PostgreSQL Source Code  git master
analyze.c
Go to the documentation of this file.
1 /*-------------------------------------------------------------------------
2  *
3  * analyze.c
4  * the Postgres statistics generator
5  *
6  * Portions Copyright (c) 1996-2021, 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/sysattr.h"
24 #include "access/table.h"
25 #include "access/tableam.h"
26 #include "access/transam.h"
27 #include "access/tupconvert.h"
28 #include "access/visibilitymap.h"
29 #include "access/xact.h"
30 #include "catalog/catalog.h"
31 #include "catalog/index.h"
32 #include "catalog/indexing.h"
33 #include "catalog/pg_collation.h"
34 #include "catalog/pg_inherits.h"
35 #include "catalog/pg_namespace.h"
37 #include "commands/dbcommands.h"
38 #include "commands/progress.h"
39 #include "commands/tablecmds.h"
40 #include "commands/vacuum.h"
41 #include "executor/executor.h"
42 #include "foreign/fdwapi.h"
43 #include "miscadmin.h"
44 #include "nodes/nodeFuncs.h"
45 #include "parser/parse_oper.h"
46 #include "parser/parse_relation.h"
47 #include "pgstat.h"
48 #include "postmaster/autovacuum.h"
50 #include "statistics/statistics.h"
51 #include "storage/bufmgr.h"
52 #include "storage/lmgr.h"
53 #include "storage/proc.h"
54 #include "storage/procarray.h"
55 #include "utils/acl.h"
56 #include "utils/attoptcache.h"
57 #include "utils/builtins.h"
58 #include "utils/datum.h"
59 #include "utils/fmgroids.h"
60 #include "utils/guc.h"
61 #include "utils/lsyscache.h"
62 #include "utils/memutils.h"
63 #include "utils/pg_rusage.h"
64 #include "utils/sampling.h"
65 #include "utils/sortsupport.h"
66 #include "utils/spccache.h"
67 #include "utils/syscache.h"
68 #include "utils/timestamp.h"
69 
70 
71 /* Per-index data for ANALYZE */
72 typedef struct AnlIndexData
73 {
74  IndexInfo *indexInfo; /* BuildIndexInfo result */
75  double tupleFract; /* fraction of rows for partial index */
76  VacAttrStats **vacattrstats; /* index attrs to analyze */
77  int attr_cnt;
78 } AnlIndexData;
79 
80 
81 /* Default statistics target (GUC parameter) */
83 
84 /* A few variables that don't seem worth passing around as parameters */
85 static MemoryContext anl_context = NULL;
87 
88 
89 static void do_analyze_rel(Relation onerel,
90  VacuumParams *params, List *va_cols,
91  AcquireSampleRowsFunc acquirefunc, BlockNumber relpages,
92  bool inh, bool in_outer_xact, int elevel);
93 static void compute_index_stats(Relation onerel, double totalrows,
94  AnlIndexData *indexdata, int nindexes,
95  HeapTuple *rows, int numrows,
96  MemoryContext col_context);
97 static VacAttrStats *examine_attribute(Relation onerel, int attnum,
98  Node *index_expr);
99 static int acquire_sample_rows(Relation onerel, int elevel,
100  HeapTuple *rows, int targrows,
101  double *totalrows, double *totaldeadrows);
102 static int compare_rows(const void *a, const void *b);
103 static int acquire_inherited_sample_rows(Relation onerel, int elevel,
104  HeapTuple *rows, int targrows,
105  double *totalrows, double *totaldeadrows);
106 static void update_attstats(Oid relid, bool inh,
107  int natts, VacAttrStats **vacattrstats);
108 static Datum std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
109 static Datum ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
110 
111 
112 /*
113  * analyze_rel() -- analyze one relation
114  *
115  * relid identifies the relation to analyze. If relation is supplied, use
116  * the name therein for reporting any failure to open/lock the rel; do not
117  * use it once we've successfully opened the rel, since it might be stale.
118  */
119 void
120 analyze_rel(Oid relid, RangeVar *relation,
121  VacuumParams *params, List *va_cols, bool in_outer_xact,
122  BufferAccessStrategy bstrategy)
123 {
124  Relation onerel;
125  int elevel;
126  AcquireSampleRowsFunc acquirefunc = NULL;
127  BlockNumber relpages = 0;
128 
129  /* Select logging level */
130  if (params->options & VACOPT_VERBOSE)
131  elevel = INFO;
132  else
133  elevel = DEBUG2;
134 
135  /* Set up static variables */
136  vac_strategy = bstrategy;
137 
138  /*
139  * Check for user-requested abort.
140  */
142 
143  /*
144  * Open the relation, getting ShareUpdateExclusiveLock to ensure that two
145  * ANALYZEs don't run on it concurrently. (This also locks out a
146  * concurrent VACUUM, which doesn't matter much at the moment but might
147  * matter if we ever try to accumulate stats on dead tuples.) If the rel
148  * has been dropped since we last saw it, we don't need to process it.
149  *
150  * Make sure to generate only logs for ANALYZE in this case.
151  */
152  onerel = vacuum_open_relation(relid, relation, params->options & ~(VACOPT_VACUUM),
153  params->log_min_duration >= 0,
155 
156  /* leave if relation could not be opened or locked */
157  if (!onerel)
158  return;
159 
160  /*
161  * Check if relation needs to be skipped based on ownership. This check
162  * happens also when building the relation list to analyze for a manual
163  * operation, and needs to be done additionally here as ANALYZE could
164  * happen across multiple transactions where relation ownership could have
165  * changed in-between. Make sure to generate only logs for ANALYZE in
166  * this case.
167  */
169  onerel->rd_rel,
170  params->options & VACOPT_ANALYZE))
171  {
173  return;
174  }
175 
176  /*
177  * Silently ignore tables that are temp tables of other backends ---
178  * trying to analyze these is rather pointless, since their contents are
179  * probably not up-to-date on disk. (We don't throw a warning here; it
180  * would just lead to chatter during a database-wide ANALYZE.)
181  */
182  if (RELATION_IS_OTHER_TEMP(onerel))
183  {
185  return;
186  }
187 
188  /*
189  * We can ANALYZE any table except pg_statistic. See update_attstats
190  */
191  if (RelationGetRelid(onerel) == StatisticRelationId)
192  {
194  return;
195  }
196 
197  /*
198  * Check that it's of an analyzable relkind, and set up appropriately.
199  */
200  if (onerel->rd_rel->relkind == RELKIND_RELATION ||
201  onerel->rd_rel->relkind == RELKIND_MATVIEW)
202  {
203  /* Regular table, so we'll use the regular row acquisition function */
204  acquirefunc = acquire_sample_rows;
205  /* Also get regular table's size */
206  relpages = RelationGetNumberOfBlocks(onerel);
207  }
208  else if (onerel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
209  {
210  /*
211  * For a foreign table, call the FDW's hook function to see whether it
212  * supports analysis.
213  */
214  FdwRoutine *fdwroutine;
215  bool ok = false;
216 
217  fdwroutine = GetFdwRoutineForRelation(onerel, false);
218 
219  if (fdwroutine->AnalyzeForeignTable != NULL)
220  ok = fdwroutine->AnalyzeForeignTable(onerel,
221  &acquirefunc,
222  &relpages);
223 
224  if (!ok)
225  {
227  (errmsg("skipping \"%s\" --- cannot analyze this foreign table",
228  RelationGetRelationName(onerel))));
230  return;
231  }
232  }
233  else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
234  {
235  /*
236  * For partitioned tables, we want to do the recursive ANALYZE below.
237  */
238  }
239  else
240  {
241  /* No need for a WARNING if we already complained during VACUUM */
242  if (!(params->options & VACOPT_VACUUM))
244  (errmsg("skipping \"%s\" --- cannot analyze non-tables or special system tables",
245  RelationGetRelationName(onerel))));
247  return;
248  }
249 
250  /*
251  * OK, let's do it. First, initialize progress reporting.
252  */
254  RelationGetRelid(onerel));
255 
256  /*
257  * Do the normal non-recursive ANALYZE. We can skip this for partitioned
258  * tables, which don't contain any rows.
259  */
260  if (onerel->rd_rel->relkind != RELKIND_PARTITIONED_TABLE)
261  do_analyze_rel(onerel, params, va_cols, acquirefunc,
262  relpages, false, in_outer_xact, elevel);
263 
264  /*
265  * If there are child tables, do recursive ANALYZE.
266  */
267  if (onerel->rd_rel->relhassubclass)
268  do_analyze_rel(onerel, params, va_cols, acquirefunc, relpages,
269  true, in_outer_xact, elevel);
270 
271  /*
272  * Close source relation now, but keep lock so that no one deletes it
273  * before we commit. (If someone did, they'd fail to clean up the entries
274  * we made in pg_statistic. Also, releasing the lock before commit would
275  * expose us to concurrent-update failures in update_attstats.)
276  */
277  relation_close(onerel, NoLock);
278 
280 }
281 
282 /*
283  * do_analyze_rel() -- analyze one relation, recursively or not
284  *
285  * Note that "acquirefunc" is only relevant for the non-inherited case.
286  * For the inherited case, acquire_inherited_sample_rows() determines the
287  * appropriate acquirefunc for each child table.
288  */
289 static void
291  List *va_cols, AcquireSampleRowsFunc acquirefunc,
292  BlockNumber relpages, bool inh, bool in_outer_xact,
293  int elevel)
294 {
295  int attr_cnt,
296  tcnt,
297  i,
298  ind;
299  Relation *Irel;
300  int nindexes;
301  bool hasindex;
303  AnlIndexData *indexdata;
304  int targrows,
305  numrows,
306  minrows;
307  double totalrows,
308  totaldeadrows;
309  HeapTuple *rows;
310  PGRUsage ru0;
311  TimestampTz starttime = 0;
312  MemoryContext caller_context;
313  Oid save_userid;
314  int save_sec_context;
315  int save_nestlevel;
316  int64 AnalyzePageHit = VacuumPageHit;
317  int64 AnalyzePageMiss = VacuumPageMiss;
318  int64 AnalyzePageDirty = VacuumPageDirty;
319  PgStat_Counter startreadtime = 0;
320  PgStat_Counter startwritetime = 0;
321 
322  if (inh)
323  ereport(elevel,
324  (errmsg("analyzing \"%s.%s\" inheritance tree",
326  RelationGetRelationName(onerel))));
327  else
328  ereport(elevel,
329  (errmsg("analyzing \"%s.%s\"",
331  RelationGetRelationName(onerel))));
332 
333  /*
334  * Set up a working context so that we can easily free whatever junk gets
335  * created.
336  */
338  "Analyze",
340  caller_context = MemoryContextSwitchTo(anl_context);
341 
342  /*
343  * Switch to the table owner's userid, so that any index functions are run
344  * as that user. Also lock down security-restricted operations and
345  * arrange to make GUC variable changes local to this command.
346  */
347  GetUserIdAndSecContext(&save_userid, &save_sec_context);
348  SetUserIdAndSecContext(onerel->rd_rel->relowner,
349  save_sec_context | SECURITY_RESTRICTED_OPERATION);
350  save_nestlevel = NewGUCNestLevel();
351 
352  /* measure elapsed time iff autovacuum logging requires it */
353  if (IsAutoVacuumWorkerProcess() && params->log_min_duration >= 0)
354  {
355  if (track_io_timing)
356  {
357  startreadtime = pgStatBlockReadTime;
358  startwritetime = pgStatBlockWriteTime;
359  }
360 
361  pg_rusage_init(&ru0);
362  if (params->log_min_duration >= 0)
363  starttime = GetCurrentTimestamp();
364  }
365 
366  /*
367  * Determine which columns to analyze
368  *
369  * Note that system attributes are never analyzed, so we just reject them
370  * at the lookup stage. We also reject duplicate column mentions. (We
371  * could alternatively ignore duplicates, but analyzing a column twice
372  * won't work; we'd end up making a conflicting update in pg_statistic.)
373  */
374  if (va_cols != NIL)
375  {
376  Bitmapset *unique_cols = NULL;
377  ListCell *le;
378 
379  vacattrstats = (VacAttrStats **) palloc(list_length(va_cols) *
380  sizeof(VacAttrStats *));
381  tcnt = 0;
382  foreach(le, va_cols)
383  {
384  char *col = strVal(lfirst(le));
385 
386  i = attnameAttNum(onerel, col, false);
387  if (i == InvalidAttrNumber)
388  ereport(ERROR,
389  (errcode(ERRCODE_UNDEFINED_COLUMN),
390  errmsg("column \"%s\" of relation \"%s\" does not exist",
391  col, RelationGetRelationName(onerel))));
392  if (bms_is_member(i, unique_cols))
393  ereport(ERROR,
394  (errcode(ERRCODE_DUPLICATE_COLUMN),
395  errmsg("column \"%s\" of relation \"%s\" appears more than once",
396  col, RelationGetRelationName(onerel))));
397  unique_cols = bms_add_member(unique_cols, i);
398 
399  vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
400  if (vacattrstats[tcnt] != NULL)
401  tcnt++;
402  }
403  attr_cnt = tcnt;
404  }
405  else
406  {
407  attr_cnt = onerel->rd_att->natts;
408  vacattrstats = (VacAttrStats **)
409  palloc(attr_cnt * sizeof(VacAttrStats *));
410  tcnt = 0;
411  for (i = 1; i <= attr_cnt; i++)
412  {
413  vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
414  if (vacattrstats[tcnt] != NULL)
415  tcnt++;
416  }
417  attr_cnt = tcnt;
418  }
419 
420  /*
421  * Open all indexes of the relation, and see if there are any analyzable
422  * columns in the indexes. We do not analyze index columns if there was
423  * an explicit column list in the ANALYZE command, however.
424  *
425  * If we are doing a recursive scan, we don't want to touch the parent's
426  * indexes at all. If we're processing a partitioned table, we need to
427  * know if there are any indexes, but we don't want to process them.
428  */
429  if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
430  {
431  List *idxs = RelationGetIndexList(onerel);
432 
433  Irel = NULL;
434  nindexes = 0;
435  hasindex = idxs != NIL;
436  list_free(idxs);
437  }
438  else if (!inh)
439  {
440  vac_open_indexes(onerel, AccessShareLock, &nindexes, &Irel);
441  hasindex = nindexes > 0;
442  }
443  else
444  {
445  Irel = NULL;
446  nindexes = 0;
447  hasindex = false;
448  }
449  indexdata = NULL;
450  if (nindexes > 0)
451  {
452  indexdata = (AnlIndexData *) palloc0(nindexes * sizeof(AnlIndexData));
453  for (ind = 0; ind < nindexes; ind++)
454  {
455  AnlIndexData *thisdata = &indexdata[ind];
457 
458  thisdata->indexInfo = indexInfo = BuildIndexInfo(Irel[ind]);
459  thisdata->tupleFract = 1.0; /* fix later if partial */
460  if (indexInfo->ii_Expressions != NIL && va_cols == NIL)
461  {
462  ListCell *indexpr_item = list_head(indexInfo->ii_Expressions);
463 
464  thisdata->vacattrstats = (VacAttrStats **)
465  palloc(indexInfo->ii_NumIndexAttrs * sizeof(VacAttrStats *));
466  tcnt = 0;
467  for (i = 0; i < indexInfo->ii_NumIndexAttrs; i++)
468  {
469  int keycol = indexInfo->ii_IndexAttrNumbers[i];
470 
471  if (keycol == 0)
472  {
473  /* Found an index expression */
474  Node *indexkey;
475 
476  if (indexpr_item == NULL) /* shouldn't happen */
477  elog(ERROR, "too few entries in indexprs list");
478  indexkey = (Node *) lfirst(indexpr_item);
479  indexpr_item = lnext(indexInfo->ii_Expressions,
480  indexpr_item);
481  thisdata->vacattrstats[tcnt] =
482  examine_attribute(Irel[ind], i + 1, indexkey);
483  if (thisdata->vacattrstats[tcnt] != NULL)
484  tcnt++;
485  }
486  }
487  thisdata->attr_cnt = tcnt;
488  }
489  }
490  }
491 
492  /*
493  * Determine how many rows we need to sample, using the worst case from
494  * all analyzable columns. We use a lower bound of 100 rows to avoid
495  * possible overflow in Vitter's algorithm. (Note: that will also be the
496  * target in the corner case where there are no analyzable columns.)
