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) 1996-2024, PostgreSQL Global Development Group
7  * Portions Copyright (c) 1994, Regents of the University of California
8  *
9  *
10  * IDENTIFICATION
11  * src/backend/commands/analyze.c
12  *
13  *-------------------------------------------------------------------------
14  */
15 #include "postgres.h"
16 
17 #include <math.h>
18 
19 #include "access/detoast.h"
20 #include "access/genam.h"
21 #include "access/multixact.h"
22 #include "access/relation.h"
23 #include "access/table.h"
24 #include "access/tableam.h"
25 #include "access/transam.h"
26 #include "access/tupconvert.h"
27 #include "access/visibilitymap.h"
28 #include "access/xact.h"
29 #include "catalog/index.h"
30 #include "catalog/indexing.h"
31 #include "catalog/pg_inherits.h"
32 #include "commands/dbcommands.h"
33 #include "commands/progress.h"
34 #include "commands/tablecmds.h"
35 #include "commands/vacuum.h"
36 #include "common/pg_prng.h"
37 #include "executor/executor.h"
38 #include "foreign/fdwapi.h"
39 #include "miscadmin.h"
40 #include "nodes/nodeFuncs.h"
41 #include "parser/parse_oper.h"
42 #include "parser/parse_relation.h"
43 #include "pgstat.h"
44 #include "postmaster/autovacuum.h"
46 #include "statistics/statistics.h"
47 #include "storage/bufmgr.h"
48 #include "storage/procarray.h"
49 #include "utils/attoptcache.h"
50 #include "utils/datum.h"
51 #include "utils/guc.h"
52 #include "utils/lsyscache.h"
53 #include "utils/memutils.h"
54 #include "utils/pg_rusage.h"
55 #include "utils/sampling.h"
56 #include "utils/sortsupport.h"
57 #include "utils/spccache.h"
58 #include "utils/syscache.h"
59 #include "utils/timestamp.h"
60 
61 
62 /* Per-index data for ANALYZE */
63 typedef struct AnlIndexData
64 {
65  IndexInfo *indexInfo; /* BuildIndexInfo result */
66  double tupleFract; /* fraction of rows for partial index */
67  VacAttrStats **vacattrstats; /* index attrs to analyze */
68  int attr_cnt;
70 
71 
72 /* Default statistics target (GUC parameter) */
74 
75 /* A few variables that don't seem worth passing around as parameters */
76 static MemoryContext anl_context = NULL;
78 
79 
80 static void do_analyze_rel(Relation onerel,
81  VacuumParams *params, List *va_cols,
82  AcquireSampleRowsFunc acquirefunc, BlockNumber relpages,
83  bool inh, bool in_outer_xact, int elevel);
84 static void compute_index_stats(Relation onerel, double totalrows,
85  AnlIndexData *indexdata, int nindexes,
86  HeapTuple *rows, int numrows,
87  MemoryContext col_context);
88 static VacAttrStats *examine_attribute(Relation onerel, int attnum,
89  Node *index_expr);
90 static int acquire_sample_rows(Relation onerel, int elevel,
91  HeapTuple *rows, int targrows,
92  double *totalrows, double *totaldeadrows);
93 static int compare_rows(const void *a, const void *b, void *arg);
94 static int acquire_inherited_sample_rows(Relation onerel, int elevel,
95  HeapTuple *rows, int targrows,
96  double *totalrows, double *totaldeadrows);
97 static void update_attstats(Oid relid, bool inh,
98  int natts, VacAttrStats **vacattrstats);
99 static Datum std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
100 static Datum ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
101 
102 
103 /*
104  * analyze_rel() -- analyze one relation
105  *
106  * relid identifies the relation to analyze. If relation is supplied, use
107  * the name therein for reporting any failure to open/lock the rel; do not
108  * use it once we've successfully opened the rel, since it might be stale.
109  */
110 void
111 analyze_rel(Oid relid, RangeVar *relation,
112  VacuumParams *params, List *va_cols, bool in_outer_xact,
113  BufferAccessStrategy bstrategy)
114 {
115  Relation onerel;
116  int elevel;
117  AcquireSampleRowsFunc acquirefunc = NULL;
118  BlockNumber relpages = 0;
119 
120  /* Select logging level */
121  if (params->options & VACOPT_VERBOSE)
122  elevel = INFO;
123  else
124  elevel = DEBUG2;
125 
126  /* Set up static variables */
127  vac_strategy = bstrategy;
128 
129  /*
130  * Check for user-requested abort.
131  */
133 
134  /*
135  * Open the relation, getting ShareUpdateExclusiveLock to ensure that two
136  * ANALYZEs don't run on it concurrently. (This also locks out a
137  * concurrent VACUUM, which doesn't matter much at the moment but might
138  * matter if we ever try to accumulate stats on dead tuples.) If the rel
139  * has been dropped since we last saw it, we don't need to process it.
140  *
141  * Make sure to generate only logs for ANALYZE in this case.
142  */
143  onerel = vacuum_open_relation(relid, relation, params->options & ~(VACOPT_VACUUM),
144  params->log_min_duration >= 0,
146 
147  /* leave if relation could not be opened or locked */
148  if (!onerel)
149  return;
150 
151  /*
152  * Check if relation needs to be skipped based on privileges. This check
153  * happens also when building the relation list to analyze for a manual
154  * operation, and needs to be done additionally here as ANALYZE could
155  * happen across multiple transactions where privileges could have changed
156  * in-between. Make sure to generate only logs for ANALYZE in this case.
157  */
159  onerel->rd_rel,
160  params->options & ~VACOPT_VACUUM))
161  {
163  return;
164  }
165 
166  /*
167  * Silently ignore tables that are temp tables of other backends ---
168  * trying to analyze these is rather pointless, since their contents are
169  * probably not up-to-date on disk. (We don't throw a warning here; it
170  * would just lead to chatter during a database-wide ANALYZE.)
171  */
172  if (RELATION_IS_OTHER_TEMP(onerel))
173  {
175  return;
176  }
177 
178  /*
179  * We can ANALYZE any table except pg_statistic. See update_attstats
180  */
181  if (RelationGetRelid(onerel) == StatisticRelationId)
182  {
184  return;
185  }
186 
187  /*
188  * Check that it's of an analyzable relkind, and set up appropriately.
189  */
190  if (onerel->rd_rel->relkind == RELKIND_RELATION ||
191  onerel->rd_rel->relkind == RELKIND_MATVIEW)
192  {
193  /* Regular table, so we'll use the regular row acquisition function */
194  acquirefunc = acquire_sample_rows;
195  /* Also get regular table's size */
196  relpages = RelationGetNumberOfBlocks(onerel);
197  }
198  else if (onerel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
199  {
200  /*
201  * For a foreign table, call the FDW's hook function to see whether it
202  * supports analysis.
203  */
204  FdwRoutine *fdwroutine;
205  bool ok = false;
206 
207  fdwroutine = GetFdwRoutineForRelation(onerel, false);
208 
209  if (fdwroutine->AnalyzeForeignTable != NULL)
210  ok = fdwroutine->AnalyzeForeignTable(onerel,
211  &acquirefunc,
212  &relpages);
213 
214  if (!ok)
215  {
217  (errmsg("skipping \"%s\" --- cannot analyze this foreign table",
218  RelationGetRelationName(onerel))));
220  return;
221  }
222  }
223  else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
224  {
225  /*
226  * For partitioned tables, we want to do the recursive ANALYZE below.
227  */
228  }
229  else
230  {
231  /* No need for a WARNING if we already complained during VACUUM */
232  if (!(params->options & VACOPT_VACUUM))
234  (errmsg("skipping \"%s\" --- cannot analyze non-tables or special system tables",
235  RelationGetRelationName(onerel))));
237  return;
238  }
239 
240  /*
241  * OK, let's do it. First, initialize progress reporting.
242  */
244  RelationGetRelid(onerel));
245 
246  /*
247  * Do the normal non-recursive ANALYZE. We can skip this for partitioned
248  * tables, which don't contain any rows.
249  */
250  if (onerel->rd_rel->relkind != RELKIND_PARTITIONED_TABLE)
251  do_analyze_rel(onerel, params, va_cols, acquirefunc,
252  relpages, false, in_outer_xact, elevel);
253 
254  /*
255  * If there are child tables, do recursive ANALYZE.
256  */
257  if (onerel->rd_rel->relhassubclass)
258  do_analyze_rel(onerel, params, va_cols, acquirefunc, relpages,
259  true, in_outer_xact, elevel);
260 
261  /*
262  * Close source relation now, but keep lock so that no one deletes it
263  * before we commit. (If someone did, they'd fail to clean up the entries
264  * we made in pg_statistic. Also, releasing the lock before commit would
265  * expose us to concurrent-update failures in update_attstats.)
266  */
267  relation_close(onerel, NoLock);
268 
270 }
271 
272 /*
273  * do_analyze_rel() -- analyze one relation, recursively or not
274  *
275  * Note that "acquirefunc" is only relevant for the non-inherited case.
276  * For the inherited case, acquire_inherited_sample_rows() determines the
277  * appropriate acquirefunc for each child table.
278  */
279 static void
281  List *va_cols, AcquireSampleRowsFunc acquirefunc,
282  BlockNumber relpages, bool inh, bool in_outer_xact,
283  int elevel)
284 {
285  int attr_cnt,
286  tcnt,
287  i,
288  ind;
289  Relation *Irel;
290  int nindexes;
291  bool hasindex;
292  VacAttrStats **vacattrstats;
293  AnlIndexData *indexdata;
294  int targrows,
295  numrows,
296  minrows;
297  double totalrows,
298  totaldeadrows;
299  HeapTuple *rows;
300  PGRUsage ru0;
301  TimestampTz starttime = 0;
302  MemoryContext caller_context;
303  Oid save_userid;
304  int save_sec_context;
305  int save_nestlevel;
306  int64 AnalyzePageHit = VacuumPageHit;
307  int64 AnalyzePageMiss = VacuumPageMiss;
308  int64 AnalyzePageDirty = VacuumPageDirty;
309  PgStat_Counter startreadtime = 0;
310  PgStat_Counter startwritetime = 0;
311 
312  if (inh)
313  ereport(elevel,
314  (errmsg("analyzing \"%s.%s\" inheritance tree",
316  RelationGetRelationName(onerel))));
317  else
318  ereport(elevel,
319  (errmsg("analyzing \"%s.%s\"",
321  RelationGetRelationName(onerel))));
322 
323  /*
324  * Set up a working context so that we can easily free whatever junk gets
325  * created.
326  */
328  "Analyze",
330  caller_context = MemoryContextSwitchTo(anl_context);
331 
332  /*
333  * Switch to the table owner's userid, so that any index functions are run
334  * as that user. Also lock down security-restricted operations and
335  * arrange to make GUC variable changes local to this command.
336  */
337  GetUserIdAndSecContext(&save_userid, &save_sec_context);
338  SetUserIdAndSecContext(onerel->rd_rel->relowner,
339  save_sec_context | SECURITY_RESTRICTED_OPERATION);
340  save_nestlevel = NewGUCNestLevel();
342 
343  /* measure elapsed time iff autovacuum logging requires it */
344  if (AmAutoVacuumWorkerProcess() && params->log_min_duration >= 0)
345  {
346  if (track_io_timing)
347  {
348  startreadtime = pgStatBlockReadTime;
349  startwritetime = pgStatBlockWriteTime;
350  }
351 
352  pg_rusage_init(&ru0);
353  starttime = GetCurrentTimestamp();
354  }
355 
356  /*
357  * Determine which columns to analyze
358  *
359  * Note that system attributes are never analyzed, so we just reject them
360  * at the lookup stage. We also reject duplicate column mentions. (We
361  * could alternatively ignore duplicates, but analyzing a column twice
362  * won't work; we'd end up making a conflicting update in pg_statistic.)
363  */
364  if (va_cols != NIL)
365  {
366  Bitmapset *unique_cols = NULL;
367  ListCell *le;
368 
369  vacattrstats = (VacAttrStats **) palloc(list_length(va_cols) *
370  sizeof(VacAttrStats *));
371  tcnt = 0;
372  foreach(le, va_cols)
373  {
374  char *col = strVal(lfirst(le));
375 
376  i = attnameAttNum(onerel, col, false);
377  if (i == InvalidAttrNumber)
378  ereport(ERROR,
379  (errcode(ERRCODE_UNDEFINED_COLUMN),
380  errmsg("column \"%s\" of relation \"%s\" does not exist",
381  col, RelationGetRelationName(onerel))));
382  if (bms_is_member(i, unique_cols))
383  ereport(ERROR,
384  (errcode(ERRCODE_DUPLICATE_COLUMN),
385  errmsg("column \"%s\" of relation \"%s\" appears more than once",
386  col, RelationGetRelationName(onerel))));
387  unique_cols = bms_add_member(unique_cols, i);
388 
389  vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
390  if (vacattrstats[tcnt] != NULL)
391  tcnt++;
392  }
393  attr_cnt = tcnt;
394  }
395  else
396  {
397  attr_cnt = onerel->rd_att->natts;
398  vacattrstats = (VacAttrStats **)
399  palloc(attr_cnt * sizeof(VacAttrStats *));
400  tcnt = 0;
401  for (i = 1; i <= attr_cnt; i++)
402  {
403  vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
404  if (vacattrstats[tcnt] != NULL)
405  tcnt++;
406  }
407  attr_cnt = tcnt;
408  }
409 
410  /*
411  * Open all indexes of the relation, and see if there are any analyzable
412  * columns in the indexes. We do not analyze index columns if there was
413  * an explicit column list in the ANALYZE command, however.
414  *
415  * If we are doing a recursive scan, we don't want to touch the parent's
416  * indexes at all. If we're processing a partitioned table, we need to
417  * know if there are any indexes, but we don't want to process them.
418  */
419  if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
420  {
421  List *idxs = RelationGetIndexList(onerel);
422 
423  Irel = NULL;
424  nindexes = 0;
425  hasindex = idxs != NIL;
426  list_free(idxs);
427  }
428  else if (!inh)
429  {
430  vac_open_indexes(onerel, AccessShareLock, &nindexes, &Irel);
431  hasindex = nindexes > 0;
432  }
433  else
434  {
435  Irel = NULL;
436  nindexes = 0;
437  hasindex = false;
438  }
439  indexdata = NULL;
440  if (nindexes > 0)
441  {
442  indexdata = (AnlIndexData *) palloc0(nindexes * sizeof(AnlIndexData));
443  for (ind = 0; ind < nindexes; ind++)
444  {
445  AnlIndexData *thisdata = &indexdata[ind];
446  IndexInfo *indexInfo;
447 
448  thisdata->indexInfo = indexInfo = BuildIndexInfo(Irel[ind]);
449  thisdata->tupleFract = 1.0; /* fix later if partial */
450  if (indexInfo->ii_Expressions != NIL && va_cols == NIL)
451  {
452  ListCell *indexpr_item = list_head(indexInfo->ii_Expressions);
453 
454  thisdata->vacattrstats = (VacAttrStats **)
455  palloc(indexInfo->ii_NumIndexAttrs * sizeof(VacAttrStats *));
456  tcnt = 0;
457  for (i = 0; i < indexInfo->ii_NumIndexAttrs; i++)
458  {
459  int keycol = indexInfo->ii_IndexAttrNumbers[i];
460 
461  if (keycol == 0)
462  {
463  /* Found an index expression */
464  Node *indexkey;
465 
466  if (indexpr_item == NULL) /* shouldn't happen */
467  elog(ERROR, "too few entries in indexprs list");
468  indexkey = (Node *) lfirst(indexpr_item);
469  indexpr_item = lnext(indexInfo->ii_Expressions,
470  indexpr_item);
471  thisdata->vacattrstats[tcnt] =
472  examine_attribute(Irel[ind], i + 1, indexkey);
473  if (thisdata->vacattrstats[tcnt] != NULL)
474  tcnt++;
475  }
476  }
477  thisdata->attr_cnt = tcnt;
478  }
479  }
480  }
481 
482  /*
483  * Determine how many rows we need to sample, using the worst case from
484  * all analyzable columns. We use a lower bound of 100 rows to avoid
485  * possible overflow in Vitter's algorithm. (Note: that will also be the
486  * target in the corner case where there are no analyzable columns.)
