PostgreSQL Source Code  git master
analyze.c File Reference
#include "postgres.h"
#include <math.h>
#include "access/detoast.h"
#include "access/genam.h"
#include "access/multixact.h"
#include "access/relation.h"
#include "access/table.h"
#include "access/tableam.h"
#include "access/transam.h"
#include "access/tupconvert.h"
#include "access/visibilitymap.h"
#include "access/xact.h"
#include "catalog/index.h"
#include "catalog/indexing.h"
#include "catalog/pg_inherits.h"
#include "commands/dbcommands.h"
#include "commands/progress.h"
#include "commands/tablecmds.h"
#include "commands/vacuum.h"
#include "common/pg_prng.h"
#include "executor/executor.h"
#include "foreign/fdwapi.h"
#include "miscadmin.h"
#include "nodes/nodeFuncs.h"
#include "parser/parse_oper.h"
#include "parser/parse_relation.h"
#include "pgstat.h"
#include "postmaster/autovacuum.h"
#include "statistics/extended_stats_internal.h"
#include "statistics/statistics.h"
#include "storage/bufmgr.h"
#include "storage/procarray.h"
#include "utils/attoptcache.h"
#include "utils/datum.h"
#include "utils/guc.h"
#include "utils/lsyscache.h"
#include "utils/memutils.h"
#include "utils/pg_rusage.h"
#include "utils/sampling.h"
#include "utils/sortsupport.h"
#include "utils/spccache.h"
#include "utils/syscache.h"
#include "utils/timestamp.h"
Include dependency graph for analyze.c:

Go to the source code of this file.

Data Structures

struct  AnlIndexData
 
struct  ScalarMCVItem
 
struct  CompareScalarsContext
 

Macros

#define WIDTH_THRESHOLD   1024
 
#define swapInt(a, b)   do {int _tmp; _tmp=a; a=b; b=_tmp;} while(0)
 
#define swapDatum(a, b)   do {Datum _tmp; _tmp=a; a=b; b=_tmp;} while(0)
 

Typedefs

typedef struct AnlIndexData AnlIndexData
 

Functions

static void do_analyze_rel (Relation onerel, VacuumParams *params, List *va_cols, AcquireSampleRowsFunc acquirefunc, BlockNumber relpages, bool inh, bool in_outer_xact, int elevel)
 
static void compute_index_stats (Relation onerel, double totalrows, AnlIndexData *indexdata, int nindexes, HeapTuple *rows, int numrows, MemoryContext col_context)
 
static VacAttrStatsexamine_attribute (Relation onerel, int attnum, Node *index_expr)
 
static int acquire_sample_rows (Relation onerel, int elevel, HeapTuple *rows, int targrows, double *totalrows, double *totaldeadrows)
 
static int compare_rows (const void *a, const void *b, void *arg)
 
static int acquire_inherited_sample_rows (Relation onerel, int elevel, HeapTuple *rows, int targrows, double *totalrows, double *totaldeadrows)
 
static void update_attstats (Oid relid, bool inh, int natts, VacAttrStats **vacattrstats)
 
static Datum std_fetch_func (VacAttrStatsP stats, int rownum, bool *isNull)
 
static Datum ind_fetch_func (VacAttrStatsP stats, int rownum, bool *isNull)
 
void analyze_rel (Oid relid, RangeVar *relation, VacuumParams *params, List *va_cols, bool in_outer_xact, BufferAccessStrategy bstrategy)
 
static void compute_trivial_stats (VacAttrStatsP stats, AnalyzeAttrFetchFunc fetchfunc, int samplerows, double totalrows)
 
static void compute_distinct_stats (VacAttrStatsP stats, AnalyzeAttrFetchFunc fetchfunc, int samplerows, double totalrows)
 
static void compute_scalar_stats (VacAttrStatsP stats, AnalyzeAttrFetchFunc fetchfunc, int samplerows, double totalrows)
 
static int compare_scalars (const void *a, const void *b, void *arg)
 
static int compare_mcvs (const void *a, const void *b, void *arg)
 
static int analyze_mcv_list (int *mcv_counts, int num_mcv, double stadistinct, double stanullfrac, int samplerows, double totalrows)
 
bool std_typanalyze (VacAttrStats *stats)
 

