850 : num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2698 24850 : stats->stadistinct,
2699 24850 : stats->stanullfrac,
2700 : samplerows, totalrows);
2701 : }
2702 : }
2703 :
2704 : /* Generate MCV slot entry */
2705 72878 : if (num_mcv > 0)
2706 : {
2707 : MemoryContext old_context;
2708 : Datum *mcv_values;
2709 : float4 *mcv_freqs;
2710 :
2711 : /* Must copy the target values into anl_context */
2712 53800 : old_context = MemoryContextSwitchTo(stats->anl_context);
2713 53800 : mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2714 53800 : mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2715 689100 : for (i = 0; i < num_mcv; i++)
2716 : {
2717 1270600 : mcv_values[i] = datumCopy(values[track[i].first].value,
2718 635300 : stats->attrtype->typbyval,
2719 635300 : stats->attrtype->typlen);
2720 635300 : mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2721 : }
2722 53800 : MemoryContextSwitchTo(old_context);
2723 :
2724 53800 : stats->stakind[slot_idx] = STATISTIC_KIND_MCV;
2725 53800 : stats->staop[slot_idx] = mystats->eqopr;
2726 53800 : stats->stacoll[slot_idx] = stats->attrcollid;
2727 53800 : stats->stanumbers[slot_idx] = mcv_freqs;
2728 53800 : stats->numnumbers[slot_idx] = num_mcv;
2729 53800 : stats->stavalues[slot_idx] = mcv_values;
2730 53800 : stats->numvalues[slot_idx] = num_mcv;
2731 :
2732 : /*
2733 : * Accept the defaults for stats->statypid and others. They have
2734 : * been set before we were called (see vacuum.h)
2735 : */
2736 53800 : slot_idx++;
2737 : }
2738 :
2739 : /*
2740 : * Generate a histogram slot entry if there are at least two distinct
2741 : * values not accounted for in the MCV list. (This ensures the
2742 : * histogram won't collapse to empty or a singleton.)
2743 : */
2744 72878 : num_hist = ndistinct - num_mcv;
2745 72878 : if (num_hist > num_bins)
2746 11994 : num_hist = num_bins + 1;
2747 72878 : if (num_hist >= 2)
2748 : {
2749 : MemoryContext old_context;
2750 : Datum *hist_values;
2751 : int nvals;
2752 : int pos,
2753 : posfrac,
2754 : delta,
2755 : deltafrac;
2756 :
2757 : /* Sort the MCV items into position order to speed next loop */
2758 32224 : qsort_interruptible(track, num_mcv, sizeof(ScalarMCVItem),
2759 : compare_mcvs, NULL);
2760 :
2761 : /*
2762 : * Collapse out the MCV items from the values[] array.
2763 : *
2764 : * Note we destroy the values[] array here... but we don't need it
2765 : * for anything more. We do, however, still need values_cnt.
2766 : * nvals will be the number of remaining entries in values[].
2767 : */
2768 32224 : if (num_mcv > 0)
2769 : {
2770 : int src,
2771 : dest;
2772 : int j;
2773 :
2774 17432 : src = dest = 0;
2775 17432 : j = 0; /* index of next interesting MCV item */
2776 623708 : while (src < values_cnt)
2777 : {
2778 : int ncopy;
2779 :
2780 606276 : if (j < num_mcv)
2781 : {
2782 592570 : int first = track[j].first;
2783 :
2784 592570 : if (src >= first)
2785 : {
2786 : /* advance past this MCV item */
2787 432978 : src = first + track[j].count;
2788 432978 : j++;
2789 432978 : continue;
2790 : }
2791 159592 : ncopy = first - src;
2792 : }
2793 : else
2794 13706 : ncopy = values_cnt - src;
2795 173298 : memmove(&values[dest], &values[src],
2796 : ncopy * sizeof(ScalarItem));
2797 173298 : src += ncopy;
2798 173298 : dest += ncopy;
2799 : }
2800 17432 : nvals = dest;
2801 : }
2802 : else
2803 14792 : nvals = values_cnt;
2804 : Assert(nvals >= num_hist);
2805 :
2806 : /* Must copy the target values into anl_context */
2807 32224 : old_context = MemoryContextSwitchTo(stats->anl_context);
2808 32224 : hist_values = (Datum *) palloc(num_hist * sizeof(Datum));
2809 :
2810 : /*
2811 : * The object of this loop is to copy the first and last values[]
2812 : * entries along with evenly-spaced values in between. So the
2813 : * i'th value is values[(i * (nvals - 1)) / (num_hist - 1)]. But
2814 : * computing that subscript directly risks integer overflow when
2815 : * the stats target is more than a couple thousand. Instead we
2816 : * add (nvals - 1) / (num_hist - 1) to pos at each step, tracking
2817 : * the integral and fractional parts of the sum separately.
