1
/*-------------------------------------------------------------------------
4
* the Postgres statistics generator
6
* Portions Copyright (c) 1996-2018, PostgreSQL Global Development Group
7
* Portions Copyright (c) 1994, Regents of the University of California
11
* src/backend/commands/analyze.c
13
*-------------------------------------------------------------------------
19
#include "access/multixact.h"
20
#include "access/sysattr.h"
21
#include "access/transam.h"
22
#include "access/tupconvert.h"
23
#include "access/tuptoaster.h"
24
#include "access/visibilitymap.h"
25
#include "access/xact.h"
26
#include "catalog/catalog.h"
27
#include "catalog/index.h"
28
#include "catalog/indexing.h"
29
#include "catalog/pg_collation.h"
30
#include "catalog/pg_inherits.h"
31
#include "catalog/pg_namespace.h"
32
#include "catalog/pg_statistic_ext.h"
33
#include "commands/dbcommands.h"
34
#include "commands/tablecmds.h"
35
#include "commands/vacuum.h"
36
#include "executor/executor.h"
37
#include "foreign/fdwapi.h"
38
#include "miscadmin.h"
39
#include "nodes/nodeFuncs.h"
40
#include "parser/parse_oper.h"
41
#include "parser/parse_relation.h"
43
#include "postmaster/autovacuum.h"
44
#include "statistics/extended_stats_internal.h"
45
#include "statistics/statistics.h"
46
#include "storage/bufmgr.h"
47
#include "storage/lmgr.h"
48
#include "storage/proc.h"
49
#include "storage/procarray.h"
50
#include "utils/acl.h"
51
#include "utils/attoptcache.h"
52
#include "utils/builtins.h"
53
#include "utils/datum.h"
54
#include "utils/fmgroids.h"
55
#include "utils/guc.h"
56
#include "utils/lsyscache.h"
57
#include "utils/memutils.h"
58
#include "utils/pg_rusage.h"
59
#include "utils/sampling.h"
60
#include "utils/sortsupport.h"
61
#include "utils/syscache.h"
62
#include "utils/timestamp.h"
63
#include "utils/tqual.h"
66
/* Per-index data for ANALYZE */
67
typedef struct AnlIndexData
69
IndexInfo *indexInfo; /* BuildIndexInfo result */
70
double tupleFract; /* fraction of rows for partial index */
71
VacAttrStats **vacattrstats; /* index attrs to analyze */
76
/* Default statistics target (GUC parameter) */
77
int default_statistics_target = 100;
79
/* A few variables that don't seem worth passing around as parameters */
80
static MemoryContext anl_context = NULL;
81
static BufferAccessStrategy vac_strategy;
84
static void do_analyze_rel(Relation onerel, int options,
85
VacuumParams *params, List *va_cols,
86
AcquireSampleRowsFunc acquirefunc, BlockNumber relpages,
87
bool inh, bool in_outer_xact, int elevel);
88
static void compute_index_stats(Relation onerel, double totalrows,
89
AnlIndexData *indexdata, int nindexes,
90
HeapTuple *rows, int numrows,
91
MemoryContext col_context);
92
static VacAttrStats *examine_attribute(Relation onerel, int attnum,
94
static int acquire_sample_rows(Relation onerel, int elevel,
95
HeapTuple *rows, int targrows,
96
double *totalrows, double *totaldeadrows);
97
static int compare_rows(const void *a, const void *b);
98
static int acquire_inherited_sample_rows(Relation onerel, int elevel,
99
HeapTuple *rows, int targrows,
100
double *totalrows, double *totaldeadrows);
101
static void update_attstats(Oid relid, bool inh,
102
int natts, VacAttrStats **vacattrstats);
103
static Datum std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
104
static Datum ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
108
* analyze_rel() -- analyze one relation
110
* relid identifies the relation to analyze. If relation is supplied, use
111
* the name therein for reporting any failure to open/lock the rel; do not
112
* use it once we've successfully opened the rel, since it might be stale.
115
analyze_rel(Oid relid, RangeVar *relation, int options,
116
VacuumParams *params, List *va_cols, bool in_outer_xact,
117
BufferAccessStrategy bstrategy)
121
AcquireSampleRowsFunc acquirefunc = NULL;
122
BlockNumber relpages = 0;
123
bool rel_lock = true;
125
/* Select logging level */
126
if (options & VACOPT_VERBOSE)
131
/* Set up static variables */
132
vac_strategy = bstrategy;
135
* Check for user-requested abort.
137
CHECK_FOR_INTERRUPTS();
140
* Open the relation, getting ShareUpdateExclusiveLock to ensure that two
141
* ANALYZEs don't run on it concurrently. (This also locks out a
142
* concurrent VACUUM, which doesn't matter much at the moment but might
143
* matter if we ever try to accumulate stats on dead tuples.) If the rel
144
* has been dropped since we last saw it, we don't need to process it.
146
if (!(options & VACOPT_NOWAIT))
147
onerel = try_relation_open(relid, ShareUpdateExclusiveLock);
148
else if (ConditionalLockRelationOid(relid, ShareUpdateExclusiveLock))
149
onerel = try_relation_open(relid, NoLock);
157
* If we failed to open or lock the relation, emit a log message before
163
* If the RangeVar is not defined, we do not have enough information
164
* to provide a meaningful log statement. Chances are that
165
* analyze_rel's caller has intentionally not provided this
166
* information so that this logging is skipped, anyway.
168
if (relation == NULL)
172
* Determine the log level. For autovacuum logs, we emit a LOG if
173
* log_autovacuum_min_duration is not disabled. For manual ANALYZE,
174
* we emit a WARNING to match the log statements in the permissions
177
if (!IsAutoVacuumWorkerProcess())
179
else if (params->log_min_duration >= 0)
186
(errcode(ERRCODE_LOCK_NOT_AVAILABLE),
187
errmsg("skipping analyze of \"%s\" --- lock not available",
188
relation->relname)));
191
(errcode(ERRCODE_UNDEFINED_TABLE),
192
errmsg("skipping analyze of \"%s\" --- relation no longer exists",
193
relation->relname)));
199
* Check permissions --- this should match vacuum's check!
201
if (!(pg_class_ownercheck(RelationGetRelid(onerel), GetUserId()) ||
202
(pg_database_ownercheck(MyDatabaseId, GetUserId()) && !onerel->rd_rel->relisshared)))
204
/* No need for a WARNING if we already complained during VACUUM */
205
if (!(options & VACOPT_VACUUM))
207
if (onerel->rd_rel->relisshared)
209
(errmsg("skipping \"%s\" --- only superuser can analyze it",
210
RelationGetRelationName(onerel))));
211
else if (onerel->rd_rel->relnamespace == PG_CATALOG_NAMESPACE)
213
(errmsg("skipping \"%s\" --- only superuser or database owner can analyze it",
214
RelationGetRelationName(onerel))));
217
(errmsg("skipping \"%s\" --- only table or database owner can analyze it",
218
RelationGetRelationName(onerel))));
220
relation_close(onerel, ShareUpdateExclusiveLock);
225
* Silently ignore tables that are temp tables of other backends ---
226
* trying to analyze these is rather pointless, since their contents are
227
* probably not up-to-date on disk. (We don't throw a warning here; it
228
* would just lead to chatter during a database-wide ANALYZE.)
230
if (RELATION_IS_OTHER_TEMP(onerel))
232
relation_close(onerel, ShareUpdateExclusiveLock);
237
* We can ANALYZE any table except pg_statistic. See update_attstats
239
if (RelationGetRelid(onerel) == StatisticRelationId)
241
relation_close(onerel, ShareUpdateExclusiveLock);
246
* Check that it's of an analyzable relkind, and set up appropriately.
248
if (onerel->rd_rel->relkind == RELKIND_RELATION ||
249
onerel->rd_rel->relkind == RELKIND_MATVIEW)
251
/* Regular table, so we'll use the regular row acquisition function */
252
acquirefunc = acquire_sample_rows;
253
/* Also get regular table's size */
254
relpages = RelationGetNumberOfBlocks(onerel);
256
else if (onerel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
259
* For a foreign table, call the FDW's hook function to see whether it
262
FdwRoutine *fdwroutine;
265
fdwroutine = GetFdwRoutineForRelation(onerel, false);
267
if (fdwroutine->AnalyzeForeignTable != NULL)
268
ok = fdwroutine->AnalyzeForeignTable(onerel,
275
(errmsg("skipping \"%s\" --- cannot analyze this foreign table",
276
RelationGetRelationName(onerel))));
277
relation_close(onerel, ShareUpdateExclusiveLock);
281
else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
284
* For partitioned tables, we want to do the recursive ANALYZE below.
289
/* No need for a WARNING if we already complained during VACUUM */
290
if (!(options & VACOPT_VACUUM))
292
(errmsg("skipping \"%s\" --- cannot analyze non-tables or special system tables",
293
RelationGetRelationName(onerel))));
294
relation_close(onerel, ShareUpdateExclusiveLock);
299
* OK, let's do it. First let other backends know I'm in ANALYZE.
301
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
302
MyPgXact->vacuumFlags |= PROC_IN_ANALYZE;
303
LWLockRelease(ProcArrayLock);
306
* Do the normal non-recursive ANALYZE. We can skip this for partitioned
307
* tables, which don't contain any rows.
309
if (onerel->rd_rel->relkind != RELKIND_PARTITIONED_TABLE)
310
do_analyze_rel(onerel, options, params, va_cols, acquirefunc,
311
relpages, false, in_outer_xact, elevel);
314
* If there are child tables, do recursive ANALYZE.
316
if (onerel->rd_rel->relhassubclass)
317
do_analyze_rel(onerel, options, params, va_cols, acquirefunc, relpages,
318
true, in_outer_xact, elevel);
321
* Close source relation now, but keep lock so that no one deletes it
322
* before we commit. (If someone did, they'd fail to clean up the entries
323
* we made in pg_statistic. Also, releasing the lock before commit would
324
* expose us to concurrent-update failures in update_attstats.)
326
relation_close(onerel, NoLock);
329
* Reset my PGXACT flag. Note: we need this here, and not in vacuum_rel,
330
* because the vacuum flag is cleared by the end-of-xact code.
332
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
333
MyPgXact->vacuumFlags &= ~PROC_IN_ANALYZE;
334
LWLockRelease(ProcArrayLock);
338
* do_analyze_rel() -- analyze one relation, recursively or not
340
* Note that "acquirefunc" is only relevant for the non-inherited case.
341
* For the inherited case, acquire_inherited_sample_rows() determines the
342
* appropriate acquirefunc for each child table.