Variables

int default_statistics_target = 100
 
static MemoryContext anl_context = NULL
 
static BufferAccessStrategy vac_strategy
 

Macro Definition Documentation

◆ swapDatum

#define swapDatum (   a,
  b 
)    do {Datum _tmp; _tmp=a; a=b; b=_tmp;} while(0)

Definition at line 1842 of file analyze.c.

◆ swapInt

#define swapInt (   a,
  b 
)    do {int _tmp; _tmp=a; a=b; b=_tmp;} while(0)

Definition at line 1841 of file analyze.c.

◆ WIDTH_THRESHOLD

#define WIDTH_THRESHOLD   1024

Definition at line 1839 of file analyze.c.

Typedef Documentation

◆ AnlIndexData

typedef struct AnlIndexData AnlIndexData

Function Documentation

◆ acquire_inherited_sample_rows()

static int acquire_inherited_sample_rows ( Relation  onerel,
int  elevel,
HeapTuple rows,
int  targrows,
double *  totalrows,
double *  totaldeadrows 
)
static

Definition at line 1386 of file analyze.c.

1389 {
1390  List *tableOIDs;
1391  Relation *rels;
1392  AcquireSampleRowsFunc *acquirefuncs;
1393  double *relblocks;
1394  double totalblocks;
1395  int numrows,
1396  nrels,
1397  i;
1398  ListCell *lc;
1399  bool has_child;
1400 
1401  /* Initialize output parameters to zero now, in case we exit early */
1402  *totalrows = 0;
1403  *totaldeadrows = 0;
1404 
1405  /*
1406  * Find all members of inheritance set. We only need AccessShareLock on
1407  * the children.
1408  */
1409  tableOIDs =
1411 
1412  /*
1413  * Check that there's at least one descendant, else fail. This could
1414  * happen despite analyze_rel's relhassubclass check, if table once had a
1415  * child but no longer does. In that case, we can clear the
1416  * relhassubclass field so as not to make the same mistake again later.
1417  * (This is safe because we hold ShareUpdateExclusiveLock.)
1418  */
1419  if (list_length(tableOIDs) < 2)
1420  {
1421  /* CCI because we already updated the pg_class row in this command */
1423  SetRelationHasSubclass(RelationGetRelid(onerel), false);
1424  ereport(elevel,
1425  (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no child tables",
1427  RelationGetRelationName(onerel))));
1428  return 0;
1429  }
1430 
1431  /*
1432  * Identify acquirefuncs to use, and count blocks in all the relations.
1433  * The result could overflow BlockNumber, so we use double arithmetic.
1434  */
1435  rels = (Relation *) palloc(list_length(tableOIDs) * sizeof(Relation));
1436  acquirefuncs = (AcquireSampleRowsFunc *)
1437  palloc(list_length(tableOIDs) * sizeof(AcquireSampleRowsFunc));
1438  relblocks = (double *) palloc(list_length(tableOIDs) * sizeof(double));
1439  totalblocks = 0;
1440  nrels = 0;
1441  has_child = false;
1442  foreach(lc, tableOIDs)
1443  {
1444  Oid childOID = lfirst_oid(lc);
1445  Relation childrel;
1446  AcquireSampleRowsFunc acquirefunc = NULL;
1447  BlockNumber relpages = 0;
1448 
1449  /* We already got the needed lock */
1450  childrel = table_open(childOID, NoLock);
1451 
1452  /* Ignore if temp table of another backend */
1453  if (RELATION_IS_OTHER_TEMP(childrel))
1454  {
1455  /* ... but release the lock on it */
1456  Assert(childrel != onerel);
1457  table_close(childrel, AccessShareLock);
1458  continue;
1459  }
1460 
1461  /* Check table type (MATVIEW can't happen, but might as well allow) */
1462  if (childrel->rd_rel->relkind == RELKIND_RELATION ||
1463  childrel->rd_rel->relkind == RELKIND_MATVIEW)
1464  {
1465  /* Regular table, so use the regular row acquisition function */
1466  acquirefunc = acquire_sample_rows;