487  */
488  targrows = 100;
489  for (i = 0; i < attr_cnt; i++)
490  {
491  if (targrows < vacattrstats[i]->minrows)
492  targrows = vacattrstats[i]->minrows;
493  }
494  for (ind = 0; ind < nindexes; ind++)
495  {
496  AnlIndexData *thisdata = &indexdata[ind];
497 
498  for (i = 0; i < thisdata->attr_cnt; i++)
499  {
500  if (targrows < thisdata->vacattrstats[i]->minrows)
501  targrows = thisdata->vacattrstats[i]->minrows;
502  }
503  }
504 
505  /*
506  * Look at extended statistics objects too, as those may define custom
507  * statistics target. So we may need to sample more rows and then build
508  * the statistics with enough detail.
509  */
510  minrows = ComputeExtStatisticsRows(onerel, attr_cnt, vacattrstats);
511 
512  if (targrows < minrows)
513  targrows = minrows;
514 
515  /*
516  * Acquire the sample rows
517  */
518  rows = (HeapTuple *) palloc(targrows * sizeof(HeapTuple));
522  if (inh)
523  numrows = acquire_inherited_sample_rows(onerel, elevel,
524  rows, targrows,
525  &totalrows, &totaldeadrows);
526  else
527  numrows = (*acquirefunc) (onerel, elevel,
528  rows, targrows,
529  &totalrows, &totaldeadrows);
530 
531  /*
532  * Compute the statistics. Temporary results during the calculations for
533  * each column are stored in a child context. The calc routines are
534  * responsible to make sure that whatever they store into the VacAttrStats
535  * structure is allocated in anl_context.
536  */
537  if (numrows > 0)
538  {
539  MemoryContext col_context,
540  old_context;
541 
544 
545  col_context = AllocSetContextCreate(anl_context,
546  "Analyze Column",
548  old_context = MemoryContextSwitchTo(col_context);
549 
550  for (i = 0; i < attr_cnt; i++)
551  {
552  VacAttrStats *stats = vacattrstats[i];
553  AttributeOpts *aopt;
554 
555  stats->rows = rows;
556  stats->tupDesc = onerel->rd_att;
557  stats->compute_stats(stats,
559  numrows,
560  totalrows);
561 
562  /*
563  * If the appropriate flavor of the n_distinct option is
564  * specified, override with the corresponding value.
565  */
566  aopt = get_attribute_options(onerel->rd_id, stats->tupattnum);
567  if (aopt != NULL)
568  {
569  float8 n_distinct;
570 
571  n_distinct = inh ? aopt->n_distinct_inherited : aopt->n_distinct;
572  if (n_distinct != 0.0)
573  stats->stadistinct = n_distinct;
574  }
575 
576  MemoryContextReset(col_context);
577  }
578 
579  if (nindexes > 0)
580  compute_index_stats(onerel, totalrows,
581  indexdata, nindexes,
582  rows, numrows,
583  col_context);
584 
585  MemoryContextSwitchTo(old_context);
586  MemoryContextDelete(col_context);
587 
588  /*
589  * Emit the completed stats rows into pg_statistic, replacing any
590  * previous statistics for the target columns. (If there are stats in
591  * pg_statistic for columns we didn't process, we leave them alone.)
592  */
593  update_attstats(RelationGetRelid(onerel), inh,
594  attr_cnt, vacattrstats);
595 
596  for (ind = 0; ind < nindexes; ind++)
597  {
598  AnlIndexData *thisdata = &indexdata[ind];
599 
600  update_attstats(RelationGetRelid(Irel[ind]), false,
601  thisdata->attr_cnt, thisdata->vacattrstats);
602  }
603 
604  /* Build extended statistics (if there are any). */
605  BuildRelationExtStatistics(onerel, inh, totalrows, numrows, rows,
606  attr_cnt, vacattrstats);
607  }
608 
611 
612  /*
613  * Update pages/tuples stats in pg_class ... but not if we're doing
614  * inherited stats.
615  *
616  * We assume that VACUUM hasn't set pg_class.reltuples already, even
617  * during a VACUUM ANALYZE. Although VACUUM often updates pg_class,
618  * exceptions exist. A "VACUUM (ANALYZE, INDEX_CLEANUP OFF)" command will
619  * never update pg_class entries for index relations. It's also possible
620  * that an individual index's pg_class entry won't be updated during
621  * VACUUM if the index AM returns NULL from its amvacuumcleanup() routine.
622  */
623  if (!inh)
624  {
625  BlockNumber relallvisible;
626 
627  if (RELKIND_HAS_STORAGE(onerel->rd_rel->relkind))
628  visibilitymap_count(onerel, &relallvisible, NULL);
629  else
630  relallvisible = 0;
631 
632  /*
633  * Update pg_class for table relation. CCI first, in case acquirefunc
634  * updated pg_class.
635  */
637  vac_update_relstats(onerel,
638  relpages,
639  totalrows,
640  relallvisible,
641  hasindex,
644  NULL, NULL,
645  in_outer_xact);
646 
647  /* Same for indexes */
648  for (ind = 0; ind < nindexes; ind++)
649  {
650  AnlIndexData *thisdata = &indexdata[ind];
651  double totalindexrows;
652 
653  totalindexrows = ceil(thisdata->tupleFract * totalrows);
654  vac_update_relstats(Irel[ind],
656  totalindexrows,
657  0,
658  false,
661  NULL, NULL,
662  in_outer_xact);
663  }
664  }
665  else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
666  {
667  /*
668  * Partitioned tables don't have storage, so we don't set any fields
669  * in their pg_class entries except for reltuples and relhasindex.
670  */
672  vac_update_relstats(onerel, -1, totalrows,
673  0, hasindex, InvalidTransactionId,
675  NULL, NULL,
676  in_outer_xact);
677  }
678 
679  /*
680  * Now report ANALYZE to the cumulative stats system. For regular tables,
681  * we do it only if not doing inherited stats. For partitioned tables, we
682  * only do it for inherited stats. (We're never called for not-inherited
683  * stats on partitioned tables anyway.)
684  *
685  * Reset the changes_since_analyze counter only if we analyzed all
686  * columns; otherwise, there is still work for auto-analyze to do.
687  */
688  if (!inh)
689  pgstat_report_analyze(onerel, totalrows, totaldeadrows,
690  (va_cols == NIL));
691  else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
692  pgstat_report_analyze(onerel, 0, 0, (va_cols == NIL));
693 
694  /*
695  * If this isn't part of VACUUM ANALYZE, let index AMs do cleanup.
696  *
697  * Note that most index AMs perform a no-op as a matter of policy for
698  * amvacuumcleanup() when called in ANALYZE-only mode. The only exception
699  * among core index AMs is GIN/ginvacuumcleanup().
700  */
701  if (!(params->options & VACOPT_VACUUM))
702  {
703  for (ind = 0; ind < nindexes; ind++)
704  {
705  IndexBulkDeleteResult *stats;
706  IndexVacuumInfo ivinfo;
707 
708  ivinfo.index = Irel[ind];
709  ivinfo.heaprel = onerel;
710  ivinfo.analyze_only = true;
711  ivinfo.estimated_count = true;
712  ivinfo.message_level = elevel;
713  ivinfo.num_heap_tuples = onerel->rd_rel->reltuples;
714  ivinfo.strategy = vac_strategy;
715 
716  stats = index_vacuum_cleanup(&ivinfo, NULL);
717 
718  if (stats)
719  pfree(stats);
720  }
721  }
722 
723  /* Done with indexes */
724  vac_close_indexes(nindexes, Irel, NoLock);
725 
726  /* Log the action if appropriate */
727  if (AmAutoVacuumWorkerProcess() && params->log_min_duration >= 0)
728  {
729  TimestampTz endtime = GetCurrentTimestamp();
730 
731  if (params->log_min_duration == 0 ||
732  TimestampDifferenceExceeds(starttime, endtime,
733  params->log_min_duration))
734  {
735  long delay_in_ms;
736  double read_rate = 0;
737  double write_rate = 0;
739 
740  /*
741  * Calculate the difference in the Page Hit/Miss/Dirty that
742  * happened as part of the analyze by subtracting out the
743  * pre-analyze values which we saved above.
744  */
745  AnalyzePageHit = VacuumPageHit - AnalyzePageHit;
746  AnalyzePageMiss = VacuumPageMiss - AnalyzePageMiss;
747  AnalyzePageDirty = VacuumPageDirty - AnalyzePageDirty;
748 
749  /*
750  * We do not expect an analyze to take > 25 days and it simplifies
751  * things a bit to use TimestampDifferenceMilliseconds.
752  */
753  delay_in_ms = TimestampDifferenceMilliseconds(starttime, endtime);
754 
755  /*
756  * Note that we are reporting these read/write rates in the same
757  * manner as VACUUM does, which means that while the 'average read
758  * rate' here actually corresponds to page misses and resulting
759  * reads which are also picked up by track_io_timing, if enabled,
760  * the 'average write rate' is actually talking about the rate of
761  * pages being dirtied, not being written out, so it's typical to
762  * have a non-zero 'avg write rate' while I/O timings only reports
763  * reads.
764  *
765  * It's not clear that an ANALYZE will ever result in
766  * FlushBuffer() being called, but we track and support reporting
767  * on I/O write time in case that changes as it's practically free
768  * to do so anyway.
769  */
770 
771  if (delay_in_ms > 0)
772  {
773  read_rate = (double) BLCKSZ * AnalyzePageMiss / (1024 * 1024) /
774  (delay_in_ms / 1000.0);
775  write_rate = (double) BLCKSZ * AnalyzePageDirty / (1024 * 1024) /
776  (delay_in_ms / 1000.0);
777  }
778 
779  /*
780  * We split this up so we don't emit empty I/O timing values when
781  * track_io_timing isn't enabled.
782  */
783 
785  appendStringInfo(&buf, _("automatic analyze of table \"%s.%s.%s\"\n"),
788  RelationGetRelationName(onerel));
789  if (track_io_timing)
790  {
791  double read_ms = (double) (pgStatBlockReadTime - startreadtime) / 1000;
792  double write_ms = (double) (pgStatBlockWriteTime - startwritetime) / 1000;
793 
794  appendStringInfo(&buf, _("I/O timings: read: %.3f ms, write: %.3f ms\n"),
795  read_ms, write_ms);
796  }
797  appendStringInfo(&buf, _("avg read rate: %.3f MB/s, avg write rate: %.3f MB/s\n"),
798  read_rate, write_rate);
799  appendStringInfo(&buf, _("buffer usage: %lld hits, %lld misses, %lld dirtied\n"),
800  (long long) AnalyzePageHit,
801  (long long) AnalyzePageMiss,
802  (long long) AnalyzePageDirty);
803  appendStringInfo(&buf, _("system usage: %s"), pg_rusage_show(&ru0));
804 
805  ereport(LOG,
806  (errmsg_internal("%s", buf.data)));
807 
808  pfree(buf.data);
809  }
810  }
811 
812  /* Roll back any GUC changes executed by index functions */
813  AtEOXact_GUC(false, save_nestlevel);
814 
815  /* Restore userid and security context */
816  SetUserIdAndSecContext(save_userid, save_sec_context);
817 
818  /* Restore current context and release memory */
819  MemoryContextSwitchTo(caller_context);
821  anl_context = NULL;
822 }
823 
824 /*
825  * Compute statistics about indexes of a relation
826  */
827 static void
828 compute_index_stats(Relation onerel, double totalrows,
829  AnlIndexData *indexdata, int nindexes,
830  HeapTuple *rows, int numrows,
831  MemoryContext col_context)
832 {
833  MemoryContext ind_context,
834  old_context;
836  bool isnull[INDEX_MAX_KEYS];
837  int ind,
838  i;
839 
840  ind_context = AllocSetContextCreate(anl_context,
841  "Analyze Index",
843  old_context = MemoryContextSwitchTo(ind_context);
844 
845  for (ind = 0; ind < nindexes; ind++)
846  {
847  AnlIndexData *thisdata = &indexdata[ind];
848  IndexInfo *indexInfo = thisdata->indexInfo;
849  int attr_cnt = thisdata->attr_cnt;
850  TupleTableSlot *slot;
851  EState *estate;
852  ExprContext *econtext;
853  ExprState *predicate;
854  Datum *exprvals;
855  bool *exprnulls;
856  int numindexrows,
857  tcnt,
858  rowno;
859  double totalindexrows;
860 
861  /* Ignore index if no columns to analyze and not partial */
862  if (attr_cnt == 0 && indexInfo->ii_Predicate == NIL)
863  continue;
864 
865  /*
866  * Need an EState for evaluation of index expressions and
867  * partial-index predicates. Create it in the per-index context to be
868  * sure it gets cleaned up at the bottom of the loop.
869  */
870  estate = CreateExecutorState();
871  econtext = GetPerTupleExprContext(estate);
872  /* Need a slot to hold the current heap tuple, too */
874  &TTSOpsHeapTuple);
875 
876  /* Arrange for econtext's scan tuple to be the tuple under test */
877  econtext->ecxt_scantuple = slot;
878 
879  /* Set up execution state for predicate. */
880  predicate = ExecPrepareQual(indexInfo->ii_Predicate, estate);
881 
882  /* Compute and save index expression values */
883  exprvals = (Datum *) palloc(numrows * attr_cnt * sizeof(Datum));
884  exprnulls = (bool *) palloc(numrows * attr_cnt * sizeof(bool));
885  numindexrows = 0;
886  tcnt = 0;
887  for (rowno = 0; rowno < numrows; rowno++)
888  {
889  HeapTuple heapTuple = rows[rowno];
890 
892 
893  /*
894  * Reset the per-tuple context each time, to reclaim any cruft
895  * left behind by evaluating the predicate or index expressions.
896  */
897  ResetExprContext(econtext);
898 
899  /* Set up for predicate or expression evaluation */
900  ExecStoreHeapTuple(heapTuple, slot, false);
901 
902  /* If index is partial, check predicate */
903  if (predicate != NULL)
904  {
905  if (!ExecQual(predicate, econtext))
906  continue;
907  }
908  numindexrows++;
909 
910  if (attr_cnt > 0)
911  {
912  /*
913  * Evaluate the index row to compute expression values. We
914  * could do this by hand, but FormIndexDatum is convenient.
915  */
916  FormIndexDatum(indexInfo,
917  slot,
918  estate,
919  values,
920  isnull);
921 
922  /*
923  * Save just the columns we care about. We copy the values
924  * into ind_context from the estate's per-tuple context.
925  */
926  for (i = 0; i < attr_cnt; i++)
927  {
928  VacAttrStats *stats = thisdata->vacattrstats[i];
929  int attnum = stats->tupattnum;
930 
931  if (isnull[attnum - 1])
932  {
933  exprvals[tcnt] = (Datum) 0;
934  exprnulls[tcnt] = true;
935  }
936  else
937  {
938  exprvals[tcnt] = datumCopy(values[attnum - 1],
939  stats->attrtype->typbyval,
940  stats->attrtype->typlen);
941  exprnulls[tcnt] = false;
942  }
943  tcnt++;
944  }
945  }
946  }
947 
948  /*
949  * Having counted the number of rows that pass the predicate in the
950  * sample, we can estimate the total number of rows in the index.
951  */
952  thisdata->tupleFract = (double) numindexrows / (double) numrows;
953  totalindexrows = ceil(thisdata->tupleFract * totalrows);
954 
955  /*
956  * Now we can compute the statistics for the expression columns.