345
do_analyze_rel(Relation onerel, int options, VacuumParams *params,
346
List *va_cols, AcquireSampleRowsFunc acquirefunc,
347
BlockNumber relpages, bool inh, bool in_outer_xact,
357
VacAttrStats **vacattrstats;
358
AnlIndexData *indexdata;
365
TimestampTz starttime = 0;
366
MemoryContext caller_context;
368
int save_sec_context;
373
(errmsg("analyzing \"%s.%s\" inheritance tree",
374
get_namespace_name(RelationGetNamespace(onerel)),
375
RelationGetRelationName(onerel))));
378
(errmsg("analyzing \"%s.%s\"",
379
get_namespace_name(RelationGetNamespace(onerel)),
380
RelationGetRelationName(onerel))));
383
* Set up a working context so that we can easily free whatever junk gets
386
anl_context = AllocSetContextCreate(CurrentMemoryContext,
388
ALLOCSET_DEFAULT_SIZES);
389
caller_context = MemoryContextSwitchTo(anl_context);
392
* Switch to the table owner's userid, so that any index functions are run
393
* as that user. Also lock down security-restricted operations and
394
* arrange to make GUC variable changes local to this command.
396
GetUserIdAndSecContext(&save_userid, &save_sec_context);
397
SetUserIdAndSecContext(onerel->rd_rel->relowner,
398
save_sec_context | SECURITY_RESTRICTED_OPERATION);
399
save_nestlevel = NewGUCNestLevel();
401
/* measure elapsed time iff autovacuum logging requires it */
402
if (IsAutoVacuumWorkerProcess() && params->log_min_duration >= 0)
404
pg_rusage_init(&ru0);
405
if (params->log_min_duration > 0)
406
starttime = GetCurrentTimestamp();
410
* Determine which columns to analyze
412
* Note that system attributes are never analyzed, so we just reject them
413
* at the lookup stage. We also reject duplicate column mentions. (We
414
* could alternatively ignore duplicates, but analyzing a column twice
415
* won't work; we'd end up making a conflicting update in pg_statistic.)
419
Bitmapset *unique_cols = NULL;
422
vacattrstats = (VacAttrStats **) palloc(list_length(va_cols) *
423
sizeof(VacAttrStats *));
427
char *col = strVal(lfirst(le));
429
i = attnameAttNum(onerel, col, false);
430
if (i == InvalidAttrNumber)
432
(errcode(ERRCODE_UNDEFINED_COLUMN),
433
errmsg("column \"%s\" of relation \"%s\" does not exist",
434
col, RelationGetRelationName(onerel))));
435
if (bms_is_member(i, unique_cols))
437
(errcode(ERRCODE_DUPLICATE_COLUMN),
438
errmsg("column \"%s\" of relation \"%s\" appears more than once",
439
col, RelationGetRelationName(onerel))));
440
unique_cols = bms_add_member(unique_cols, i);
442
vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
443
if (vacattrstats[tcnt] != NULL)
450
attr_cnt = onerel->rd_att->natts;
451
vacattrstats = (VacAttrStats **)
452
palloc(attr_cnt * sizeof(VacAttrStats *));
454
for (i = 1; i <= attr_cnt; i++)
456
vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
457
if (vacattrstats[tcnt] != NULL)
464
* Open all indexes of the relation, and see if there are any analyzable
465
* columns in the indexes. We do not analyze index columns if there was
466
* an explicit column list in the ANALYZE command, however. If we are
467
* doing a recursive scan, we don't want to touch the parent's indexes at
471
vac_open_indexes(onerel, AccessShareLock, &nindexes, &Irel);
477
hasindex = (nindexes > 0);
481
indexdata = (AnlIndexData *) palloc0(nindexes * sizeof(AnlIndexData));
482
for (ind = 0; ind < nindexes; ind++)
484
AnlIndexData *thisdata = &indexdata[ind];
485
IndexInfo *indexInfo;
487
thisdata->indexInfo = indexInfo = BuildIndexInfo(Irel[ind]);
488
thisdata->tupleFract = 1.0; /* fix later if partial */
489
if (indexInfo->ii_Expressions != NIL && va_cols == NIL)
491
ListCell *indexpr_item = list_head(indexInfo->ii_Expressions);
493
thisdata->vacattrstats = (VacAttrStats **)
494
palloc(indexInfo->ii_NumIndexAttrs * sizeof(VacAttrStats *));
496
for (i = 0; i < indexInfo->ii_NumIndexAttrs; i++)
498
int keycol = indexInfo->ii_IndexAttrNumbers[i];
502
/* Found an index expression */
505
if (indexpr_item == NULL) /* shouldn't happen */
506
elog(ERROR, "too few entries in indexprs list");
507
indexkey = (Node *) lfirst(indexpr_item);
508
indexpr_item = lnext(indexpr_item);
509
thisdata->vacattrstats[tcnt] =
510
examine_attribute(Irel[ind], i + 1, indexkey);
511
if (thisdata->vacattrstats[tcnt] != NULL)
515
thisdata->attr_cnt = tcnt;
521
* Determine how many rows we need to sample, using the worst case from
522
* all analyzable columns. We use a lower bound of 100 rows to avoid
523
* possible overflow in Vitter's algorithm. (Note: that will also be the
524
* target in the corner case where there are no analyzable columns.)
527
for (i = 0; i < attr_cnt; i++)
529
if (targrows < vacattrstats[i]->minrows)
530
targrows = vacattrstats[i]->minrows;
532
for (ind = 0; ind < nindexes; ind++)
534
AnlIndexData *thisdata = &indexdata[ind];
536
for (i = 0; i < thisdata->attr_cnt; i++)
538
if (targrows < thisdata->vacattrstats[i]->minrows)
539
targrows = thisdata->vacattrstats[i]->minrows;
544
* Acquire the sample rows
546
rows = (HeapTuple *) palloc(targrows * sizeof(HeapTuple));
548
numrows = acquire_inherited_sample_rows(onerel, elevel,
550
&totalrows, &totaldeadrows);
552
numrows = (*acquirefunc) (onerel, elevel,
554
&totalrows, &totaldeadrows);
557
* Compute the statistics. Temporary results during the calculations for
558
* each column are stored in a child context. The calc routines are
559
* responsible to make sure that whatever they store into the VacAttrStats
560
* structure is allocated in anl_context.
564
MemoryContext col_context,
567
col_context = AllocSetContextCreate(anl_context,
569
ALLOCSET_DEFAULT_SIZES);
570
old_context = MemoryContextSwitchTo(col_context);
572
for (i = 0; i < attr_cnt; i++)
574
VacAttrStats *stats = vacattrstats[i];
578
stats->tupDesc = onerel->rd_att;
579
stats->compute_stats(stats,
585
* If the appropriate flavor of the n_distinct option is
586
* specified, override with the corresponding value.
588
aopt = get_attribute_options(onerel->rd_id, stats->attr->attnum);
593
n_distinct = inh ? aopt->n_distinct_inherited : aopt->n_distinct;
594
if (n_distinct != 0.0)
595
stats->stadistinct = n_distinct;
598
MemoryContextResetAndDeleteChildren(col_context);
602
compute_index_stats(onerel, totalrows,
607
MemoryContextSwitchTo(old_context);
608
MemoryContextDelete(col_context);
611
* Emit the completed stats rows into pg_statistic, replacing any
612
* previous statistics for the target columns. (If there are stats in
613
* pg_statistic for columns we didn't process, we leave them alone.)
615
update_attstats(RelationGetRelid(onerel), inh,
616
attr_cnt, vacattrstats);
618
for (ind = 0; ind < nindexes; ind++)
620
AnlIndexData *thisdata = &indexdata[ind];
622
update_attstats(RelationGetRelid(Irel[ind]), false,
623
thisdata->attr_cnt, thisdata->vacattrstats);
626
/* Build extended statistics (if there are any). */
627
BuildRelationExtStatistics(onerel, totalrows, numrows, rows, attr_cnt,
632
* Update pages/tuples stats in pg_class ... but not if we're doing
637
BlockNumber relallvisible;
639
visibilitymap_count(onerel, &relallvisible, NULL);
641
vac_update_relstats(onerel,
646
InvalidTransactionId,
652
* Same for indexes. Vacuum always scans all indexes, so if we're part of
653
* VACUUM ANALYZE, don't overwrite the accurate count already inserted by
656
if (!inh && !(options & VACOPT_VACUUM))
658
for (ind = 0; ind < nindexes; ind++)
660
AnlIndexData *thisdata = &indexdata[ind];
661
double totalindexrows;
663
totalindexrows = ceil(thisdata->tupleFract * totalrows);
664
vac_update_relstats(Irel[ind],
665
RelationGetNumberOfBlocks(Irel[ind]),
669
InvalidTransactionId,
676
* Report ANALYZE to the stats collector, too. However, if doing
677
* inherited stats we shouldn't report, because the stats collector only
678
* tracks per-table stats. Reset the changes_since_analyze counter only
679
* if we analyzed all columns; otherwise, there is still work for
680
* auto-analyze to do.
683
pgstat_report_analyze(onerel, totalrows, totaldeadrows,
686
/* If this isn't part of VACUUM ANALYZE, let index AMs do cleanup */
687
if (!(options & VACOPT_VACUUM))
689
for (ind = 0; ind < nindexes; ind++)
691
IndexBulkDeleteResult *stats;
692
IndexVacuumInfo ivinfo;
694
ivinfo.index = Irel[ind];
695
ivinfo.analyze_only = true;
696
ivinfo.estimated_count = true;
697
ivinfo.message_level = elevel;
698
ivinfo.num_heap_tuples = onerel->rd_rel->reltuples;
699
ivinfo.strategy = vac_strategy;
701
stats = index_vacuum_cleanup(&ivinfo, NULL);
708
/* Done with indexes */
709
vac_close_indexes(nindexes, Irel, NoLock);
711
/* Log the action if appropriate */
712
if (IsAutoVacuumWorkerProcess() && params->log_min_duration >= 0)
714
if (params->log_min_duration == 0 ||
715
TimestampDifferenceExceeds(starttime, GetCurrentTimestamp(),
716
params->log_min_duration))
718
(errmsg("automatic analyze of table \"%s.%s.%s\" system usage: %s",
719
get_database_name(MyDatabaseId),
720
get_namespace_name(RelationGetNamespace(onerel)),
721
RelationGetRelationName(onerel),
722
pg_rusage_show(&ru0))));
725
/* Roll back any GUC changes executed by index functions */
726
AtEOXact_GUC(false, save_nestlevel);
728
/* Restore userid and security context */
729
SetUserIdAndSecContext(save_userid, save_sec_context);
731
/* Restore current context and release memory */
732
MemoryContextSwitchTo(caller_context);
733
MemoryContextDelete(anl_context);
738
* Compute statistics about indexes of a relation
741
compute_index_stats(Relation onerel, double totalrows,
742
AnlIndexData *indexdata, int nindexes,
743
HeapTuple *rows, int numrows,
744
MemoryContext col_context)
746
MemoryContext ind_context,
748
Datum values[INDEX_MAX_KEYS];
749
bool isnull[INDEX_MAX_KEYS];
753
ind_context = AllocSetContextCreate(anl_context,
755
ALLOCSET_DEFAULT_SIZES);
756
old_context = MemoryContextSwitchTo(ind_context);
758
for (ind = 0; ind < nindexes; ind++)
760
AnlIndexData *thisdata = &indexdata[ind];
761
IndexInfo *indexInfo = thisdata->indexInfo;
762
int attr_cnt = thisdata->attr_cnt;
763
TupleTableSlot *slot;
765
ExprContext *econtext;
766
ExprState *predicate;
772
double totalindexrows;
774
/* Ignore index if no columns to analyze and not partial */
775
if (attr_cnt == 0 && indexInfo->ii_Predicate == NIL)
779
* Need an EState for evaluation of index expressions and
780
* partial-index predicates. Create it in the per-index context to be
781
* sure it gets cleaned up at the bottom of the loop.