497  */
498  targrows = 100;
499  for (i = 0; i < attr_cnt; i++)
500  {
501  if (targrows < vacattrstats[i]->minrows)
502  targrows = vacattrstats[i]->minrows;
503  }
504  for (ind = 0; ind < nindexes; ind++)
505  {
506  AnlIndexData *thisdata = &indexdata[ind];
507 
508  for (i = 0; i < thisdata->attr_cnt; i++)
509  {
510  if (targrows < thisdata->vacattrstats[i]->minrows)
511  targrows = thisdata->vacattrstats[i]->minrows;
512  }
513  }
514 
515  /*
516  * Look at extended statistics objects too, as those may define custom
517  * statistics target. So we may need to sample more rows and then build
518  * the statistics with enough detail.
519  */
520  minrows = ComputeExtStatisticsRows(onerel, attr_cnt, vacattrstats);
521 
522  if (targrows < minrows)
523  targrows = minrows;
524 
525  /*
526  * Acquire the sample rows
527  */
528  rows = (HeapTuple *) palloc(targrows * sizeof(HeapTuple));
532  if (inh)
533  numrows = acquire_inherited_sample_rows(onerel, elevel,
534  rows, targrows,
535  &totalrows, &totaldeadrows);
536  else
537  numrows = (*acquirefunc) (onerel, elevel,
538  rows, targrows,
539  &totalrows, &totaldeadrows);
540 
541  /*
542  * Compute the statistics. Temporary results during the calculations for
543  * each column are stored in a child context. The calc routines are
544  * responsible to make sure that whatever they store into the VacAttrStats
545  * structure is allocated in anl_context.
546  */
547  if (numrows > 0)
548  {
549  MemoryContext col_context,
550  old_context;
551 
554 
555  col_context = AllocSetContextCreate(anl_context,
556  "Analyze Column",
558  old_context = MemoryContextSwitchTo(col_context);
559 
560  for (i = 0; i < attr_cnt; i++)
561  {
562  VacAttrStats *stats = vacattrstats[i];
563  AttributeOpts *aopt;
564 
565  stats->rows = rows;
566  stats->tupDesc = onerel->rd_att;
567  stats->compute_stats(stats,
569  numrows,
570  totalrows);
571 
572  /*
573  * If the appropriate flavor of the n_distinct option is
574  * specified, override with the corresponding value.
575  */
576  aopt = get_attribute_options(onerel->rd_id, stats->attr->attnum);
577  if (aopt != NULL)
578  {
579  float8 n_distinct;
580 
581  n_distinct = inh ? aopt->n_distinct_inherited : aopt->n_distinct;
582  if (n_distinct != 0.0)
583  stats->stadistinct = n_distinct;
584  }
585 
587  }
588 
589  if (nindexes > 0)
590  compute_index_stats(onerel, totalrows,
591  indexdata, nindexes,
592  rows, numrows,
593  col_context);
594 
595  MemoryContextSwitchTo(old_context);
596  MemoryContextDelete(col_context);
597 
598  /*
599  * Emit the completed stats rows into pg_statistic, replacing any
600  * previous statistics for the target columns. (If there are stats in
601  * pg_statistic for columns we didn't process, we leave them alone.)
602  */
603  update_attstats(RelationGetRelid(onerel), inh,
604  attr_cnt, vacattrstats);
605 
606  for (ind = 0; ind < nindexes; ind++)
607  {
608  AnlIndexData *thisdata = &indexdata[ind];
609 
610  update_attstats(RelationGetRelid(Irel[ind]), false,
611  thisdata->attr_cnt, thisdata->vacattrstats);
612  }
613 
614  /*
615  * Build extended statistics (if there are any).
616  *
617  * For now we only build extended statistics on individual relations,
618  * not for relations representing inheritance trees.
619  */
620  if (!inh)
621  BuildRelationExtStatistics(onerel, totalrows, numrows, rows,
622  attr_cnt, vacattrstats);
623  }
624 
627 
628  /*
629  * Update pages/tuples stats in pg_class ... but not if we're doing
630  * inherited stats.
631  *
632  * We assume that VACUUM hasn't set pg_class.reltuples already, even
633  * during a VACUUM ANALYZE. Although VACUUM often updates pg_class,
634  * exceptions exist. A "VACUUM (ANALYZE, INDEX_CLEANUP OFF)" command will
635  * never update pg_class entries for index relations. It's also possible
636  * that an individual index's pg_class entry won't be updated during
637  * VACUUM if the index AM returns NULL from its amvacuumcleanup() routine.
638  */
639  if (!inh)
640  {
641  BlockNumber relallvisible;
642 
643  visibilitymap_count(onerel, &relallvisible, NULL);
644 
645  /* Update pg_class for table relation */
646  vac_update_relstats(onerel,
647  relpages,
648  totalrows,
649  relallvisible,
650  hasindex,
653  in_outer_xact);
654 
655  /* Same for indexes */
656  for (ind = 0; ind < nindexes; ind++)
657  {
658  AnlIndexData *thisdata = &indexdata[ind];
659  double totalindexrows;
660 
661  totalindexrows = ceil(thisdata->tupleFract * totalrows);
662  vac_update_relstats(Irel[ind],
663  RelationGetNumberOfBlocks(Irel[ind]),
664  totalindexrows,
665  0,
666  false,
669  in_outer_xact);
670  }
671  }
672  else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
673  {
674  /*
675  * Partitioned tables don't have storage, so we don't set any fields
676  * in their pg_class entries except for reltuples and relhasindex.
677  */
678  vac_update_relstats(onerel, -1, totalrows,
679  0, hasindex, InvalidTransactionId,
681  in_outer_xact);
682  }
683 
684  /*
685  * Now report ANALYZE to the stats collector. For regular tables, we do
686  * it only if not doing inherited stats. For partitioned tables, we only
687  * do it for inherited stats. (We're never called for not-inherited stats
688  * on partitioned tables anyway.)
689  *
690  * Reset the changes_since_analyze counter only if we analyzed all
691  * columns; otherwise, there is still work for auto-analyze to do.
692  */
693  if (!inh)
694  pgstat_report_analyze(onerel, totalrows, totaldeadrows,
695  (va_cols == NIL));
696  else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
697  pgstat_report_analyze(onerel, 0, 0, (va_cols == NIL));
698 
699  /*
700  * If this isn't part of VACUUM ANALYZE, let index AMs do cleanup.
701  *
702  * Note that most index AMs perform a no-op as a matter of policy for
703  * amvacuumcleanup() when called in ANALYZE-only mode. The only exception
704  * among core index AMs is GIN/ginvacuumcleanup().
705  */
706  if (!(params->options & VACOPT_VACUUM))
707  {
708  for (ind = 0; ind < nindexes; ind++)
709  {
710  IndexBulkDeleteResult *stats;
711  IndexVacuumInfo ivinfo;
712 
713  ivinfo.index = Irel[ind];
714  ivinfo.analyze_only = true;
715  ivinfo.estimated_count = true;
716  ivinfo.message_level = elevel;
717  ivinfo.num_heap_tuples = onerel->rd_rel->reltuples;
718  ivinfo.strategy = vac_strategy;
719 
720  stats = index_vacuum_cleanup(&ivinfo, NULL);
721 
722  if (stats)
723  pfree(stats);
724  }
725  }
726 
727  /* Done with indexes */
728  vac_close_indexes(nindexes, Irel, NoLock);
729 
730  /* Log the action if appropriate */
731  if (IsAutoVacuumWorkerProcess() && params->log_min_duration >= 0)
732  {
733  TimestampTz endtime = GetCurrentTimestamp();
734 
735  if (params->log_min_duration == 0 ||
736  TimestampDifferenceExceeds(starttime, endtime,
737  params->log_min_duration))
738  {
739  long delay_in_ms;
740  double read_rate = 0;
741  double write_rate = 0;
743 
744  /*
745  * Calculate the difference in the Page Hit/Miss/Dirty that
746  * happened as part of the analyze by subtracting out the
747  * pre-analyze values which we saved above.
748  */
749  AnalyzePageHit = VacuumPageHit - AnalyzePageHit;
750  AnalyzePageMiss = VacuumPageMiss - AnalyzePageMiss;
751  AnalyzePageDirty = VacuumPageDirty - AnalyzePageDirty;
752 
753  /*
754  * We do not expect an analyze to take > 25 days and it simplifies
755  * things a bit to use TimestampDifferenceMilliseconds.
756  */
757  delay_in_ms = TimestampDifferenceMilliseconds(starttime, endtime);
758 
759  /*
760  * Note that we are reporting these read/write rates in the same
761  * manner as VACUUM does, which means that while the 'average read
762  * rate' here actually corresponds to page misses and resulting
763  * reads which are also picked up by track_io_timing, if enabled,
764  * the 'average write rate' is actually talking about the rate of
765  * pages being dirtied, not being written out, so it's typical to
766  * have a non-zero 'avg write rate' while I/O timings only reports
767  * reads.
768  *
769  * It's not clear that an ANALYZE will ever result in
770  * FlushBuffer() being called, but we track and support reporting
771  * on I/O write time in case that changes as it's practically free
772  * to do so anyway.
773  */
774 
775  if (delay_in_ms > 0)
776  {
777  read_rate = (double) BLCKSZ * AnalyzePageMiss / (1024 * 1024) /
778  (delay_in_ms / 1000.0);
779  write_rate = (double) BLCKSZ * AnalyzePageDirty / (1024 * 1024) /
780  (delay_in_ms / 1000.0);
781  }
782 
783  /*
784  * We split this up so we don't emit empty I/O timing values when
785  * track_io_timing isn't enabled.
786  */
787 
788  initStringInfo(&buf);
789  appendStringInfo(&buf, _("automatic analyze of table \"%s.%s.%s\"\n"),
792  RelationGetRelationName(onerel));
793  if (track_io_timing)
794  {
795  double read_ms = (double) (pgStatBlockReadTime - startreadtime) / 1000;
796  double write_ms = (double) (pgStatBlockWriteTime - startwritetime) / 1000;
797 
798  appendStringInfo(&buf, _("I/O timings: read: %.3f ms, write: %.3f ms\n"),
799  read_ms, write_ms);
800  }
801  appendStringInfo(&buf, _("avg read rate: %.3f MB/s, avg write rate: %.3f MB/s\n"),
802  read_rate, write_rate);
803  appendStringInfo(&buf, _("buffer usage: %lld hits, %lld misses, %lld dirtied\n"),
804  (long long) AnalyzePageHit,
805  (long long) AnalyzePageMiss,
806  (long long) AnalyzePageDirty);
807  appendStringInfo(&buf, _("system usage: %s"), pg_rusage_show(&ru0));
808 
809  ereport(LOG,
810  (errmsg_internal("%s", buf.data)));
811 
812  pfree(buf.data);
813  }
814  }
815 
816  /* Roll back any GUC changes executed by index functions */
817  AtEOXact_GUC(false, save_nestlevel);
818 
819  /* Restore userid and security context */
820  SetUserIdAndSecContext(save_userid, save_sec_context);
821 
822  /* Restore current context and release memory */
823  MemoryContextSwitchTo(caller_context);
824  MemoryContextDelete(anl_context);
825  anl_context = NULL;
826 }
827 
828 /*
829  * Compute statistics about indexes of a relation
830  */
831 static void
832 compute_index_stats(Relation onerel, double totalrows,
833  AnlIndexData *indexdata, int nindexes,
834  HeapTuple *rows, int numrows,
835  MemoryContext col_context)
836 {
837  MemoryContext ind_context,
838  old_context;
840  bool isnull[INDEX_MAX_KEYS];
841  int ind,
842  i;
843 
844  ind_context = AllocSetContextCreate(anl_context,
845  "Analyze Index",
847  old_context = MemoryContextSwitchTo(ind_context);
848 
849  for (ind = 0; ind < nindexes; ind++)
850  {
851  AnlIndexData *thisdata = &indexdata[ind];
852  IndexInfo *indexInfo = thisdata->indexInfo;
853  int attr_cnt = thisdata->attr_cnt;
854  TupleTableSlot *slot;
855  EState *estate;
856  ExprContext *econtext;
857  ExprState *predicate;
858  Datum *exprvals;
859  bool *exprnulls;
860  int numindexrows,
861  tcnt,
862  rowno;
863  double totalindexrows;
864 
865  /* Ignore index if no columns to analyze and not partial */
866  if (attr_cnt == 0 && indexInfo->ii_Predicate == NIL)
867  continue;
868 
869  /*
870  * Need an EState for evaluation of index expressions and
871  * partial-index predicates. Create it in the per-index context to be
872  * sure it gets cleaned up at the bottom of the loop.
873  */
874  estate = CreateExecutorState();
875  econtext = GetPerTupleExprContext(estate);
876  /* Need a slot to hold the current heap tuple, too */
878  &TTSOpsHeapTuple);
879 
880  /* Arrange for econtext's scan tuple to be the tuple under test */
881  econtext->ecxt_scantuple = slot;
882 
883  /* Set up execution state for predicate. */
884  predicate = ExecPrepareQual(indexInfo->ii_Predicate, estate);
885 
886  /* Compute and save index expression values */
887  exprvals = (Datum *) palloc(numrows * attr_cnt * sizeof(Datum));
888  exprnulls = (bool *) palloc(numrows * attr_cnt * sizeof(bool));
889  numindexrows = 0;
890  tcnt = 0;
891  for (rowno = 0; rowno < numrows; rowno++)
892  {
893  HeapTuple heapTuple = rows[rowno];
894 
896 
897  /*
898  * Reset the per-tuple context each time, to reclaim any cruft
899  * left behind by evaluating the predicate or index expressions.
900  */
901  ResetExprContext(econtext);
902 
903  /* Set up for predicate or expression evaluation */
904  ExecStoreHeapTuple(heapTuple, slot, false);
905 
906  /* If index is partial, check predicate */
907  if (predicate != NULL)
908  {
909  if (!ExecQual(predicate, econtext))
910  continue;
911  }
912  numindexrows++;
913 
914  if (attr_cnt > 0)
915  {
916  /*
917  * Evaluate the index row to compute expression values. We
918  * could do this by hand, but FormIndexDatum is convenient.
919  */
920  FormIndexDatum(indexInfo,
921  slot,
922  estate,
923  values,
924  isnull);
925 
926  /*
927  * Save just the columns we care about. We copy the values
928  * into ind_context from the estate's per-tuple context.
929  */
930  for (i = 0; i < attr_cnt; i++)
931  {
932  VacAttrStats *stats = thisdata->vacattrstats[i];
933  int attnum = stats->attr->attnum;
934 
935  if (isnull[attnum - 1])
936  {
937  exprvals[tcnt] = (Datum) 0;
938  exprnulls[tcnt] = true;
939  }
940  else
941  {
942  exprvals[tcnt] = datumCopy(values[attnum - 1],
943  stats->attrtype->typbyval,
944  stats->attrtype->typlen);
945  exprnulls[tcnt] = false;
946  }
947  tcnt++;
948  }
949  }
950  }
951 
952  /*
953  * Having counted the number of rows that pass the predicate in the
954  * sample, we can estimate the total number of rows in the index.
955  */
956  thisdata->tupleFract = (double) numindexrows / (double) numrows;
957  totalindexrows = ceil(thisdata->tupleFract * totalrows);
958 
959  /*
960  * Now we can compute the statistics for the expression columns.
961  */
962  if (numindexrows > 0)
963  {
964  MemoryContextSwitchTo(col_context);
965  for (i = 0; i < attr_cnt; i++)
966  {
967  VacAttrStats *stats = thisdata->vacattrstats[i];
968  AttributeOpts *aopt =
969  get_attribute_options(stats->attr->attrelid,
970  stats->attr->attnum);
971 
972  stats->exprvals = exprvals + i;
973  stats->exprnulls = exprnulls + i;
974  stats->rowstride = attr_cnt;
975  stats->compute_stats(stats,
977  numindexrows,
978  totalindexrows);
979 
980  /*
981  * If the n_distinct option is specified, it overrides the
982  * above computation. For indices, we always use just
983  * n_distinct, not n_distinct_inherited.