957  */
958  if (numindexrows > 0)
959  {
960  MemoryContextSwitchTo(col_context);
961  for (i = 0; i < attr_cnt; i++)
962  {
963  VacAttrStats *stats = thisdata->vacattrstats[i];
964 
965  stats->exprvals = exprvals + i;
966  stats->exprnulls = exprnulls + i;
967  stats->rowstride = attr_cnt;
968  stats->compute_stats(stats,
970  numindexrows,
971  totalindexrows);
972 
973  MemoryContextReset(col_context);
974  }
975  }
976 
977  /* And clean up */
978  MemoryContextSwitchTo(ind_context);
979 
981  FreeExecutorState(estate);
982  MemoryContextReset(ind_context);
983  }
984 
985  MemoryContextSwitchTo(old_context);
986  MemoryContextDelete(ind_context);
987 }
988 
989 /*
990  * examine_attribute -- pre-analysis of a single column
991  *
992  * Determine whether the column is analyzable; if so, create and initialize
993  * a VacAttrStats struct for it. If not, return NULL.
994  *
995  * If index_expr isn't NULL, then we're trying to analyze an expression index,
996  * and index_expr is the expression tree representing the column's data.
997  */
998 static VacAttrStats *
999 examine_attribute(Relation onerel, int attnum, Node *index_expr)
1000 {
1001  Form_pg_attribute attr = TupleDescAttr(onerel->rd_att, attnum - 1);
1002  int attstattarget;
1003  HeapTuple atttuple;
1004  Datum dat;
1005  bool isnull;
1006  HeapTuple typtuple;
1007  VacAttrStats *stats;
1008  int i;
1009  bool ok;
1010 
1011  /* Never analyze dropped columns */
1012  if (attr->attisdropped)
1013  return NULL;
1014 
1015  /*
1016  * Get attstattarget value. Set to -1 if null. (Analyze functions expect
1017  * -1 to mean use default_statistics_target; see for example
1018  * std_typanalyze.)
1019  */
1020  atttuple = SearchSysCache2(ATTNUM, ObjectIdGetDatum(RelationGetRelid(onerel)), Int16GetDatum(attnum));
1021  if (!HeapTupleIsValid(atttuple))
1022  elog(ERROR, "cache lookup failed for attribute %d of relation %u",
1023  attnum, RelationGetRelid(onerel));
1024  dat = SysCacheGetAttr(ATTNUM, atttuple, Anum_pg_attribute_attstattarget, &isnull);
1025  attstattarget = isnull ? -1 : DatumGetInt16(dat);
1026  ReleaseSysCache(atttuple);
1027 
1028  /* Don't analyze column if user has specified not to */
1029  if (attstattarget == 0)
1030  return NULL;
1031 
1032  /*
1033  * Create the VacAttrStats struct.
1034  */
1035  stats = (VacAttrStats *) palloc0(sizeof(VacAttrStats));
1036  stats->attstattarget = attstattarget;
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);
1104  return NULL;
1105  }
1106 
1107  return stats;
1108 }
1109 
1110 /*
1111  * Read stream callback returning the next BlockNumber as chosen by the
1112  * BlockSampling algorithm.
1113  */
1114 static BlockNumber
1116  void *callback_private_data,
1117  void *per_buffer_data)
1118 {
1119  BlockSamplerData *bs = callback_private_data;
1120 
1122 }
1123 
1124 /*
1125  * acquire_sample_rows -- acquire a random sample of rows from the table
1126  *
1127  * Selected rows are returned in the caller-allocated array rows[], which
1128  * must have at least targrows entries.
1129  * The actual number of rows selected is returned as the function result.
1130  * We also estimate the total numbers of live and dead rows in the table,
1131  * and return them into *totalrows and *totaldeadrows, respectively.
1132  *
1133  * The returned list of tuples is in order by physical position in the table.
1134  * (We will rely on this later to derive correlation estimates.)
1135  *
1136  * As of May 2004 we use a new two-stage method: Stage one selects up
1137  * to targrows random blocks (or all blocks, if there aren't so many).
1138  * Stage two scans these blocks and uses the Vitter algorithm to create
1139  * a random sample of targrows rows (or less, if there are less in the
1140  * sample of blocks). The two stages are executed simultaneously: each
1141  * block is processed as soon as stage one returns its number and while
1142  * the rows are read stage two controls which ones are to be inserted
1143  * into the sample.
1144  *
1145  * Although every row has an equal chance of ending up in the final
1146  * sample, this sampling method is not perfect: not every possible
1147  * sample has an equal chance of being selected. For large relations
1148  * the number of different blocks represented by the sample tends to be
1149  * too small. We can live with that for now. Improvements are welcome.
1150  *
1151  * An important property of this sampling method is that because we do
1152  * look at a statistically unbiased set of blocks, we should get
1153  * unbiased estimates of the average numbers of live and dead rows per
1154  * block. The previous sampling method put too much credence in the row
1155  * density near the start of the table.
1156  */
1157 static int
1158 acquire_sample_rows(Relation onerel, int elevel,
1159  HeapTuple *rows, int targrows,
1160  double *totalrows, double *totaldeadrows)
1161 {
1162  int numrows = 0; /* # rows now in reservoir */
1163  double samplerows = 0; /* total # rows collected */
1164  double liverows = 0; /* # live rows seen */
1165  double deadrows = 0; /* # dead rows seen */
1166  double rowstoskip = -1; /* -1 means not set yet */
1167  uint32 randseed; /* Seed for block sampler(s) */
1168  BlockNumber totalblocks;
1169  TransactionId OldestXmin;
1170  BlockSamplerData bs;
1171  ReservoirStateData rstate;
1172  TupleTableSlot *slot;
1173  TableScanDesc scan;
1174  BlockNumber nblocks;
1175  BlockNumber blksdone = 0;
1176  ReadStream *stream;
1177 
1178  Assert(targrows > 0);
1179 
1180  totalblocks = RelationGetNumberOfBlocks(onerel);
1181 
1182  /* Need a cutoff xmin for HeapTupleSatisfiesVacuum */
1183  OldestXmin = GetOldestNonRemovableTransactionId(onerel);
1184 
1185  /* Prepare for sampling block numbers */
1186  randseed = pg_prng_uint32(&pg_global_prng_state);
1187  nblocks = BlockSampler_Init(&bs, totalblocks, targrows, randseed);
1188 
1189  /* Report sampling block numbers */
1191  nblocks);
1192 
1193  /* Prepare for sampling rows */
1194  reservoir_init_selection_state(&rstate, targrows);
1195 
1196  scan = table_beginscan_analyze(onerel);
1197  slot = table_slot_create(onerel, NULL);
1198 
1200  vac_strategy,
1201  scan->rs_rd,
1202  MAIN_FORKNUM,
1204  &bs,
1205  0);
1206 
1207  /* Outer loop over blocks to sample */
1208  while (table_scan_analyze_next_block(scan, stream))
1209  {
1211 
1212  while (table_scan_analyze_next_tuple(scan, OldestXmin, &liverows, &deadrows, slot))
1213  {
1214  /*
1215  * The first targrows sample rows are simply copied into the
1216  * reservoir. Then we start replacing tuples in the sample until
1217  * we reach the end of the relation. This algorithm is from Jeff
1218  * Vitter's paper (see full citation in utils/misc/sampling.c). It
1219  * works by repeatedly computing the number of tuples to skip
1220  * before selecting a tuple, which replaces a randomly chosen
1221  * element of the reservoir (current set of tuples). At all times
1222  * the reservoir is a true random sample of the tuples we've
1223  * passed over so far, so when we fall off the end of the relation
1224  * we're done.
1225  */
1226  if (numrows < targrows)
1227  rows[numrows++] = ExecCopySlotHeapTuple(slot);
1228  else
1229  {
1230  /*
1231  * t in Vitter's paper is the number of records already
1232  * processed. If we need to compute a new S value, we must
1233  * use the not-yet-incremented value of samplerows as t.
1234  */
1235  if (rowstoskip < 0)
1236  rowstoskip = reservoir_get_next_S(&rstate, samplerows, targrows);
1237 
1238  if (rowstoskip <= 0)
1239  {
1240  /*
1241  * Found a suitable tuple, so save it, replacing one old
1242  * tuple at random
1243  */
1244  int k = (int) (targrows * sampler_random_fract(&rstate.randstate));
1245 
1246  Assert(k >= 0 && k < targrows);
1247  heap_freetuple(rows[k]);
1248  rows[k] = ExecCopySlotHeapTuple(slot);
1249  }
1250 
1251  rowstoskip -= 1;
1252  }
1253 
1254  samplerows += 1;
1255  }
1256 
1258  ++blksdone);
1259  }
1260 
1261  read_stream_end(stream);
1262 
1264  table_endscan(scan);
1265 
1266  /*
1267  * If we didn't find as many tuples as we wanted then we're done. No sort
1268  * is needed, since they're already in order.
1269  *
1270  * Otherwise we need to sort the collected tuples by position
1271  * (itempointer). It's not worth worrying about corner cases where the
1272  * tuples are already sorted.
1273  */
1274  if (numrows == targrows)
1275  qsort_interruptible(rows, numrows, sizeof(HeapTuple),
1276  compare_rows, NULL);
1277 
1278  /*
1279  * Estimate total numbers of live and dead rows in relation, extrapolating
1280  * on the assumption that the average tuple density in pages we didn't
1281  * scan is the same as in the pages we did scan. Since what we scanned is
1282  * a random sample of the pages in the relation, this should be a good
1283  * assumption.
1284  */
1285  if (bs.m > 0)
1286  {
1287  *totalrows = floor((liverows / bs.m) * totalblocks + 0.5);
1288  *totaldeadrows = floor((deadrows / bs.m) * totalblocks + 0.5);
1289  }
1290  else
1291  {
1292  *totalrows = 0.0;
1293  *totaldeadrows = 0.0;
1294  }
1295 
1296  /*
1297  * Emit some interesting relation info
1298  */
1299  ereport(elevel,
1300  (errmsg("\"%s\": scanned %d of %u pages, "
1301  "containing %.0f live rows and %.0f dead rows; "
1302  "%d rows in sample, %.0f estimated total rows",
1303  RelationGetRelationName(onerel),
1304  bs.m, totalblocks,
1305  liverows, deadrows,
1306  numrows, *totalrows)));
1307 
1308  return numrows;
1309 }
1310 
1311 /*
1312  * Comparator for sorting rows[] array
1313  */
1314 static int
1315 compare_rows(const void *a, const void *b, void *arg)
1316 {
1317  HeapTuple ha = *(const HeapTuple *) a;
1318  HeapTuple hb = *(const HeapTuple *) b;
1323 
1324  if (ba < bb)
1325  return -1;
1326  if (ba > bb)
1327  return 1;
1328  if (oa < ob)
1329  return -1;
1330  if (oa > ob)
1331  return 1;
1332  return 0;
1333 }
1334 
1335 
1336 /*
1337  * acquire_inherited_sample_rows -- acquire sample rows from inheritance tree
1338  *
1339  * This has the same API as acquire_sample_rows, except that rows are
1340  * collected from all inheritance children as well as the specified table.
1341  * We fail and return zero if there are no inheritance children, or if all
1342  * children are foreign tables that don't support ANALYZE.
1343  */
1344 static int
1346  HeapTuple *rows, int targrows,
1347  double *totalrows, double *totaldeadrows)
1348 {
1349  List *tableOIDs;
1350  Relation *rels;
1351  AcquireSampleRowsFunc *acquirefuncs;
1352  double *relblocks;
1353  double totalblocks;
1354  int numrows,
1355  nrels,
1356  i;
1357  ListCell *lc;
1358  bool has_child;
1359 
1360  /* Initialize output parameters to zero now, in case we exit early */
1361  *totalrows = 0;
1362  *totaldeadrows = 0;
1363 
1364  /*
1365  * Find all members of inheritance set. We only need AccessShareLock on
1366  * the children.
1367  */
1368  tableOIDs =
1370 
1371  /*
1372  * Check that there's at least one descendant, else fail. This could
1373  * happen despite analyze_rel's relhassubclass check, if table once had a
1374  * child but no longer does. In that case, we can clear the
1375  * relhassubclass field so as not to make the same mistake again later.
1376  * (This is safe because we hold ShareUpdateExclusiveLock.)
1377  */
1378  if (list_length(tableOIDs) < 2)
1379  {
1380  /* CCI because we already updated the pg_class row in this command */
1382  SetRelationHasSubclass(RelationGetRelid(onerel), false);
1383  ereport(elevel,
1384  (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no child tables",
1386  RelationGetRelationName(onerel))));
1387  return 0;
1388  }
1389 
1390  /*
1391  * Identify acquirefuncs to use, and count blocks in all the relations.
1392  * The result could overflow BlockNumber, so we use double arithmetic.
1393  */
1394  rels = (Relation *) palloc(list_length(tableOIDs) * sizeof(Relation));
1395  acquirefuncs = (AcquireSampleRowsFunc *)
1396  palloc(list_length(tableOIDs) * sizeof(AcquireSampleRowsFunc));
1397  relblocks = (double *) palloc(list_length(tableOIDs) * sizeof(double));
1398  totalblocks = 0;
1399  nrels = 0;
1400  has_child = false;
1401  foreach(lc, tableOIDs)
1402  {
1403  Oid childOID = lfirst_oid(lc);
1404  Relation childrel;
1405  AcquireSampleRowsFunc acquirefunc = NULL;
1406  BlockNumber relpages = 0;
1407 
1408  /* We already got the needed lock */
1409  childrel = table_open(childOID, NoLock);
1410 
1411  /* Ignore if temp table of another backend */
1412  if (RELATION_IS_OTHER_TEMP(childrel))
1413  {
1414  /* ... but release the lock on it */
1415  Assert(childrel != onerel);
1416  table_close(childrel, AccessShareLock);
1417  continue;
1418  }
1419 
1420  /* Check table type (MATVIEW can't happen, but might as well allow) */
1421  if (childrel->rd_rel->relkind == RELKIND_RELATION ||
1422  childrel->rd_rel->relkind == RELKIND_MATVIEW)
1423  {
1424  /* Regular table, so use the regular row acquisition function */
1425  acquirefunc = acquire_sample_rows;
1426  relpages = RelationGetNumberOfBlocks(childrel);
1427  }
1428  else if (childrel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
1429  {
1430  /*
1431  * For a foreign table, call the FDW's hook function to see
1432  * whether it supports analysis.
1433  */
1434  FdwRoutine *fdwroutine;
1435  bool ok = false;
1436 
1437  fdwroutine = GetFdwRoutineForRelation(childrel, false);
1438 
1439  if (fdwroutine->AnalyzeForeignTable != NULL)
1440  ok = fdwroutine->AnalyzeForeignTable(childrel,
1441  &acquirefunc,
1442  &relpages);
1443 
1444  if (!ok)
1445  {
1446  /* ignore, but release the lock on it */
1447  Assert(childrel != onerel);
1448  table_close(childrel, AccessShareLock);
1449  continue;
1450  }
1451  }
1452  else
1453  {
1454  /*
1455  * ignore, but release the lock on it. don't try to unlock the
1456  * passed-in relation
1457  */
1458  Assert(childrel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE);
1459  if (childrel != onerel)
1460  table_close(childrel, AccessShareLock);
1461  else
1462  table_close(childrel, NoLock);
1463  continue;
1464  }
1465 
1466  /* OK, we'll process this child */
1467  has_child = true;
1468  rels[nrels] = childrel;
1469  acquirefuncs[nrels] = acquirefunc;
1470  relblocks[nrels] = (double) relpages;
1471  totalblocks += (double) relpages;
1472  nrels++;
1473  }
1474 
1475  /*
1476  * If we don't have at least one child table to consider, fail. If the
1477  * relation is a partitioned table, it's not counted as a child table.
1478  */
1479  if (!has_child)
1480  {
1481  ereport(elevel,
1482  (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no analyzable child tables",
1484  RelationGetRelationName(onerel))));
1485  return 0;
1486  }
1487 
1488  /*
1489  * Now sample rows from each relation, proportionally to its fraction of
1490  * the total block count. (This might be less than desirable if the child
1491  * rels have radically different free-space percentages, but it's not
1492  * clear that it's worth working harder.)