783
estate = CreateExecutorState();
784
econtext = GetPerTupleExprContext(estate);
785
/* Need a slot to hold the current heap tuple, too */
786
slot = MakeSingleTupleTableSlot(RelationGetDescr(onerel));
788
/* Arrange for econtext's scan tuple to be the tuple under test */
789
econtext->ecxt_scantuple = slot;
791
/* Set up execution state for predicate. */
792
predicate = ExecPrepareQual(indexInfo->ii_Predicate, estate);
794
/* Compute and save index expression values */
795
exprvals = (Datum *) palloc(numrows * attr_cnt * sizeof(Datum));
796
exprnulls = (bool *) palloc(numrows * attr_cnt * sizeof(bool));
799
for (rowno = 0; rowno < numrows; rowno++)
801
HeapTuple heapTuple = rows[rowno];
803
vacuum_delay_point();
806
* Reset the per-tuple context each time, to reclaim any cruft
807
* left behind by evaluating the predicate or index expressions.
809
ResetExprContext(econtext);
811
/* Set up for predicate or expression evaluation */
812
ExecStoreTuple(heapTuple, slot, InvalidBuffer, false);
814
/* If index is partial, check predicate */
815
if (predicate != NULL)
817
if (!ExecQual(predicate, econtext))
825
* Evaluate the index row to compute expression values. We
826
* could do this by hand, but FormIndexDatum is convenient.
828
FormIndexDatum(indexInfo,
835
* Save just the columns we care about. We copy the values
836
* into ind_context from the estate's per-tuple context.
838
for (i = 0; i < attr_cnt; i++)
840
VacAttrStats *stats = thisdata->vacattrstats[i];
841
int attnum = stats->attr->attnum;
843
if (isnull[attnum - 1])
845
exprvals[tcnt] = (Datum) 0;
846
exprnulls[tcnt] = true;
850
exprvals[tcnt] = datumCopy(values[attnum - 1],
851
stats->attrtype->typbyval,
852
stats->attrtype->typlen);
853
exprnulls[tcnt] = false;
861
* Having counted the number of rows that pass the predicate in the
862
* sample, we can estimate the total number of rows in the index.
864
thisdata->tupleFract = (double) numindexrows / (double) numrows;
865
totalindexrows = ceil(thisdata->tupleFract * totalrows);
868
* Now we can compute the statistics for the expression columns.
870
if (numindexrows > 0)
872
MemoryContextSwitchTo(col_context);
873
for (i = 0; i < attr_cnt; i++)
875
VacAttrStats *stats = thisdata->vacattrstats[i];
876
AttributeOpts *aopt =
877
get_attribute_options(stats->attr->attrelid,
878
stats->attr->attnum);
880
stats->exprvals = exprvals + i;
881
stats->exprnulls = exprnulls + i;
882
stats->rowstride = attr_cnt;
883
stats->compute_stats(stats,
889
* If the n_distinct option is specified, it overrides the
890
* above computation. For indices, we always use just
891
* n_distinct, not n_distinct_inherited.
893
if (aopt != NULL && aopt->n_distinct != 0.0)
894
stats->stadistinct = aopt->n_distinct;
896
MemoryContextResetAndDeleteChildren(col_context);
901
MemoryContextSwitchTo(ind_context);
903
ExecDropSingleTupleTableSlot(slot);
904
FreeExecutorState(estate);
905
MemoryContextResetAndDeleteChildren(ind_context);
908
MemoryContextSwitchTo(old_context);
909
MemoryContextDelete(ind_context);
913
* examine_attribute -- pre-analysis of a single column
915
* Determine whether the column is analyzable; if so, create and initialize
916
* a VacAttrStats struct for it. If not, return NULL.
918
* If index_expr isn't NULL, then we're trying to analyze an expression index,
919
* and index_expr is the expression tree representing the column's data.
921
static VacAttrStats *
922
examine_attribute(Relation onerel, int attnum, Node *index_expr)
924
Form_pg_attribute attr = TupleDescAttr(onerel->rd_att, attnum - 1);
930
/* Never analyze dropped columns */
931
if (attr->attisdropped)
934
/* Don't analyze column if user has specified not to */
935
if (attr->attstattarget == 0)
939
* Create the VacAttrStats struct. Note that we only have a copy of the
940
* fixed fields of the pg_attribute tuple.
942
stats = (VacAttrStats *) palloc0(sizeof(VacAttrStats));
943
stats->attr = (Form_pg_attribute) palloc(ATTRIBUTE_FIXED_PART_SIZE);
944
memcpy(stats->attr, attr, ATTRIBUTE_FIXED_PART_SIZE);
947
* When analyzing an expression index, believe the expression tree's type
948
* not the column datatype --- the latter might be the opckeytype storage
949
* type of the opclass, which is not interesting for our purposes. (Note:
950
* if we did anything with non-expression index columns, we'd need to
951
* figure out where to get the correct type info from, but for now that's
952
* not a problem.) It's not clear whether anyone will care about the
953
* typmod, but we store that too just in case.
957
stats->attrtypid = exprType(index_expr);
958
stats->attrtypmod = exprTypmod(index_expr);
962
stats->attrtypid = attr->atttypid;
963
stats->attrtypmod = attr->atttypmod;
966
typtuple = SearchSysCacheCopy1(TYPEOID,
967
ObjectIdGetDatum(stats->attrtypid));
968
if (!HeapTupleIsValid(typtuple))
969
elog(ERROR, "cache lookup failed for type %u", stats->attrtypid);
970
stats->attrtype = (Form_pg_type) GETSTRUCT(typtuple);
971
stats->anl_context = anl_context;
972
stats->tupattnum = attnum;
975
* The fields describing the stats->stavalues[n] element types default to
976
* the type of the data being analyzed, but the type-specific typanalyze
977
* function can change them if it wants to store something else.
979
for (i = 0; i < STATISTIC_NUM_SLOTS; i++)
981
stats->statypid[i] = stats->attrtypid;
982
stats->statyplen[i] = stats->attrtype->typlen;
983
stats->statypbyval[i] = stats->attrtype->typbyval;
984
stats->statypalign[i] = stats->attrtype->typalign;
988
* Call the type-specific typanalyze function. If none is specified, use
991
if (OidIsValid(stats->attrtype->typanalyze))
992
ok = DatumGetBool(OidFunctionCall1(stats->attrtype->typanalyze,
993
PointerGetDatum(stats)));
995
ok = std_typanalyze(stats);
997
if (!ok || stats->compute_stats == NULL || stats->minrows <= 0)
999
heap_freetuple(typtuple);
1009
* acquire_sample_rows -- acquire a random sample of rows from the table
1011
* Selected rows are returned in the caller-allocated array rows[], which
1012
* must have at least targrows entries.
1013
* The actual number of rows selected is returned as the function result.
1014
* We also estimate the total numbers of live and dead rows in the table,
1015
* and return them into *totalrows and *totaldeadrows, respectively.
1017
* The returned list of tuples is in order by physical position in the table.
1018
* (We will rely on this later to derive correlation estimates.)
1020
* As of May 2004 we use a new two-stage method: Stage one selects up
1021
* to targrows random blocks (or all blocks, if there aren't so many).
1022
* Stage two scans these blocks and uses the Vitter algorithm to create
1023
* a random sample of targrows rows (or less, if there are less in the
1024
* sample of blocks). The two stages are executed simultaneously: each
1025
* block is processed as soon as stage one returns its number and while
1026
* the rows are read stage two controls which ones are to be inserted
1029
* Although every row has an equal chance of ending up in the final
1030
* sample, this sampling method is not perfect: not every possible
1031
* sample has an equal chance of being selected. For large relations
1032
* the number of different blocks represented by the sample tends to be
1033
* too small. We can live with that for now. Improvements are welcome.
1035
* An important property of this sampling method is that because we do
1036
* look at a statistically unbiased set of blocks, we should get
1037
* unbiased estimates of the average numbers of live and dead rows per
1038
* block. The previous sampling method put too much credence in the row
1039
* density near the start of the table.
1042
acquire_sample_rows(Relation onerel, int elevel,
1043
HeapTuple *rows, int targrows,
1044
double *totalrows, double *totaldeadrows)
1046
int numrows = 0; /* # rows now in reservoir */
1047
double samplerows = 0; /* total # rows collected */
1048
double liverows = 0; /* # live rows seen */
1049
double deadrows = 0; /* # dead rows seen */
1050
double rowstoskip = -1; /* -1 means not set yet */
1051
BlockNumber totalblocks;
1052
TransactionId OldestXmin;
1053
BlockSamplerData bs;
1054
ReservoirStateData rstate;
1056
Assert(targrows > 0);
1058
totalblocks = RelationGetNumberOfBlocks(onerel);
1060
/* Need a cutoff xmin for HeapTupleSatisfiesVacuum */
1061
OldestXmin = GetOldestXmin(onerel, PROCARRAY_FLAGS_VACUUM);
1063
/* Prepare for sampling block numbers */
1064
BlockSampler_Init(&bs, totalblocks, targrows, random());
1065
/* Prepare for sampling rows */
1066
reservoir_init_selection_state(&rstate, targrows);
1068
/* Outer loop over blocks to sample */
1069
while (BlockSampler_HasMore(&bs))
1071
BlockNumber targblock = BlockSampler_Next(&bs);
1074
OffsetNumber targoffset,
1077
vacuum_delay_point();
1080
* We must maintain a pin on the target page's buffer to ensure that
1081
* the maxoffset value stays good (else concurrent VACUUM might delete
1082
* tuples out from under us). Hence, pin the page until we are done
1083
* looking at it. We also choose to hold sharelock on the buffer
1084
* throughout --- we could release and re-acquire sharelock for each
1085
* tuple, but since we aren't doing much work per tuple, the extra
1086
* lock traffic is probably better avoided.