1467  relpages = RelationGetNumberOfBlocks(childrel);
1468  }
1469  else if (childrel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
1470  {
1471  /*
1472  * For a foreign table, call the FDW's hook function to see
1473  * whether it supports analysis.
1474  */
1475  FdwRoutine *fdwroutine;
1476  bool ok = false;
1477 
1478  fdwroutine = GetFdwRoutineForRelation(childrel, false);
1479 
1480  if (fdwroutine->AnalyzeForeignTable != NULL)
1481  ok = fdwroutine->AnalyzeForeignTable(childrel,
1482  &acquirefunc,
1483  &relpages);
1484 
1485  if (!ok)
1486  {
1487  /* ignore, but release the lock on it */
1488  Assert(childrel != onerel);
1489  table_close(childrel, AccessShareLock);
1490  continue;
1491  }
1492  }
1493  else
1494  {
1495  /*
1496  * ignore, but release the lock on it. don't try to unlock the
1497  * passed-in relation
1498  */
1499  Assert(childrel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE);
1500  if (childrel != onerel)
1501  table_close(childrel, AccessShareLock);
1502  else
1503  table_close(childrel, NoLock);
1504  continue;
1505  }
1506 
1507  /* OK, we'll process this child */
1508  has_child = true;
1509  rels[nrels] = childrel;
1510  acquirefuncs[nrels] = acquirefunc;
1511  relblocks[nrels] = (double) relpages;
1512  totalblocks += (double) relpages;
1513  nrels++;
1514  }
1515 
1516  /*
1517  * If we don't have at least one child table to consider, fail. If the
1518  * relation is a partitioned table, it's not counted as a child table.
1519  */
1520  if (!has_child)
1521  {
1522  ereport(elevel,
1523  (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no analyzable child tables",
1525  RelationGetRelationName(onerel))));
1526  return 0;
1527  }
1528 
1529  /*
1530  * Now sample rows from each relation, proportionally to its fraction of
1531  * the total block count. (This might be less than desirable if the child
1532  * rels have radically different free-space percentages, but it's not
1533  * clear that it's worth working harder.)
1534  */
1536  nrels);
1537  numrows = 0;
1538  for (i = 0; i < nrels; i++)
1539  {
1540  Relation childrel = rels[i];
1541  AcquireSampleRowsFunc acquirefunc = acquirefuncs[i];
1542  double childblocks = relblocks[i];
1543 
1544  /*
1545  * Report progress. The sampling function will normally report blocks
1546  * done/total, but we need to reset them to 0 here, so that they don't
1547  * show an old value until that.
1548  */
1549  {
1550  const int progress_index[] = {
1554  };
1555  const int64 progress_vals[] = {
1556  RelationGetRelid(childrel),
1557  0,
1558  0,
1559  };
1560 
1561  pgstat_progress_update_multi_param(3, progress_index, progress_vals);
1562  }
1563 
1564  if (childblocks > 0)
1565  {
1566  int childtargrows;
1567 
1568  childtargrows = (int) rint(targrows * childblocks / totalblocks);
1569  /* Make sure we don't overrun due to roundoff error */
1570  childtargrows = Min(childtargrows, targrows - numrows);
1571  if (childtargrows > 0)
1572  {
1573  int childrows;
1574  double trows,
1575  tdrows;
1576 
1577  /* Fetch a random sample of the child's rows */
1578  childrows = (*acquirefunc) (childrel, elevel,
1579  rows + numrows, childtargrows,
1580  &trows, &tdrows);
1581 
1582  /* We may need to convert from child's rowtype to parent's */
1583  if (childrows > 0 &&
1584  !equalRowTypes(RelationGetDescr(childrel),
1585  RelationGetDescr(onerel)))
1586  {
1587  TupleConversionMap *map;
1588 