984  */
985  if (aopt != NULL && aopt->n_distinct != 0.0)
986  stats->stadistinct = aopt->n_distinct;
987 
989  }
990  }
991 
992  /* And clean up */
993  MemoryContextSwitchTo(ind_context);
994 
996  FreeExecutorState(estate);
998  }
999 
1000  MemoryContextSwitchTo(old_context);
1001  MemoryContextDelete(ind_context);
1002 }
1003 
1004 /*
1005  * examine_attribute -- pre-analysis of a single column
1006  *
1007  * Determine whether the column is analyzable; if so, create and initialize
1008  * a VacAttrStats struct for it. If not, return NULL.
1009  *
1010  * If index_expr isn't NULL, then we're trying to analyze an expression index,
1011  * and index_expr is the expression tree representing the column's data.
1012  */
1013 static VacAttrStats *
1014 examine_attribute(Relation onerel, int attnum, Node *index_expr)
1015 {
1016  Form_pg_attribute attr = TupleDescAttr(onerel->rd_att, attnum - 1);
1017  HeapTuple typtuple;
1018  VacAttrStats *stats;
1019  int i;
1020  bool ok;
1021 
1022  /* Never analyze dropped columns */
1023  if (attr->attisdropped)
1024  return NULL;
1025 
1026  /* Don't analyze column if user has specified not to */
1027  if (attr->attstattarget == 0)
1028  return NULL;
1029 
1030  /*
1031  * Create the VacAttrStats struct. Note that we only have a copy of the
1032  * fixed fields of the pg_attribute tuple.
1033  */
1034  stats = (VacAttrStats *) palloc0(sizeof(VacAttrStats));
1036  memcpy(stats->attr, attr, ATTRIBUTE_FIXED_PART_SIZE);
1037 
1038  /*
1039  * When analyzing an expression index, believe the expression tree's type
1040  * not the column datatype --- the latter might be the opckeytype storage
1041  * type of the opclass, which is not interesting for our purposes. (Note:
1042  * if we did anything with non-expression index columns, we'd need to
1043  * figure out where to get the correct type info from, but for now that's
1044  * not a problem.) It's not clear whether anyone will care about the
1045  * typmod, but we store that too just in case.
1046  */
1047  if (index_expr)
1048  {
1049  stats->attrtypid = exprType(index_expr);
1050  stats->attrtypmod = exprTypmod(index_expr);
1051 
1052  /*
1053  * If a collation has been specified for the index column, use that in
1054  * preference to anything else; but if not, fall back to whatever we
1055  * can get from the expression.
1056  */
1057  if (OidIsValid(onerel->rd_indcollation[attnum - 1]))
1058  stats->attrcollid = onerel->rd_indcollation[attnum - 1];
1059  else
1060  stats->attrcollid = exprCollation(index_expr);
1061  }
1062  else
1063  {
1064  stats->attrtypid = attr->atttypid;
1065  stats->attrtypmod = attr->atttypmod;
1066  stats->attrcollid = attr->attcollation;
1067  }
1068 
1069  typtuple = SearchSysCacheCopy1(TYPEOID,
1070  ObjectIdGetDatum(stats->attrtypid));
1071  if (!HeapTupleIsValid(typtuple))
1072  elog(ERROR, "cache lookup failed for type %u", stats->attrtypid);
1073  stats->attrtype = (Form_pg_type) GETSTRUCT(typtuple);
1074  stats->anl_context = anl_context;
1075  stats->tupattnum = attnum;
1076 
1077  /*
1078  * The fields describing the stats->stavalues[n] element types default to
1079  * the type of the data being analyzed, but the type-specific typanalyze
1080  * function can change them if it wants to store something else.
1081  */
1082  for (i = 0; i < STATISTIC_NUM_SLOTS; i++)
1083  {
1084  stats->statypid[i] = stats->attrtypid;
1085  stats->statyplen[i] = stats->attrtype->typlen;
1086  stats->statypbyval[i] = stats->attrtype->typbyval;
1087  stats->statypalign[i] = stats->attrtype->typalign;
1088  }
1089 
1090  /*
1091  * Call the type-specific typanalyze function. If none is specified, use
1092  * std_typanalyze().
1093  */
1094  if (OidIsValid(stats->attrtype->typanalyze))
1095  ok = DatumGetBool(OidFunctionCall1(stats->attrtype->typanalyze,
1096  PointerGetDatum(stats)));
1097  else
1098  ok = std_typanalyze(stats);
1099 
1100  if (!ok || stats->compute_stats == NULL || stats->minrows <= 0)
1101  {
1102  heap_freetuple(typtuple);
1103  pfree(stats->attr);
1104  pfree(stats);
1105  return NULL;
1106  }
1107 
1108  return stats;
1109 }
1110 
1111 /*
1112  * acquire_sample_rows -- acquire a random sample of rows from the table
1113  *
1114  * Selected rows are returned in the caller-allocated array rows[], which
1115  * must have at least targrows entries.
1116  * The actual number of rows selected is returned as the function result.
1117  * We also estimate the total numbers of live and dead rows in the table,
1118  * and return them into *totalrows and *totaldeadrows, respectively.
1119  *
1120  * The returned list of tuples is in order by physical position in the table.
1121  * (We will rely on this later to derive correlation estimates.)
1122  *
1123  * As of May 2004 we use a new two-stage method: Stage one selects up
1124  * to targrows random blocks (or all blocks, if there aren't so many).
1125  * Stage two scans these blocks and uses the Vitter algorithm to create
1126  * a random sample of targrows rows (or less, if there are less in the
1127  * sample of blocks). The two stages are executed simultaneously: each
1128  * block is processed as soon as stage one returns its number and while
1129  * the rows are read stage two controls which ones are to be inserted
1130  * into the sample.
1131  *
1132  * Although every row has an equal chance of ending up in the final
1133  * sample, this sampling method is not perfect: not every possible
1134  * sample has an equal chance of being selected. For large relations
1135  * the number of different blocks represented by the sample tends to be
1136  * too small. We can live with that for now. Improvements are welcome.
1137  *
1138  * An important property of this sampling method is that because we do
1139  * look at a statistically unbiased set of blocks, we should get
1140  * unbiased estimates of the average numbers of live and dead rows per
1141  * block. The previous sampling method put too much credence in the row
1142  * density near the start of the table.
1143  */
1144 static int
1146  HeapTuple *rows, int targrows,
1147  double *totalrows, double *totaldeadrows)
1148 {
1149  int numrows = 0; /* # rows now in reservoir */
1150  double samplerows = 0; /* total # rows collected */
1151  double liverows = 0; /* # live rows seen */
1152  double deadrows = 0; /* # dead rows seen */
1153  double rowstoskip = -1; /* -1 means not set yet */
1154  long randseed; /* Seed for block sampler(s) */
1155  BlockNumber totalblocks;
1156  TransactionId OldestXmin;
1157  BlockSamplerData bs;
1158  ReservoirStateData rstate;
1159  TupleTableSlot *slot;
1160  TableScanDesc scan;
1161  BlockNumber nblocks;
1162  BlockNumber blksdone = 0;
1163 #ifdef USE_PREFETCH
1164  int prefetch_maximum = 0; /* blocks to prefetch if enabled */
1165  BlockSamplerData prefetch_bs;
1166 #endif
1167 
1168  Assert(targrows > 0);
1169 
1170  totalblocks = RelationGetNumberOfBlocks(onerel);
1171 
1172  /* Need a cutoff xmin for HeapTupleSatisfiesVacuum */
1173  OldestXmin = GetOldestNonRemovableTransactionId(onerel);
1174 
1175  /* Prepare for sampling block numbers */
1176  randseed = random();
1177  nblocks = BlockSampler_Init(&bs, totalblocks, targrows, randseed);
1178 
1179 #ifdef USE_PREFETCH
1180  prefetch_maximum = get_tablespace_maintenance_io_concurrency(onerel->rd_rel->reltablespace);
1181  /* Create another BlockSampler, using the same seed, for prefetching */
1182  if (prefetch_maximum)
1183  (void) BlockSampler_Init(&prefetch_bs, totalblocks, targrows, randseed);
1184 #endif
1185 
1186  /* Report sampling block numbers */
1188  nblocks);
1189 
1190  /* Prepare for sampling rows */
1191  reservoir_init_selection_state(&rstate, targrows);
1192 
1193  scan = table_beginscan_analyze(onerel);
1194  slot = table_slot_create(onerel, NULL);
1195 
1196 #ifdef USE_PREFETCH
1197 
1198  /*
1199  * If we are doing prefetching, then go ahead and tell the kernel about
1200  * the first set of pages we are going to want. This also moves our
1201  * iterator out ahead of the main one being used, where we will keep it so
1202  * that we're always pre-fetching out prefetch_maximum number of blocks
1203  * ahead.
1204  */
1205  if (prefetch_maximum)
1206  {
1207  for (int i = 0; i < prefetch_maximum; i++)
1208  {
1209  BlockNumber prefetch_block;
1210 
1211  if (!BlockSampler_HasMore(&prefetch_bs))
1212  break;
1213 
1214  prefetch_block = BlockSampler_Next(&prefetch_bs);
1215  PrefetchBuffer(scan->rs_rd, MAIN_FORKNUM, prefetch_block);
1216  }
1217  }
1218 #endif
1219 
1220  /* Outer loop over blocks to sample */
1221  while (BlockSampler_HasMore(&bs))
1222  {
1223  bool block_accepted;
1224  BlockNumber targblock = BlockSampler_Next(&bs);
1225 #ifdef USE_PREFETCH
1226  BlockNumber prefetch_targblock = InvalidBlockNumber;
1227 
1228  /*
1229  * Make sure that every time the main BlockSampler is moved forward
1230  * that our prefetch BlockSampler also gets moved forward, so that we
1231  * always stay out ahead.
1232  */
1233  if (prefetch_maximum && BlockSampler_HasMore(&prefetch_bs))
1234  prefetch_targblock = BlockSampler_Next(&prefetch_bs);
1235 #endif
1236 
1238 
1239  block_accepted = table_scan_analyze_next_block(scan, targblock, vac_strategy);
1240 
1241 #ifdef USE_PREFETCH
1242 
1243  /*
1244  * When pre-fetching, after we get a block, tell the kernel about the
1245  * next one we will want, if there's any left.
1246  *
1247  * We want to do this even if the table_scan_analyze_next_block() call
1248  * above decides against analyzing the block it picked.
1249  */
1250  if (prefetch_maximum && prefetch_targblock != InvalidBlockNumber)
1251  PrefetchBuffer(scan->rs_rd, MAIN_FORKNUM, prefetch_targblock);
1252 #endif
1253 
1254  /*
1255  * Don't analyze if table_scan_analyze_next_block() indicated this
1256  * block is unsuitable for analyzing.
1257  */
1258  if (!block_accepted)
1259  continue;
1260 
1261  while (table_scan_analyze_next_tuple(scan, OldestXmin, &liverows, &deadrows, slot))
1262  {
1263  /*
1264  * The first targrows sample rows are simply copied into the
1265  * reservoir. Then we start replacing tuples in the sample until
1266  * we reach the end of the relation. This algorithm is from Jeff
1267  * Vitter's paper (see full citation in utils/misc/sampling.c). It
1268  * works by repeatedly computing the number of tuples to skip
1269  * before selecting a tuple, which replaces a randomly chosen
1270  * element of the reservoir (current set of tuples). At all times
1271  * the reservoir is a true random sample of the tuples we've
1272  * passed over so far, so when we fall off the end of the relation
1273  * we're done.
1274  */
1275  if (numrows < targrows)
1276  rows[numrows++] = ExecCopySlotHeapTuple(slot);
1277  else
1278  {
1279  /*
1280  * t in Vitter's paper is the number of records already
1281  * processed. If we need to compute a new S value, we must
1282  * use the not-yet-incremented value of samplerows as t.
1283  */
1284  if (rowstoskip < 0)
1285  rowstoskip = reservoir_get_next_S(&rstate, samplerows, targrows);
1286 
1287  if (rowstoskip <= 0)
1288  {
1289  /*
1290  * Found a suitable tuple, so save it, replacing one old
1291  * tuple at random
1292  */
1293  int k = (int) (targrows * sampler_random_fract(rstate.randstate));
1294 
1295  Assert(k >= 0 && k < targrows);
1296  heap_freetuple(rows[k]);
1297  rows[k] = ExecCopySlotHeapTuple(slot);
1298  }
1299 
1300  rowstoskip -= 1;
1301  }
1302 
1303  samplerows += 1;
1304  }
1305 
1307  ++blksdone);
1308  }
1309 
1311  table_endscan(scan);
1312 
1313  /*
1314  * If we didn't find as many tuples as we wanted then we're done. No sort
1315  * is needed, since they're already in order.
1316  *
1317  * Otherwise we need to sort the collected tuples by position
1318  * (itempointer). It's not worth worrying about corner cases where the
1319  * tuples are already sorted.
1320  */
1321  if (numrows == targrows)
1322  qsort((void *) rows, numrows, sizeof(HeapTuple), compare_rows);
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  if (bs.m > 0)
1332  {
1333  *totalrows = floor((liverows / bs.m) * totalblocks + 0.5);
1334  *totaldeadrows = floor((deadrows / bs.m) * totalblocks + 0.5);
1335  }
1336  else
1337  {
1338  *totalrows = 0.0;
1339  *totaldeadrows = 0.0;
1340  }
1341 
1342  /*
1343  * Emit some interesting relation info
1344  */
1345  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  return numrows;
1355 }
1356 
1357 /*
1358  * qsort comparator for sorting rows[] array
1359  */
1360 static int
1361 compare_rows(const void *a, const void *b)
1362 {
1363  HeapTuple ha = *(const HeapTuple *) a;
1364  HeapTuple hb = *(const HeapTuple *) b;
1369 
1370  if (ba < bb)
1371  return -1;
1372  if (ba > bb)
1373  return 1;
1374  if (oa < ob)
1375  return -1;
1376  if (oa > ob)
1377  return 1;
1378  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
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  /*
1407  * Find all members of inheritance set. We only need AccessShareLock on
1408  * the children.
1409  */
1410  tableOIDs =
1412 
1413  /*
1414  * Check that there's at least one descendant, else fail. This could
1415  * happen despite analyze_rel's relhassubclass check, if table once had a
1416  * child but no longer does. In that case, we can clear the
1417  * relhassubclass field so as not to make the same mistake again later.
1418  * (This is safe because we hold ShareUpdateExclusiveLock.)
1419  */
1420  if (list_length(tableOIDs) < 2)
1421  {
1422  /* CCI because we already updated the pg_class row in this command */
1424  SetRelationHasSubclass(RelationGetRelid(onerel), false);
1425  ereport(elevel,
1426  (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no child tables",
1428  RelationGetRelationName(onerel))));
1429  return 0;
1430  }
1431 
1432  /*
1433  * Identify acquirefuncs to use, and count blocks in all the relations.
1434  * The result could overflow BlockNumber, so we use double arithmetic.
1435  */
1436  rels = (Relation *) palloc(list_length(tableOIDs) * sizeof(Relation));
1437  acquirefuncs = (AcquireSampleRowsFunc *)
1438  palloc(list_length(tableOIDs) * sizeof(AcquireSampleRowsFunc));
1439  relblocks = (double *) palloc(list_length(tableOIDs) * sizeof(double));
1440  totalblocks = 0;
1441  nrels = 0;
1442  has_child = false;
1443  foreach(lc, tableOIDs)
1444  {
1445  Oid childOID = lfirst_oid(lc);
1446  Relation childrel;
1447  AcquireSampleRowsFunc acquirefunc = NULL;
1448  BlockNumber relpages = 0;
1449 
1450  /* We already got the needed lock */
1451  childrel = table_open(childOID, NoLock);
1452 
1453  /* Ignore if temp table of another backend */
1454  if (RELATION_IS_OTHER_TEMP(childrel))
1455  {
1456  /* ... but release the lock on it */
1457  Assert(childrel != onerel);
1458  table_close(childrel, AccessShareLock);
1459  continue;
1460  }
1461 
1462  /* Check table type (MATVIEW can't happen, but might as well allow) */
1463  if (childrel->rd_rel->relkind == RELKIND_RELATION ||
1464  childrel->rd_rel->relkind == RELKIND_MATVIEW)
1465  {
1466  /* Regular table, so use the regular row acquisition function */
1467  acquirefunc = acquire_sample_rows;
1468  relpages = RelationGetNumberOfBlocks(childrel);
1469  }
1470  else if (childrel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
1471  {
1472  /*
1473  * For a foreign table, call the FDW's hook function to see
1474  * whether it supports analysis.