1088
targbuffer = ReadBufferExtended(onerel, MAIN_FORKNUM, targblock,
1089
RBM_NORMAL, vac_strategy);
1090
LockBuffer(targbuffer, BUFFER_LOCK_SHARE);
1091
targpage = BufferGetPage(targbuffer);
1092
maxoffset = PageGetMaxOffsetNumber(targpage);
1094
/* Inner loop over all tuples on the selected page */
1095
for (targoffset = FirstOffsetNumber; targoffset <= maxoffset; targoffset++)
1098
HeapTupleData targtuple;
1099
bool sample_it = false;
1101
itemid = PageGetItemId(targpage, targoffset);
1104
* We ignore unused and redirect line pointers. DEAD line
1105
* pointers should be counted as dead, because we need vacuum to
1106
* run to get rid of them. Note that this rule agrees with the
1107
* way that heap_page_prune() counts things.
1109
if (!ItemIdIsNormal(itemid))
1111
if (ItemIdIsDead(itemid))
1116
ItemPointerSet(&targtuple.t_self, targblock, targoffset);
1118
targtuple.t_tableOid = RelationGetRelid(onerel);
1119
targtuple.t_data = (HeapTupleHeader) PageGetItem(targpage, itemid);
1120
targtuple.t_len = ItemIdGetLength(itemid);
1122
switch (HeapTupleSatisfiesVacuum(&targtuple,
1126
case HEAPTUPLE_LIVE:
1131
case HEAPTUPLE_DEAD:
1132
case HEAPTUPLE_RECENTLY_DEAD:
1133
/* Count dead and recently-dead rows */
1137
case HEAPTUPLE_INSERT_IN_PROGRESS:
1140
* Insert-in-progress rows are not counted. We assume
1141
* that when the inserting transaction commits or aborts,
1142
* it will send a stats message to increment the proper
1143
* count. This works right only if that transaction ends
1144
* after we finish analyzing the table; if things happen
1145
* in the other order, its stats update will be
1146
* overwritten by ours. However, the error will be large
1147
* only if the other transaction runs long enough to
1148
* insert many tuples, so assuming it will finish after us
1149
* is the safer option.
1151
* A special case is that the inserting transaction might
1152
* be our own. In this case we should count and sample
1153
* the row, to accommodate users who load a table and
1154
* analyze it in one transaction. (pgstat_report_analyze
1155
* has to adjust the numbers we send to the stats
1156
* collector to make this come out right.)
1158
if (TransactionIdIsCurrentTransactionId(HeapTupleHeaderGetXmin(targtuple.t_data)))
1165
case HEAPTUPLE_DELETE_IN_PROGRESS:
1168
* We count delete-in-progress rows as still live, using
1169
* the same reasoning given above; but we don't bother to
1170
* include them in the sample.
1172
* If the delete was done by our own transaction, however,
1173
* we must count the row as dead to make
1174
* pgstat_report_analyze's stats adjustments come out
1175
* right. (Note: this works out properly when the row was
1176
* both inserted and deleted in our xact.)
1178
if (TransactionIdIsCurrentTransactionId(HeapTupleHeaderGetUpdateXid(targtuple.t_data)))
1185
elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result");
1192
* The first targrows sample rows are simply copied into the
1193
* reservoir. Then we start replacing tuples in the sample
1194
* until we reach the end of the relation. This algorithm is
1195
* from Jeff Vitter's paper (see full citation below). It
1196
* works by repeatedly computing the number of tuples to skip
1197
* before selecting a tuple, which replaces a randomly chosen
1198
* element of the reservoir (current set of tuples). At all
1199
* times the reservoir is a true random sample of the tuples
1200
* we've passed over so far, so when we fall off the end of
1201
* the relation we're done.
1203
if (numrows < targrows)
1204
rows[numrows++] = heap_copytuple(&targtuple);
1208
* t in Vitter's paper is the number of records already
1209
* processed. If we need to compute a new S value, we
1210
* must use the not-yet-incremented value of samplerows as
1214
rowstoskip = reservoir_get_next_S(&rstate, samplerows, targrows);
1216
if (rowstoskip <= 0)
1219
* Found a suitable tuple, so save it, replacing one
1220
* old tuple at random
1222
int k = (int) (targrows * sampler_random_fract(rstate.randstate));
1224
Assert(k >= 0 && k < targrows);
1225
heap_freetuple(rows[k]);
1226
rows[k] = heap_copytuple(&targtuple);
1236
/* Now release the lock and pin on the page */
1237
UnlockReleaseBuffer(targbuffer);
1241
* If we didn't find as many tuples as we wanted then we're done. No sort
1242
* is needed, since they're already in order.
1244
* Otherwise we need to sort the collected tuples by position
1245
* (itempointer). It's not worth worrying about corner cases where the
1246
* tuples are already sorted.
1248
if (numrows == targrows)
1249
qsort((void *) rows, numrows, sizeof(HeapTuple), compare_rows);
1252
* Estimate total numbers of live and dead rows in relation, extrapolating
1253
* on the assumption that the average tuple density in pages we didn't
1254
* scan is the same as in the pages we did scan. Since what we scanned is
1255
* a random sample of the pages in the relation, this should be a good
1260
*totalrows = floor((liverows / bs.m) * totalblocks + 0.5);
1261
*totaldeadrows = floor((deadrows / bs.m) * totalblocks + 0.5);
1266
*totaldeadrows = 0.0;
1270
* Emit some interesting relation info
1273
(errmsg("\"%s\": scanned %d of %u pages, "
1274
"containing %.0f live rows and %.0f dead rows; "
1275
"%d rows in sample, %.0f estimated total rows",
1276
RelationGetRelationName(onerel),
1279
numrows, *totalrows)));
1285
* qsort comparator for sorting rows[] array
1288
compare_rows(const void *a, const void *b)
1290
HeapTuple ha = *(const HeapTuple *) a;
1291
HeapTuple hb = *(const HeapTuple *) b;
1292
BlockNumber ba = ItemPointerGetBlockNumber(&ha->t_self);
1293
OffsetNumber oa = ItemPointerGetOffsetNumber(&ha->t_self);
1294
BlockNumber bb = ItemPointerGetBlockNumber(&hb->t_self);
1295
OffsetNumber ob = ItemPointerGetOffsetNumber(&hb->t_self);
1310
* acquire_inherited_sample_rows -- acquire sample rows from inheritance tree
1312
* This has the same API as acquire_sample_rows, except that rows are
1313
* collected from all inheritance children as well as the specified table.
1314
* We fail and return zero if there are no inheritance children, or if all
1315
* children are foreign tables that don't support ANALYZE.
1318
acquire_inherited_sample_rows(Relation onerel, int elevel,
1319
HeapTuple *rows, int targrows,
1320
double *totalrows, double *totaldeadrows)
1324
AcquireSampleRowsFunc *acquirefuncs;
1334
* Find all members of inheritance set. We only need AccessShareLock on
1338
find_all_inheritors(RelationGetRelid(onerel), AccessShareLock, NULL);
1341
* Check that there's at least one descendant, else fail. This could
1342
* happen despite analyze_rel's relhassubclass check, if table once had a
1343
* child but no longer does. In that case, we can clear the
1344
* relhassubclass field so as not to make the same mistake again later.
1345
* (This is safe because we hold ShareUpdateExclusiveLock.)
1347
if (list_length(tableOIDs) < 2)
1349
/* CCI because we already updated the pg_class row in this command */
1350
CommandCounterIncrement();
1351
SetRelationHasSubclass(RelationGetRelid(onerel), false);
1353
(errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no child tables",
1354
get_namespace_name(RelationGetNamespace(onerel)),
1355
RelationGetRelationName(onerel))));
1360
* Identify acquirefuncs to use, and count blocks in all the relations.
1361
* The result could overflow BlockNumber, so we use double arithmetic.
1363
rels = (Relation *) palloc(list_length(tableOIDs) * sizeof(Relation));
1364
acquirefuncs = (AcquireSampleRowsFunc *)
1365
palloc(list_length(tableOIDs) * sizeof(AcquireSampleRowsFunc));
1366
relblocks = (double *) palloc(list_length(tableOIDs) * sizeof(double));
1370
foreach(lc, tableOIDs)
1372
Oid childOID = lfirst_oid(lc);
1374
AcquireSampleRowsFunc acquirefunc = NULL;
1375
BlockNumber relpages = 0;
1377
/* We already got the needed lock */
1378
childrel = heap_open(childOID, NoLock);
1380
/* Ignore if temp table of another backend */
1381
if (RELATION_IS_OTHER_TEMP(childrel))
1383
/* ... but release the lock on it */
1384
Assert(childrel != onerel);
1385
heap_close(childrel, AccessShareLock);
1389
/* Check table type (MATVIEW can't happen, but might as well allow) */
1390
if (childrel->rd_rel->relkind == RELKIND_RELATION ||
1391
childrel->rd_rel->relkind == RELKIND_MATVIEW)
1393
/* Regular table, so use the regular row acquisition function */
1394
acquirefunc = acquire_sample_rows;
1395
relpages = RelationGetNumberOfBlocks(childrel);
1397
else if (childrel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
1400
* For a foreign table, call the FDW's hook function to see
1401
* whether it supports analysis.
1403
FdwRoutine *fdwroutine;
1406
fdwroutine = GetFdwRoutineForRelation(childrel, false);
1408
if (fdwroutine->AnalyzeForeignTable != NULL)
1409
ok = fdwroutine->AnalyzeForeignTable(childrel,
1415
/* ignore, but release the lock on it */
1416
Assert(childrel != onerel);
1417
heap_close(childrel, AccessShareLock);
1424
* ignore, but release the lock on it. don't try to unlock the
1425
* passed-in relation
1427
Assert(childrel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE);
1428
if (childrel != onerel)
1429
heap_close(childrel, AccessShareLock);
1431
heap_close(childrel, NoLock);
1435
/* OK, we'll process this child */
1437
rels[nrels] = childrel;
1438
acquirefuncs[nrels] = acquirefunc;
1439
relblocks[nrels] = (double) relpages;
1440
totalblocks += (double) relpages;
1445
* If we don't have at least one child table to consider, fail. If the
1446
* relation is a partitioned table, it's not counted as a child table.