1589  map = convert_tuples_by_name(RelationGetDescr(childrel),
1590  RelationGetDescr(onerel));
1591  if (map != NULL)
1592  {
1593  int j;
1594 
1595  for (j = 0; j < childrows; j++)
1596  {
1597  HeapTuple newtup;
1598 
1599  newtup = execute_attr_map_tuple(rows[numrows + j], map);
1600  heap_freetuple(rows[numrows + j]);
1601  rows[numrows + j] = newtup;
1602  }
1603  free_conversion_map(map);
1604  }
1605  }
1606 
1607  /* And add to counts */
1608  numrows += childrows;
1609  *totalrows += trows;
1610  *totaldeadrows += tdrows;
1611  }
1612  }
1613 
1614  /*
1615  * Note: we cannot release the child-table locks, since we may have
1616  * pointers to their TOAST tables in the sampled rows.
1617  */
1618  table_close(childrel, NoLock);
1620  i + 1);
1621  }
1622 
1623  return numrows;
1624 }
void pgstat_progress_update_param(int index, int64 val)
void pgstat_progress_update_multi_param(int nparam, const int *index, const int64 *val)
uint32 BlockNumber
Definition: block.h:31
#define RelationGetNumberOfBlocks(reln)
Definition: bufmgr.h:229
#define Min(x, y)
Definition: c.h:991
static int acquire_sample_rows(Relation onerel, int elevel, HeapTuple *rows, int targrows, double *totalrows, double *totaldeadrows)
Definition: analyze.c:1139
int errmsg(const char *fmt,...)
Definition: elog.c:1072
#define ereport(elevel,...)
Definition: elog.h:149
int(* AcquireSampleRowsFunc)(Relation relation, int elevel, HeapTuple *rows, int targrows, double *totalrows, double *totaldeadrows)
Definition: fdwapi.h:151
FdwRoutine * GetFdwRoutineForRelation(Relation relation, bool makecopy)
Definition: foreign.c:432
void heap_freetuple(HeapTuple htup)
Definition: heaptuple.c:1434
int j
Definition: isn.c:74
int i
Definition: isn.c:73
Assert(fmt[strlen(fmt) - 1] !='\n')
#define NoLock
Definition: lockdefs.h:34
#define AccessShareLock
Definition: lockdefs.h:36
char * get_namespace_name(Oid nspid)
Definition: lsyscache.c:3344
void * palloc(Size size)
Definition: mcxt.c:1304
List * find_all_inheritors(Oid parentrelId, LOCKMODE lockmode, List **numparents)
Definition: pg_inherits.c:255
static int list_length(const List *l)
Definition: pg_list.h:152
#define lfirst_oid(lc)
Definition: pg_list.h:174
unsigned int Oid
Definition: postgres_ext.h:31
#define PROGRESS_ANALYZE_BLOCKS_DONE
Definition: progress.h:42
#define PROGRESS_ANALYZE_CHILD_TABLES_TOTAL
Definition: progress.h:45
#define PROGRESS_ANALYZE_BLOCKS_TOTAL
Definition: progress.h:41
#define PROGRESS_ANALYZE_CHILD_TABLES_DONE
Definition: progress.h:46
#define PROGRESS_ANALYZE_CURRENT_CHILD_TABLE_RELID
Definition: progress.h:47
#define RelationGetRelid(relation)
Definition: rel.h:507
#define RelationGetDescr(relation)
Definition: rel.h:533
#define RelationGetRelationName(relation)
Definition: rel.h:541
#define RELATION_IS_OTHER_TEMP(relation)
Definition: rel.h:660
#define RelationGetNamespace(relation)
Definition: rel.h:548
struct RelationData * Relation
Definition: relcache.h:27
AnalyzeForeignTable_function AnalyzeForeignTable
Definition: fdwapi.h:257
Definition: pg_list.h:54
Form_pg_class rd_rel
Definition: rel.h:111
void table_close(Relation relation, LOCKMODE lockmode)
Definition: table.c:126
Relation table_open(Oid relationId, LOCKMODE lockmode)
Definition: table.c:40
void SetRelationHasSubclass(Oid relationId, bool relhassubclass)
Definition: tablecmds.c:3578
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
void CommandCounterIncrement(void)
Definition: xact.c:1079