1475  */
1476  FdwRoutine *fdwroutine;
1477  bool ok = false;
1478 
1479  fdwroutine = GetFdwRoutineForRelation(childrel, false);
1480 
1481  if (fdwroutine->AnalyzeForeignTable != NULL)
1482  ok = fdwroutine->AnalyzeForeignTable(childrel,
1483  &acquirefunc,
1484  &relpages);
1485 
1486  if (!ok)
1487  {
1488  /* ignore, but release the lock on it */
1489  Assert(childrel != onerel);
1490  table_close(childrel, AccessShareLock);
1491  continue;
1492  }
1493  }
1494  else
1495  {
1496  /*
1497  * ignore, but release the lock on it. don't try to unlock the
1498  * passed-in relation
1499  */
1500  Assert(childrel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE);
1501  if (childrel != onerel)
1502  table_close(childrel, AccessShareLock);
1503  else
1504  table_close(childrel, NoLock);
1505  continue;
1506  }
1507 
1508  /* OK, we'll process this child */
1509  has_child = true;
1510  rels[nrels] = childrel;
1511  acquirefuncs[nrels] = acquirefunc;
1512  relblocks[nrels] = (double) relpages;
1513  totalblocks += (double) relpages;
1514  nrels++;
1515  }
1516 
1517  /*
1518  * If we don't have at least one child table to consider, fail. If the
1519  * relation is a partitioned table, it's not counted as a child table.
1520  */
1521  if (!has_child)
1522  {
1523  ereport(elevel,
1524  (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no analyzable child tables",
1526  RelationGetRelationName(onerel))));
1527  return 0;
1528  }
1529 
1530  /*
1531  * Now sample rows from each relation, proportionally to its fraction of
1532  * the total block count. (This might be less than desirable if the child
1533  * rels have radically different free-space percentages, but it's not
1534  * clear that it's worth working harder.)
1535  */
1537  nrels);
1538  numrows = 0;
1539  *totalrows = 0;
1540  *totaldeadrows = 0;
1541  for (i = 0; i < nrels; i++)
1542  {
1543  Relation childrel = rels[i];
1544  AcquireSampleRowsFunc acquirefunc = acquirefuncs[i];
1545  double childblocks = relblocks[i];
1546 
1548  RelationGetRelid(childrel));
1549 
1550  if (childblocks > 0)
1551  {
1552  int childtargrows;
1553 
1554  childtargrows = (int) rint(targrows * childblocks / totalblocks);
1555  /* Make sure we don't overrun due to roundoff error */
1556  childtargrows = Min(childtargrows, targrows - numrows);
1557  if (childtargrows > 0)
1558  {
1559  int childrows;
1560  double trows,
1561  tdrows;
1562 
1563  /* Fetch a random sample of the child's rows */
1564  childrows = (*acquirefunc) (childrel, elevel,
1565  rows + numrows, childtargrows,
1566  &trows, &tdrows);
1567 
1568  /* We may need to convert from child's rowtype to parent's */
1569  if (childrows > 0 &&
1570  !equalTupleDescs(RelationGetDescr(childrel),
1571  RelationGetDescr(onerel)))
1572  {
1573  TupleConversionMap *map;
1574 
1575  map = convert_tuples_by_name(RelationGetDescr(childrel),
1576  RelationGetDescr(onerel));
1577  if (map != NULL)
1578  {
1579  int j;
1580 
1581  for (j = 0; j < childrows; j++)
1582  {
1583  HeapTuple newtup;
1584 
1585  newtup = execute_attr_map_tuple(rows[numrows + j], map);
1586  heap_freetuple(rows[numrows + j]);
1587  rows[numrows + j] = newtup;
1588  }
1589  free_conversion_map(map);
1590  }
1591  }
1592 
1593  /* And add to counts */
1594  numrows += childrows;
1595  *totalrows += trows;
1596  *totaldeadrows += tdrows;
1597  }
1598  }
1599 
1600  /*
1601  * Note: we cannot release the child-table locks, since we may have
1602  * pointers to their TOAST tables in the sampled rows.
1603  */
1604  table_close(childrel, NoLock);
1606  i + 1);
1607  }
1608 
1609  return numrows;
1610 }
1611 
1612 
1613 /*
1614  * update_attstats() -- update attribute statistics for one relation
1615  *
1616  * Statistics are stored in several places: the pg_class row for the
1617  * relation has stats about the whole relation, and there is a
1618  * pg_statistic row for each (non-system) attribute that has ever
1619  * been analyzed. The pg_class values are updated by VACUUM, not here.
1620  *
1621  * pg_statistic rows are just added or updated normally. This means
1622  * that pg_statistic will probably contain some deleted rows at the
1623  * completion of a vacuum cycle, unless it happens to get vacuumed last.
1624  *
1625  * To keep things simple, we punt for pg_statistic, and don't try
1626  * to compute or store rows for pg_statistic itself in pg_statistic.
1627  * This could possibly be made to work, but it's not worth the trouble.
1628  * Note analyze_rel() has seen to it that we won't come here when
1629  * vacuuming pg_statistic itself.
1630  *
1631  * Note: there would be a race condition here if two backends could
1632  * ANALYZE the same table concurrently. Presently, we lock that out
1633  * by taking a self-exclusive lock on the relation in analyze_rel().
1634  */
1635 static void
1636 update_attstats(Oid relid, bool inh, int natts, VacAttrStats **vacattrstats)
1637 {
1638  Relation sd;
1639  int attno;
1640 
1641  if (natts <= 0)
1642  return; /* nothing to do */
1643 
1644  sd = table_open(StatisticRelationId, RowExclusiveLock);
1645 
1646  for (attno = 0; attno < natts; attno++)
1647  {
1648  VacAttrStats *stats = vacattrstats[attno];
1649  HeapTuple stup,
1650  oldtup;
1651  int i,
1652  k,
1653  n;
1654  Datum values[Natts_pg_statistic];
1655  bool nulls[Natts_pg_statistic];
1656  bool replaces[Natts_pg_statistic];
1657 
1658  /* Ignore attr if we weren't able to collect stats */
1659  if (!stats->stats_valid)
1660  continue;
1661 
1662  /*
1663  * Construct a new pg_statistic tuple
1664  */
1665  for (i = 0; i < Natts_pg_statistic; ++i)
1666  {
1667  nulls[i] = false;
1668  replaces[i] = true;
1669  }
1670 
1671  values[Anum_pg_statistic_starelid - 1] = ObjectIdGetDatum(relid);
1672  values[Anum_pg_statistic_staattnum - 1] = Int16GetDatum(stats->attr->attnum);
1673  values[Anum_pg_statistic_stainherit - 1] = BoolGetDatum(inh);
1674  values[Anum_pg_statistic_stanullfrac - 1] = Float4GetDatum(stats->stanullfrac);
1675  values[Anum_pg_statistic_stawidth - 1] = Int32GetDatum(stats->stawidth);
1676  values[Anum_pg_statistic_stadistinct - 1] = Float4GetDatum(stats->stadistinct);
1677  i = Anum_pg_statistic_stakind1 - 1;
1678  for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1679  {
1680  values[i++] = Int16GetDatum(stats->stakind[k]); /* stakindN */
1681  }
1682  i = Anum_pg_statistic_staop1 - 1;
1683  for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1684  {
1685  values[i++] = ObjectIdGetDatum(stats->staop[k]); /* staopN */
1686  }
1687  i = Anum_pg_statistic_stacoll1 - 1;
1688  for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1689  {
1690  values[i++] = ObjectIdGetDatum(stats->stacoll[k]); /* stacollN */
1691  }
1692  i = Anum_pg_statistic_stanumbers1 - 1;
1693  for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1694  {
1695  int nnum = stats->numnumbers[k];
1696 
1697  if (nnum > 0)
1698  {
1699  Datum *numdatums = (Datum *) palloc(nnum * sizeof(Datum));
1700  ArrayType *arry;
1701 
1702  for (n = 0; n < nnum; n++)
1703  numdatums[n] = Float4GetDatum(stats->stanumbers[k][n]);
1704  /* XXX knows more than it should about type float4: */
1705  arry = construct_array(numdatums, nnum,
1706  FLOAT4OID,
1707  sizeof(float4), true, TYPALIGN_INT);
1708  values[i++] = PointerGetDatum(arry); /* stanumbersN */
1709  }
1710  else
1711  {
1712  nulls[i] = true;
1713  values[i++] = (Datum) 0;
1714  }
1715  }
1716  i = Anum_pg_statistic_stavalues1 - 1;
1717  for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1718  {
1719  if (stats->numvalues[k] > 0)
1720  {
1721  ArrayType *arry;
1722 
1723  arry = construct_array(stats->stavalues[k],
1724  stats->numvalues[k],
1725  stats->statypid[k],
1726  stats->statyplen[k],
1727  stats->statypbyval[k],
1728  stats->statypalign[k]);
1729  values[i++] = PointerGetDatum(arry); /* stavaluesN */
1730  }
1731  else
1732  {
1733  nulls[i] = true;
1734  values[i++] = (Datum) 0;
1735  }
1736  }
1737 
1738  /* Is there already a pg_statistic tuple for this attribute? */
1739  oldtup = SearchSysCache3(STATRELATTINH,
1740  ObjectIdGetDatum(relid),
1741  Int16GetDatum(stats->attr->attnum),
1742  BoolGetDatum(inh));
1743 
1744  if (HeapTupleIsValid(oldtup))
1745  {
1746  /* Yes, replace it */
1747  stup = heap_modify_tuple(oldtup,
1748  RelationGetDescr(sd),
1749  values,
1750  nulls,
1751  replaces);
1752  ReleaseSysCache(oldtup);
1753  CatalogTupleUpdate(sd, &stup->t_self, stup);
1754  }
1755  else
1756  {
1757  /* No, insert new tuple */
1758  stup = heap_form_tuple(RelationGetDescr(sd), values, nulls);
1759  CatalogTupleInsert(sd, stup);
1760  }
1761 
1762  heap_freetuple(stup);
1763  }
1764 
1766 }
1767 
1768 /*
1769  * Standard fetch function for use by compute_stats subroutines.
1770  *
1771  * This exists to provide some insulation between compute_stats routines
1772  * and the actual storage of the sample data.
1773  */
1774 static Datum
1775 std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1776 {
1777  int attnum = stats->tupattnum;
1778  HeapTuple tuple = stats->rows[rownum];
1779  TupleDesc tupDesc = stats->tupDesc;
1780 
1781  return heap_getattr(tuple, attnum, tupDesc, isNull);
1782 }
1783 
1784 /*
1785  * Fetch function for analyzing index expressions.
1786  *
1787  * We have not bothered to construct index tuples, instead the data is
1788  * just in Datum arrays.
1789  */
1790 static Datum
1791 ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1792 {
1793  int i;
1794 
1795  /* exprvals and exprnulls are already offset for proper column */
1796  i = rownum * stats->rowstride;
1797  *isNull = stats->exprnulls[i];
1798  return stats->exprvals[i];
1799 }
1800 
1801 
1802 /*==========================================================================
1803  *
1804  * Code below this point represents the "standard" type-specific statistics
1805  * analysis algorithms. This code can be replaced on a per-data-type basis
1806  * by setting a nonzero value in pg_type.typanalyze.
1807  *
1808  *==========================================================================
1809  */
1810 
1811 
1812 /*
1813  * To avoid consuming too much memory during analysis and/or too much space
1814  * in the resulting pg_statistic rows, we ignore varlena datums that are wider
1815  * than WIDTH_THRESHOLD (after detoasting!). This is legitimate for MCV
1816  * and distinct-value calculations since a wide value is unlikely to be
1817  * duplicated at all, much less be a most-common value. For the same reason,
1818  * ignoring wide values will not affect our estimates of histogram bin
1819  * boundaries very much.
1820  */
1821 #define WIDTH_THRESHOLD 1024
1822 
1823 #define swapInt(a,b) do {int _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1824 #define swapDatum(a,b) do {Datum _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1825 
1826 /*
1827  * Extra information used by the default analysis routines
1828  */
1829 typedef struct
1830 {
1831  int count; /* # of duplicates */
1832  int first; /* values[] index of first occurrence */
1833 } ScalarMCVItem;
1834 
1835 typedef struct
1836 {
1840 
1841 
1842 static void compute_trivial_stats(VacAttrStatsP stats,
1843  AnalyzeAttrFetchFunc fetchfunc,
1844  int samplerows,
1845  double totalrows);
1846 static void compute_distinct_stats(VacAttrStatsP stats,
1847  AnalyzeAttrFetchFunc fetchfunc,
1848  int samplerows,
1849  double totalrows);
1850 static void compute_scalar_stats(VacAttrStatsP stats,
1851  AnalyzeAttrFetchFunc fetchfunc,
1852  int samplerows,
1853  double totalrows);
1854 static int compare_scalars(const void *a, const void *b, void *arg);
1855 static int compare_mcvs(const void *a, const void *b);
1856 static int analyze_mcv_list(int *mcv_counts,
1857  int num_mcv,
1858  double stadistinct,
1859  double stanullfrac,
1860  int samplerows,
1861  double totalrows);
1862 
1863 
1864 /*
1865  * std_typanalyze -- the default type-specific typanalyze function
1866  */
1867 bool
1869 {
1870  Form_pg_attribute attr = stats->attr;
1871  Oid ltopr;
1872  Oid eqopr;
1873  StdAnalyzeData *mystats;
1874 
1875  /* If the attstattarget column is negative, use the default value */
1876  /* NB: it is okay to scribble on stats->attr since it's a copy */
1877  if (attr->attstattarget < 0)
1878  attr->attstattarget = default_statistics_target;
1879 
1880  /* Look for default "<" and "=" operators for column's type */
1882  false, false, false,
1883  &ltopr, &eqopr, NULL,
1884  NULL);
1885 
1886  /* Save the operator info for compute_stats routines */
1887  mystats = (StdAnalyzeData *) palloc(sizeof(StdAnalyzeData));
1888  mystats->eqopr = eqopr;
1889  mystats->eqfunc = OidIsValid(eqopr) ? get_opcode(eqopr) : InvalidOid;
1890  mystats->ltopr = ltopr;
1891  stats->extra_data = mystats;
1892 
1893  /*
1894  * Determine which standard statistics algorithm to use
1895  */
1896  if (OidIsValid(eqopr) && OidIsValid(ltopr))
1897  {
1898  /* Seems to be a scalar datatype */
1900  /*--------------------
1901  * The following choice of minrows is based on the paper
1902  * "Random sampling for histogram construction: how much is enough?"
1903  * by Surajit Chaudhuri, Rajeev Motwani and Vivek Narasayya, in
1904  * Proceedings of ACM SIGMOD International Conference on Management
1905  * of Data, 1998, Pages 436-447. Their Corollary 1 to Theorem 5
1906  * says that for table size n, histogram size k, maximum relative
1907  * error in bin size f, and error probability gamma, the minimum
1908  * random sample size is
1909  * r = 4 * k * ln(2*n/gamma) / f^2
1910  * Taking f = 0.5, gamma = 0.01, n = 10^6 rows, we obtain
1911  * r = 305.82 * k
1912  * Note that because of the log function, the dependence on n is
1913  * quite weak; even at n = 10^12, a 300*k sample gives <= 0.66
1914  * bin size error with probability 0.99. So there's no real need to
1915  * scale for n, which is a good thing because we don't necessarily
1916  * know it at this point.