1451
(errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no analyzable child tables",
1452
get_namespace_name(RelationGetNamespace(onerel)),
1453
RelationGetRelationName(onerel))));
1458
* Now sample rows from each relation, proportionally to its fraction of
1459
* the total block count. (This might be less than desirable if the child
1460
* rels have radically different free-space percentages, but it's not
1461
* clear that it's worth working harder.)
1466
for (i = 0; i < nrels; i++)
1468
Relation childrel = rels[i];
1469
AcquireSampleRowsFunc acquirefunc = acquirefuncs[i];
1470
double childblocks = relblocks[i];
1472
if (childblocks > 0)
1476
childtargrows = (int) rint(targrows * childblocks / totalblocks);
1477
/* Make sure we don't overrun due to roundoff error */
1478
childtargrows = Min(childtargrows, targrows - numrows);
1479
if (childtargrows > 0)
1485
/* Fetch a random sample of the child's rows */
1486
childrows = (*acquirefunc) (childrel, elevel,
1487
rows + numrows, childtargrows,
1490
/* We may need to convert from child's rowtype to parent's */
1491
if (childrows > 0 &&
1492
!equalTupleDescs(RelationGetDescr(childrel),
1493
RelationGetDescr(onerel)))
1495
TupleConversionMap *map;
1497
map = convert_tuples_by_name(RelationGetDescr(childrel),
1498
RelationGetDescr(onerel),
1499
gettext_noop("could not convert row type"));
1504
for (j = 0; j < childrows; j++)
1508
newtup = do_convert_tuple(rows[numrows + j], map);
1509
heap_freetuple(rows[numrows + j]);
1510
rows[numrows + j] = newtup;
1512
free_conversion_map(map);
1516
/* And add to counts */
1517
numrows += childrows;
1518
*totalrows += trows;
1519
*totaldeadrows += tdrows;
1524
* Note: we cannot release the child-table locks, since we may have
1525
* pointers to their TOAST tables in the sampled rows.
1527
heap_close(childrel, NoLock);
1535
* update_attstats() -- update attribute statistics for one relation
1537
* Statistics are stored in several places: the pg_class row for the
1538
* relation has stats about the whole relation, and there is a
1539
* pg_statistic row for each (non-system) attribute that has ever
1540
* been analyzed. The pg_class values are updated by VACUUM, not here.
1542
* pg_statistic rows are just added or updated normally. This means
1543
* that pg_statistic will probably contain some deleted rows at the
1544
* completion of a vacuum cycle, unless it happens to get vacuumed last.
1546
* To keep things simple, we punt for pg_statistic, and don't try
1547
* to compute or store rows for pg_statistic itself in pg_statistic.
1548
* This could possibly be made to work, but it's not worth the trouble.
1549
* Note analyze_rel() has seen to it that we won't come here when
1550
* vacuuming pg_statistic itself.
1552
* Note: there would be a race condition here if two backends could
1553
* ANALYZE the same table concurrently. Presently, we lock that out
1554
* by taking a self-exclusive lock on the relation in analyze_rel().
1557
update_attstats(Oid relid, bool inh, int natts, VacAttrStats **vacattrstats)
1563
return; /* nothing to do */
1565
sd = heap_open(StatisticRelationId, RowExclusiveLock);
1567
for (attno = 0; attno < natts; attno++)
1569
VacAttrStats *stats = vacattrstats[attno];
1575
Datum values[Natts_pg_statistic];
1576
bool nulls[Natts_pg_statistic];
1577
bool replaces[Natts_pg_statistic];
1579
/* Ignore attr if we weren't able to collect stats */
1580
if (!stats->stats_valid)
1584
* Construct a new pg_statistic tuple
1586
for (i = 0; i < Natts_pg_statistic; ++i)
1592
values[Anum_pg_statistic_starelid - 1] = ObjectIdGetDatum(relid);
1593
values[Anum_pg_statistic_staattnum - 1] = Int16GetDatum(stats->attr->attnum);
1594
values[Anum_pg_statistic_stainherit - 1] = BoolGetDatum(inh);
1595
values[Anum_pg_statistic_stanullfrac - 1] = Float4GetDatum(stats->stanullfrac);
1596
values[Anum_pg_statistic_stawidth - 1] = Int32GetDatum(stats->stawidth);
1597
values[Anum_pg_statistic_stadistinct - 1] = Float4GetDatum(stats->stadistinct);
1598
i = Anum_pg_statistic_stakind1 - 1;
1599
for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1601
values[i++] = Int16GetDatum(stats->stakind[k]); /* stakindN */
1603
i = Anum_pg_statistic_staop1 - 1;
1604
for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1606
values[i++] = ObjectIdGetDatum(stats->staop[k]); /* staopN */
1608
i = Anum_pg_statistic_stanumbers1 - 1;
1609
for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1611
int nnum = stats->numnumbers[k];
1615
Datum *numdatums = (Datum *) palloc(nnum * sizeof(Datum));
1618
for (n = 0; n < nnum; n++)
1619
numdatums[n] = Float4GetDatum(stats->stanumbers[k][n]);
1620
/* XXX knows more than it should about type float4: */
1621
arry = construct_array(numdatums, nnum,
1623
sizeof(float4), FLOAT4PASSBYVAL, 'i');
1624
values[i++] = PointerGetDatum(arry); /* stanumbersN */
1629
values[i++] = (Datum) 0;
1632
i = Anum_pg_statistic_stavalues1 - 1;
1633
for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1635
if (stats->numvalues[k] > 0)
1639
arry = construct_array(stats->stavalues[k],
1640
stats->numvalues[k],
1642
stats->statyplen[k],
1643
stats->statypbyval[k],
1644
stats->statypalign[k]);
1645
values[i++] = PointerGetDatum(arry); /* stavaluesN */
1650
values[i++] = (Datum) 0;
1654
/* Is there already a pg_statistic tuple for this attribute? */
1655
oldtup = SearchSysCache3(STATRELATTINH,
1656
ObjectIdGetDatum(relid),
1657
Int16GetDatum(stats->attr->attnum),
1660
if (HeapTupleIsValid(oldtup))
1662
/* Yes, replace it */
1663
stup = heap_modify_tuple(oldtup,
1664
RelationGetDescr(sd),
1668
ReleaseSysCache(oldtup);
1669
CatalogTupleUpdate(sd, &stup->t_self, stup);
1673
/* No, insert new tuple */
1674
stup = heap_form_tuple(RelationGetDescr(sd), values, nulls);
1675
CatalogTupleInsert(sd, stup);
1678
heap_freetuple(stup);
1681
heap_close(sd, RowExclusiveLock);
1685
* Standard fetch function for use by compute_stats subroutines.
1687
* This exists to provide some insulation between compute_stats routines
1688
* and the actual storage of the sample data.
1691
std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1693
int attnum = stats->tupattnum;
1694
HeapTuple tuple = stats->rows[rownum];
1695
TupleDesc tupDesc = stats->tupDesc;
1697
return heap_getattr(tuple, attnum, tupDesc, isNull);
1701
* Fetch function for analyzing index expressions.
1703
* We have not bothered to construct index tuples, instead the data is
1704
* just in Datum arrays.
1707
ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1711
/* exprvals and exprnulls are already offset for proper column */
1712
i = rownum * stats->rowstride;
1713
*isNull = stats->exprnulls[i];
1714
return stats->exprvals[i];
1718
/*==========================================================================
1720
* Code below this point represents the "standard" type-specific statistics
1721
* analysis algorithms. This code can be replaced on a per-data-type basis
1722
* by setting a nonzero value in pg_type.typanalyze.
1724
*==========================================================================
1729
* To avoid consuming too much memory during analysis and/or too much space
1730
* in the resulting pg_statistic rows, we ignore varlena datums that are wider
1731
* than WIDTH_THRESHOLD (after detoasting!). This is legitimate for MCV
1732
* and distinct-value calculations since a wide value is unlikely to be
1733
* duplicated at all, much less be a most-common value. For the same reason,
1734
* ignoring wide values will not affect our estimates of histogram bin
1735
* boundaries very much.
1737
#define WIDTH_THRESHOLD 1024
1739
#define swapInt(a,b) do {int _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1740
#define swapDatum(a,b) do {Datum _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1743
* Extra information used by the default analysis routines
1747
int count; /* # of duplicates */
1748
int first; /* values[] index of first occurrence */
1755
} CompareScalarsContext;
1758
static void compute_trivial_stats(VacAttrStatsP stats,
1759
AnalyzeAttrFetchFunc fetchfunc,
1762
static void compute_distinct_stats(VacAttrStatsP stats,
1763
AnalyzeAttrFetchFunc fetchfunc,
1766
static void compute_scalar_stats(VacAttrStatsP stats,
1767
AnalyzeAttrFetchFunc fetchfunc,
1770
static int compare_scalars(const void *a, const void *b, void *arg);
1771
static int compare_mcvs(const void *a, const void *b);
1772
static int analyze_mcv_list(int *mcv_counts,
1781
* std_typanalyze -- the default type-specific typanalyze function
1784
std_typanalyze(VacAttrStats *stats)
1786
Form_pg_attribute attr = stats->attr;
1789
StdAnalyzeData *mystats;
1791
/* If the attstattarget column is negative, use the default value */
1792
/* NB: it is okay to scribble on stats->attr since it's a copy */
1793
if (attr->attstattarget < 0)
1794
attr->attstattarget = default_statistics_target;
1796
/* Look for default "<" and "=" operators for column's type */
1797
get_sort_group_operators(stats->attrtypid,
1798
false, false, false,
1799
<opr, &eqopr, NULL,
1802
/* Save the operator info for compute_stats routines */
1803
mystats = (StdAnalyzeData *) palloc(sizeof(StdAnalyzeData));
1804
mystats->eqopr = eqopr;
1805
mystats->eqfunc = OidIsValid(eqopr) ? get_opcode(eqopr) : InvalidOid;
1806
mystats->ltopr = ltopr;
1807
stats->extra_data = mystats;
1810
* Determine which standard statistics algorithm to use
1812
if (OidIsValid(eqopr) && OidIsValid(ltopr))
1814
/* Seems to be a scalar datatype */
1815
stats->compute_stats = compute_scalar_stats;
1816
/*--------------------
1817
* The following choice of minrows is based on the paper
1818
* "Random sampling for histogram construction: how much is enough?"
1819
* by Surajit Chaudhuri, Rajeev Motwani and Vivek Narasayya, in
1820
* Proceedings of ACM SIGMOD International Conference on Management
1821
* of Data, 1998, Pages 436-447. Their Corollary 1 to Theorem 5
1822
* says that for table size n, histogram size k, maximum relative
1823
* error in bin size f, and error probability gamma, the minimum
1824
* random sample size is
1825
* r = 4 * k * ln(2*n/gamma) / f^2
1826
* Taking f = 0.5, gamma = 0.01, n = 10^6 rows, we obtain
1828
* Note that because of the log function, the dependence on n is
1829
* quite weak; even at n = 10^12, a 300*k sample gives <= 0.66
1830
* bin size error with probability 0.99. So there's no real need to
1831
* scale for n, which is a good thing because we don't necessarily
1832
* know it at this point.