1493  */
1495  nrels);
1496  numrows = 0;
1497  for (i = 0; i < nrels; i++)
1498  {
1499  Relation childrel = rels[i];
1500  AcquireSampleRowsFunc acquirefunc = acquirefuncs[i];
1501  double childblocks = relblocks[i];
1502 
1503  /*
1504  * Report progress. The sampling function will normally report blocks
1505  * done/total, but we need to reset them to 0 here, so that they don't
1506  * show an old value until that.
1507  */
1508  {
1509  const int progress_index[] = {
1513  };
1514  const int64 progress_vals[] = {
1515  RelationGetRelid(childrel),
1516  0,
1517  0,
1518  };
1519 
1520  pgstat_progress_update_multi_param(3, progress_index, progress_vals);
1521  }
1522 
1523  if (childblocks > 0)
1524  {
1525  int childtargrows;
1526 
1527  childtargrows = (int) rint(targrows * childblocks / totalblocks);
1528  /* Make sure we don't overrun due to roundoff error */
1529  childtargrows = Min(childtargrows, targrows - numrows);
1530  if (childtargrows > 0)
1531  {
1532  int childrows;
1533  double trows,
1534  tdrows;
1535 
1536  /* Fetch a random sample of the child's rows */
1537  childrows = (*acquirefunc) (childrel, elevel,
1538  rows + numrows, childtargrows,
1539  &trows, &tdrows);
1540 
1541  /* We may need to convert from child's rowtype to parent's */
1542  if (childrows > 0 &&
1543  !equalRowTypes(RelationGetDescr(childrel),
1544  RelationGetDescr(onerel)))
1545  {
1546  TupleConversionMap *map;
1547 
1548  map = convert_tuples_by_name(RelationGetDescr(childrel),
1549  RelationGetDescr(onerel));
1550  if (map != NULL)
1551  {
1552  int j;
1553 
1554  for (j = 0; j < childrows; j++)
1555  {
1556  HeapTuple newtup;
1557 
1558  newtup = execute_attr_map_tuple(rows[numrows + j], map);
1559  heap_freetuple(rows[numrows + j]);
1560  rows[numrows + j] = newtup;
1561  }
1562  free_conversion_map(map);
1563  }
1564  }
1565 
1566  /* And add to counts */
1567  numrows += childrows;
1568  *totalrows += trows;
1569  *totaldeadrows += tdrows;
1570  }
1571  }
1572 
1573  /*
1574  * Note: we cannot release the child-table locks, since we may have
1575  * pointers to their TOAST tables in the sampled rows.
1576  */
1577  table_close(childrel, NoLock);
1579  i + 1);
1580  }
1581 
1582  return numrows;
1583 }
1584 
1585 
1586 /*
1587  * update_attstats() -- update attribute statistics for one relation
1588  *
1589  * Statistics are stored in several places: the pg_class row for the
1590  * relation has stats about the whole relation, and there is a
1591  * pg_statistic row for each (non-system) attribute that has ever
1592  * been analyzed. The pg_class values are updated by VACUUM, not here.
1593  *
1594  * pg_statistic rows are just added or updated normally. This means
1595  * that pg_statistic will probably contain some deleted rows at the
1596  * completion of a vacuum cycle, unless it happens to get vacuumed last.
1597  *
1598  * To keep things simple, we punt for pg_statistic, and don't try
1599  * to compute or store rows for pg_statistic itself in pg_statistic.
1600  * This could possibly be made to work, but it's not worth the trouble.
1601  * Note analyze_rel() has seen to it that we won't come here when
1602  * vacuuming pg_statistic itself.
1603  *
1604  * Note: there would be a race condition here if two backends could
1605  * ANALYZE the same table concurrently. Presently, we lock that out
1606  * by taking a self-exclusive lock on the relation in analyze_rel().
1607  */
1608 static void
1609 update_attstats(Oid relid, bool inh, int natts, VacAttrStats **vacattrstats)
1610 {
1611  Relation sd;
1612  int attno;
1613  CatalogIndexState indstate = NULL;
1614 
1615  if (natts <= 0)
1616  return; /* nothing to do */
1617 
1618  sd = table_open(StatisticRelationId, RowExclusiveLock);
1619 
1620  for (attno = 0; attno < natts; attno++)
1621  {
1622  VacAttrStats *stats = vacattrstats[attno];
1623  HeapTuple stup,
1624  oldtup;
1625  int i,
1626  k,
1627  n;
1628  Datum values[Natts_pg_statistic];
1629  bool nulls[Natts_pg_statistic];
1630  bool replaces[Natts_pg_statistic];
1631 
1632  /* Ignore attr if we weren't able to collect stats */
1633  if (!stats->stats_valid)
1634  continue;
1635 
1636  /*
1637  * Construct a new pg_statistic tuple
1638  */
1639  for (i = 0; i < Natts_pg_statistic; ++i)
1640  {
1641  nulls[i] = false;
1642  replaces[i] = true;
1643  }
1644 
1645  values[Anum_pg_statistic_starelid - 1] = ObjectIdGetDatum(relid);
1646  values[Anum_pg_statistic_staattnum - 1] = Int16GetDatum(stats->tupattnum);
1647  values[Anum_pg_statistic_stainherit - 1] = BoolGetDatum(inh);
1648  values[Anum_pg_statistic_stanullfrac - 1] = Float4GetDatum(stats->stanullfrac);
1649  values[Anum_pg_statistic_stawidth - 1] = Int32GetDatum(stats->stawidth);
1650  values[Anum_pg_statistic_stadistinct - 1] = Float4GetDatum(stats->stadistinct);
1651  i = Anum_pg_statistic_stakind1 - 1;
1652  for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1653  {
1654  values[i++] = Int16GetDatum(stats->stakind[k]); /* stakindN */
1655  }
1656  i = Anum_pg_statistic_staop1 - 1;
1657  for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1658  {
1659  values[i++] = ObjectIdGetDatum(stats->staop[k]); /* staopN */
1660  }
1661  i = Anum_pg_statistic_stacoll1 - 1;
1662  for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1663  {
1664  values[i++] = ObjectIdGetDatum(stats->stacoll[k]); /* stacollN */
1665  }
1666  i = Anum_pg_statistic_stanumbers1 - 1;
1667  for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1668  {
1669  int nnum = stats->numnumbers[k];
1670 
1671  if (nnum > 0)
1672  {
1673  Datum *numdatums = (Datum *) palloc(nnum * sizeof(Datum));
1674  ArrayType *arry;
1675 
1676  for (n = 0; n < nnum; n++)
1677  numdatums[n] = Float4GetDatum(stats->stanumbers[k][n]);
1678  arry = construct_array_builtin(numdatums, nnum, FLOAT4OID);
1679  values[i++] = PointerGetDatum(arry); /* stanumbersN */
1680  }
1681  else
1682  {
1683  nulls[i] = true;
1684  values[i++] = (Datum) 0;
1685  }
1686  }
1687  i = Anum_pg_statistic_stavalues1 - 1;
1688  for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1689  {
1690  if (stats->numvalues[k] > 0)
1691  {
1692  ArrayType *arry;
1693 
1694  arry = construct_array(stats->stavalues[k],
1695  stats->numvalues[k],
1696  stats->statypid[k],
1697  stats->statyplen[k],
1698  stats->statypbyval[k],
1699  stats->statypalign[k]);
1700  values[i++] = PointerGetDatum(arry); /* stavaluesN */
1701  }
1702  else
1703  {
1704  nulls[i] = true;
1705  values[i++] = (Datum) 0;
1706  }
1707  }
1708 
1709  /* Is there already a pg_statistic tuple for this attribute? */
1710  oldtup = SearchSysCache3(STATRELATTINH,
1711  ObjectIdGetDatum(relid),
1712  Int16GetDatum(stats->tupattnum),
1713  BoolGetDatum(inh));
1714 
1715  /* Open index information when we know we need it */
1716  if (indstate == NULL)
1717  indstate = CatalogOpenIndexes(sd);
1718 
1719  if (HeapTupleIsValid(oldtup))
1720  {
1721  /* Yes, replace it */
1722  stup = heap_modify_tuple(oldtup,
1723  RelationGetDescr(sd),
1724  values,
1725  nulls,
1726  replaces);
1727  ReleaseSysCache(oldtup);
1728  CatalogTupleUpdateWithInfo(sd, &stup->t_self, stup, indstate);
1729  }
1730  else
1731  {
1732  /* No, insert new tuple */
1733  stup = heap_form_tuple(RelationGetDescr(sd), values, nulls);
1734  CatalogTupleInsertWithInfo(sd, stup, indstate);
1735  }
1736 
1737  heap_freetuple(stup);
1738  }
1739 
1740  if (indstate != NULL)
1741  CatalogCloseIndexes(indstate);
1743 }
1744 
1745 /*
1746  * Standard fetch function for use by compute_stats subroutines.
1747  *
1748  * This exists to provide some insulation between compute_stats routines
1749  * and the actual storage of the sample data.
1750  */
1751 static Datum
1752 std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1753 {
1754  int attnum = stats->tupattnum;
1755  HeapTuple tuple = stats->rows[rownum];
1756  TupleDesc tupDesc = stats->tupDesc;
1757 
1758  return heap_getattr(tuple, attnum, tupDesc, isNull);
1759 }
1760 
1761 /*
1762  * Fetch function for analyzing index expressions.
1763  *
1764  * We have not bothered to construct index tuples, instead the data is
1765  * just in Datum arrays.
1766  */
1767 static Datum
1768 ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1769 {
1770  int i;
1771 
1772  /* exprvals and exprnulls are already offset for proper column */
1773  i = rownum * stats->rowstride;
1774  *isNull = stats->exprnulls[i];
1775  return stats->exprvals[i];
1776 }
1777 
1778 
1779 /*==========================================================================
1780  *
1781  * Code below this point represents the "standard" type-specific statistics
1782  * analysis algorithms. This code can be replaced on a per-data-type basis
1783  * by setting a nonzero value in pg_type.typanalyze.
1784  *
1785  *==========================================================================
1786  */
1787 
1788 
1789 /*
1790  * To avoid consuming too much memory during analysis and/or too much space
1791  * in the resulting pg_statistic rows, we ignore varlena datums that are wider
1792  * than WIDTH_THRESHOLD (after detoasting!). This is legitimate for MCV
1793  * and distinct-value calculations since a wide value is unlikely to be
1794  * duplicated at all, much less be a most-common value. For the same reason,
1795  * ignoring wide values will not affect our estimates of histogram bin
1796  * boundaries very much.
1797  */
1798 #define WIDTH_THRESHOLD 1024
1799 
1800 #define swapInt(a,b) do {int _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1801 #define swapDatum(a,b) do {Datum _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1802 
1803 /*
1804  * Extra information used by the default analysis routines
1805  */
1806 typedef struct
1807 {
1808  int count; /* # of duplicates */
1809  int first; /* values[] index of first occurrence */
1810 } ScalarMCVItem;
1811 
1812 typedef struct
1813 {
1817 
1818 
1819 static void compute_trivial_stats(VacAttrStatsP stats,
1820  AnalyzeAttrFetchFunc fetchfunc,
1821  int samplerows,
1822  double totalrows);
1823 static void compute_distinct_stats(VacAttrStatsP stats,
1824  AnalyzeAttrFetchFunc fetchfunc,
1825  int samplerows,
1826  double totalrows);
1827 static void compute_scalar_stats(VacAttrStatsP stats,
1828  AnalyzeAttrFetchFunc fetchfunc,
1829  int samplerows,
1830  double totalrows);
1831 static int compare_scalars(const void *a, const void *b, void *arg);
1832 static int compare_mcvs(const void *a, const void *b, void *arg);
1833 static int analyze_mcv_list(int *mcv_counts,
1834  int num_mcv,
1835  double stadistinct,
1836  double stanullfrac,
1837  int samplerows,
1838  double totalrows);
1839 
1840 
1841 /*
1842  * std_typanalyze -- the default type-specific typanalyze function
1843  */
1844 bool
1846 {
1847  Oid ltopr;
1848  Oid eqopr;
1849  StdAnalyzeData *mystats;
1850 
1851  /* If the attstattarget column is negative, use the default value */
1852  if (stats->attstattarget < 0)
1854 
1855  /* Look for default "<" and "=" operators for column's type */
1857  false, false, false,
1858  &ltopr, &eqopr, NULL,
1859  NULL);
1860 
1861  /* Save the operator info for compute_stats routines */
1862  mystats = (StdAnalyzeData *) palloc(sizeof(StdAnalyzeData));
1863  mystats->eqopr = eqopr;
1864  mystats->eqfunc = OidIsValid(eqopr) ? get_opcode(eqopr) : InvalidOid;
1865  mystats->ltopr = ltopr;
1866  stats->extra_data = mystats;
1867 
1868  /*
1869  * Determine which standard statistics algorithm to use
1870  */
1871  if (OidIsValid(eqopr) && OidIsValid(ltopr))
1872  {
1873  /* Seems to be a scalar datatype */
1875  /*--------------------
1876  * The following choice of minrows is based on the paper
1877  * "Random sampling for histogram construction: how much is enough?"
1878  * by Surajit Chaudhuri, Rajeev Motwani and Vivek Narasayya, in
1879  * Proceedings of ACM SIGMOD International Conference on Management
1880  * of Data, 1998, Pages 436-447. Their Corollary 1 to Theorem 5
1881  * says that for table size n, histogram size k, maximum relative
1882  * error in bin size f, and error probability gamma, the minimum
1883  * random sample size is
1884  * r = 4 * k * ln(2*n/gamma) / f^2
1885  * Taking f = 0.5, gamma = 0.01, n = 10^6 rows, we obtain
1886  * r = 305.82 * k
1887  * Note that because of the log function, the dependence on n is
1888  * quite weak; even at n = 10^12, a 300*k sample gives <= 0.66
1889  * bin size error with probability 0.99. So there's no real need to
1890  * scale for n, which is a good thing because we don't necessarily
1891  * know it at this point.
1892  *--------------------
1893  */
1894  stats->minrows = 300 * stats->attstattarget;
1895  }
1896  else if (OidIsValid(eqopr))
1897  {
1898  /* We can still recognize distinct values */
1900  /* Might as well use the same minrows as above */
1901  stats->minrows = 300 * stats->attstattarget;
1902  }
1903  else
1904  {
1905  /* Can't do much but the trivial stuff */
1907  /* Might as well use the same minrows as above */
1908  stats->minrows = 300 * stats->attstattarget;
1909  }
1910 
1911  return true;
1912 }
1913 
1914 
1915 /*
1916  * compute_trivial_stats() -- compute very basic column statistics
1917  *
1918  * We use this when we cannot find a hash "=" operator for the datatype.
1919  *
1920  * We determine the fraction of non-null rows and the average datum width.
1921  */
1922 static void
1924  AnalyzeAttrFetchFunc fetchfunc,
1925  int samplerows,
1926  double totalrows)
1927 {
1928  int i;
1929  int null_cnt = 0;
1930  int nonnull_cnt = 0;
1931  double total_width = 0;
1932  bool is_varlena = (!stats->attrtype->typbyval &&
1933  stats->attrtype->typlen == -1);
1934  bool is_varwidth = (!stats->attrtype->typbyval &&
1935  stats->attrtype->typlen < 0);
1936 
1937  for (i = 0; i < samplerows; i++)
1938  {
1939  Datum value;
1940  bool isnull;
1941 
1943 
1944  value = fetchfunc(stats, i, &isnull);
1945 
1946  /* Check for null/nonnull */
1947  if (isnull)
1948  {
1949  null_cnt++;
1950  continue;
1951  }
1952  nonnull_cnt++;
1953 
1954  /*
1955  * If it's a variable-width field, add up widths for average width
1956  * calculation. Note that if the value is toasted, we use the toasted
1957  * width. We don't bother with this calculation if it's a fixed-width
1958  * type.