1917  *--------------------
1918  */
1919  stats->minrows = 300 * attr->attstattarget;
1920  }
1921  else if (OidIsValid(eqopr))
1922  {
1923  /* We can still recognize distinct values */
1925  /* Might as well use the same minrows as above */
1926  stats->minrows = 300 * attr->attstattarget;
1927  }
1928  else
1929  {
1930  /* Can't do much but the trivial stuff */
1932  /* Might as well use the same minrows as above */
1933  stats->minrows = 300 * attr->attstattarget;
1934  }
1935 
1936  return true;
1937 }
1938 
1939 
1940 /*
1941  * compute_trivial_stats() -- compute very basic column statistics
1942  *
1943  * We use this when we cannot find a hash "=" operator for the datatype.
1944  *
1945  * We determine the fraction of non-null rows and the average datum width.
1946  */
1947 static void
1949  AnalyzeAttrFetchFunc fetchfunc,
1950  int samplerows,
1951  double totalrows)
1952 {
1953  int i;
1954  int null_cnt = 0;
1955  int nonnull_cnt = 0;
1956  double total_width = 0;
1957  bool is_varlena = (!stats->attrtype->typbyval &&
1958  stats->attrtype->typlen == -1);
1959  bool is_varwidth = (!stats->attrtype->typbyval &&
1960  stats->attrtype->typlen < 0);
1961 
1962  for (i = 0; i < samplerows; i++)
1963  {
1964  Datum value;
1965  bool isnull;
1966 
1968 
1969  value = fetchfunc(stats, i, &isnull);
1970 
1971  /* Check for null/nonnull */
1972  if (isnull)
1973  {
1974  null_cnt++;
1975  continue;
1976  }
1977  nonnull_cnt++;
1978 
1979  /*
1980  * If it's a variable-width field, add up widths for average width
1981  * calculation. Note that if the value is toasted, we use the toasted
1982  * width. We don't bother with this calculation if it's a fixed-width
1983  * type.
1984  */
1985  if (is_varlena)
1986  {
1987  total_width += VARSIZE_ANY(DatumGetPointer(value));
1988  }
1989  else if (is_varwidth)
1990  {
1991  /* must be cstring */
1992  total_width += strlen(DatumGetCString(value)) + 1;
1993  }
1994  }
1995 
1996  /* We can only compute average width if we found some non-null values. */
1997  if (nonnull_cnt > 0)
1998  {
1999  stats->stats_valid = true;
2000  /* Do the simple null-frac and width stats */
2001  stats->stanullfrac = (double) null_cnt / (double) samplerows;
2002  if (is_varwidth)
2003  stats->stawidth = total_width / (double) nonnull_cnt;
2004  else
2005  stats->stawidth = stats->attrtype->typlen;
2006  stats->stadistinct = 0.0; /* "unknown" */
2007  }
2008  else if (null_cnt > 0)
2009  {
2010  /* We found only nulls; assume the column is entirely null */
2011  stats->stats_valid = true;
2012  stats->stanullfrac = 1.0;
2013  if (is_varwidth)
2014  stats->stawidth = 0; /* "unknown" */
2015  else
2016  stats->stawidth = stats->attrtype->typlen;
2017  stats->stadistinct = 0.0; /* "unknown" */
2018  }
2019 }
2020 
2021 
2022 /*
2023  * compute_distinct_stats() -- compute column statistics including ndistinct
2024  *
2025  * We use this when we can find only an "=" operator for the datatype.
2026  *
2027  * We determine the fraction of non-null rows, the average width, the
2028  * most common values, and the (estimated) number of distinct values.
2029  *
2030  * The most common values are determined by brute force: we keep a list
2031  * of previously seen values, ordered by number of times seen, as we scan
2032  * the samples. A newly seen value is inserted just after the last
2033  * multiply-seen value, causing the bottommost (oldest) singly-seen value
2034  * to drop off the list. The accuracy of this method, and also its cost,
2035  * depend mainly on the length of the list we are willing to keep.
2036  */
2037 static void
2039  AnalyzeAttrFetchFunc fetchfunc,
2040  int samplerows,
2041  double totalrows)
2042 {
2043  int i;
2044  int null_cnt = 0;
2045  int nonnull_cnt = 0;
2046  int toowide_cnt = 0;
2047  double total_width = 0;
2048  bool is_varlena = (!stats->attrtype->typbyval &&
2049  stats->attrtype->typlen == -1);
2050  bool is_varwidth = (!stats->attrtype->typbyval &&
2051  stats->attrtype->typlen < 0);
2052  FmgrInfo f_cmpeq;
2053  typedef struct
2054  {
2055  Datum value;
2056  int count;
2057  } TrackItem;
2058  TrackItem *track;
2059  int track_cnt,
2060  track_max;
2061  int num_mcv = stats->attr->attstattarget;
2062  StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2063 
2064  /*
2065  * We track up to 2*n values for an n-element MCV list; but at least 10
2066  */
2067  track_max = 2 * num_mcv;
2068  if (track_max < 10)
2069  track_max = 10;
2070  track = (TrackItem *) palloc(track_max * sizeof(TrackItem));
2071  track_cnt = 0;
2072 
2073  fmgr_info(mystats->eqfunc, &f_cmpeq);
2074 
2075  for (i = 0; i < samplerows; i++)
2076  {
2077  Datum value;
2078  bool isnull;
2079  bool match;
2080  int firstcount1,
2081  j;
2082 
2084 
2085  value = fetchfunc(stats, i, &isnull);
2086 
2087  /* Check for null/nonnull */
2088  if (isnull)
2089  {
2090  null_cnt++;
2091  continue;
2092  }
2093  nonnull_cnt++;
2094 
2095  /*
2096  * If it's a variable-width field, add up widths for average width
2097  * calculation. Note that if the value is toasted, we use the toasted
2098  * width. We don't bother with this calculation if it's a fixed-width
2099  * type.
2100  */
2101  if (is_varlena)
2102  {
2103  total_width += VARSIZE_ANY(DatumGetPointer(value));
2104 
2105  /*
2106  * If the value is toasted, we want to detoast it just once to
2107  * avoid repeated detoastings and resultant excess memory usage
2108  * during the comparisons. Also, check to see if the value is
2109  * excessively wide, and if so don't detoast at all --- just
2110  * ignore the value.
2111  */
2113  {
2114  toowide_cnt++;
2115  continue;
2116  }
2117  value = PointerGetDatum(PG_DETOAST_DATUM(value));
2118  }
2119  else if (is_varwidth)
2120  {
2121  /* must be cstring */
2122  total_width += strlen(DatumGetCString(value)) + 1;
2123  }
2124 
2125  /*
2126  * See if the value matches anything we're already tracking.
2127  */
2128  match = false;
2129  firstcount1 = track_cnt;
2130  for (j = 0; j < track_cnt; j++)
2131  {
2132  if (DatumGetBool(FunctionCall2Coll(&f_cmpeq,
2133  stats->attrcollid,
2134  value, track[j].value)))
2135  {
2136  match = true;
2137  break;
2138  }
2139  if (j < firstcount1 && track[j].count == 1)
2140  firstcount1 = j;
2141  }
2142 
2143  if (match)
2144  {
2145  /* Found a match */
2146  track[j].count++;
2147  /* This value may now need to "bubble up" in the track list */
2148  while (j > 0 && track[j].count > track[j - 1].count)
2149  {
2150  swapDatum(track[j].value, track[j - 1].value);
2151  swapInt(track[j].count, track[j - 1].count);
2152  j--;
2153  }
2154  }
2155  else
2156  {
2157  /* No match. Insert at head of count-1 list */
2158  if (track_cnt < track_max)
2159  track_cnt++;
2160  for (j = track_cnt - 1; j > firstcount1; j--)
2161  {
2162  track[j].value = track[j - 1].value;
2163  track[j].count = track[j - 1].count;
2164  }
2165  if (firstcount1 < track_cnt)
2166  {
2167  track[firstcount1].value = value;
2168  track[firstcount1].count = 1;
2169  }
2170  }
2171  }
2172 
2173  /* We can only compute real stats if we found some non-null values. */
2174  if (nonnull_cnt > 0)
2175  {
2176  int nmultiple,
2177  summultiple;
2178 
2179  stats->stats_valid = true;
2180  /* Do the simple null-frac and width stats */
2181  stats->stanullfrac = (double) null_cnt / (double) samplerows;
2182  if (is_varwidth)
2183  stats->stawidth = total_width / (double) nonnull_cnt;
2184  else
2185  stats->stawidth = stats->attrtype->typlen;
2186 
2187  /* Count the number of values we found multiple times */
2188  summultiple = 0;
2189  for (nmultiple = 0; nmultiple < track_cnt; nmultiple++)
2190  {
2191  if (track[nmultiple].count == 1)
2192  break;
2193  summultiple += track[nmultiple].count;
2194  }
2195 
2196  if (nmultiple == 0)
2197  {
2198  /*
2199  * If we found no repeated non-null values, assume it's a unique
2200  * column; but be sure to discount for any nulls we found.
2201  */
2202  stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2203  }
2204  else if (track_cnt < track_max && toowide_cnt == 0 &&
2205  nmultiple == track_cnt)
2206  {
2207  /*
2208  * Our track list includes every value in the sample, and every
2209  * value appeared more than once. Assume the column has just
2210  * these values. (This case is meant to address columns with
2211  * small, fixed sets of possible values, such as boolean or enum
2212  * columns. If there are any values that appear just once in the
2213  * sample, including too-wide values, we should assume that that's
2214  * not what we're dealing with.)
2215  */
2216  stats->stadistinct = track_cnt;
2217  }
2218  else
2219  {
2220  /*----------
2221  * Estimate the number of distinct values using the estimator
2222  * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2223  * n*d / (n - f1 + f1*n/N)
2224  * where f1 is the number of distinct values that occurred
2225  * exactly once in our sample of n rows (from a total of N),
2226  * and d is the total number of distinct values in the sample.
2227  * This is their Duj1 estimator; the other estimators they
2228  * recommend are considerably more complex, and are numerically
2229  * very unstable when n is much smaller than N.
2230  *
2231  * In this calculation, we consider only non-nulls. We used to
2232  * include rows with null values in the n and N counts, but that
2233  * leads to inaccurate answers in columns with many nulls, and
2234  * it's intuitively bogus anyway considering the desired result is
2235  * the number of distinct non-null values.
2236  *
2237  * We assume (not very reliably!) that all the multiply-occurring
2238  * values are reflected in the final track[] list, and the other
2239  * nonnull values all appeared but once. (XXX this usually
2240  * results in a drastic overestimate of ndistinct. Can we do
2241  * any better?)
2242  *----------
2243  */
2244  int f1 = nonnull_cnt - summultiple;
2245  int d = f1 + nmultiple;
2246  double n = samplerows - null_cnt;
2247  double N = totalrows * (1.0 - stats->stanullfrac);
2248  double stadistinct;
2249 
2250  /* N == 0 shouldn't happen, but just in case ... */
2251  if (N > 0)
2252  stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2253  else
2254  stadistinct = 0;
2255 
2256  /* Clamp to sane range in case of roundoff error */
2257  if (stadistinct < d)
2258  stadistinct = d;
2259  if (stadistinct > N)
2260  stadistinct = N;
2261  /* And round to integer */
2262  stats->stadistinct = floor(stadistinct + 0.5);
2263  }
2264 
2265  /*
2266  * If we estimated the number of distinct values at more than 10% of
2267  * the total row count (a very arbitrary limit), then assume that
2268  * stadistinct should scale with the row count rather than be a fixed
2269  * value.
2270  */
2271  if (stats->stadistinct > 0.1 * totalrows)
2272  stats->stadistinct = -(stats->stadistinct / totalrows);
2273 
2274  /*
2275  * Decide how many values are worth storing as most-common values. If
2276  * we are able to generate a complete MCV list (all the values in the
2277  * sample will fit, and we think these are all the ones in the table),
2278  * then do so. Otherwise, store only those values that are
2279  * significantly more common than the values not in the list.
2280  *
2281  * Note: the first of these cases is meant to address columns with
2282  * small, fixed sets of possible values, such as boolean or enum
2283  * columns. If we can *completely* represent the column population by
2284  * an MCV list that will fit into the stats target, then we should do
2285  * so and thus provide the planner with complete information. But if
2286  * the MCV list is not complete, it's generally worth being more
2287  * selective, and not just filling it all the way up to the stats
2288  * target.
2289  */
2290  if (track_cnt < track_max && toowide_cnt == 0 &&
2291  stats->stadistinct > 0 &&
2292  track_cnt <= num_mcv)
2293  {
2294  /* Track list includes all values seen, and all will fit */
2295  num_mcv = track_cnt;
2296  }
2297  else
2298  {
2299  int *mcv_counts;
2300 
2301  /* Incomplete list; decide how many values are worth keeping */
2302  if (num_mcv > track_cnt)
2303  num_mcv = track_cnt;
2304 
2305  if (num_mcv > 0)
2306  {
2307  mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2308  for (i = 0; i < num_mcv; i++)
2309  mcv_counts[i] = track[i].count;
2310 
2311  num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2312  stats->stadistinct,
2313  stats->stanullfrac,
2314  samplerows, totalrows);
2315  }
2316  }
2317 
2318  /* Generate MCV slot entry */
2319  if (num_mcv > 0)
2320  {
2321  MemoryContext old_context;
2322  Datum *mcv_values;
2323  float4 *mcv_freqs;
2324 
2325  /* Must copy the target values into anl_context */
2326  old_context = MemoryContextSwitchTo(stats->anl_context);
2327  mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2328  mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2329  for (i = 0; i < num_mcv; i++)
2330  {
2331  mcv_values[i] = datumCopy(track[i].value,
2332  stats->attrtype->typbyval,
2333  stats->attrtype->typlen);
2334  mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2335  }
2336  MemoryContextSwitchTo(old_context);
2337 
2338  stats->stakind[0] = STATISTIC_KIND_MCV;
2339  stats->staop[0] = mystats->eqopr;
2340  stats->stacoll[0] = stats->attrcollid;
2341  stats->stanumbers[0] = mcv_freqs;
2342  stats->numnumbers[0] = num_mcv;
2343  stats->stavalues[0] = mcv_values;
2344  stats->numvalues[0] = num_mcv;
2345 
2346  /*
2347  * Accept the defaults for stats->statypid and others. They have
2348  * been set before we were called (see vacuum.h)
2349  */
2350  }
2351  }
2352  else if (null_cnt > 0)
2353  {
2354  /* We found only nulls; assume the column is entirely null */
2355  stats->stats_valid = true;
2356  stats->stanullfrac = 1.0;
2357  if (is_varwidth)
2358  stats->stawidth = 0; /* "unknown" */
2359  else
2360  stats->stawidth = stats->attrtype->typlen;
2361  stats->stadistinct = 0.0; /* "unknown" */
2362  }
2363 
2364  /* We don't need to bother cleaning up any of our temporary palloc's */
2365 }
2366 
2367 
2368 /*
2369  * compute_scalar_stats() -- compute column statistics
2370  *
2371  * We use this when we can find "=" and "<" operators for the datatype.
2372  *
2373  * We determine the fraction of non-null rows, the average width, the
2374  * most common values, the (estimated) number of distinct values, the
2375  * distribution histogram, and the correlation of physical to logical order.
2376  *
2377  * The desired stats can be determined fairly easily after sorting the
2378  * data values into order.