2818 : */
2819 32224 : delta = (nvals - 1) / (num_hist - 1);
2820 32224 : deltafrac = (nvals - 1) % (num_hist - 1);
2821 32224 : pos = posfrac = 0;
2822 :
2823 1691464 : for (i = 0; i < num_hist; i++)
2824 : {
2825 3318480 : hist_values[i] = datumCopy(values[pos].value,
2826 1659240 : stats->attrtype->typbyval,
2827 1659240 : stats->attrtype->typlen);
2828 1659240 : pos += delta;
2829 1659240 : posfrac += deltafrac;
2830 1659240 : if (posfrac >= (num_hist - 1))
2831 : {
2832 : /* fractional part exceeds 1, carry to integer part */
2833 564112 : pos++;
2834 564112 : posfrac -= (num_hist - 1);
2835 : }
2836 : }
2837 :
2838 32224 : MemoryContextSwitchTo(old_context);
2839 :
2840 32224 : stats->stakind[slot_idx] = STATISTIC_KIND_HISTOGRAM;
2841 32224 : stats->staop[slot_idx] = mystats->ltopr;
2842 32224 : stats->stacoll[slot_idx] = stats->attrcollid;
2843 32224 : stats->stavalues[slot_idx] = hist_values;
2844 32224 : stats->numvalues[slot_idx] = num_hist;
2845 :
2846 : /*
2847 : * Accept the defaults for stats->statypid and others. They have
2848 : * been set before we were called (see vacuum.h)
2849 : */
2850 32224 : slot_idx++;
2851 : }
2852 :
2853 : /* Generate a correlation entry if there are multiple values */
2854 72878 : if (values_cnt > 1)
2855 : {
2856 : MemoryContext old_context;
2857 : float4 *corrs;
2858 : double corr_xsum,
2859 : corr_x2sum;
2860 :
2861 : /* Must copy the target values into anl_context */
2862 68592 : old_context = MemoryContextSwitchTo(stats->anl_context);
2863 68592 : corrs = (float4 *) palloc(sizeof(float4));
2864 68592 : MemoryContextSwitchTo(old_context);
2865 :
2866 : /*----------
2867 : * Since we know the x and y value sets are both
2868 : * 0, 1, ..., values_cnt-1
2869 : * we have sum(x) = sum(y) =
2870 : * (values_cnt-1)*values_cnt / 2
2871 : * and sum(x^2) = sum(y^2) =
2872 : * (values_cnt-1)*values_cnt*(2*values_cnt-1) / 6.