1959  */
1960  if (is_varlena)
1961  {
1962  total_width += VARSIZE_ANY(DatumGetPointer(value));
1963  }
1964  else if (is_varwidth)
1965  {
1966  /* must be cstring */
1967  total_width += strlen(DatumGetCString(value)) + 1;
1968  }
1969  }
1970 
1971  /* We can only compute average width if we found some non-null values. */
1972  if (nonnull_cnt > 0)
1973  {
1974  stats->stats_valid = true;
1975  /* Do the simple null-frac and width stats */
1976  stats->stanullfrac = (double) null_cnt / (double) samplerows;
1977  if (is_varwidth)
1978  stats->stawidth = total_width / (double) nonnull_cnt;
1979  else
1980  stats->stawidth = stats->attrtype->typlen;
1981  stats->stadistinct = 0.0; /* "unknown" */
1982  }
1983  else if (null_cnt > 0)
1984  {
1985  /* We found only nulls; assume the column is entirely null */
1986  stats->stats_valid = true;
1987  stats->stanullfrac = 1.0;
1988  if (is_varwidth)
1989  stats->stawidth = 0; /* "unknown" */
1990  else
1991  stats->stawidth = stats->attrtype->typlen;
1992  stats->stadistinct = 0.0; /* "unknown" */
1993  }
1994 }
1995 
1996 
1997 /*
1998  * compute_distinct_stats() -- compute column statistics including ndistinct
1999  *
2000  * We use this when we can find only an "=" operator for the datatype.
2001  *
2002  * We determine the fraction of non-null rows, the average width, the
2003  * most common values, and the (estimated) number of distinct values.
2004  *
2005  * The most common values are determined by brute force: we keep a list
2006  * of previously seen values, ordered by number of times seen, as we scan
2007  * the samples. A newly seen value is inserted just after the last
2008  * multiply-seen value, causing the bottommost (oldest) singly-seen value
2009  * to drop off the list. The accuracy of this method, and also its cost,
2010  * depend mainly on the length of the list we are willing to keep.
2011  */
2012 static void
2014  AnalyzeAttrFetchFunc fetchfunc,
2015  int samplerows,
2016  double totalrows)
2017 {
2018  int i;
2019  int null_cnt = 0;
2020  int nonnull_cnt = 0;
2021  int toowide_cnt = 0;
2022  double total_width = 0;
2023  bool is_varlena = (!stats->attrtype->typbyval &&
2024  stats->attrtype->typlen == -1);
2025  bool is_varwidth = (!stats->attrtype->typbyval &&
2026  stats->attrtype->typlen < 0);
2027  FmgrInfo f_cmpeq;
2028  typedef struct
2029  {
2030  Datum value;
2031  int count;
2032  } TrackItem;
2033  TrackItem *track;
2034  int track_cnt,
2035  track_max;
2036  int num_mcv = stats->attstattarget;
2037  StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2038 
2039  /*
2040  * We track up to 2*n values for an n-element MCV list; but at least 10
2041  */
2042  track_max = 2 * num_mcv;
2043  if (track_max < 10)
2044  track_max = 10;
2045  track = (TrackItem *) palloc(track_max * sizeof(TrackItem));
2046  track_cnt = 0;
2047 
2048  fmgr_info(mystats->eqfunc, &f_cmpeq);
2049 
2050  for (i = 0; i < samplerows; i++)
2051  {
2052  Datum value;
2053  bool isnull;
2054  bool match;
2055  int firstcount1,
2056  j;
2057 
2059 
2060  value = fetchfunc(stats, i, &isnull);
2061 
2062  /* Check for null/nonnull */
2063  if (isnull)
2064  {
2065  null_cnt++;
2066  continue;
2067  }
2068  nonnull_cnt++;
2069 
2070  /*
2071  * If it's a variable-width field, add up widths for average width
2072  * calculation. Note that if the value is toasted, we use the toasted
2073  * width. We don't bother with this calculation if it's a fixed-width
2074  * type.
2075  */
2076  if (is_varlena)
2077  {
2078  total_width += VARSIZE_ANY(DatumGetPointer(value));
2079 
2080  /*
2081  * If the value is toasted, we want to detoast it just once to
2082  * avoid repeated detoastings and resultant excess memory usage
2083  * during the comparisons. Also, check to see if the value is
2084  * excessively wide, and if so don't detoast at all --- just
2085  * ignore the value.
2086  */
2088  {
2089  toowide_cnt++;
2090  continue;
2091  }
2093  }
2094  else if (is_varwidth)
2095  {
2096  /* must be cstring */
2097  total_width += strlen(DatumGetCString(value)) + 1;
2098  }
2099 
2100  /*
2101  * See if the value matches anything we're already tracking.
2102  */
2103  match = false;
2104  firstcount1 = track_cnt;
2105  for (j = 0; j < track_cnt; j++)
2106  {
2107  if (DatumGetBool(FunctionCall2Coll(&f_cmpeq,
2108  stats->attrcollid,
2109  value, track[j].value)))
2110  {
2111  match = true;
2112  break;
2113  }
2114  if (j < firstcount1 && track[j].count == 1)
2115  firstcount1 = j;
2116  }
2117 
2118  if (match)
2119  {
2120  /* Found a match */
2121  track[j].count++;
2122  /* This value may now need to "bubble up" in the track list */
2123  while (j > 0 && track[j].count > track[j - 1].count)
2124  {
2125  swapDatum(track[j].value, track[j - 1].value);
2126  swapInt(track[j].count, track[j - 1].count);
2127  j--;
2128  }
2129  }
2130  else
2131  {
2132  /* No match. Insert at head of count-1 list */
2133  if (track_cnt < track_max)
2134  track_cnt++;
2135  for (j = track_cnt - 1; j > firstcount1; j--)
2136  {
2137  track[j].value = track[j - 1].value;
2138  track[j].count = track[j - 1].count;
2139  }
2140  if (firstcount1 < track_cnt)
2141  {
2142  track[firstcount1].value = value;
2143  track[firstcount1].count = 1;
2144  }
2145  }
2146  }
2147 
2148  /* We can only compute real stats if we found some non-null values. */
2149  if (nonnull_cnt > 0)
2150  {
2151  int nmultiple,
2152  summultiple;
2153 
2154  stats->stats_valid = true;
2155  /* Do the simple null-frac and width stats */
2156  stats->stanullfrac = (double) null_cnt / (double) samplerows;
2157  if (is_varwidth)
2158  stats->stawidth = total_width / (double) nonnull_cnt;
2159  else
2160  stats->stawidth = stats->attrtype->typlen;
2161 
2162  /* Count the number of values we found multiple times */
2163  summultiple = 0;
2164  for (nmultiple = 0; nmultiple < track_cnt; nmultiple++)
2165  {
2166  if (track[nmultiple].count == 1)
2167  break;
2168  summultiple += track[nmultiple].count;
2169  }
2170 
2171  if (nmultiple == 0)
2172  {
2173  /*
2174  * If we found no repeated non-null values, assume it's a unique
2175  * column; but be sure to discount for any nulls we found.
2176  */
2177  stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2178  }
2179  else if (track_cnt < track_max && toowide_cnt == 0 &&
2180  nmultiple == track_cnt)
2181  {
2182  /*
2183  * Our track list includes every value in the sample, and every
2184  * value appeared more than once. Assume the column has just
2185  * these values. (This case is meant to address columns with
2186  * small, fixed sets of possible values, such as boolean or enum
2187  * columns. If there are any values that appear just once in the
2188  * sample, including too-wide values, we should assume that that's
2189  * not what we're dealing with.)
2190  */
2191  stats->stadistinct = track_cnt;
2192  }
2193  else
2194  {
2195  /*----------
2196  * Estimate the number of distinct values using the estimator
2197  * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2198  * n*d / (n - f1 + f1*n/N)
2199  * where f1 is the number of distinct values that occurred
2200  * exactly once in our sample of n rows (from a total of N),
2201  * and d is the total number of distinct values in the sample.
2202  * This is their Duj1 estimator; the other estimators they
2203  * recommend are considerably more complex, and are numerically
2204  * very unstable when n is much smaller than N.
2205  *
2206  * In this calculation, we consider only non-nulls. We used to
2207  * include rows with null values in the n and N counts, but that
2208  * leads to inaccurate answers in columns with many nulls, and
2209  * it's intuitively bogus anyway considering the desired result is
2210  * the number of distinct non-null values.
2211  *
2212  * We assume (not very reliably!) that all the multiply-occurring
2213  * values are reflected in the final track[] list, and the other
2214  * nonnull values all appeared but once. (XXX this usually
2215  * results in a drastic overestimate of ndistinct. Can we do
2216  * any better?)
2217  *----------
2218  */
2219  int f1 = nonnull_cnt - summultiple;
2220  int d = f1 + nmultiple;
2221  double n = samplerows - null_cnt;
2222  double N = totalrows * (1.0 - stats->stanullfrac);
2223  double stadistinct;
2224 
2225  /* N == 0 shouldn't happen, but just in case ... */
2226  if (N > 0)
2227  stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2228  else
2229  stadistinct = 0;
2230 
2231  /* Clamp to sane range in case of roundoff error */
2232  if (stadistinct < d)
2233  stadistinct = d;
2234  if (stadistinct > N)
2235  stadistinct = N;
2236  /* And round to integer */
2237  stats->stadistinct = floor(stadistinct + 0.5);
2238  }
2239 
2240  /*
2241  * If we estimated the number of distinct values at more than 10% of
2242  * the total row count (a very arbitrary limit), then assume that
2243  * stadistinct should scale with the row count rather than be a fixed
2244  * value.
2245  */
2246  if (stats->stadistinct > 0.1 * totalrows)
2247  stats->stadistinct = -(stats->stadistinct / totalrows);
2248 
2249  /*
2250  * Decide how many values are worth storing as most-common values. If
2251  * we are able to generate a complete MCV list (all the values in the
2252  * sample will fit, and we think these are all the ones in the table),
2253  * then do so. Otherwise, store only those values that are
2254  * significantly more common than the values not in the list.
2255  *
2256  * Note: the first of these cases is meant to address columns with
2257  * small, fixed sets of possible values, such as boolean or enum
2258  * columns. If we can *completely* represent the column population by
2259  * an MCV list that will fit into the stats target, then we should do
2260  * so and thus provide the planner with complete information. But if
2261  * the MCV list is not complete, it's generally worth being more
2262  * selective, and not just filling it all the way up to the stats
2263  * target.
2264  */
2265  if (track_cnt < track_max && toowide_cnt == 0 &&
2266  stats->stadistinct > 0 &&
2267  track_cnt <= num_mcv)
2268  {
2269  /* Track list includes all values seen, and all will fit */
2270  num_mcv = track_cnt;
2271  }
2272  else
2273  {
2274  int *mcv_counts;
2275 
2276  /* Incomplete list; decide how many values are worth keeping */
2277  if (num_mcv > track_cnt)
2278  num_mcv = track_cnt;
2279 
2280  if (num_mcv > 0)
2281  {
2282  mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2283  for (i = 0; i < num_mcv; i++)
2284  mcv_counts[i] = track[i].count;
2285 
2286  num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2287  stats->stadistinct,
2288  stats->stanullfrac,
2289  samplerows, totalrows);
2290  }
2291  }
2292 
2293  /* Generate MCV slot entry */
2294  if (num_mcv > 0)
2295  {
2296  MemoryContext old_context;
2297  Datum *mcv_values;
2298  float4 *mcv_freqs;
2299 
2300  /* Must copy the target values into anl_context */
2301  old_context = MemoryContextSwitchTo(stats->anl_context);
2302  mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2303  mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2304  for (i = 0; i < num_mcv; i++)
2305  {
2306  mcv_values[i] = datumCopy(track[i].value,
2307  stats->attrtype->typbyval,
2308  stats->attrtype->typlen);
2309  mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2310  }
2311  MemoryContextSwitchTo(old_context);
2312 
2313  stats->stakind[0] = STATISTIC_KIND_MCV;
2314  stats->staop[0] = mystats->eqopr;
2315  stats->stacoll[0] = stats->attrcollid;
2316  stats->stanumbers[0] = mcv_freqs;
2317  stats->numnumbers[0] = num_mcv;
2318  stats->stavalues[0] = mcv_values;
2319  stats->numvalues[0] = num_mcv;
2320 
2321  /*
2322  * Accept the defaults for stats->statypid and others. They have
2323  * been set before we were called (see vacuum.h)
2324  */
2325  }
2326  }
2327  else if (null_cnt > 0)
2328  {
2329  /* We found only nulls; assume the column is entirely null */
2330  stats->stats_valid = true;
2331  stats->stanullfrac = 1.0;
2332  if (is_varwidth)
2333  stats->stawidth = 0; /* "unknown" */
2334  else
2335  stats->stawidth = stats->attrtype->typlen;
2336  stats->stadistinct = 0.0; /* "unknown" */
2337  }
2338 
2339  /* We don't need to bother cleaning up any of our temporary palloc's */
2340 }
2341 
2342 
2343 /*
2344  * compute_scalar_stats() -- compute column statistics
2345  *
2346  * We use this when we can find "=" and "<" operators for the datatype.
2347  *
2348  * We determine the fraction of non-null rows, the average width, the
2349  * most common values, the (estimated) number of distinct values, the
2350  * distribution histogram, and the correlation of physical to logical order.
2351  *
2352  * The desired stats can be determined fairly easily after sorting the
2353  * data values into order.
2354  */
2355 static void
2357  AnalyzeAttrFetchFunc fetchfunc,
2358  int samplerows,
2359  double totalrows)
2360 {
2361  int i;
2362  int null_cnt = 0;
2363  int nonnull_cnt = 0;
2364  int toowide_cnt = 0;
2365  double total_width = 0;
2366  bool is_varlena = (!stats->attrtype->typbyval &&
2367  stats->attrtype->typlen == -1);
2368  bool is_varwidth = (!stats->attrtype->typbyval &&
2369  stats->attrtype->typlen < 0);
2370  double corr_xysum;
2371  SortSupportData ssup;
2372  ScalarItem *values;
2373  int values_cnt = 0;
2374  int *tupnoLink;
2375  ScalarMCVItem *track;
2376  int track_cnt = 0;
2377  int num_mcv = stats->attstattarget;
2378  int num_bins = stats->attstattarget;
2379  StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2380 
2381  values = (ScalarItem *) palloc(samplerows * sizeof(ScalarItem));
2382  tupnoLink = (int *) palloc(samplerows * sizeof(int));
2383  track = (ScalarMCVItem *) palloc(num_mcv * sizeof(ScalarMCVItem));
2384 
2385  memset(&ssup, 0, sizeof(ssup));
2387  ssup.ssup_collation = stats->attrcollid;
2388  ssup.ssup_nulls_first = false;
2389 
2390  /*
2391  * For now, don't perform abbreviated key conversion, because full values
2392  * are required for MCV slot generation. Supporting that optimization
2393  * would necessitate teaching compare_scalars() to call a tie-breaker.
2394  */
2395  ssup.abbreviate = false;
2396 
2397  PrepareSortSupportFromOrderingOp(mystats->ltopr, &ssup);
2398 
2399  /* Initial scan to find sortable values */
2400  for (i = 0; i < samplerows; i++)
2401  {
2402  Datum value;
2403  bool isnull;
2404 
2406 
2407  value = fetchfunc(stats, i, &isnull);
2408 
2409  /* Check for null/nonnull */
2410  if (isnull)
2411  {
2412  null_cnt++;
2413  continue;
2414  }
2415  nonnull_cnt++;
2416 
2417  /*
2418  * If it's a variable-width field, add up widths for average width
2419  * calculation. Note that if the value is toasted, we use the toasted
2420  * width. We don't bother with this calculation if it's a fixed-width
2421  * type.