2379  */
2380 static void
2382  AnalyzeAttrFetchFunc fetchfunc,
2383  int samplerows,
2384  double totalrows)
2385 {
2386  int i;
2387  int null_cnt = 0;
2388  int nonnull_cnt = 0;
2389  int toowide_cnt = 0;
2390  double total_width = 0;
2391  bool is_varlena = (!stats->attrtype->typbyval &&
2392  stats->attrtype->typlen == -1);
2393  bool is_varwidth = (!stats->attrtype->typbyval &&
2394  stats->attrtype->typlen < 0);
2395  double corr_xysum;
2396  SortSupportData ssup;
2397  ScalarItem *values;
2398  int values_cnt = 0;
2399  int *tupnoLink;
2400  ScalarMCVItem *track;
2401  int track_cnt = 0;
2402  int num_mcv = stats->attr->attstattarget;
2403  int num_bins = stats->attr->attstattarget;
2404  StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2405 
2406  values = (ScalarItem *) palloc(samplerows * sizeof(ScalarItem));
2407  tupnoLink = (int *) palloc(samplerows * sizeof(int));
2408  track = (ScalarMCVItem *) palloc(num_mcv * sizeof(ScalarMCVItem));
2409 
2410  memset(&ssup, 0, sizeof(ssup));
2412  ssup.ssup_collation = stats->attrcollid;
2413  ssup.ssup_nulls_first = false;
2414 
2415  /*
2416  * For now, don't perform abbreviated key conversion, because full values
2417  * are required for MCV slot generation. Supporting that optimization
2418  * would necessitate teaching compare_scalars() to call a tie-breaker.
2419  */
2420  ssup.abbreviate = false;
2421 
2422  PrepareSortSupportFromOrderingOp(mystats->ltopr, &ssup);
2423 
2424  /* Initial scan to find sortable values */
2425  for (i = 0; i < samplerows; i++)
2426  {
2427  Datum value;
2428  bool isnull;
2429 
2431 
2432  value = fetchfunc(stats, i, &isnull);
2433 
2434  /* Check for null/nonnull */
2435  if (isnull)
2436  {
2437  null_cnt++;
2438  continue;
2439  }
2440  nonnull_cnt++;
2441 
2442  /*
2443  * If it's a variable-width field, add up widths for average width
2444  * calculation. Note that if the value is toasted, we use the toasted
2445  * width. We don't bother with this calculation if it's a fixed-width
2446  * type.
2447  */
2448  if (is_varlena)
2449  {
2450  total_width += VARSIZE_ANY(DatumGetPointer(value));
2451 
2452  /*
2453  * If the value is toasted, we want to detoast it just once to
2454  * avoid repeated detoastings and resultant excess memory usage
2455  * during the comparisons. Also, check to see if the value is
2456  * excessively wide, and if so don't detoast at all --- just
2457  * ignore the value.
2458  */
2460  {
2461  toowide_cnt++;
2462  continue;
2463  }
2464  value = PointerGetDatum(PG_DETOAST_DATUM(value));
2465  }
2466  else if (is_varwidth)
2467  {
2468  /* must be cstring */
2469  total_width += strlen(DatumGetCString(value)) + 1;
2470  }
2471 
2472  /* Add it to the list to be sorted */
2473  values[values_cnt].value = value;
2474  values[values_cnt].tupno = values_cnt;
2475  tupnoLink[values_cnt] = values_cnt;
2476  values_cnt++;
2477  }
2478 
2479  /* We can only compute real stats if we found some sortable values. */
2480  if (values_cnt > 0)
2481  {
2482  int ndistinct, /* # distinct values in sample */
2483  nmultiple, /* # that appear multiple times */
2484  num_hist,
2485  dups_cnt;
2486  int slot_idx = 0;
2488 
2489  /* Sort the collected values */
2490  cxt.ssup = &ssup;
2491  cxt.tupnoLink = tupnoLink;
2492  qsort_arg((void *) values, values_cnt, sizeof(ScalarItem),
2493  compare_scalars, (void *) &cxt);
2494 
2495  /*
2496  * Now scan the values in order, find the most common ones, and also
2497  * accumulate ordering-correlation statistics.
2498  *
2499  * To determine which are most common, we first have to count the
2500  * number of duplicates of each value. The duplicates are adjacent in
2501  * the sorted list, so a brute-force approach is to compare successive
2502  * datum values until we find two that are not equal. However, that
2503  * requires N-1 invocations of the datum comparison routine, which are
2504  * completely redundant with work that was done during the sort. (The
2505  * sort algorithm must at some point have compared each pair of items
2506  * that are adjacent in the sorted order; otherwise it could not know
2507  * that it's ordered the pair correctly.) We exploit this by having
2508  * compare_scalars remember the highest tupno index that each
2509  * ScalarItem has been found equal to. At the end of the sort, a
2510  * ScalarItem's tupnoLink will still point to itself if and only if it
2511  * is the last item of its group of duplicates (since the group will
2512  * be ordered by tupno).
2513  */
2514  corr_xysum = 0;
2515  ndistinct = 0;
2516  nmultiple = 0;
2517  dups_cnt = 0;
2518  for (i = 0; i < values_cnt; i++)
2519  {
2520  int tupno = values[i].tupno;
2521 
2522  corr_xysum += ((double) i) * ((double) tupno);
2523  dups_cnt++;
2524  if (tupnoLink[tupno] == tupno)
2525  {
2526  /* Reached end of duplicates of this value */
2527  ndistinct++;
2528  if (dups_cnt > 1)
2529  {
2530  nmultiple++;
2531  if (track_cnt < num_mcv ||
2532  dups_cnt > track[track_cnt - 1].count)
2533  {
2534  /*
2535  * Found a new item for the mcv list; find its
2536  * position, bubbling down old items if needed. Loop
2537  * invariant is that j points at an empty/ replaceable
2538  * slot.
2539  */
2540  int j;
2541 
2542  if (track_cnt < num_mcv)
2543  track_cnt++;
2544  for (j = track_cnt - 1; j > 0; j--)
2545  {
2546  if (dups_cnt <= track[j - 1].count)
2547  break;
2548  track[j].count = track[j - 1].count;
2549  track[j].first = track[j - 1].first;
2550  }
2551  track[j].count = dups_cnt;
2552  track[j].first = i + 1 - dups_cnt;
2553  }
2554  }
2555  dups_cnt = 0;
2556  }
2557  }
2558 
2559  stats->stats_valid = true;
2560  /* Do the simple null-frac and width stats */
2561  stats->stanullfrac = (double) null_cnt / (double) samplerows;
2562  if (is_varwidth)
2563  stats->stawidth = total_width / (double) nonnull_cnt;
2564  else
2565  stats->stawidth = stats->attrtype->typlen;
2566 
2567  if (nmultiple == 0)
2568  {
2569  /*
2570  * If we found no repeated non-null values, assume it's a unique
2571  * column; but be sure to discount for any nulls we found.
2572  */
2573  stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2574  }
2575  else if (toowide_cnt == 0 && nmultiple == ndistinct)
2576  {
2577  /*
2578  * Every value in the sample appeared more than once. Assume the
2579  * column has just these values. (This case is meant to address
2580  * columns with small, fixed sets of possible values, such as
2581  * boolean or enum columns. If there are any values that appear
2582  * just once in the sample, including too-wide values, we should
2583  * assume that that's not what we're dealing with.)
2584  */
2585  stats->stadistinct = ndistinct;
2586  }
2587  else
2588  {
2589  /*----------
2590  * Estimate the number of distinct values using the estimator
2591  * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2592  * n*d / (n - f1 + f1*n/N)
2593  * where f1 is the number of distinct values that occurred
2594  * exactly once in our sample of n rows (from a total of N),
2595  * and d is the total number of distinct values in the sample.
2596  * This is their Duj1 estimator; the other estimators they
2597  * recommend are considerably more complex, and are numerically
2598  * very unstable when n is much smaller than N.
2599  *
2600  * In this calculation, we consider only non-nulls. We used to
2601  * include rows with null values in the n and N counts, but that
2602  * leads to inaccurate answers in columns with many nulls, and
2603  * it's intuitively bogus anyway considering the desired result is
2604  * the number of distinct non-null values.
2605  *
2606  * Overwidth values are assumed to have been distinct.
2607  *----------
2608  */
2609  int f1 = ndistinct - nmultiple + toowide_cnt;
2610  int d = f1 + nmultiple;
2611  double n = samplerows - null_cnt;
2612  double N = totalrows * (1.0 - stats->stanullfrac);
2613  double stadistinct;
2614 
2615  /* N == 0 shouldn't happen, but just in case ... */
2616  if (N > 0)
2617  stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2618  else
2619  stadistinct = 0;
2620 
2621  /* Clamp to sane range in case of roundoff error */
2622  if (stadistinct < d)
2623  stadistinct = d;
2624  if (stadistinct > N)
2625  stadistinct = N;
2626  /* And round to integer */
2627  stats->stadistinct = floor(stadistinct + 0.5);
2628  }
2629 
2630  /*
2631  * If we estimated the number of distinct values at more than 10% of
2632  * the total row count (a very arbitrary limit), then assume that
2633  * stadistinct should scale with the row count rather than be a fixed
2634  * value.
2635  */
2636  if (stats->stadistinct > 0.1 * totalrows)
2637  stats->stadistinct = -(stats->stadistinct / totalrows);
2638 
2639  /*
2640  * Decide how many values are worth storing as most-common values. If
2641  * we are able to generate a complete MCV list (all the values in the
2642  * sample will fit, and we think these are all the ones in the table),
2643  * then do so. Otherwise, store only those values that are
2644  * significantly more common than the values not in the list.
2645  *
2646  * Note: the first of these cases is meant to address columns with
2647  * small, fixed sets of possible values, such as boolean or enum
2648  * columns. If we can *completely* represent the column population by
2649  * an MCV list that will fit into the stats target, then we should do
2650  * so and thus provide the planner with complete information. But if
2651  * the MCV list is not complete, it's generally worth being more
2652  * selective, and not just filling it all the way up to the stats
2653  * target.
2654  */
2655  if (track_cnt == ndistinct && toowide_cnt == 0 &&
2656  stats->stadistinct > 0 &&
2657  track_cnt <= num_mcv)
2658  {
2659  /* Track list includes all values seen, and all will fit */
2660  num_mcv = track_cnt;
2661  }
2662  else
2663  {
2664  int *mcv_counts;
2665 
2666  /* Incomplete list; decide how many values are worth keeping */
2667  if (num_mcv > track_cnt)
2668  num_mcv = track_cnt;
2669 
2670  if (num_mcv > 0)
2671  {
2672  mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2673  for (i = 0; i < num_mcv; i++)
2674  mcv_counts[i] = track[i].count;
2675 
2676  num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2677  stats->stadistinct,
2678  stats->stanullfrac,
2679  samplerows, totalrows);
2680  }
2681  }
2682 
2683  /* Generate MCV slot entry */
2684  if (num_mcv > 0)
2685  {
2686  MemoryContext old_context;
2687  Datum *mcv_values;
2688  float4 *mcv_freqs;
2689 
2690  /* Must copy the target values into anl_context */
2691  old_context = MemoryContextSwitchTo(stats->anl_context);
2692  mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2693  mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2694  for (i = 0; i < num_mcv; i++)
2695  {
2696  mcv_values[i] = datumCopy(values[track[i].first].value,
2697  stats->attrtype->typbyval,
2698  stats->attrtype->typlen);
2699  mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2700  }
2701  MemoryContextSwitchTo(old_context);
2702 
2703  stats->stakind[slot_idx] = STATISTIC_KIND_MCV;
2704  stats->staop[slot_idx] = mystats->eqopr;
2705  stats->stacoll[slot_idx] = stats->attrcollid;
2706  stats->stanumbers[slot_idx] = mcv_freqs;
2707  stats->numnumbers[slot_idx] = num_mcv;
2708  stats->stavalues[slot_idx] = mcv_values;
2709  stats->numvalues[slot_idx] = num_mcv;
2710 
2711  /*
2712  * Accept the defaults for stats->statypid and others. They have
2713  * been set before we were called (see vacuum.h)
2714  */
2715  slot_idx++;
2716  }
2717 
2718  /*
2719  * Generate a histogram slot entry if there are at least two distinct
2720  * values not accounted for in the MCV list. (This ensures the
2721  * histogram won't collapse to empty or a singleton.)
2722  */
2723  num_hist = ndistinct - num_mcv;
2724  if (num_hist > num_bins)
2725  num_hist = num_bins + 1;
2726  if (num_hist >= 2)
2727  {
2728  MemoryContext old_context;
2729  Datum *hist_values;
2730  int nvals;
2731  int pos,
2732  posfrac,
2733  delta,
2734  deltafrac;
2735 
2736  /* Sort the MCV items into position order to speed next loop */
2737  qsort((void *) track, num_mcv,
2738  sizeof(ScalarMCVItem), compare_mcvs);
2739 
2740  /*
2741  * Collapse out the MCV items from the values[] array.
2742  *
2743  * Note we destroy the values[] array here... but we don't need it
2744  * for anything more. We do, however, still need values_cnt.
2745  * nvals will be the number of remaining entries in values[].
2746  */
2747  if (num_mcv > 0)
2748  {
2749  int src,
2750  dest;
2751  int j;
2752 
2753  src = dest = 0;
2754  j = 0; /* index of next interesting MCV item */
2755  while (src < values_cnt)
2756  {
2757  int ncopy;
2758 
2759  if (j < num_mcv)
2760  {
2761  int first = track[j].first;
2762 
2763  if (src >= first)
2764  {
2765  /* advance past this MCV item */
2766  src = first + track[j].count;
2767  j++;
2768  continue;
2769  }
2770  ncopy = first - src;
2771  }
2772  else
2773  ncopy = values_cnt - src;
2774  memmove(&values[dest], &values[src],
2775  ncopy * sizeof(ScalarItem));
2776  src += ncopy;
2777  dest += ncopy;
2778  }
2779  nvals = dest;
2780  }
2781  else
2782  nvals = values_cnt;
2783  Assert(nvals >= num_hist);
2784 
2785  /* Must copy the target values into anl_context */
2786  old_context = MemoryContextSwitchTo(stats->anl_context);
2787  hist_values = (Datum *) palloc(num_hist * sizeof(Datum));
2788 
2789  /*
2790  * The object of this loop is to copy the first and last values[]
2791  * entries along with evenly-spaced values in between. So the
2792  * i'th value is values[(i * (nvals - 1)) / (num_hist - 1)]. But
2793  * computing that subscript directly risks integer overflow when
2794  * the stats target is more than a couple thousand. Instead we
2795  * add (nvals - 1) / (num_hist - 1) to pos at each step, tracking
2796  * the integral and fractional parts of the sum separately.
2797  */
2798  delta = (nvals - 1) / (num_hist - 1);
2799  deltafrac = (nvals - 1) % (num_hist - 1);
2800  pos = posfrac = 0;
2801 
2802  for (i = 0; i < num_hist; i++)
2803  {
2804  hist_values[i] = datumCopy(values[pos].value,
2805  stats->attrtype->typbyval,
2806  stats->attrtype->typlen);
2807  pos += delta;
2808  posfrac += deltafrac;
2809  if (posfrac >= (num_hist - 1))
2810  {
2811  /* fractional part exceeds 1, carry to integer part */
2812  pos++;
2813  posfrac -= (num_hist - 1);
2814  }
2815  }
2816 
2817  MemoryContextSwitchTo(old_context);
2818 
2819  stats->stakind[slot_idx] = STATISTIC_KIND_HISTOGRAM;
2820  stats->staop[slot_idx] = mystats->ltopr;
2821  stats->stacoll[slot_idx] = stats->attrcollid;
2822  stats->stavalues[slot_idx] = hist_values;
2823  stats->numvalues[slot_idx] = num_hist;
2824 
2825  /*
2826  * Accept the defaults for stats->statypid and others. They have
2827  * been set before we were called (see vacuum.h)
2828  */
2829  slot_idx++;
2830  }
2831 
2832  /* Generate a correlation entry if there are multiple values */
2833  if (values_cnt > 1)
2834  {
2835  MemoryContext old_context;
2836  float4 *corrs;
2837  double corr_xsum,
2838  corr_x2sum;
2839 
2840  /* Must copy the target values into anl_context */
2841  old_context = MemoryContextSwitchTo(stats->anl_context);
2842  corrs = (float4 *) palloc(sizeof(float4));
2843  MemoryContextSwitchTo(old_context);
2844 
2845  /*----------
2846  * Since we know the x and y value sets are both
2847  * 0, 1, ..., values_cnt-1
2848  * we have sum(x) = sum(y) =
2849  * (values_cnt-1)*values_cnt / 2
2850  * and sum(x^2) = sum(y^2) =
2851  * (values_cnt-1)*values_cnt*(2*values_cnt-1) / 6.