1833
*--------------------
1835
stats->minrows = 300 * attr->attstattarget;
1837
else if (OidIsValid(eqopr))
1839
/* We can still recognize distinct values */
1840
stats->compute_stats = compute_distinct_stats;
1841
/* Might as well use the same minrows as above */
1842
stats->minrows = 300 * attr->attstattarget;
1846
/* Can't do much but the trivial stuff */
1847
stats->compute_stats = compute_trivial_stats;
1848
/* Might as well use the same minrows as above */
1849
stats->minrows = 300 * attr->attstattarget;
1857
* compute_trivial_stats() -- compute very basic column statistics
1859
* We use this when we cannot find a hash "=" operator for the datatype.
1861
* We determine the fraction of non-null rows and the average datum width.
1864
compute_trivial_stats(VacAttrStatsP stats,
1865
AnalyzeAttrFetchFunc fetchfunc,
1871
int nonnull_cnt = 0;
1872
double total_width = 0;
1873
bool is_varlena = (!stats->attrtype->typbyval &&
1874
stats->attrtype->typlen == -1);
1875
bool is_varwidth = (!stats->attrtype->typbyval &&
1876
stats->attrtype->typlen < 0);
1878
for (i = 0; i < samplerows; i++)
1883
vacuum_delay_point();
1885
value = fetchfunc(stats, i, &isnull);
1887
/* Check for null/nonnull */
1896
* If it's a variable-width field, add up widths for average width
1897
* calculation. Note that if the value is toasted, we use the toasted
1898
* width. We don't bother with this calculation if it's a fixed-width
1903
total_width += VARSIZE_ANY(DatumGetPointer(value));
1905
else if (is_varwidth)
1907
/* must be cstring */
1908
total_width += strlen(DatumGetCString(value)) + 1;
1912
/* We can only compute average width if we found some non-null values. */
1913
if (nonnull_cnt > 0)
1915
stats->stats_valid = true;
1916
/* Do the simple null-frac and width stats */
1917
stats->stanullfrac = (double) null_cnt / (double) samplerows;
1919
stats->stawidth = total_width / (double) nonnull_cnt;
1921
stats->stawidth = stats->attrtype->typlen;
1922
stats->stadistinct = 0.0; /* "unknown" */
1924
else if (null_cnt > 0)
1926
/* We found only nulls; assume the column is entirely null */
1927
stats->stats_valid = true;
1928
stats->stanullfrac = 1.0;
1930
stats->stawidth = 0; /* "unknown" */
1932
stats->stawidth = stats->attrtype->typlen;
1933
stats->stadistinct = 0.0; /* "unknown" */
1939
* compute_distinct_stats() -- compute column statistics including ndistinct
1941
* We use this when we can find only an "=" operator for the datatype.
1943
* We determine the fraction of non-null rows, the average width, the
1944
* most common values, and the (estimated) number of distinct values.
1946
* The most common values are determined by brute force: we keep a list
1947
* of previously seen values, ordered by number of times seen, as we scan
1948
* the samples. A newly seen value is inserted just after the last
1949
* multiply-seen value, causing the bottommost (oldest) singly-seen value
1950
* to drop off the list. The accuracy of this method, and also its cost,
1951
* depend mainly on the length of the list we are willing to keep.
1954
compute_distinct_stats(VacAttrStatsP stats,
1955
AnalyzeAttrFetchFunc fetchfunc,
1961
int nonnull_cnt = 0;
1962
int toowide_cnt = 0;
1963
double total_width = 0;
1964
bool is_varlena = (!stats->attrtype->typbyval &&
1965
stats->attrtype->typlen == -1);
1966
bool is_varwidth = (!stats->attrtype->typbyval &&
1967
stats->attrtype->typlen < 0);
1977
int num_mcv = stats->attr->attstattarget;
1978
StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
1981
* We track up to 2*n values for an n-element MCV list; but at least 10
1983
track_max = 2 * num_mcv;
1986
track = (TrackItem *) palloc(track_max * sizeof(TrackItem));
1989
fmgr_info(mystats->eqfunc, &f_cmpeq);
1991
for (i = 0; i < samplerows; i++)
1999
vacuum_delay_point();
2001
value = fetchfunc(stats, i, &isnull);
2003
/* Check for null/nonnull */
2012
* If it's a variable-width field, add up widths for average width
2013
* calculation. Note that if the value is toasted, we use the toasted
2014
* width. We don't bother with this calculation if it's a fixed-width
2019
total_width += VARSIZE_ANY(DatumGetPointer(value));
2022
* If the value is toasted, we want to detoast it just once to
2023
* avoid repeated detoastings and resultant excess memory usage
2024
* during the comparisons. Also, check to see if the value is
2025
* excessively wide, and if so don't detoast at all --- just
2028
if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
2033
value = PointerGetDatum(PG_DETOAST_DATUM(value));
2035
else if (is_varwidth)
2037
/* must be cstring */
2038
total_width += strlen(DatumGetCString(value)) + 1;
2042
* See if the value matches anything we're already tracking.
2045
firstcount1 = track_cnt;
2046
for (j = 0; j < track_cnt; j++)
2048
/* We always use the default collation for statistics */
2049
if (DatumGetBool(FunctionCall2Coll(&f_cmpeq,
2050
DEFAULT_COLLATION_OID,
2051
value, track[j].value)))
2056
if (j < firstcount1 && track[j].count == 1)
2064
/* This value may now need to "bubble up" in the track list */
2065
while (j > 0 && track[j].count > track[j - 1].count)
2067
swapDatum(track[j].value, track[j - 1].value);
2068
swapInt(track[j].count, track[j - 1].count);
2074
/* No match. Insert at head of count-1 list */
2075
if (track_cnt < track_max)
2077
for (j = track_cnt - 1; j > firstcount1; j--)
2079
track[j].value = track[j - 1].value;
2080
track[j].count = track[j - 1].count;
2082
if (firstcount1 < track_cnt)
2084
track[firstcount1].value = value;
2085
track[firstcount1].count = 1;
2090
/* We can only compute real stats if we found some non-null values. */
2091
if (nonnull_cnt > 0)
2096
stats->stats_valid = true;
2097
/* Do the simple null-frac and width stats */
2098
stats->stanullfrac = (double) null_cnt / (double) samplerows;
2100
stats->stawidth = total_width / (double) nonnull_cnt;
2102
stats->stawidth = stats->attrtype->typlen;
2104
/* Count the number of values we found multiple times */
2106
for (nmultiple = 0; nmultiple < track_cnt; nmultiple++)
2108
if (track[nmultiple].count == 1)
2110
summultiple += track[nmultiple].count;
2116
* If we found no repeated non-null values, assume it's a unique
2117
* column; but be sure to discount for any nulls we found.
2119
stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2121
else if (track_cnt < track_max && toowide_cnt == 0 &&
2122
nmultiple == track_cnt)
2125
* Our track list includes every value in the sample, and every
2126
* value appeared more than once. Assume the column has just
2127
* these values. (This case is meant to address columns with
2128
* small, fixed sets of possible values, such as boolean or enum
2129
* columns. If there are any values that appear just once in the
2130
* sample, including too-wide values, we should assume that that's
2131
* not what we're dealing with.)
2133
stats->stadistinct = track_cnt;
2138
* Estimate the number of distinct values using the estimator
2139
* proposed by Haas and Stokes in IBM Research Report RJ 10025:
2140
* n*d / (n - f1 + f1*n/N)
2141
* where f1 is the number of distinct values that occurred
2142
* exactly once in our sample of n rows (from a total of N),
2143
* and d is the total number of distinct values in the sample.
2144
* This is their Duj1 estimator; the other estimators they
2145
* recommend are considerably more complex, and are numerically
2146
* very unstable when n is much smaller than N.
2148
* In this calculation, we consider only non-nulls. We used to
2149
* include rows with null values in the n and N counts, but that
2150
* leads to inaccurate answers in columns with many nulls, and
2151
* it's intuitively bogus anyway considering the desired result is
2152
* the number of distinct non-null values.
2154
* We assume (not very reliably!) that all the multiply-occurring
2155
* values are reflected in the final track[] list, and the other
2156
* nonnull values all appeared but once. (XXX this usually
2157
* results in a drastic overestimate of ndistinct. Can we do
2161
int f1 = nonnull_cnt - summultiple;
2162
int d = f1 + nmultiple;
2163
double n = samplerows - null_cnt;
2164
double N = totalrows * (1.0 - stats->stanullfrac);
2167
/* N == 0 shouldn't happen, but just in case ... */
2169
stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2173
/* Clamp to sane range in case of roundoff error */
2174
if (stadistinct < d)
2176
if (stadistinct > N)
2178
/* And round to integer */
2179
stats->stadistinct = floor(stadistinct + 0.5);
2183
* If we estimated the number of distinct values at more than 10% of
2184
* the total row count (a very arbitrary limit), then assume that
2185
* stadistinct should scale with the row count rather than be a fixed
2188
if (stats->stadistinct > 0.1 * totalrows)
2189
stats->stadistinct = -(stats->stadistinct / totalrows);
2192
* Decide how many values are worth storing as most-common values. If
2193
* we are able to generate a complete MCV list (all the values in the
2194
* sample will fit, and we think these are all the ones in the table),
2195
* then do so. Otherwise, store only those values that are
2196
* significantly more common than the values not in the list.