2422  */
2423  if (is_varlena)
2424  {
2425  total_width += VARSIZE_ANY(DatumGetPointer(value));
2426 
2427  /*
2428  * If the value is toasted, we want to detoast it just once to
2429  * avoid repeated detoastings and resultant excess memory usage
2430  * during the comparisons. Also, check to see if the value is
2431  * excessively wide, and if so don't detoast at all --- just
2432  * ignore the value.
2433  */
2435  {
2436  toowide_cnt++;
2437  continue;
2438  }
2440  }
2441  else if (is_varwidth)
2442  {
2443  /* must be cstring */
2444  total_width += strlen(DatumGetCString(value)) + 1;
2445  }
2446 
2447  /* Add it to the list to be sorted */
2448  values[values_cnt].value = value;
2449  values[values_cnt].tupno = values_cnt;
2450  tupnoLink[values_cnt] = values_cnt;
2451  values_cnt++;
2452  }
2453 
2454  /* We can only compute real stats if we found some sortable values. */
2455  if (values_cnt > 0)
2456  {
2457  int ndistinct, /* # distinct values in sample */
2458  nmultiple, /* # that appear multiple times */
2459  num_hist,
2460  dups_cnt;
2461  int slot_idx = 0;
2463 
2464  /* Sort the collected values */
2465  cxt.ssup = &ssup;
2466  cxt.tupnoLink = tupnoLink;
2467  qsort_interruptible(values, values_cnt, sizeof(ScalarItem),
2468  compare_scalars, &cxt);
2469 
2470  /*
2471  * Now scan the values in order, find the most common ones, and also
2472  * accumulate ordering-correlation statistics.
2473  *
2474  * To determine which are most common, we first have to count the
2475  * number of duplicates of each value. The duplicates are adjacent in
2476  * the sorted list, so a brute-force approach is to compare successive
2477  * datum values until we find two that are not equal. However, that
2478  * requires N-1 invocations of the datum comparison routine, which are
2479  * completely redundant with work that was done during the sort. (The
2480  * sort algorithm must at some point have compared each pair of items
2481  * that are adjacent in the sorted order; otherwise it could not know
2482  * that it's ordered the pair correctly.) We exploit this by having
2483  * compare_scalars remember the highest tupno index that each
2484  * ScalarItem has been found equal to. At the end of the sort, a
2485  * ScalarItem's tupnoLink will still point to itself if and only if it
2486  * is the last item of its group of duplicates (since the group will
2487  * be ordered by tupno).
2488  */
2489  corr_xysum = 0;
2490  ndistinct = 0;
2491  nmultiple = 0;
2492  dups_cnt = 0;
2493  for (i = 0; i < values_cnt; i++)
2494  {
2495  int tupno = values[i].tupno;
2496 
2497  corr_xysum += ((double) i) * ((double) tupno);
2498  dups_cnt++;
2499  if (tupnoLink[tupno] == tupno)
2500  {
2501  /* Reached end of duplicates of this value */
2502  ndistinct++;
2503  if (dups_cnt > 1)
2504  {
2505  nmultiple++;
2506  if (track_cnt < num_mcv ||
2507  dups_cnt > track[track_cnt - 1].count)
2508  {
2509  /*
2510  * Found a new item for the mcv list; find its
2511  * position, bubbling down old items if needed. Loop
2512  * invariant is that j points at an empty/ replaceable
2513  * slot.
2514  */
2515  int j;
2516 
2517  if (track_cnt < num_mcv)
2518  track_cnt++;
2519  for (j = track_cnt - 1; j > 0; j--)
2520  {
2521  if (dups_cnt <= track[j - 1].count)
2522  break;
2523  track[j].count = track[j - 1].count;
2524  track[j].first = track[j - 1].first;
2525  }
2526  track[j].count = dups_cnt;
2527  track[j].first = i + 1 - dups_cnt;
2528  }
2529  }
2530  dups_cnt = 0;
2531  }
2532  }
2533 
2534  stats->stats_valid = true;
2535  /* Do the simple null-frac and width stats */
2536  stats->stanullfrac = (double) null_cnt / (double) samplerows;
2537  if (is_varwidth)
2538  stats->stawidth = total_width / (double) nonnull_cnt;
2539  else
2540  stats->stawidth = stats->attrtype->typlen;
2541 
2542  if (nmultiple == 0)
2543  {
2544  /*
2545  * If we found no repeated non-null values, assume it's a unique
2546  * column; but be sure to discount for any nulls we found.
2547  */
2548  stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2549  }
2550  else if (toowide_cnt == 0 && nmultiple == ndistinct)
2551  {
2552  /*
2553  * Every value in the sample appeared more than once. Assume the
2554  * column has just these values. (This case is meant to address
2555  * columns with small, fixed sets of possible values, such as
2556  * boolean or enum columns. If there are any values that appear
2557  * just once in the sample, including too-wide values, we should
2558  * assume that that's not what we're dealing with.)
2559  */
2560  stats->stadistinct = ndistinct;
2561  }
2562  else
2563  {
2564  /*----------
2565  * Estimate the number of distinct values using the estimator
2566  * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2567  * n*d / (n - f1 + f1*n/N)
2568  * where f1 is the number of distinct values that occurred
2569  * exactly once in our sample of n rows (from a total of N),
2570  * and d is the total number of distinct values in the sample.
2571  * This is their Duj1 estimator; the other estimators they
2572  * recommend are considerably more complex, and are numerically
2573  * very unstable when n is much smaller than N.
2574  *
2575  * In this calculation, we consider only non-nulls. We used to
2576  * include rows with null values in the n and N counts, but that
2577  * leads to inaccurate answers in columns with many nulls, and
2578  * it's intuitively bogus anyway considering the desired result is
2579  * the number of distinct non-null values.
2580  *
2581  * Overwidth values are assumed to have been distinct.
2582  *----------
2583  */
2584  int f1 = ndistinct - nmultiple + toowide_cnt;
2585  int d = f1 + nmultiple;
2586  double n = samplerows - null_cnt;
2587  double N = totalrows * (1.0 - stats->stanullfrac);
2588  double stadistinct;
2589 
2590  /* N == 0 shouldn't happen, but just in case ... */
2591  if (N > 0)
2592  stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2593  else
2594  stadistinct = 0;
2595 
2596  /* Clamp to sane range in case of roundoff error */
2597  if (stadistinct < d)
2598  stadistinct = d;
2599  if (stadistinct > N)
2600  stadistinct = N;
2601  /* And round to integer */
2602  stats->stadistinct = floor(stadistinct + 0.5);
2603  }
2604 
2605  /*
2606  * If we estimated the number of distinct values at more than 10% of
2607  * the total row count (a very arbitrary limit), then assume that
2608  * stadistinct should scale with the row count rather than be a fixed
2609  * value.
2610  */
2611  if (stats->stadistinct > 0.1 * totalrows)
2612  stats->stadistinct = -(stats->stadistinct / totalrows);
2613 
2614  /*
2615  * Decide how many values are worth storing as most-common values. If
2616  * we are able to generate a complete MCV list (all the values in the
2617  * sample will fit, and we think these are all the ones in the table),
2618  * then do so. Otherwise, store only those values that are
2619  * significantly more common than the values not in the list.
2620  *
2621  * Note: the first of these cases is meant to address columns with
2622  * small, fixed sets of possible values, such as boolean or enum
2623  * columns. If we can *completely* represent the column population by
2624  * an MCV list that will fit into the stats target, then we should do
2625  * so and thus provide the planner with complete information. But if
2626  * the MCV list is not complete, it's generally worth being more
2627  * selective, and not just filling it all the way up to the stats
2628  * target.
2629  */
2630  if (track_cnt == ndistinct && toowide_cnt == 0 &&
2631  stats->stadistinct > 0 &&
2632  track_cnt <= num_mcv)
2633  {
2634  /* Track list includes all values seen, and all will fit */
2635  num_mcv = track_cnt;
2636  }
2637  else
2638  {
2639  int *mcv_counts;
2640 
2641  /* Incomplete list; decide how many values are worth keeping */
2642  if (num_mcv > track_cnt)
2643  num_mcv = track_cnt;
2644 
2645  if (num_mcv > 0)
2646  {
2647  mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2648  for (i = 0; i < num_mcv; i++)
2649  mcv_counts[i] = track[i].count;
2650 
2651  num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2652  stats->stadistinct,
2653  stats->stanullfrac,
2654  samplerows, totalrows);
2655  }
2656  }
2657 
2658  /* Generate MCV slot entry */
2659  if (num_mcv > 0)
2660  {
2661  MemoryContext old_context;
2662  Datum *mcv_values;
2663  float4 *mcv_freqs;
2664 
2665  /* Must copy the target values into anl_context */
2666  old_context = MemoryContextSwitchTo(stats->anl_context);
2667  mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2668  mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2669  for (i = 0; i < num_mcv; i++)
2670  {
2671  mcv_values[i] = datumCopy(values[track[i].first].value,
2672  stats->attrtype->typbyval,
2673  stats->attrtype->typlen);
2674  mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2675  }
2676  MemoryContextSwitchTo(old_context);
2677 
2678  stats->stakind[slot_idx] = STATISTIC_KIND_MCV;
2679  stats->staop[slot_idx] = mystats->eqopr;
2680  stats->stacoll[slot_idx] = stats->attrcollid;
2681  stats->stanumbers[slot_idx] = mcv_freqs;
2682  stats->numnumbers[slot_idx] = num_mcv;
2683  stats->stavalues[slot_idx] = mcv_values;
2684  stats->numvalues[slot_idx] = num_mcv;
2685 
2686  /*
2687  * Accept the defaults for stats->statypid and others. They have
2688  * been set before we were called (see vacuum.h)
2689  */
2690  slot_idx++;
2691  }
2692 
2693  /*
2694  * Generate a histogram slot entry if there are at least two distinct
2695  * values not accounted for in the MCV list. (This ensures the
2696  * histogram won't collapse to empty or a singleton.)
2697  */
2698  num_hist = ndistinct - num_mcv;
2699  if (num_hist > num_bins)
2700  num_hist = num_bins + 1;
2701  if (num_hist >= 2)
2702  {
2703  MemoryContext old_context;
2704  Datum *hist_values;
2705  int nvals;
2706  int pos,
2707  posfrac,
2708  delta,
2709  deltafrac;
2710 
2711  /* Sort the MCV items into position order to speed next loop */
2712  qsort_interruptible(track, num_mcv, sizeof(ScalarMCVItem),
2713  compare_mcvs, NULL);
2714 
2715  /*
2716  * Collapse out the MCV items from the values[] array.
2717  *
2718  * Note we destroy the values[] array here... but we don't need it
2719  * for anything more. We do, however, still need values_cnt.
2720  * nvals will be the number of remaining entries in values[].
2721  */
2722  if (num_mcv > 0)
2723  {
2724  int src,
2725  dest;
2726  int j;
2727 
2728  src = dest = 0;
2729  j = 0; /* index of next interesting MCV item */
2730  while (src < values_cnt)
2731  {
2732  int ncopy;
2733 
2734  if (j < num_mcv)
2735  {
2736  int first = track[j].first;
2737 
2738  if (src >= first)
2739  {
2740  /* advance past this MCV item */
2741  src = first + track[j].count;
2742  j++;
2743  continue;
2744  }
2745  ncopy = first - src;
2746  }
2747  else
2748  ncopy = values_cnt - src;
2749  memmove(&values[dest], &values[src],
2750  ncopy * sizeof(ScalarItem));
2751  src += ncopy;
2752  dest += ncopy;
2753  }
2754  nvals = dest;
2755  }
2756  else
2757  nvals = values_cnt;
2758  Assert(nvals >= num_hist);
2759 
2760  /* Must copy the target values into anl_context */
2761  old_context = MemoryContextSwitchTo(stats->anl_context);
2762  hist_values = (Datum *) palloc(num_hist * sizeof(Datum));
2763 
2764  /*
2765  * The object of this loop is to copy the first and last values[]
2766  * entries along with evenly-spaced values in between. So the
2767  * i'th value is values[(i * (nvals - 1)) / (num_hist - 1)]. But
2768  * computing that subscript directly risks integer overflow when
2769  * the stats target is more than a couple thousand. Instead we
2770  * add (nvals - 1) / (num_hist - 1) to pos at each step, tracking
2771  * the integral and fractional parts of the sum separately.
2772  */
2773  delta = (nvals - 1) / (num_hist - 1);
2774  deltafrac = (nvals - 1) % (num_hist - 1);
2775  pos = posfrac = 0;
2776 
2777  for (i = 0; i < num_hist; i++)
2778  {
2779  hist_values[i] = datumCopy(values[pos].value,
2780  stats->attrtype->typbyval,
2781  stats->attrtype->typlen);
2782  pos += delta;
2783  posfrac += deltafrac;
2784  if (posfrac >= (num_hist - 1))
2785  {
2786  /* fractional part exceeds 1, carry to integer part */
2787  pos++;
2788  posfrac -= (num_hist - 1);
2789  }
2790  }
2791 
2792  MemoryContextSwitchTo(old_context);
2793 
2794  stats->stakind[slot_idx] = STATISTIC_KIND_HISTOGRAM;
2795  stats->staop[slot_idx] = mystats->ltopr;
2796  stats->stacoll[slot_idx] = stats->attrcollid;
2797  stats->stavalues[slot_idx] = hist_values;
2798  stats->numvalues[slot_idx] = num_hist;
2799 
2800  /*
2801  * Accept the defaults for stats->statypid and others. They have
2802  * been set before we were called (see vacuum.h)
2803  */
2804  slot_idx++;
2805  }
2806 
2807  /* Generate a correlation entry if there are multiple values */
2808  if (values_cnt > 1)
2809  {
2810  MemoryContext old_context;
2811  float4 *corrs;
2812  double corr_xsum,
2813  corr_x2sum;
2814 
2815  /* Must copy the target values into anl_context */
2816  old_context = MemoryContextSwitchTo(stats->anl_context);
2817  corrs = (float4 *) palloc(sizeof(float4));
2818  MemoryContextSwitchTo(old_context);
2819 
2820  /*----------
2821  * Since we know the x and y value sets are both
2822  * 0, 1, ..., values_cnt-1
2823  * we have sum(x) = sum(y) =
2824  * (values_cnt-1)*values_cnt / 2
2825  * and sum(x^2) = sum(y^2) =
2826  * (values_cnt-1)*values_cnt*(2*values_cnt-1) / 6.
2827  *----------
2828  */
2829  corr_xsum = ((double) (values_cnt - 1)) *
2830  ((double) values_cnt) / 2.0;
2831  corr_x2sum = ((double) (values_cnt - 1)) *
2832  ((double) values_cnt) * (double) (2 * values_cnt - 1) / 6.0;
2833 
2834  /* And the correlation coefficient reduces to */
2835  corrs[0] = (values_cnt * corr_xysum - corr_xsum * corr_xsum) /
2836  (values_cnt * corr_x2sum - corr_xsum * corr_xsum);
2837 
2838  stats->stakind[slot_idx] = STATISTIC_KIND_CORRELATION;
2839  stats->staop[slot_idx] = mystats->ltopr;
2840  stats->stacoll[slot_idx] = stats->attrcollid;
2841  stats->stanumbers[slot_idx] = corrs;
2842  stats->numnumbers[slot_idx] = 1;
2843  slot_idx++;
2844  }
2845  }
2846  else if (nonnull_cnt > 0)
2847  {
2848  /* We found some non-null values, but they were all too wide */
2849  Assert(nonnull_cnt == toowide_cnt);
2850  stats->stats_valid = true;
2851  /* Do the simple null-frac and width stats */
2852  stats->stanullfrac = (double) null_cnt / (double) samplerows;
2853  if (is_varwidth)
2854  stats->stawidth = total_width / (double) nonnull_cnt;
2855  else
2856  stats->stawidth = stats->attrtype->typlen;
2857  /* Assume all too-wide values are distinct, so it's a unique column */
2858  stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2859  }
2860  else if (null_cnt > 0)
2861  {
2862  /* We found only nulls; assume the column is entirely null */
2863  stats->stats_valid = true;
2864  stats->stanullfrac = 1.0;
2865  if (is_varwidth)
2866  stats->stawidth = 0; /* "unknown" */
2867  else
2868  stats->stawidth = stats->attrtype->typlen;
2869  stats->stadistinct = 0.0; /* "unknown" */
2870  }
2871 
2872  /* We don't need to bother cleaning up any of our temporary palloc's */
2873 }
2874 
2875 /*
2876  * Comparator for sorting ScalarItems
2877  *
2878  * Aside from sorting the items, we update the tupnoLink[] array
2879  * whenever two ScalarItems are found to contain equal datums. The array
2880  * is indexed by tupno; for each ScalarItem, it contains the highest
2881  * tupno that that item's datum has been found to be equal to. This allows
2882  * us to avoid additional comparisons in compute_scalar_stats().