2873 : *----------
2874 : */
2875 68592 : corr_xsum = ((double) (values_cnt - 1)) *
2876 68592 : ((double) values_cnt) / 2.0;
2877 68592 : corr_x2sum = ((double) (values_cnt - 1)) *
2878 68592 : ((double) values_cnt) * (double) (2 * values_cnt - 1) / 6.0;
2879 :
2880 : /* And the correlation coefficient reduces to */
2881 68592 : corrs[0] = (values_cnt * corr_xysum - corr_xsum * corr_xsum) /
2882 68592 : (values_cnt * corr_x2sum - corr_xsum * corr_xsum);
2883 :
2884 68592 : stats->stakind[slot_idx] = STATISTIC_KIND_CORRELATION;
2885 68592 : stats->staop[slot_idx] = mystats->ltopr;
2886 68592 : stats->stacoll[slot_idx] = stats->attrcollid;
2887 68592 : stats->stanumbers[slot_idx] = corrs;
2888 68592 : stats->numnumbers[slot_idx] = 1;
2889 68592 : slot_idx++;
2890 : }
2891 : }
2892 5178 : else if (nonnull_cnt > 0)
2893 : {
2894 : /* We found some non-null values, but they were all too wide */
2895 : Assert(nonnull_cnt == toowide_cnt);
2896 330 : stats->stats_valid = true;
2897 : /* Do the simple null-frac and width stats */
2898 330 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
2899 330 : if (is_varwidth)
2900 330 : stats->stawidth = total_width / (double) nonnull_cnt;
2901 : else
2902 0 : stats->stawidth = stats->attrtype->typlen;
2903 : /* Assume all too-wide values are distinct, so it's a unique column */
2904 330 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2905 : }
2906 4848 : else if (null_cnt > 0)
2907 : {
2908 : /* We found only nulls; assume the column is entirely null */
2909 4848 : stats->stats_valid = true;
2910 4848 : stats->stanullfrac = 1.0;
2911 4848 : if (is_varwidth)
2912 4212 : stats->stawidth = 0; /* "unknown" */
2913 : else
2914 636 : stats->stawidth = stats->attrtype->typlen;
2915 4848 : stats->stadistinct = 0.0; /* "unknown" */
2916 : }
2917 :
2918 : /* We don't need to bother cleaning up any of our temporary palloc's */
2919 78056 : }
2920 :
2921 : /*
2922 : * Comparator for sorting ScalarItems
2923 : *
2924 : * Aside from sorting the items, we update the tupnoLink[] array
2925 : * whenever two ScalarItems are found to contain equal datums. The array
2926 : * is indexed by tupno; for each ScalarItem, it contains the highest
2927 : * tupno that that item's datum has been found to be equal to. This allows
2928 : * us to avoid additional comparisons in compute_scalar_stats().
2929 : */
2930 : static int
2931 554398202 : compare_scalars(const void *a, const void *b, void *arg)
2932 : {
2933 554398202 : Datum da = ((const ScalarItem *) a)->value;
2934 554398202 : int ta = ((const ScalarItem *) a)->tupno;
2935 554398202 : Datum db = ((const ScalarItem *) b)->value;
2936 554398202 : int tb = ((const ScalarItem *) b)->tupno;
2937 554398202 : CompareScalarsContext *cxt = (CompareScalarsContext *) arg;
2938 : int compare;
2939 :
2940 554398202 : compare = ApplySortComparator(da, false, db, false, cxt->ssup);
2941 554398202 : if (compare != 0)
2942 216300366 : return compare;
2943 :
2944 : /*
2945 : * The two datums are equal, so update cxt->tupnoLink[].
2946 : */
2947 338097836 : if (cxt->tupnoLink[ta] < tb)
2948 49579592 : cxt->tupnoLink[ta] = tb;
2949 338097836 : if (cxt->tupnoLink[tb] < ta)
2950 3555208 : cxt->tupnoLink[tb] = ta;
2951 :
2952 : /*
2953 : * For equal datums, sort by tupno
2954 : */
2955 338097836 : return ta - tb;
2956 : }
2957 :
2958 : /*
2959 : * Comparator for sorting ScalarMCVItems by position
2960 : */
2961 : static int
2962 2228276 : compare_mcvs(const void *a, const void *b, void *arg)
2963 : {
2964 2228276 : int da = ((const ScalarMCVItem *) a)->first;
2965 2228276 : int db = ((const ScalarMCVItem *) b)->first;
2966 :
2967 2228276 : return da - db;
2968 : }
2969 :
2970 : /*
2971 : * Analyze the list of common values in the sample and decide how many are
2972 : * worth storing in the table's MCV list.
2973 : *
2974 : * mcv_counts is assumed to be a list of the counts of the most common values
2975 : * seen in the sample, starting with the most common. The return value is the
2976 : * number that are significantly more common than the values not in the list,
2977 : * and which are therefore deemed worth storing in the table's MCV list.
2978 : */
2979 : static int
2980 25420 : analyze_mcv_list(int *mcv_counts,
2981 : int num_mcv,
2982 : double stadistinct,
2983 : double stanullfrac,
2984 : int samplerows,
2985 : double totalrows)
2986 : {
2987 : double ndistinct_table;
2988 : double sumcount;
2989 : int i;
2990 :
2991 : /*
2992 : * If the entire table was sampled, keep the whole list. This also
2993 : * protects us against division by zero in the code below.