2198
* Note: the first of these cases is meant to address columns with
2199
* small, fixed sets of possible values, such as boolean or enum
2200
* columns. If we can *completely* represent the column population by
2201
* an MCV list that will fit into the stats target, then we should do
2202
* so and thus provide the planner with complete information. But if
2203
* the MCV list is not complete, it's generally worth being more
2204
* selective, and not just filling it all the way up to the stats
2207
if (track_cnt < track_max && toowide_cnt == 0 &&
2208
stats->stadistinct > 0 &&
2209
track_cnt <= num_mcv)
2211
/* Track list includes all values seen, and all will fit */
2212
num_mcv = track_cnt;
2218
/* Incomplete list; decide how many values are worth keeping */
2219
if (num_mcv > track_cnt)
2220
num_mcv = track_cnt;
2224
mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2225
for (i = 0; i < num_mcv; i++)
2226
mcv_counts[i] = track[i].count;
2228
num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2231
samplerows, totalrows);
2235
/* Generate MCV slot entry */
2238
MemoryContext old_context;
2242
/* Must copy the target values into anl_context */
2243
old_context = MemoryContextSwitchTo(stats->anl_context);
2244
mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2245
mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2246
for (i = 0; i < num_mcv; i++)
2248
mcv_values[i] = datumCopy(track[i].value,
2249
stats->attrtype->typbyval,
2250
stats->attrtype->typlen);
2251
mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2253
MemoryContextSwitchTo(old_context);
2255
stats->stakind[0] = STATISTIC_KIND_MCV;
2256
stats->staop[0] = mystats->eqopr;
2257
stats->stanumbers[0] = mcv_freqs;
2258
stats->numnumbers[0] = num_mcv;
2259
stats->stavalues[0] = mcv_values;
2260
stats->numvalues[0] = num_mcv;
2263
* Accept the defaults for stats->statypid and others. They have
2264
* been set before we were called (see vacuum.h)
2268
else if (null_cnt > 0)
2270
/* We found only nulls; assume the column is entirely null */
2271
stats->stats_valid = true;
2272
stats->stanullfrac = 1.0;
2274
stats->stawidth = 0; /* "unknown" */
2276
stats->stawidth = stats->attrtype->typlen;
2277
stats->stadistinct = 0.0; /* "unknown" */
2280
/* We don't need to bother cleaning up any of our temporary palloc's */
2285
* compute_scalar_stats() -- compute column statistics
2287
* We use this when we can find "=" and "<" operators for the datatype.
2289
* We determine the fraction of non-null rows, the average width, the
2290
* most common values, the (estimated) number of distinct values, the
2291
* distribution histogram, and the correlation of physical to logical order.
2293
* The desired stats can be determined fairly easily after sorting the
2294
* data values into order.
2297
compute_scalar_stats(VacAttrStatsP stats,
2298
AnalyzeAttrFetchFunc fetchfunc,
2304
int nonnull_cnt = 0;
2305
int toowide_cnt = 0;
2306
double total_width = 0;
2307
bool is_varlena = (!stats->attrtype->typbyval &&
2308
stats->attrtype->typlen == -1);
2309
bool is_varwidth = (!stats->attrtype->typbyval &&
2310
stats->attrtype->typlen < 0);
2312
SortSupportData ssup;
2316
ScalarMCVItem *track;
2318
int num_mcv = stats->attr->attstattarget;
2319
int num_bins = stats->attr->attstattarget;
2320
StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2322
values = (ScalarItem *) palloc(samplerows * sizeof(ScalarItem));
2323
tupnoLink = (int *) palloc(samplerows * sizeof(int));
2324
track = (ScalarMCVItem *) palloc(num_mcv * sizeof(ScalarMCVItem));
2326
memset(&ssup, 0, sizeof(ssup));
2327
ssup.ssup_cxt = CurrentMemoryContext;
2328
/* We always use the default collation for statistics */
2329
ssup.ssup_collation = DEFAULT_COLLATION_OID;
2330
ssup.ssup_nulls_first = false;
2333
* For now, don't perform abbreviated key conversion, because full values
2334
* are required for MCV slot generation. Supporting that optimization
2335
* would necessitate teaching compare_scalars() to call a tie-breaker.
2337
ssup.abbreviate = false;
2339
PrepareSortSupportFromOrderingOp(mystats->ltopr, &ssup);
2341
/* Initial scan to find sortable values */
2342
for (i = 0; i < samplerows; i++)
2347
vacuum_delay_point();
2349
value = fetchfunc(stats, i, &isnull);
2351
/* Check for null/nonnull */
2360
* If it's a variable-width field, add up widths for average width
2361
* calculation. Note that if the value is toasted, we use the toasted
2362
* width. We don't bother with this calculation if it's a fixed-width
2367
total_width += VARSIZE_ANY(DatumGetPointer(value));
2370
* If the value is toasted, we want to detoast it just once to
2371
* avoid repeated detoastings and resultant excess memory usage
2372
* during the comparisons. Also, check to see if the value is
2373
* excessively wide, and if so don't detoast at all --- just
2376
if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
2381
value = PointerGetDatum(PG_DETOAST_DATUM(value));
2383
else if (is_varwidth)
2385
/* must be cstring */
2386
total_width += strlen(DatumGetCString(value)) + 1;
2389
/* Add it to the list to be sorted */
2390
values[values_cnt].value = value;
2391
values[values_cnt].tupno = values_cnt;
2392
tupnoLink[values_cnt] = values_cnt;
2396
/* We can only compute real stats if we found some sortable values. */
2399
int ndistinct, /* # distinct values in sample */
2400
nmultiple, /* # that appear multiple times */
2404
CompareScalarsContext cxt;
2406
/* Sort the collected values */
2408
cxt.tupnoLink = tupnoLink;
2409
qsort_arg((void *) values, values_cnt, sizeof(ScalarItem),
2410
compare_scalars, (void *) &cxt);
2413
* Now scan the values in order, find the most common ones, and also
2414
* accumulate ordering-correlation statistics.
2416
* To determine which are most common, we first have to count the
2417
* number of duplicates of each value. The duplicates are adjacent in
2418
* the sorted list, so a brute-force approach is to compare successive
2419
* datum values until we find two that are not equal. However, that
2420
* requires N-1 invocations of the datum comparison routine, which are
2421
* completely redundant with work that was done during the sort. (The
2422
* sort algorithm must at some point have compared each pair of items
2423
* that are adjacent in the sorted order; otherwise it could not know
2424
* that it's ordered the pair correctly.) We exploit this by having
2425
* compare_scalars remember the highest tupno index that each
2426
* ScalarItem has been found equal to. At the end of the sort, a
2427
* ScalarItem's tupnoLink will still point to itself if and only if it
2428
* is the last item of its group of duplicates (since the group will
2429
* be ordered by tupno).
2435
for (i = 0; i < values_cnt; i++)
2437
int tupno = values[i].tupno;
2439
corr_xysum += ((double) i) * ((double) tupno);
2441
if (tupnoLink[tupno] == tupno)
2443
/* Reached end of duplicates of this value */
2448
if (track_cnt < num_mcv ||
2449
dups_cnt > track[track_cnt - 1].count)
2452
* Found a new item for the mcv list; find its
2453
* position, bubbling down old items if needed. Loop
2454
* invariant is that j points at an empty/ replaceable
2459
if (track_cnt < num_mcv)
2461
for (j = track_cnt - 1; j > 0; j--)
2463
if (dups_cnt <= track[j - 1].count)
2465
track[j].count = track[j - 1].count;
2466
track[j].first = track[j - 1].first;
2468
track[j].count = dups_cnt;
2469
track[j].first = i + 1 - dups_cnt;
2476
stats->stats_valid = true;
2477
/* Do the simple null-frac and width stats */
2478
stats->stanullfrac = (double) null_cnt / (double) samplerows;
2480
stats->stawidth = total_width / (double) nonnull_cnt;
2482
stats->stawidth = stats->attrtype->typlen;
2487
* If we found no repeated non-null values, assume it's a unique
2488
* column; but be sure to discount for any nulls we found.
2490
stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2492
else if (toowide_cnt == 0 && nmultiple == ndistinct)
2495
* Every value in the sample appeared more than once. Assume the
2496
* column has just these values. (This case is meant to address
2497
* columns with small, fixed sets of possible values, such as
2498
* boolean or enum columns. If there are any values that appear
2499
* just once in the sample, including too-wide values, we should
2500
* assume that that's not what we're dealing with.)
2502
stats->stadistinct = ndistinct;
2507
* Estimate the number of distinct values using the estimator
2508
* proposed by Haas and Stokes in IBM Research Report RJ 10025:
2509
* n*d / (n - f1 + f1*n/N)
2510
* where f1 is the number of distinct values that occurred
2511
* exactly once in our sample of n rows (from a total of N),
2512
* and d is the total number of distinct values in the sample.
2513
* This is their Duj1 estimator; the other estimators they
2514
* recommend are considerably more complex, and are numerically
2515
* very unstable when n is much smaller than N.
2517
* In this calculation, we consider only non-nulls. We used to
2518
* include rows with null values in the n and N counts, but that
2519
* leads to inaccurate answers in columns with many nulls, and
2520
* it's intuitively bogus anyway considering the desired result is
2521
* the number of distinct non-null values.
2523
* Overwidth values are assumed to have been distinct.
2526
int f1 = ndistinct - nmultiple + toowide_cnt;
2527
int d = f1 + nmultiple;
2528
double n = samplerows - null_cnt;
2529
double N = totalrows * (1.0 - stats->stanullfrac);
2532
/* N == 0 shouldn't happen, but just in case ... */
2534
stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2538
/* Clamp to sane range in case of roundoff error */
2539
if (stadistinct < d)
2541
if (stadistinct > N)
2543
/* And round to integer */
2544
stats->stadistinct = floor(stadistinct + 0.5);
2548
* If we estimated the number of distinct values at more than 10% of
2549
* the total row count (a very arbitrary limit), then assume that
2550
* stadistinct should scale with the row count rather than be a fixed
2553
if (stats->stadistinct > 0.1 * totalrows)
2554
stats->stadistinct = -(stats->stadistinct / totalrows);
2557
* Decide how many values are worth storing as most-common values. If
2558
* we are able to generate a complete MCV list (all the values in the
2559
* sample will fit, and we think these are all the ones in the table),
2560
* then do so. Otherwise, store only those values that are
2561
* significantly more common than the values not in the list.
2563
* Note: the first of these cases is meant to address columns with
2564
* small, fixed sets of possible values, such as boolean or enum
2565
* columns. If we can *completely* represent the column population by
2566
* an MCV list that will fit into the stats target, then we should do
2567
* so and thus provide the planner with complete information. But if
2568
* the MCV list is not complete, it's generally worth being more
2569
* selective, and not just filling it all the way up to the stats
2572
if (track_cnt == ndistinct && toowide_cnt == 0 &&
2573
stats->stadistinct > 0 &&
2574
track_cnt <= num_mcv)
2576
/* Track list includes all values seen, and all will fit */
2577
num_mcv = track_cnt;
2583
/* Incomplete list; decide how many values are worth keeping */
2584
if (num_mcv > track_cnt)
2585
num_mcv = track_cnt;
2589
mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2590
for (i = 0; i < num_mcv; i++)
2591
mcv_counts[i] = track[i].count;
2593
num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2596
samplerows, totalrows);
2600
/* Generate MCV slot entry */
2603
MemoryContext old_context;
2607
/* Must copy the target values into anl_context */
2608
old_context = MemoryContextSwitchTo(stats->anl_context);
2609
mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2610
mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2611
for (i = 0; i < num_mcv; i++)
2613
mcv_values[i] = datumCopy(values[track[i].first].value,
2614
stats->attrtype->typbyval,
2615
stats->attrtype->typlen);
2616
mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2618
MemoryContextSwitchTo(old_context);
2620
stats->stakind[slot_idx] = STATISTIC_KIND_MCV;
2621
stats->staop[slot_idx] = mystats->eqopr;
2622
stats->stanumbers[slot_idx] = mcv_freqs;
2623
stats->numnumbers[slot_idx] = num_mcv;
2624
stats->stavalues[slot_idx] = mcv_values;
2625
stats->numvalues[slot_idx] = num_mcv;
2628
* Accept the defaults for stats->statypid and others. They have
2629
* been set before we were called (see vacuum.h)
2635
* Generate a histogram slot entry if there are at least two distinct
2636
* values not accounted for in the MCV list. (This ensures the
2637
* histogram won't collapse to empty or a singleton.)