2883  */
2884 static int
2885 compare_scalars(const void *a, const void *b, void *arg)
2886 {
2887  Datum da = ((const ScalarItem *) a)->value;
2888  int ta = ((const ScalarItem *) a)->tupno;
2889  Datum db = ((const ScalarItem *) b)->value;
2890  int tb = ((const ScalarItem *) b)->tupno;
2892  int compare;
2893 
2894  compare = ApplySortComparator(da, false, db, false, cxt->ssup);
2895  if (compare != 0)
2896  return compare;
2897 
2898  /*
2899  * The two datums are equal, so update cxt->tupnoLink[].
2900  */
2901  if (cxt->tupnoLink[ta] < tb)
2902  cxt->tupnoLink[ta] = tb;
2903  if (cxt->tupnoLink[tb] < ta)
2904  cxt->tupnoLink[tb] = ta;
2905 
2906  /*
2907  * For equal datums, sort by tupno
2908  */
2909  return ta - tb;
2910 }
2911 
2912 /*
2913  * Comparator for sorting ScalarMCVItems by position
2914  */
2915 static int
2916 compare_mcvs(const void *a, const void *b, void *arg)
2917 {
2918  int da = ((const ScalarMCVItem *) a)->first;
2919  int db = ((const ScalarMCVItem *) b)->first;
2920 
2921  return da - db;
2922 }
2923 
2924 /*
2925  * Analyze the list of common values in the sample and decide how many are
2926  * worth storing in the table's MCV list.
2927  *
2928  * mcv_counts is assumed to be a list of the counts of the most common values
2929  * seen in the sample, starting with the most common. The return value is the
2930  * number that are significantly more common than the values not in the list,
2931  * and which are therefore deemed worth storing in the table's MCV list.
2932  */
2933 static int
2934 analyze_mcv_list(int *mcv_counts,
2935  int num_mcv,
2936  double stadistinct,
2937  double stanullfrac,
2938  int samplerows,
2939  double totalrows)
2940 {
2941  double ndistinct_table;
2942  double sumcount;
2943  int i;
2944 
2945  /*
2946  * If the entire table was sampled, keep the whole list. This also
2947  * protects us against division by zero in the code below.
2948  */
2949  if (samplerows == totalrows || totalrows <= 1.0)
2950  return num_mcv;
2951 
2952  /* Re-extract the estimated number of distinct nonnull values in table */
2953  ndistinct_table = stadistinct;
2954  if (ndistinct_table < 0)
2955  ndistinct_table = -ndistinct_table * totalrows;
2956 
2957  /*
2958  * Exclude the least common values from the MCV list, if they are not
2959  * significantly more common than the estimated selectivity they would
2960  * have if they weren't in the list. All non-MCV values are assumed to be
2961  * equally common, after taking into account the frequencies of all the
2962  * values in the MCV list and the number of nulls (c.f. eqsel()).
2963  *
2964  * Here sumcount tracks the total count of all but the last (least common)
2965  * value in the MCV list, allowing us to determine the effect of excluding
2966  * that value from the list.
2967  *
2968  * Note that we deliberately do this by removing values from the full
2969  * list, rather than starting with an empty list and adding values,
2970  * because the latter approach can fail to add any values if all the most
2971  * common values have around the same frequency and make up the majority
2972  * of the table, so that the overall average frequency of all values is
2973  * roughly the same as that of the common values. This would lead to any
2974  * uncommon values being significantly overestimated.
2975  */
2976  sumcount = 0.0;
2977  for (i = 0; i < num_mcv - 1; i++)
2978  sumcount += mcv_counts[i];
2979 
2980  while (num_mcv > 0)
2981  {
2982  double selec,
2983  otherdistinct,
2984  N,
2985  n,
2986  K,
2987  variance,
2988  stddev;
2989 
2990  /*
2991  * Estimated selectivity the least common value would have if it
2992  * wasn't in the MCV list (c.f. eqsel()).
2993  */
2994  selec = 1.0 - sumcount / samplerows - stanullfrac;
2995  if (selec < 0.0)
2996  selec = 0.0;
2997  if (selec > 1.0)
2998  selec = 1.0;
2999  otherdistinct = ndistinct_table - (num_mcv - 1);
3000  if (otherdistinct > 1)
3001  selec /= otherdistinct;
3002 
3003  /*
3004  * If the value is kept in the MCV list, its population frequency is
3005  * assumed to equal its sample frequency. We use the lower end of a
3006  * textbook continuity-corrected Wald-type confidence interval to
3007  * determine if that is significantly more common than the non-MCV
3008  * frequency --- specifically we assume the population frequency is
3009  * highly likely to be within around 2 standard errors of the sample
3010  * frequency, which equates to an interval of 2 standard deviations
3011  * either side of the sample count, plus an additional 0.5 for the
3012  * continuity correction. Since we are sampling without replacement,
3013  * this is a hypergeometric distribution.
3014  *
3015  * XXX: Empirically, this approach seems to work quite well, but it
3016  * may be worth considering more advanced techniques for estimating
3017  * the confidence interval of the hypergeometric distribution.
3018  */
3019  N = totalrows;
3020  n = samplerows;
3021  K = N * mcv_counts[num_mcv - 1] / n;
3022  variance = n * K * (N - K) * (N - n) / (N * N * (N - 1));
3023  stddev = sqrt(variance);
3024 
3025  if (mcv_counts[num_mcv - 1] > selec * samplerows + 2 * stddev + 0.5)
3026  {
3027  /*
3028  * The value is significantly more common than the non-MCV
3029  * selectivity would suggest. Keep it, and all the other more
3030  * common values in the list.
3031  */
3032  break;
3033  }
3034  else
3035  {
3036  /* Discard this value and consider the next least common value */
3037  num_mcv--;
3038  if (num_mcv == 0)
3039  break;
3040  sumcount -= mcv_counts[num_mcv - 1];
3041  }
3042  }
3043  return num_mcv;
3044 }
ArrayType * construct_array_builtin(Datum *elems, int nelems, Oid elmtype)
Definition: arrayfuncs.c:3374
ArrayType * construct_array(Datum *elems, int nelems, Oid elmtype, int elmlen, bool elmbyval, char elmalign)
Definition: arrayfuncs.c:3354
#define InvalidAttrNumber
Definition: attnum.h:23
AttributeOpts * get_attribute_options(Oid attrelid, int attnum)
Definition: attoptcache.c:104
long TimestampDifferenceMilliseconds(TimestampTz start_time, TimestampTz stop_time)
Definition: timestamp.c:1767
bool TimestampDifferenceExceeds(TimestampTz start_time, TimestampTz stop_time, int msec)
Definition: timestamp.c:1791
TimestampTz GetCurrentTimestamp(void)
Definition: timestamp.c:1655
void pgstat_progress_start_command(ProgressCommandType cmdtype, Oid relid)
void pgstat_progress_update_param(int index, int64 val)
void pgstat_progress_update_multi_param(int nparam, const int *index, const int64 *val)
void pgstat_progress_end_command(void)
@ PROGRESS_COMMAND_ANALYZE
bool bms_is_member(int x, const Bitmapset *a)
Definition: bitmapset.c:510
Bitmapset * bms_add_member(Bitmapset *a, int x)
Definition: bitmapset.c:815
uint32 BlockNumber
Definition: block.h:31
#define InvalidBlockNumber
Definition: block.h:33
static Datum values[MAXATTR]
Definition: bootstrap.c:152
bool track_io_timing
Definition: bufmgr.c:142
#define RelationGetNumberOfBlocks(reln)
Definition: bufmgr.h:277
unsigned int uint32
Definition: c.h:506
#define Min(x, y)
Definition: c.h:1004
#define Assert(condition)
Definition: c.h:858
double float8
Definition: c.h:630
float float4
Definition: c.h:629
uint32 TransactionId
Definition: c.h:652
#define OidIsValid(objectId)
Definition: c.h:775
static Datum std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
Definition: analyze.c:1752
#define swapInt(a, b)
Definition: analyze.c:1800
static void compute_scalar_stats(VacAttrStatsP stats, AnalyzeAttrFetchFunc fetchfunc, int samplerows, double totalrows)
Definition: analyze.c:2356
#define swapDatum(a, b)
Definition: analyze.c:1801
#define WIDTH_THRESHOLD
Definition: analyze.c:1798
static void update_attstats(Oid relid, bool inh, int natts, VacAttrStats **vacattrstats)
Definition: analyze.c:1609
int default_statistics_target
Definition: analyze.c:73
static void compute_distinct_stats(VacAttrStatsP stats, AnalyzeAttrFetchFunc fetchfunc, int samplerows, double totalrows)
Definition: analyze.c:2013
static MemoryContext anl_context
Definition: analyze.c:76
bool std_typanalyze(VacAttrStats *stats)
Definition: analyze.c:1845
static int analyze_mcv_list(int *mcv_counts, int num_mcv, double stadistinct, double stanullfrac, int samplerows, double totalrows)
Definition: analyze.c:2934
static int acquire_inherited_sample_rows(Relation onerel, int elevel, HeapTuple *rows, int targrows, double *totalrows, double *totaldeadrows)
Definition: analyze.c:1345
static BufferAccessStrategy vac_strategy
Definition: analyze.c:77
static BlockNumber block_sampling_read_stream_next(ReadStream *stream, void *callback_private_data, void *per_buffer_data)
Definition: analyze.c:1115
static int compare_mcvs(const void *a, const void *b, void *arg)
Definition: analyze.c:2916
struct AnlIndexData AnlIndexData
static int acquire_sample_rows(Relation onerel, int elevel, HeapTuple *rows, int targrows, double *totalrows, double *totaldeadrows)
Definition: analyze.c:1158
static void compute_trivial_stats(VacAttrStatsP stats, AnalyzeAttrFetchFunc fetchfunc, int samplerows, double totalrows)
Definition: analyze.c:1923
static int compare_scalars(const void *a, const void *b, void *arg)
Definition: analyze.c:2885
static Datum ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
Definition: analyze.c:1768
void analyze_rel(Oid relid, RangeVar *relation, VacuumParams *params, List *va_cols, bool in_outer_xact, BufferAccessStrategy bstrategy)
Definition: analyze.c:111
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:280
static VacAttrStats * examine_attribute(Relation onerel, int attnum, Node *index_expr)
Definition: analyze.c:999
static int compare_rows(const void *a, const void *b, void *arg)
Definition: analyze.c:1315
static void compute_index_stats(Relation onerel, double totalrows, AnlIndexData *indexdata, int nindexes, HeapTuple *rows, int numrows, MemoryContext col_context)
Definition: analyze.c:828
int64 TimestampTz
Definition: timestamp.h:39
Datum datumCopy(Datum value, bool typByVal, int typLen)
Definition: datum.c:132
char * get_database_name(Oid dbid)
Definition: dbcommands.c:3166
Size toast_raw_datum_size(Datum value)
Definition: detoast.c:545
int errmsg_internal(const char *fmt,...)
Definition: elog.c:1155
int errcode(int sqlerrcode)
Definition: elog.c:855
int errmsg(const char *fmt,...)
Definition: elog.c:1068
#define _(x)
Definition: elog.c:90
#define LOG
Definition: elog.h:31
#define WARNING
Definition: elog.h:36
#define DEBUG2
Definition: elog.h:29
#define ERROR
Definition: elog.h:39
#define elog(elevel,...)
Definition: elog.h:224
#define INFO
Definition: elog.h:34
#define ereport(elevel,...)