2994 : */
2995 25420 : if (samplerows == totalrows || totalrows <= 1.0)
2996 24592 : return num_mcv;
2997 :
2998 : /* Re-extract the estimated number of distinct nonnull values in table */
2999 828 : ndistinct_table = stadistinct;
3000 828 : if (ndistinct_table < 0)
3001 150 : ndistinct_table = -ndistinct_table * totalrows;
3002 :
3003 : /*
3004 : * Exclude the least common values from the MCV list, if they are not
3005 : * significantly more common than the estimated selectivity they would
3006 : * have if they weren't in the list. All non-MCV values are assumed to be
3007 : * equally common, after taking into account the frequencies of all the
3008 : * values in the MCV list and the number of nulls (c.f. eqsel()).
3009 : *
3010 : * Here sumcount tracks the total count of all but the last (least common)
3011 : * value in the MCV list, allowing us to determine the effect of excluding
3012 : * that value from the list.
3013 : *
3014 : * Note that we deliberately do this by removing values from the full
3015 : * list, rather than starting with an empty list and adding values,
3016 : * because the latter approach can fail to add any values if all the most
3017 : * common values have around the same frequency and make up the majority
3018 : * of the table, so that the overall average frequency of all values is
3019 : * roughly the same as that of the common values. This would lead to any
3020 : * uncommon values being significantly overestimated.
3021 : */
3022 828 : sumcount = 0.0;
3023 1738 : for (i = 0; i < num_mcv - 1; i++)
3024 910 : sumcount += mcv_counts[i];
3025 :
3026 990 : while (num_mcv > 0)
3027 : {
3028 : double selec,
3029 : otherdistinct,
3030 : N,
3031 : n,
3032 : K,
3033 : variance,
3034 : stddev;
3035 :
3036 : /*
3037 : * Estimated selectivity the least common value would have if it
3038 : * wasn't in the MCV list (c.f. eqsel()).
3039 : */
3040 990 : selec = 1.0 - sumcount / samplerows - stanullfrac;
3041 990 : if (selec < 0.0)
3042 0 : selec = 0.0;
3043 990 : if (selec > 1.0)
3044 0 : selec = 1.0;
3045 990 : otherdistinct = ndistinct_table - (num_mcv - 1);
3046 990 : if (otherdistinct > 1)
3047 990 : selec /= otherdistinct;
3048 :
3049 : /*
3050 : * If the value is kept in the MCV list, its population frequency is
3051 : * assumed to equal its sample frequency. We use the lower end of a
3052 : * textbook continuity-corrected Wald-type confidence interval to
3053 : * determine if that is significantly more common than the non-MCV
3054 : * frequency --- specifically we assume the population frequency is
3055 : * highly likely to be within around 2 standard errors of the sample
3056 : * frequency, which equates to an interval of 2 standard deviations
3057 : * either side of the sample count, plus an additional 0.5 for the
3058 : * continuity correction. Since we are sampling without replacement,
3059 : * this is a hypergeometric distribution.
3060 : *
3061 : * XXX: Empirically, this approach seems to work quite well, but it
3062 : * may be worth considering more advanced techniques for estimating
3063 : * the confidence interval of the hypergeometric distribution.
3064 : */
3065 990 : N = totalrows;
3066 990 : n = samplerows;
3067 990 : K = N * mcv_counts[num_mcv - 1] / n;
3068 990 : variance = n * K * (N - K) * (N - n) / (N * N * (N - 1));
3069 990 : stddev = sqrt(variance);
3070 :
3071 990 : if (mcv_counts[num_mcv - 1] > selec * samplerows + 2 * stddev + 0.5)
3072 : {
3073 : /*
3074 : * The value is significantly more common than the non-MCV
3075 : * selectivity would suggest. Keep it, and all the other more
3076 : * common values in the list.
3077 : */
3078 780 : break;
3079 : }
3080 : else
3081 : {
3082 : /* Discard this value and consider the next least common value */
3083 210 : num_mcv--;
3084 210 : if (num_mcv == 0)
3085 48 : break;
3086 162 : sumcount -= mcv_counts[num_mcv - 1];
3087 : }
3088 : }
3089 828 : return num_mcv;
3090 : }
|