2639
num_hist = ndistinct - num_mcv;
2640
if (num_hist > num_bins)
2641
num_hist = num_bins + 1;
2644
MemoryContext old_context;
2652
/* Sort the MCV items into position order to speed next loop */
2653
qsort((void *) track, num_mcv,
2654
sizeof(ScalarMCVItem), compare_mcvs);
2657
* Collapse out the MCV items from the values[] array.
2659
* Note we destroy the values[] array here... but we don't need it
2660
* for anything more. We do, however, still need values_cnt.
2661
* nvals will be the number of remaining entries in values[].
2670
j = 0; /* index of next interesting MCV item */
2671
while (src < values_cnt)
2677
int first = track[j].first;
2681
/* advance past this MCV item */
2682
src = first + track[j].count;
2686
ncopy = first - src;
2689
ncopy = values_cnt - src;
2690
memmove(&values[dest], &values[src],
2691
ncopy * sizeof(ScalarItem));
2699
Assert(nvals >= num_hist);
2701
/* Must copy the target values into anl_context */
2702
old_context = MemoryContextSwitchTo(stats->anl_context);
2703
hist_values = (Datum *) palloc(num_hist * sizeof(Datum));
2706
* The object of this loop is to copy the first and last values[]
2707
* entries along with evenly-spaced values in between. So the
2708
* i'th value is values[(i * (nvals - 1)) / (num_hist - 1)]. But
2709
* computing that subscript directly risks integer overflow when
2710
* the stats target is more than a couple thousand. Instead we
2711
* add (nvals - 1) / (num_hist - 1) to pos at each step, tracking
2712
* the integral and fractional parts of the sum separately.
2714
delta = (nvals - 1) / (num_hist - 1);
2715
deltafrac = (nvals - 1) % (num_hist - 1);
2718
for (i = 0; i < num_hist; i++)
2720
hist_values[i] = datumCopy(values[pos].value,
2721
stats->attrtype->typbyval,
2722
stats->attrtype->typlen);
2724
posfrac += deltafrac;
2725
if (posfrac >= (num_hist - 1))
2727
/* fractional part exceeds 1, carry to integer part */
2729
posfrac -= (num_hist - 1);
2733
MemoryContextSwitchTo(old_context);
2735
stats->stakind[slot_idx] = STATISTIC_KIND_HISTOGRAM;
2736
stats->staop[slot_idx] = mystats->ltopr;
2737
stats->stavalues[slot_idx] = hist_values;
2738
stats->numvalues[slot_idx] = num_hist;
2741
* Accept the defaults for stats->statypid and others. They have
2742
* been set before we were called (see vacuum.h)
2747
/* Generate a correlation entry if there are multiple values */
2750
MemoryContext old_context;
2755
/* Must copy the target values into anl_context */
2756
old_context = MemoryContextSwitchTo(stats->anl_context);
2757
corrs = (float4 *) palloc(sizeof(float4));
2758
MemoryContextSwitchTo(old_context);
2761
* Since we know the x and y value sets are both
2762
* 0, 1, ..., values_cnt-1
2763
* we have sum(x) = sum(y) =
2764
* (values_cnt-1)*values_cnt / 2
2765
* and sum(x^2) = sum(y^2) =
2766
* (values_cnt-1)*values_cnt*(2*values_cnt-1) / 6.
2769
corr_xsum = ((double) (values_cnt - 1)) *
2770
((double) values_cnt) / 2.0;
2771
corr_x2sum = ((double) (values_cnt - 1)) *
2772
((double) values_cnt) * (double) (2 * values_cnt - 1) / 6.0;
2774
/* And the correlation coefficient reduces to */
2775
corrs[0] = (values_cnt * corr_xysum - corr_xsum * corr_xsum) /
2776
(values_cnt * corr_x2sum - corr_xsum * corr_xsum);
2778
stats->stakind[slot_idx] = STATISTIC_KIND_CORRELATION;
2779
stats->staop[slot_idx] = mystats->ltopr;
2780
stats->stanumbers[slot_idx] = corrs;
2781
stats->numnumbers[slot_idx] = 1;
2785
else if (nonnull_cnt > 0)
2787
/* We found some non-null values, but they were all too wide */
2788
Assert(nonnull_cnt == toowide_cnt);
2789
stats->stats_valid = true;
2790
/* Do the simple null-frac and width stats */
2791
stats->stanullfrac = (double) null_cnt / (double) samplerows;
2793
stats->stawidth = total_width / (double) nonnull_cnt;
2795
stats->stawidth = stats->attrtype->typlen;
2796
/* Assume all too-wide values are distinct, so it's a unique column */
2797
stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2799
else if (null_cnt > 0)
2801
/* We found only nulls; assume the column is entirely null */
2802
stats->stats_valid = true;
2803
stats->stanullfrac = 1.0;
2805
stats->stawidth = 0; /* "unknown" */
2807
stats->stawidth = stats->attrtype->typlen;
2808
stats->stadistinct = 0.0; /* "unknown" */
2811
/* We don't need to bother cleaning up any of our temporary palloc's */
2815
* qsort_arg comparator for sorting ScalarItems
2817
* Aside from sorting the items, we update the tupnoLink[] array
2818
* whenever two ScalarItems are found to contain equal datums. The array
2819
* is indexed by tupno; for each ScalarItem, it contains the highest
2820
* tupno that that item's datum has been found to be equal to. This allows
2821
* us to avoid additional comparisons in compute_scalar_stats().
2824
compare_scalars(const void *a, const void *b, void *arg)
2826
Datum da = ((const ScalarItem *) a)->value;
2827
int ta = ((const ScalarItem *) a)->tupno;
2828
Datum db = ((const ScalarItem *) b)->value;
2829
int tb = ((const ScalarItem *) b)->tupno;
2830
CompareScalarsContext *cxt = (CompareScalarsContext *) arg;
2833
compare = ApplySortComparator(da, false, db, false, cxt->ssup);
2838
* The two datums are equal, so update cxt->tupnoLink[].
2840
if (cxt->tupnoLink[ta] < tb)
2841
cxt->tupnoLink[ta] = tb;
2842
if (cxt->tupnoLink[tb] < ta)
2843
cxt->tupnoLink[tb] = ta;
2846
* For equal datums, sort by tupno
2852
* qsort comparator for sorting ScalarMCVItems by position
2855
compare_mcvs(const void *a, const void *b)
2857
int da = ((const ScalarMCVItem *) a)->first;
2858
int db = ((const ScalarMCVItem *) b)->first;
2864
* Analyze the list of common values in the sample and decide how many are
2865
* worth storing in the table's MCV list.
2867
* mcv_counts is assumed to be a list of the counts of the most common values
2868
* seen in the sample, starting with the most common. The return value is the
2869
* number that are significantly more common than the values not in the list,
2870
* and which are therefore deemed worth storing in the table's MCV list.
2873
analyze_mcv_list(int *mcv_counts,
2880
double ndistinct_table;
2885
* If the entire table was sampled, keep the whole list. This also
2886
* protects us against division by zero in the code below.
2888
if (samplerows == totalrows || totalrows <= 1.0)
2891
/* Re-extract the estimated number of distinct nonnull values in table */
2892
ndistinct_table = stadistinct;
2893
if (ndistinct_table < 0)
2894
ndistinct_table = -ndistinct_table * totalrows;
2897
* Exclude the least common values from the MCV list, if they are not
2898
* significantly more common than the estimated selectivity they would
2899
* have if they weren't in the list. All non-MCV values are assumed to be
2900
* equally common, after taking into account the frequencies of all the
2901
* the values in the MCV list and the number of nulls (c.f. eqsel()).
2903
* Here sumcount tracks the total count of all but the last (least common)
2904
* value in the MCV list, allowing us to determine the effect of excluding
2905
* that value from the list.
2907
* Note that we deliberately do this by removing values from the full
2908
* list, rather than starting with an empty list and adding values,
2909
* because the latter approach can fail to add any values if all the most
2910
* common values have around the same frequency and make up the majority
2911
* of the table, so that the overall average frequency of all values is
2912
* roughly the same as that of the common values. This would lead to any
2913
* uncommon values being significantly overestimated.
2916
for (i = 0; i < num_mcv - 1; i++)
2917
sumcount += mcv_counts[i];
2930
* Estimated selectivity the least common value would have if it
2931
* wasn't in the MCV list (c.f. eqsel()).
2933
selec = 1.0 - sumcount / samplerows - stanullfrac;
2938
otherdistinct = ndistinct_table - (num_mcv - 1);
2939
if (otherdistinct > 1)
2940
selec /= otherdistinct;
2943
* If the value is kept in the MCV list, its population frequency is
2944
* assumed to equal its sample frequency. We use the lower end of a
2945
* textbook continuity-corrected Wald-type confidence interval to
2946
* determine if that is significantly more common than the non-MCV
2947
* frequency --- specifically we assume the population frequency is
2948
* highly likely to be within around 2 standard errors of the sample
2949
* frequency, which equates to an interval of 2 standard deviations
2950
* either side of the sample count, plus an additional 0.5 for the
2951
* continuity correction. Since we are sampling without replacement,
2952
* this is a hypergeometric distribution.
2954
* XXX: Empirically, this approach seems to work quite well, but it
2955
* may be worth considering more advanced techniques for estimating
2956
* the confidence interval of the hypergeometric distribution.
2960
K = N * mcv_counts[num_mcv - 1] / n;
2961
variance = n * K * (N - K) * (N - n) / (N * N * (N - 1));
2962
stddev = sqrt(variance);
2964
if (mcv_counts[num_mcv - 1] > selec * samplerows + 2 * stddev + 0.5)
2967
* The value is significantly more common than the non-MCV
2968
* selectivity would suggest. Keep it, and all the other more
2969
* common values in the list.
2975
/* Discard this value and consider the next least common value */
2979
sumcount -= mcv_counts[num_mcv - 1];
added
removed
src/backend/commands/analyze.c