Definition: elog.h:149
ExprState * ExecPrepareQual(List *qual, EState *estate)
Definition: execExpr.c:767
void ExecDropSingleTupleTableSlot(TupleTableSlot *slot)
Definition: execTuples.c:1341
TupleTableSlot * ExecStoreHeapTuple(HeapTuple tuple, TupleTableSlot *slot, bool shouldFree)
Definition: execTuples.c:1439
const TupleTableSlotOps TTSOpsHeapTuple
Definition: execTuples.c:85
TupleTableSlot * MakeSingleTupleTableSlot(TupleDesc tupdesc, const TupleTableSlotOps *tts_ops)
Definition: execTuples.c:1325
EState * CreateExecutorState(void)
Definition: execUtils.c:88
void FreeExecutorState(EState *estate)
Definition: execUtils.c:189
#define GetPerTupleExprContext(estate)
Definition: executor.h:550
#define ResetExprContext(econtext)
Definition: executor.h:544
static bool ExecQual(ExprState *state, ExprContext *econtext)
Definition: executor.h:413
int ComputeExtStatisticsRows(Relation onerel, int natts, VacAttrStats **vacattrstats)
void BuildRelationExtStatistics(Relation onerel, bool inh, double totalrows, int numrows, HeapTuple *rows, int natts, VacAttrStats **vacattrstats)
int(* AcquireSampleRowsFunc)(Relation relation, int elevel, HeapTuple *rows, int targrows, double *totalrows, double *totaldeadrows)
Definition: fdwapi.h:151
Datum FunctionCall2Coll(FmgrInfo *flinfo, Oid collation, Datum arg1, Datum arg2)
Definition: fmgr.c:1149
void fmgr_info(Oid functionId, FmgrInfo *finfo)
Definition: fmgr.c:127
#define OidFunctionCall1(functionId, arg1)
Definition: fmgr.h:680
#define PG_DETOAST_DATUM(datum)
Definition: fmgr.h:240
FdwRoutine * GetFdwRoutineForRelation(Relation relation, bool makecopy)
Definition: foreign.c:432
static int compare(const void *arg1, const void *arg2)
Definition: geqo_pool.c:145
int64 VacuumPageHit
Definition: globals.c:155
int64 VacuumPageMiss
Definition: globals.c:156
int64 VacuumPageDirty
Definition: globals.c:157
Oid MyDatabaseId
Definition: globals.c:92
int NewGUCNestLevel(void)
Definition: guc.c:2237
void RestrictSearchPath(void)
Definition: guc.c:2248
void AtEOXact_GUC(bool isCommit, int nestLevel)
Definition: guc.c:2264
HeapTuple heap_modify_tuple(HeapTuple tuple, TupleDesc tupleDesc, const Datum *replValues, const bool *replIsnull, const bool *doReplace)
Definition: heaptuple.c:1209
HeapTuple heap_form_tuple(TupleDesc tupleDescriptor, const Datum *values, const bool *isnull)
Definition: heaptuple.c:1116
void heap_freetuple(HeapTuple htup)
Definition: heaptuple.c:1434
#define HeapTupleIsValid(tuple)
Definition: htup.h:78
static Datum heap_getattr(HeapTuple tup, int attnum, TupleDesc tupleDesc, bool *isnull)
Definition: htup_details.h:792
#define GETSTRUCT(TUP)
Definition: htup_details.h:653
void FormIndexDatum(IndexInfo *indexInfo, TupleTableSlot *slot, EState *estate, Datum *values, bool *isnull)
Definition: index.c:2702
IndexInfo * BuildIndexInfo(Relation index)
Definition: index.c:2404
IndexBulkDeleteResult * index_vacuum_cleanup(IndexVacuumInfo *info, IndexBulkDeleteResult *istat)
Definition: indexam.c:769
void CatalogTupleInsertWithInfo(Relation heapRel, HeapTuple tup, CatalogIndexState indstate)
Definition: indexing.c:256
void CatalogCloseIndexes(CatalogIndexState indstate)
Definition: indexing.c:61
CatalogIndexState CatalogOpenIndexes(Relation heapRel)
Definition: indexing.c:43
void CatalogTupleUpdateWithInfo(Relation heapRel, ItemPointer otid, HeapTuple tup, CatalogIndexState indstate)
Definition: indexing.c:337
static struct @155 value
int b
Definition: isn.c:70
int a
Definition: isn.c:69
int j
Definition: isn.c:74
int i
Definition: isn.c:73
if(TABLE==NULL||TABLE_index==NULL)
Definition: isn.c:77
static OffsetNumber ItemPointerGetOffsetNumber(const ItemPointerData *pointer)
Definition: itemptr.h:124
static BlockNumber ItemPointerGetBlockNumber(const ItemPointerData *pointer)
Definition: itemptr.h:103
void list_free(List *list)
Definition: list.c:1546
#define NoLock
Definition: lockdefs.h:34
#define AccessShareLock
Definition: lockdefs.h:36
#define ShareUpdateExclusiveLock
Definition: lockdefs.h:39
#define RowExclusiveLock
Definition: lockdefs.h:38
char * get_namespace_name(Oid nspid)
Definition: lsyscache.c:3366
RegProcedure get_opcode(Oid opno)
Definition: lsyscache.c:1285
void MemoryContextReset(MemoryContext context)
Definition: mcxt.c:383
void pfree(void *pointer)
Definition: mcxt.c:1521
void * palloc0(Size size)
Definition: mcxt.c:1347
MemoryContext CurrentMemoryContext
Definition: mcxt.c:143
void MemoryContextDelete(MemoryContext context)
Definition: mcxt.c:454
void * palloc(Size size)
Definition: mcxt.c:1317
#define AllocSetContextCreate
Definition: memutils.h:129
#define ALLOCSET_DEFAULT_SIZES
Definition: memutils.h:160
#define AmAutoVacuumWorkerProcess()
Definition: miscadmin.h:375
#define SECURITY_RESTRICTED_OPERATION
Definition: miscadmin.h:315
#define CHECK_FOR_INTERRUPTS()
Definition: miscadmin.h:122
void GetUserIdAndSecContext(Oid *userid, int *sec_context)
Definition: miscinit.c:635
void SetUserIdAndSecContext(Oid userid, int sec_context)
Definition: miscinit.c:642
#define InvalidMultiXactId
Definition: multixact.h:24
Oid exprType(const Node *expr)
Definition: nodeFuncs.c:42
int32 exprTypmod(const Node *expr)
Definition: nodeFuncs.c:298
Oid exprCollation(const Node *expr)
Definition: nodeFuncs.c:816
uint16 OffsetNumber
Definition: off.h:24
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:180
int attnameAttNum(Relation rd, const char *attname, bool sysColOK)
int16 attnum
Definition: pg_attribute.h:74
FormData_pg_attribute * Form_pg_attribute
Definition: pg_attribute.h:209
void * arg
#define INDEX_MAX_KEYS
List * find_all_inheritors(Oid parentrelId, LOCKMODE lockmode, List **numparents)
Definition: pg_inherits.c:255
#define lfirst(lc)
Definition: pg_list.h:172
static int list_length(const List *l)
Definition: pg_list.h:152
#define NIL
Definition: pg_list.h:68
static ListCell * list_head(const List *l)
Definition: pg_list.h:128
static ListCell * lnext(const List *l, const ListCell *c)
Definition: pg_list.h:343
#define lfirst_oid(lc)
Definition: pg_list.h:174
uint32 pg_prng_uint32(pg_prng_state *state)
Definition: pg_prng.c:227
pg_prng_state pg_global_prng_state
Definition: pg_prng.c:34
const char * pg_rusage_show(const PGRUsage *ru0)
Definition: pg_rusage.c:40
void pg_rusage_init(PGRUsage *ru0)
Definition: pg_rusage.c:27
#define STATISTIC_NUM_SLOTS
Definition: pg_statistic.h:127
static char * buf
Definition: pg_test_fsync.c:73
FormData_pg_type * Form_pg_type
Definition: pg_type.h:261
int64 PgStat_Counter
Definition: pgstat.h:89
PgStat_Counter pgStatBlockReadTime
PgStat_Counter pgStatBlockWriteTime
void pgstat_report_analyze(Relation rel, PgStat_Counter livetuples, PgStat_Counter deadtuples, bool resetcounter)
void qsort_interruptible(void *base, size_t nel, size_t elsize, qsort_arg_comparator cmp, void *arg)
static bool DatumGetBool(Datum X)
Definition: postgres.h:90
static Datum PointerGetDatum(const void *X)
Definition: postgres.h:322
static Datum Float4GetDatum(float4 X)
Definition: postgres.h:475
static char * DatumGetCString(Datum X)
Definition: postgres.h:335
uintptr_t Datum
Definition: postgres.h:64
static Datum Int16GetDatum(int16 X)
Definition: postgres.h:172
static Datum BoolGetDatum(bool X)
Definition: postgres.h:102
static Datum ObjectIdGetDatum(Oid X)
Definition: postgres.h:252
static Pointer DatumGetPointer(Datum X)
Definition: postgres.h:312
static Datum Int32GetDatum(int32 X)
Definition: postgres.h:212
static int16 DatumGetInt16(Datum X)
Definition: postgres.h:162
#define InvalidOid
Definition: postgres_ext.h:36
unsigned int Oid
Definition: postgres_ext.h:31
TransactionId GetOldestNonRemovableTransactionId(Relation rel)
Definition: procarray.c:2005
#define PROGRESS_ANALYZE_PHASE_FINALIZE_ANALYZE
Definition: progress.h:55
#define PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS_INH
Definition: progress.h:52
#define PROGRESS_ANALYZE_BLOCKS_DONE
Definition: progress.h:43
#define PROGRESS_ANALYZE_PHASE
Definition: progress.h:41
#define PROGRESS_ANALYZE_CHILD_TABLES_TOTAL
Definition: progress.h:46
#define PROGRESS_ANALYZE_BLOCKS_TOTAL
Definition: progress.h:42
#define PROGRESS_ANALYZE_PHASE_COMPUTE_STATS
Definition: progress.h:53
#define PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS
Definition: progress.h:51
#define PROGRESS_ANALYZE_CHILD_TABLES_DONE
Definition: progress.h:47
#define PROGRESS_ANALYZE_CURRENT_CHILD_TABLE_RELID
Definition: progress.h:48
MemoryContextSwitchTo(old_ctx)
ReadStream * read_stream_begin_relation(int flags, BufferAccessStrategy strategy, Relation rel, ForkNumber forknum, ReadStreamBlockNumberCB callback, void *callback_private_data, size_t per_buffer_data_size)
Definition: read_stream.c:410
void read_stream_end(ReadStream *stream)
Definition: read_stream.c:802
#define READ_STREAM_MAINTENANCE
Definition: read_stream.h:27
#define RelationGetRelid(relation)
Definition: rel.h:505
#define RelationGetDescr(relation)
Definition: rel.h:531
#define RelationGetRelationName(relation)
Definition: rel.h:539
#define RELATION_IS_OTHER_TEMP(relation)
Definition: rel.h:658
#define RelationGetNamespace(relation)
Definition: rel.h:546
List * RelationGetIndexList(Relation relation)
Definition: relcache.c:4801
struct RelationData * Relation
Definition: relcache.h:27
@ MAIN_FORKNUM
Definition: relpath.h:50
BlockNumber BlockSampler_Init(BlockSampler bs, BlockNumber nblocks, int samplesize, uint32 randseed)
Definition: sampling.c:39
void reservoir_init_selection_state(ReservoirState rs, int n)
Definition: sampling.c:133
double sampler_random_fract(pg_prng_state *randstate)
Definition: sampling.c:241
bool BlockSampler_HasMore(BlockSampler bs)
Definition: sampling.c:58
BlockNumber BlockSampler_Next(BlockSampler bs)
Definition: sampling.c:64
double reservoir_get_next_S(ReservoirState rs, double t, int n)
Definition: sampling.c:147
#define K(t)
Definition: sha1.c:66
void PrepareSortSupportFromOrderingOp(Oid orderingOp, SortSupport ssup)
Definition: sortsupport.c:134
static int ApplySortComparator(Datum datum1, bool isNull1, Datum datum2, bool isNull2, SortSupport ssup)
Definition: sortsupport.h:200
int f1[ARRAY_SIZE]
Definition: sql-declare.c:113
void relation_close(Relation relation, LOCKMODE lockmode)
Definition: relation.c:205
void appendStringInfo(StringInfo str, const char *fmt,...)
Definition: stringinfo.c:97
void initStringInfo(StringInfo str)
Definition: stringinfo.c:59
double tupleFract
Definition: analyze.c:66
int attr_cnt
Definition: analyze.c:68
IndexInfo * indexInfo
Definition: analyze.c:65
VacAttrStats ** vacattrstats
Definition: analyze.c:67
float8 n_distinct
Definition: attoptcache.h:22
float8 n_distinct_inherited
Definition: attoptcache.h:23
SortSupport ssup
Definition: analyze.c:1814
TupleTableSlot * ecxt_scantuple
Definition: execnodes.h:255
AnalyzeForeignTable_function AnalyzeForeignTable
Definition: fdwapi.h:257
Definition: fmgr.h:57
ItemPointerData t_self
Definition: htup.h:65
int ii_NumIndexAttrs
Definition: execnodes.h:184
List * ii_Expressions
Definition: execnodes.h:187
AttrNumber ii_IndexAttrNumbers[INDEX_MAX_KEYS]
Definition: execnodes.h:186
List * ii_Predicate
Definition: execnodes.h:189
Relation index
Definition: genam.h:46
double num_heap_tuples
Definition: genam.h:52
bool analyze_only
Definition: genam.h:48
BufferAccessStrategy strategy
Definition: genam.h:53
Relation heaprel
Definition: genam.h:47
int message_level
Definition: genam.h:51
bool estimated_count
Definition: genam.h:50
Definition: pg_list.h:54
Definition: nodes.h:129
TupleDesc rd_att
Definition: rel.h:112
Oid rd_id
Definition: rel.h:113
Oid * rd_indcollation
Definition: rel.h:217
Form_pg_class rd_rel
Definition: rel.h:111
pg_prng_state randstate
Definition: sampling.h:49
bool ssup_nulls_first
Definition: sortsupport.h:75
MemoryContext ssup_cxt
Definition: sortsupport.h:66
Relation rs_rd
Definition: relscan.h:34
int32 attrtypmod
Definition: vacuum.h:127
bool stats_valid
Definition: vacuum.h:144
float4 stanullfrac
Definition: vacuum.h:145
Form_pg_type attrtype
Definition: vacuum.h:128
int16 stakind[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:148
int tupattnum
Definition: vacuum.h:171
MemoryContext anl_context
Definition: vacuum.h:130
Oid statypid[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:162
Oid staop[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:149
Oid stacoll[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:150
char statypalign[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:165
float4 * stanumbers[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:152
int rowstride
Definition: vacuum.h:176
Oid attrtypid
Definition: vacuum.h:126
HeapTuple * rows
Definition: vacuum.h:172
int minrows
Definition: vacuum.h:137
int attstattarget
Definition: vacuum.h:125
int32 stawidth
Definition: vacuum.h:146
void * extra_data
Definition: vacuum.h:138
bool statypbyval[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:164
int16 statyplen[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:163
bool * exprnulls
Definition: vacuum.h:175
TupleDesc tupDesc
Definition: vacuum.h:173
Datum * exprvals
Definition: vacuum.h:174
int numvalues[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:153
Datum * stavalues[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:154
float4 stadistinct
Definition: vacuum.h:147
int numnumbers[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:151
AnalyzeAttrComputeStatsFunc compute_stats
Definition: vacuum.h:136
Oid attrcollid
Definition: vacuum.h:129
bits32 options
Definition: vacuum.h:219
int log_min_duration
Definition: vacuum.h:227
void ReleaseSysCache(HeapTuple tuple)
Definition: syscache.c:266
HeapTuple SearchSysCache3(int cacheId, Datum key1, Datum key2, Datum key3)
Definition: syscache.c:240
Datum SysCacheGetAttr(int cacheId, HeapTuple tup, AttrNumber attributeNumber, bool *isNull)
Definition: syscache.c:479
HeapTuple SearchSysCache2(int cacheId, Datum key1, Datum key2)
Definition: syscache.c:229
#define SearchSysCacheCopy1(cacheId, key1)
Definition: syscache.h:86
void table_close(Relation relation, LOCKMODE lockmode)
Definition: table.c:126
Relation table_open(Oid relationId, LOCKMODE lockmode)
Definition: table.c:40
TupleTableSlot * table_slot_create(Relation relation, List **reglist)
Definition: tableam.c:91
static void table_endscan(TableScanDesc scan)
Definition: tableam.h:1019
static bool table_scan_analyze_next_tuple(TableScanDesc scan, TransactionId OldestXmin, double *liverows, double *deadrows, TupleTableSlot *slot)
Definition: tableam.h:1738
static bool table_scan_analyze_next_block(TableScanDesc scan, ReadStream *stream)
Definition: tableam.h:1722
static TableScanDesc table_beginscan_analyze(Relation rel)
Definition: tableam.h:1008
void SetRelationHasSubclass(Oid relationId, bool relhassubclass)
Definition: tablecmds.c:3484
#define InvalidTransactionId
Definition: transam.h:31
void free_conversion_map(TupleConversionMap *map)
Definition: tupconvert.c:299
TupleConversionMap * convert_tuples_by_name(TupleDesc indesc, TupleDesc outdesc)
Definition: tupconvert.c:102
HeapTuple execute_attr_map_tuple(HeapTuple tuple, TupleConversionMap *map)
Definition: tupconvert.c:154
bool equalRowTypes(TupleDesc tupdesc1, TupleDesc tupdesc2)
Definition: tupdesc.c:586
#define TupleDescAttr(tupdesc, i)
Definition: tupdesc.h:92
static HeapTuple ExecCopySlotHeapTuple(TupleTableSlot *slot)
Definition: tuptable.h:481
void vac_open_indexes(Relation relation, LOCKMODE lockmode, int *nindexes, Relation **Irel)
Definition: vacuum.c:2276
void vac_update_relstats(Relation relation, BlockNumber num_pages, double num_tuples, BlockNumber num_all_visible_pages, bool hasindex, TransactionId frozenxid, MultiXactId minmulti, bool *frozenxid_updated, bool *minmulti_updated, bool in_outer_xact)
Definition: vacuum.c:1398
Relation vacuum_open_relation(Oid relid, RangeVar *relation, bits32 options, bool verbose, LOCKMODE lmode)
Definition: vacuum.c:758
void vac_close_indexes(int nindexes, Relation *Irel, LOCKMODE lockmode)
Definition: vacuum.c:2319
void vacuum_delay_point(void)
Definition: vacuum.c:2340
bool vacuum_is_permitted_for_relation(Oid relid, Form_pg_class reltuple, bits32 options)
Definition: vacuum.c:706
#define VACOPT_VACUUM
Definition: vacuum.h:180
#define VACOPT_VERBOSE
Definition: vacuum.h:182
Datum(* AnalyzeAttrFetchFunc)(VacAttrStatsP stats, int rownum, bool *isNull)
Definition: vacuum.h:108
#define strVal(v)
Definition: value.h:82
#define VARSIZE_ANY(PTR)
Definition: varatt.h:311
void visibilitymap_count(Relation rel, BlockNumber *all_visible, BlockNumber *all_frozen)
void CommandCounterIncrement(void)
Definition: xact.c:1098