2852  *----------
2853  */
2854  corr_xsum = ((double) (values_cnt - 1)) *
2855  ((double) values_cnt) / 2.0;
2856  corr_x2sum = ((double) (values_cnt - 1)) *
2857  ((double) values_cnt) * (double) (2 * values_cnt - 1) / 6.0;
2858 
2859  /* And the correlation coefficient reduces to */
2860  corrs[0] = (values_cnt * corr_xysum - corr_xsum * corr_xsum) /
2861  (values_cnt * corr_x2sum - corr_xsum * corr_xsum);
2862 
2863  stats->stakind[slot_idx] = STATISTIC_KIND_CORRELATION;
2864  stats->staop[slot_idx] = mystats->ltopr;
2865  stats->stacoll[slot_idx] = stats->attrcollid;
2866  stats->stanumbers[slot_idx] = corrs;
2867  stats->numnumbers[slot_idx] = 1;
2868  slot_idx++;
2869  }
2870  }
2871  else if (nonnull_cnt > 0)
2872  {
2873  /* We found some non-null values, but they were all too wide */
2874  Assert(nonnull_cnt == toowide_cnt);
2875  stats->stats_valid = true;
2876  /* Do the simple null-frac and width stats */
2877  stats->stanullfrac = (double) null_cnt / (double) samplerows;
2878  if (is_varwidth)
2879  stats->stawidth = total_width / (double) nonnull_cnt;
2880  else
2881  stats->stawidth = stats->attrtype->typlen;
2882  /* Assume all too-wide values are distinct, so it's a unique column */
2883  stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2884  }
2885  else if (null_cnt > 0)
2886  {
2887  /* We found only nulls; assume the column is entirely null */
2888  stats->stats_valid = true;
2889  stats->stanullfrac = 1.0;
2890  if (is_varwidth)
2891  stats->stawidth = 0; /* "unknown" */
2892  else
2893  stats->stawidth = stats->attrtype->typlen;
2894  stats->stadistinct = 0.0; /* "unknown" */
2895  }
2896 
2897  /* We don't need to bother cleaning up any of our temporary palloc's */
2898 }
2899 
2900 /*
2901  * qsort_arg comparator for sorting ScalarItems
2902  *
2903  * Aside from sorting the items, we update the tupnoLink[] array
2904  * whenever two ScalarItems are found to contain equal datums. The array
2905  * is indexed by tupno; for each ScalarItem, it contains the highest
2906  * tupno that that item's datum has been found to be equal to. This allows
2907  * us to avoid additional comparisons in compute_scalar_stats().
2908  */
2909 static int
2910 compare_scalars(const void *a, const void *b, void *arg)
2911 {
2912  Datum da = ((const ScalarItem *) a)->value;
2913  int ta = ((const ScalarItem *) a)->tupno;
2914  Datum db = ((const ScalarItem *) b)->value;
2915  int tb = ((const ScalarItem *) b)->tupno;
2917  int compare;
2918 
2919  compare = ApplySortComparator(da, false, db, false, cxt->ssup);
2920  if (compare != 0)
2921  return compare;
2922 
2923  /*
2924  * The two datums are equal, so update cxt->tupnoLink[].
2925  */
2926  if (cxt->tupnoLink[ta] < tb)
2927  cxt->tupnoLink[ta] = tb;
2928  if (cxt->tupnoLink[tb] < ta)
2929  cxt->tupnoLink[tb] = ta;
2930 
2931  /*
2932  * For equal datums, sort by tupno
2933  */
2934  return ta - tb;
2935 }
2936 
2937 /*
2938  * qsort comparator for sorting ScalarMCVItems by position
2939  */
2940 static int
2941 compare_mcvs(const void *a, const void *b)
2942 {
2943  int da = ((const ScalarMCVItem *) a)->first;
2944  int db = ((const ScalarMCVItem *) b)->first;
2945 
2946  return da - db;
2947 }
2948 
2949 /*
2950  * Analyze the list of common values in the sample and decide how many are
2951  * worth storing in the table's MCV list.
2952  *
2953  * mcv_counts is assumed to be a list of the counts of the most common values
2954  * seen in the sample, starting with the most common. The return value is the
2955  * number that are significantly more common than the values not in the list,
2956  * and which are therefore deemed worth storing in the table's MCV list.
2957  */
2958 static int
2959 analyze_mcv_list(int *mcv_counts,
2960  int num_mcv,
2961  double stadistinct,
2962  double stanullfrac,
2963  int samplerows,
2964  double totalrows)
2965 {
2966  double ndistinct_table;
2967  double sumcount;
2968  int i;
2969 
2970  /*
2971  * If the entire table was sampled, keep the whole list. This also
2972  * protects us against division by zero in the code below.
2973  */
2974  if (samplerows == totalrows || totalrows <= 1.0)
2975  return num_mcv;
2976 
2977  /* Re-extract the estimated number of distinct nonnull values in table */
2978  ndistinct_table = stadistinct;
2979  if (ndistinct_table < 0)
2980  ndistinct_table = -ndistinct_table * totalrows;
2981 
2982  /*
2983  * Exclude the least common values from the MCV list, if they are not
2984  * significantly more common than the estimated selectivity they would
2985  * have if they weren't in the list. All non-MCV values are assumed to be
2986  * equally common, after taking into account the frequencies of all the
2987  * values in the MCV list and the number of nulls (c.f. eqsel()).
2988  *
2989  * Here sumcount tracks the total count of all but the last (least common)
2990  * value in the MCV list, allowing us to determine the effect of excluding
2991  * that value from the list.
2992  *
2993  * Note that we deliberately do this by removing values from the full
2994  * list, rather than starting with an empty list and adding values,
2995  * because the latter approach can fail to add any values if all the most
2996  * common values have around the same frequency and make up the majority
2997  * of the table, so that the overall average frequency of all values is
2998  * roughly the same as that of the common values. This would lead to any
2999  * uncommon values being significantly overestimated.
3000  */
3001  sumcount = 0.0;
3002  for (i = 0; i < num_mcv - 1; i++)
3003  sumcount += mcv_counts[i];
3004 
3005  while (num_mcv > 0)
3006  {
3007  double selec,
3008  otherdistinct,
3009  N,
3010  n,
3011  K,
3012  variance,
3013  stddev;
3014 
3015  /*
3016  * Estimated selectivity the least common value would have if it
3017  * wasn't in the MCV list (c.f. eqsel()).
3018  */
3019  selec = 1.0 - sumcount / samplerows - stanullfrac;
3020  if (selec < 0.0)
3021  selec = 0.0;
3022  if (selec > 1.0)
3023  selec = 1.0;
3024  otherdistinct = ndistinct_table - (num_mcv - 1);
3025  if (otherdistinct > 1)
3026  selec /= otherdistinct;
3027 
3028  /*
3029  * If the value is kept in the MCV list, its population frequency is
3030  * assumed to equal its sample frequency. We use the lower end of a
3031  * textbook continuity-corrected Wald-type confidence interval to
3032  * determine if that is significantly more common than the non-MCV
3033  * frequency --- specifically we assume the population frequency is
3034  * highly likely to be within around 2 standard errors of the sample
3035  * frequency, which equates to an interval of 2 standard deviations
3036  * either side of the sample count, plus an additional 0.5 for the
3037  * continuity correction. Since we are sampling without replacement,
3038  * this is a hypergeometric distribution.
3039  *
3040  * XXX: Empirically, this approach seems to work quite well, but it
3041  * may be worth considering more advanced techniques for estimating
3042  * the confidence interval of the hypergeometric distribution.
3043  */
3044  N = totalrows;
3045  n = samplerows;
3046  K = N * mcv_counts[num_mcv - 1] / n;
3047  variance = n * K * (N - K) * (N - n) / (N * N * (N - 1));
3048  stddev = sqrt(variance);
3049 
3050  if (mcv_counts[num_mcv - 1] > selec * samplerows + 2 * stddev + 0.5)
3051  {
3052  /*
3053  * The value is significantly more common than the non-MCV
3054  * selectivity would suggest. Keep it, and all the other more
3055  * common values in the list.
3056  */
3057  break;
3058  }
3059  else
3060  {
3061  /* Discard this value and consider the next least common value */
3062  num_mcv--;
3063  if (num_mcv == 0)
3064  break;
3065  sumcount -= mcv_counts[num_mcv - 1];
3066  }
3067  }
3068  return num_mcv;
3069 }
TupleTableSlot * table_slot_create(Relation relation, List **reglist)
Definition: tableam.c:91
static Datum Float4GetDatum(float4 X)
Definition: postgres.h:725
AttributeOpts * get_attribute_options(Oid attrelid, int attnum)
Definition: attoptcache.c:103
bool BlockSampler_HasMore(BlockSampler bs)
Definition: sampling.c:58
int rowstride
Definition: vacuum.h:174
static void do_analyze_rel(Relation onerel, VacuumParams *params, List *va_cols, AcquireSampleRowsFunc acquirefunc, BlockNumber relpages, bool inh, bool in_outer_xact, int elevel)
Definition: analyze.c:290
void FormIndexDatum(IndexInfo *indexInfo, TupleTableSlot *slot, EState *estate, Datum *values, bool *isnull)
Definition: index.c:2662
bool ssup_nulls_first
Definition: sortsupport.h:75
#define NIL
Definition: pg_list.h:65
void analyze_rel(Oid relid, RangeVar *relation, VacuumParams *params, List *va_cols, bool in_outer_xact, BufferAccessStrategy bstrategy)
Definition: analyze.c:120
void vac_close_indexes(int nindexes, Relation *Irel, LOCKMODE lockmode)
Definition: vacuum.c:2128
int ComputeExtStatisticsRows(Relation onerel, int natts, VacAttrStats **vacattrstats)
Definition: fmgr.h:56
static int acquire_inherited_sample_rows(Relation onerel, int elevel, HeapTuple *rows, int targrows, double *totalrows, double *totaldeadrows)
Definition: analyze.c:1391
int minrows
Definition: vacuum.h:135
void MemoryContextDelete(MemoryContext context)
Definition: mcxt.c:218
#define AllocSetContextCreate
Definition: memutils.h:173
int64 VacuumPageMiss
Definition: globals.c:148
int64 PgStat_Counter
Definition: pgstat.h:92
void table_close(Relation relation, LOCKMODE lockmode)
Definition: table.c:167
#define GETSTRUCT(TUP)
Definition: htup_details.h:654
List * ii_Predicate
Definition: execnodes.h:163
#define PROGRESS_ANALYZE_BLOCKS_TOTAL
Definition: progress.h:39
#define PROGRESS_ANALYZE_PHASE
Definition: progress.h:38
uint32 TransactionId
Definition: c.h:587
#define SECURITY_RESTRICTED_OPERATION
Definition: miscadmin.h:312
static ListCell * lnext(const List *l, const ListCell *c)
Definition: pg_list.h:322
void SetUserIdAndSecContext(Oid userid, int sec_context)
Definition: miscinit.c:607
IndexBulkDeleteResult * index_vacuum_cleanup(IndexVacuumInfo *info, IndexBulkDeleteResult *istat)
Definition: indexam.c:712
#define RelationGetDescr(relation)
Definition: rel.h:503
static int compare_mcvs(const void *a, const void *b)
Definition: analyze.c:2941
TupleTableSlot * MakeSingleTupleTableSlot(TupleDesc tupdesc, const TupleTableSlotOps *tts_ops)
Definition: execTuples.c:1238
AnalyzeForeignTable_function AnalyzeForeignTable
Definition: fdwapi.h:257
TimestampTz GetCurrentTimestamp(void)
Definition: timestamp.c:1580
int32 exprTypmod(const Node *expr)
Definition: nodeFuncs.c:267
BlockNumber BlockSampler_Next(BlockSampler bs)
Definition: sampling.c:64
int64 TimestampTz
Definition: timestamp.h:39
HeapTuple * rows
Definition: vacuum.h:170
#define PointerGetDatum(X)
Definition: postgres.h:600
#define VACOPT_ANALYZE
Definition: vacuum.h:179
static int compare_scalars(const void *a, const void *b, void *arg)
Definition: analyze.c:2910
long random(void)
Definition: random.c:22
#define TupleDescAttr(tupdesc, i)
Definition: tupdesc.h:92
struct AnlIndexData AnlIndexData
bool analyze_only
Definition: genam.h:47
int64 VacuumPageHit
Definition: globals.c:147
#define Min(x, y)
Definition: c.h:986
void pgstat_report_analyze(Relation rel, PgStat_Counter livetuples, PgStat_Counter deadtuples, bool resetcounter)
Definition: pgstat.c:1606
BufferAccessStrategy strategy
Definition: genam.h:52
static int analyze_mcv_list(int *mcv_counts, int num_mcv, double stadistinct, double stanullfrac, int samplerows, double totalrows)
Definition: analyze.c:2959
void PrepareSortSupportFromOrderingOp(Oid orderingOp, SortSupport ssup)
Definition: sortsupport.c:135
Datum * stavalues[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:152
static void compute_scalar_stats(VacAttrStatsP stats, AnalyzeAttrFetchFunc fetchfunc, int samplerows, double totalrows)
Definition: analyze.c:2381
static MemoryContext MemoryContextSwitchTo(MemoryContext context)
Definition: palloc.h:109
#define Int16GetDatum(X)
Definition: postgres.h:495
double sampler_random_fract(SamplerRandomState randstate)
Definition: sampling.c:242
ArrayType * construct_array(Datum *elems, int nelems, Oid elmtype, int elmlen, bool elmbyval, char elmalign)
Definition: arrayfuncs.c:3319
#define AccessShareLock
Definition: lockdefs.h:36
Definition: nodes.h:536
#define strVal(v)
Definition: value.h:65
int tupattnum
Definition: vacuum.h:169
#define PROGRESS_ANALYZE_CURRENT_CHILD_TABLE_RELID
Definition: progress.h:45
int errcode(int sqlerrcode)
Definition: elog.c:698
Relation index
Definition: genam.h:46
#define INFO
Definition: elog.h:33
static BufferAccessStrategy vac_strategy
Definition: analyze.c:86
int64 VacuumPageDirty
Definition: globals.c:149
TupleDesc tupDesc
Definition: vacuum.h:171
uint32 BlockNumber
Definition: block.h:31
bool std_typanalyze(VacAttrStats *stats)
Definition: analyze.c:1868
bool statypbyval[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:162
void reservoir_init_selection_state(ReservoirState rs, int n)
Definition: sampling.c:133
HeapTuple heap_form_tuple(TupleDesc tupleDescriptor, Datum *values, bool *isnull)
Definition: heaptuple.c:1020
static void update_attstats(Oid relid, bool inh, int natts, VacAttrStats **vacattrstats)
Definition: analyze.c:1636
IndexInfo * BuildIndexInfo(Relation index)
Definition: index.c:2378
void visibilitymap_count(Relation rel, BlockNumber *all_visible, BlockNumber *all_frozen)
Datum FunctionCall2Coll(FmgrInfo *flinfo, Oid collation, Datum arg1, Datum arg2)
Definition: fmgr.c:1148
#define LOG
Definition: elog.h:26
Form_pg_class rd_rel
Definition: rel.h:109
void heap_freetuple(HeapTuple htup)
Definition: heaptuple.c:1338
unsigned int Oid
Definition: postgres_ext.h:31
static bool ExecQual(ExprState *state, ExprContext *econtext)
Definition: executor.h:396
int32 attrtypmod
Definition: vacuum.h:125
bool TimestampDifferenceExceeds(TimestampTz start_time, TimestampTz stop_time, int msec)
Definition: timestamp.c:1711
#define PROGRESS_ANALYZE_PHASE_COMPUTE_STATS
Definition: progress.h:50
#define OidIsValid(objectId)
Definition: c.h:710
char statypalign[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:163
void pgstat_progress_end_command(void)
float8 n_distinct
Definition: attoptcache.h:22
#define PROGRESS_ANALYZE_CHILD_TABLES_TOTAL
Definition: progress.h:43
Form_pg_attribute attr
Definition: vacuum.h:123
#define swapDatum(a, b)
Definition: analyze.c:1824
int attr_cnt
Definition: analyze.c:77
uint16 OffsetNumber
Definition: off.h:24
void pg_rusage_init(PGRUsage *ru0)
Definition: pg_rusage.c:27
void SetRelationHasSubclass(Oid relationId, bool relhassubclass)
Definition: tablecmds.c:3187
void FreeExecutorState(EState *estate)
Definition: execUtils.c:186
#define GetPerTupleExprContext(estate)
Definition: executor.h:533
struct RelationData * Relation
Definition: relcache.h:26
bits32 options
Definition: vacuum.h:211
int get_tablespace_maintenance_io_concurrency(Oid spcid)
Definition: spccache.c:228
static bool table_scan_analyze_next_block(TableScanDesc scan, BlockNumber blockno, BufferAccessStrategy bstrategy)
Definition: tableam.h:1690
void pfree(void *pointer)
Definition: mcxt.c:1169
void appendStringInfo(StringInfo str, const char *fmt,...)
Definition: stringinfo.c:91
Oid attrtypid
Definition: vacuum.h:124
static int compare(const void *arg1, const void *arg2)
Definition: geqo_pool.c:145
Oid * rd_indcollation
Definition: rel.h:212
TransactionId GetOldestNonRemovableTransactionId(Relation rel)
Definition: procarray.c:2005
#define ObjectIdGetDatum(X)
Definition: postgres.h:551
#define ERROR
Definition: elog.h:46
double float8
Definition: c.h:565
#define DatumGetCString(X)
Definition: postgres.h:610
int32 stawidth
Definition: vacuum.h:144
bool * exprnulls
Definition: vacuum.h:173
Oid stacoll[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:148
void fmgr_info(Oid functionId, FmgrInfo *finfo)
Definition: fmgr.c:126
MemoryContext ssup_cxt
Definition: sortsupport.h:66
HeapTuple SearchSysCache3(int cacheId, Datum key1, Datum key2, Datum key3)
Definition: syscache.c:1149
ItemPointerData t_self
Definition: htup.h:65
PgStat_Counter pgStatBlockWriteTime
Definition: pgstat.c:248
void BuildRelationExtStatistics(Relation onerel, double totalrows, int numrows, HeapTuple *rows, int natts, VacAttrStats **vacattrstats)
#define OidFunctionCall1(functionId, arg1)
Definition: fmgr.h:664
#define ALLOCSET_DEFAULT_SIZES
Definition: memutils.h:195
char * get_database_name(Oid dbid)
Definition: dbcommands.c:2113
#define DEBUG2
Definition: elog.h:24
static MemoryContext anl_context
Definition: analyze.c:85
SortSupport ssup
Definition: analyze.c:1837
TupleConversionMap * convert_tuples_by_name(TupleDesc indesc, TupleDesc outdesc)
Definition: tupconvert.c:102
char * get_namespace_name(Oid nspid)
Definition: lsyscache.c:3316
void pgstat_progress_start_command(ProgressCommandType cmdtype, Oid relid)
PgStat_Counter pgStatBlockReadTime
Definition: pgstat.c:247
#define NoLock
Definition: lockdefs.h:34
static char * buf
Definition: pg_test_fsync.c:68
float8 n_distinct_inherited
Definition: attoptcache.h:23
void ExecDropSingleTupleTableSlot(TupleTableSlot *slot)
Definition: execTuples.c:1254
void GetUserIdAndSecContext(Oid *userid, int *sec_context)
Definition: miscinit.c:600
int numnumbers[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:149
#define RowExclusiveLock
Definition: lockdefs.h:38
void AtEOXact_GUC(bool isCommit, int nestLevel)
Definition: guc.c:6242
#define PROGRESS_ANALYZE_BLOCKS_DONE
Definition: progress.h:40
void free_conversion_map(TupleConversionMap *map)
Definition: tupconvert.c:284
#define DatumGetBool(X)
Definition: postgres.h:437
const char * pg_rusage_show(const PGRUsage *ru0)
Definition: pg_rusage.c:40
#define InvalidTransactionId
Definition: transam.h:31
#define RelationGetRelationName(relation)
Definition: rel.h:511
ExprState * ExecPrepareQual(List *qual, EState *estate)
Definition: execExpr.c:774
static ListCell * list_head(const List *l)
Definition: pg_list.h:125
FormData_pg_attribute * Form_pg_attribute
Definition: pg_attribute.h:207
MemoryContext CurrentMemoryContext
Definition: mcxt.c:42
Size toast_raw_datum_size(Datum value)
Definition: detoast.c:545
HeapTuple execute_attr_map_tuple(HeapTuple tuple, TupleConversionMap *map)
Definition: tupconvert.c:139
#define ATTRIBUTE_FIXED_PART_SIZE
Definition: pg_attribute.h:199
bool IsAutoVacuumWorkerProcess(void)
Definition: autovacuum.c:3410
Datum datumCopy(Datum value, bool typByVal, int typLen)
Definition: datum.c:131
float4 stanullfrac
Definition: vacuum.h:143
EState * CreateExecutorState(void)
Definition: execUtils.c:90
Datum * exprvals
Definition: vacuum.h:172
#define PROGRESS_ANALYZE_PHASE_FINALIZE_ANALYZE
Definition: progress.h:52
static int compare_rows(const void *a, const void *b)
Definition: analyze.c:1361
void initStringInfo(StringInfo str)
Definition: stringinfo.c:59
void qsort_arg(void *base, size_t nel, size_t elsize, qsort_arg_comparator cmp, void *arg)
#define WARNING
Definition: elog.h:40
void vac_open_indexes(Relation relation, LOCKMODE lockmode, int *nindexes, Relation **Irel)
Definition: vacuum.c:2085
#define PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS_INH
Definition: progress.h:49
#define heap_getattr(tup, attnum, tupleDesc, isnull)
Definition: htup_details.h:761
BlockNumber BlockSampler_Init(BlockSampler bs, BlockNumber nblocks, int samplesize, long randseed)
Definition: sampling.c:39
Oid staop[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:147
bool stats_valid
Definition: vacuum.h:142
float float4
Definition: c.h:564
#define WIDTH_THRESHOLD
Definition: analyze.c:1821
static int elevel
Definition: vacuumlazy.c:401
#define MemoryContextResetAndDeleteChildren(ctx)
Definition: memutils.h:67
void * palloc0(Size size)
Definition: mcxt.c:1093
uintptr_t Datum
Definition: postgres.h:411
void CommandCounterIncrement(void)
Definition: xact.c:1021
void ReleaseSysCache(HeapTuple tuple)
Definition: syscache.c:1175
int16 stakind[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:146
int ii_NumIndexAttrs
Definition: execnodes.h:158
Oid MyDatabaseId
Definition: globals.c:88
#define InvalidMultiXactId
Definition: multixact.h:24
#define RelationGetNumberOfBlocks(reln)
Definition: bufmgr.h:212
TupleDesc rd_att
Definition: rel.h:110
#define VARSIZE_ANY(PTR)
Definition: postgres.h:348
Oid statypid[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:160
#define BoolGetDatum(X)
Definition: postgres.h:446
static void compute_distinct_stats(VacAttrStatsP stats, AnalyzeAttrFetchFunc fetchfunc, int samplerows, double totalrows)
Definition: analyze.c:2038
#define InvalidOid
Definition: postgres_ext.h:36
static Datum ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
Definition: analyze.c:1791
#define K(t)
Definition: sha1.c:66
RegProcedure get_opcode(Oid opno)
Definition: lsyscache.c:1256
int16 attnum
Definition: pg_attribute.h:83
static struct @143 value
#define STATISTIC_NUM_SLOTS
Definition: pg_statistic.h:127
#define ereport(elevel,...)
Definition: elog.h:157
static HeapTuple ExecCopySlotHeapTuple(TupleTableSlot *slot)
Definition: tuptable.h:452
#define PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS
Definition: progress.h:48
int message_level
Definition: genam.h:50
int errmsg_internal(const char *fmt,...)
Definition: elog.c:996
double num_heap_tuples
Definition: genam.h:51
#define ShareUpdateExclusiveLock
Definition: lockdefs.h:39
#define HeapTupleIsValid(tuple)
Definition: htup.h:78
int attnameAttNum(Relation rd, const char *attname, bool sysColOK)
static void compute_index_stats(Relation onerel, double totalrows, AnlIndexData *indexdata, int nindexes, HeapTuple *rows, int numrows, MemoryContext col_context)
Definition: analyze.c:832
List * ii_Expressions
Definition: execnodes.h:161
void relation_close(Relation relation, LOCKMODE lockmode)
Definition: relation.c:206
#define Assert(condition)
Definition: c.h:804
#define VACOPT_VACUUM
Definition: vacuum.h:178
#define lfirst(lc)
Definition: pg_list.h:169
void pgstat_progress_update_param(int index, int64 val)
static bool table_scan_analyze_next_tuple(TableScanDesc scan, TransactionId OldestXmin, double *liverows, double *deadrows, TupleTableSlot *slot)
Definition: tableam.h:1708
float4 * stanumbers[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:150
void get_sort_group_operators(Oid argtype, bool needLT, bool needEQ, bool needGT, Oid *ltOpr, Oid *eqOpr, Oid *gtOpr, bool *isHashable)
Definition: parse_oper.c:192
#define RELATION_IS_OTHER_TEMP(relation)
Definition: rel.h:631
Datum(* AnalyzeAttrFetchFunc)(VacAttrStatsP stats, int rownum, bool *isNull)
Definition: vacuum.h:102
#define PROGRESS_ANALYZE_CHILD_TABLES_DONE
Definition: progress.h:44
#define INDEX_MAX_KEYS
void CatalogTupleUpdate(Relation heapRel, ItemPointer otid, HeapTuple tup)
Definition: indexing.c:301
Oid exprType(const Node *expr)
Definition: nodeFuncs.c:41
#define InvalidBlockNumber
Definition: block.h:33
FormData_pg_type * Form_pg_type
Definition: pg_type.h:261
static int list_length(const List *l)
Definition: pg_list.h:149
Oid exprCollation(const Node *expr)
Definition: nodeFuncs.c:759
TupleTableSlot * ecxt_scantuple
Definition: execnodes.h:226
int log_min_duration
Definition: vacuum.h:219
#define ItemPointerGetOffsetNumber(pointer)
Definition: itemptr.h:117
#define VACOPT_VERBOSE
Definition: vacuum.h:180
Bitmapset * bms_add_member(Bitmapset *a, int x)
Definition: bitmapset.c:736
List * RelationGetIndexList(Relation relation)
Definition: relcache.c:4573
Oid attrcollid
Definition: vacuum.h:127
MemoryContext anl_context
Definition: vacuum.h:128
double tupleFract
Definition: analyze.c:75
#define InvalidAttrNumber
Definition: attnum.h:23
#define swapInt(a, b)
Definition: analyze.c:1823
#define DatumGetPointer(X)
Definition: postgres.h:593
Relation rs_rd
Definition: relscan.h:34
static void compute_trivial_stats(VacAttrStatsP stats, AnalyzeAttrFetchFunc fetchfunc, int samplerows, double totalrows)
Definition: analyze.c:1948
static void table_endscan(TableScanDesc scan)
Definition: tableam.h:991
static Datum values[MAXATTR]
Definition: bootstrap.c:156
#define SearchSysCacheCopy1(cacheId, key1)
Definition: syscache.h:175
int(* AcquireSampleRowsFunc)(Relation relation, int elevel, HeapTuple *rows, int targrows, double *totalrows, double *totaldeadrows)
Definition: fdwapi.h:151
List * find_all_inheritors(Oid parentrelId, LOCKMODE lockmode, List **numparents)
Definition: pg_inherits.c:256
#define Int32GetDatum(X)
Definition: postgres.h:523
int NewGUCNestLevel(void)
Definition: guc.c:6228
int numvalues[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:151
Form_pg_type attrtype
Definition: vacuum.h:126
void * palloc(Size size)
Definition: mcxt.c:1062
int errmsg(const char *fmt,...)
Definition: elog.c:909
FdwRoutine * GetFdwRoutineForRelation(Relation relation, bool makecopy)
Definition: foreign.c:427
VacAttrStats ** vacattrstats
Definition: analyze.c:76
int16 statyplen[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:161
void list_free(List *list)
Definition: list.c:1391
#define elog(elevel,...)
Definition: elog.h:232
int i
const TupleTableSlotOps TTSOpsHeapTuple
Definition: execTuples.c:84
AnalyzeAttrComputeStatsFunc compute_stats
Definition: vacuum.h:134
PrefetchBufferResult PrefetchBuffer(Relation reln, ForkNumber forkNum, BlockNumber blockNum)
Definition: bufmgr.c:587
bool equalTupleDescs(TupleDesc tupdesc1, TupleDesc tupdesc2)
Definition: tupdesc.c:402
void * arg
#define PG_DETOAST_DATUM(datum)
Definition: fmgr.h:240
int f1[ARRAY_SIZE]
Definition: sql-declare.c:113
#define CHECK_FOR_INTERRUPTS()
Definition: miscadmin.h:120
void * extra_data
Definition: vacuum.h:136
Relation vacuum_open_relation(Oid relid, RangeVar *relation, bits32 options, bool verbose, LOCKMODE lmode)
Definition: vacuum.c:633
bool vacuum_is_relation_owner(Oid relid, Form_pg_class reltuple, bits32 options)
Definition: vacuum.c:559
#define ItemPointerGetBlockNumber(pointer)
Definition: itemptr.h:98
#define qsort(a, b, c, d)
Definition: port.h:505
void vacuum_delay_point(void)
Definition: vacuum.c:2149
AttrNumber ii_IndexAttrNumbers[INDEX_MAX_KEYS]
Definition: execnodes.h:160
Relation table_open(Oid relationId, LOCKMODE lockmode)
Definition: table.c:39
HeapTuple heap_modify_tuple(HeapTuple tuple, TupleDesc tupleDesc, Datum *replValues, bool *replIsnull, bool *doReplace)
Definition: heaptuple.c:1113
static int ApplySortComparator(Datum datum1, bool isNull1, Datum datum2, bool isNull2, SortSupport ssup)
Definition: sortsupport.h:200
void vac_update_relstats(Relation relation, BlockNumber num_pages, double num_tuples, BlockNumber num_all_visible_pages, bool hasindex, TransactionId frozenxid, MultiXactId minmulti, bool in_outer_xact)
Definition: vacuum.c:1306
Definition: pg_list.h:50
bool bms_is_member(int x, const Bitmapset *a)
Definition: bitmapset.c:427
static VacAttrStats * examine_attribute(Relation onerel, int attnum, Node *index_expr)
Definition: analyze.c:1014
#define _(x)
Definition: elog.c:89
#define RelationGetRelid(relation)
Definition: rel.h:477
void CatalogTupleInsert(Relation heapRel, HeapTuple tup)
Definition: indexing.c:221
bool track_io_timing
Definition: bufmgr.c:135
TupleTableSlot * ExecStoreHeapTuple(HeapTuple tuple, TupleTableSlot *slot, bool shouldFree)
Definition: execTuples.c:1352
int default_statistics_target
Definition: analyze.c:82
#define ResetExprContext(econtext)
Definition: executor.h:527
#define lfirst_oid(lc)
Definition: pg_list.h:171
SamplerRandomState randstate
Definition: sampling.h:50
long TimestampDifferenceMilliseconds(TimestampTz start_time, TimestampTz stop_time)
Definition: timestamp.c:1693
double reservoir_get_next_S(ReservoirState rs, double t, int n)
Definition: sampling.c:146
static Datum std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
Definition: analyze.c:1775
static int acquire_sample_rows(Relation onerel, int elevel, HeapTuple *rows, int targrows, double *totalrows, double *totaldeadrows)
Definition: analyze.c:1145
bool estimated_count
Definition: genam.h:49
float4 stadistinct
Definition: vacuum.h:145
IndexInfo * indexInfo
Definition: analyze.c:74
static TableScanDesc table_beginscan_analyze(Relation rel)
Definition: tableam.h:980
#define RelationGetNamespace(relation)
Definition: rel.h:518