References AccessShareLock, acquire_sample_rows(), FdwRoutine::AnalyzeForeignTable, Assert(), CommandCounterIncrement(), convert_tuples_by_name(), equalRowTypes(), ereport, errmsg(), execute_attr_map_tuple(), find_all_inheritors(), free_conversion_map(), get_namespace_name(), GetFdwRoutineForRelation(), heap_freetuple(), i, j, lfirst_oid, list_length(), Min, NoLock, palloc(), pgstat_progress_update_multi_param(), pgstat_progress_update_param(), PROGRESS_ANALYZE_BLOCKS_DONE, PROGRESS_ANALYZE_BLOCKS_TOTAL, PROGRESS_ANALYZE_CHILD_TABLES_DONE, PROGRESS_ANALYZE_CHILD_TABLES_TOTAL, PROGRESS_ANALYZE_CURRENT_CHILD_TABLE_RELID, RelationData::rd_rel, RELATION_IS_OTHER_TEMP, RelationGetDescr, RelationGetNamespace, RelationGetNumberOfBlocks, RelationGetRelationName, RelationGetRelid, SetRelationHasSubclass(), table_close(), and table_open().

Referenced by do_analyze_rel().

◆ acquire_sample_rows()

static int acquire_sample_rows ( Relation  onerel,
int  elevel,
HeapTuple rows,
int  targrows,
double *  totalrows,
double *  totaldeadrows 
)
static

Definition at line 1139 of file analyze.c.

1142 {
1143  int numrows = 0; /* # rows now in reservoir */
1144  double samplerows = 0; /* total # rows collected */
1145  double liverows = 0; /* # live rows seen */
1146  double deadrows = 0; /* # dead rows seen */
1147  double rowstoskip = -1; /* -1 means not set yet */
1148  uint32 randseed; /* Seed for block sampler(s) */
1149  BlockNumber totalblocks;
1150  TransactionId OldestXmin;
1151  BlockSamplerData bs;
1152  ReservoirStateData rstate;
1153  TupleTableSlot *slot;
1154  TableScanDesc scan;
1155  BlockNumber nblocks;
1156  BlockNumber blksdone = 0;
1157 #ifdef USE_PREFETCH
1158  int prefetch_maximum = 0; /* blocks to prefetch if enabled */
1159  BlockSamplerData prefetch_bs;
1160 #endif
1161 
1162  Assert(targrows > 0);
1163 
1164  totalblocks = RelationGetNumberOfBlocks(onerel);
1165 
1166  /* Need a cutoff xmin for HeapTupleSatisfiesVacuum */
1167  OldestXmin = GetOldestNonRemovableTransactionId(onerel);
1168 
1169  /* Prepare for sampling block numbers */
1170  randseed = pg_prng_uint32(&pg_global_prng_state);
1171  nblocks = BlockSampler_Init(&bs, totalblocks, targrows, randseed);
1172 
1173 #ifdef USE_PREFETCH
1174  prefetch_maximum = get_tablespace_maintenance_io_concurrency(onerel->rd_rel->reltablespace);
1175  /* Create another BlockSampler, using the same seed, for prefetching */
1176  if (prefetch_maximum)
1177  (void) BlockSampler_Init(&prefetch_bs, totalblocks, targrows, randseed);
1178 #endif
1179 
1180  /* Report sampling block numbers */
1182  nblocks);
1183 
1184  /* Prepare for sampling rows */
1185  reservoir_init_selection_state(&rstate, targrows);
1186 
1187  scan = table_beginscan_analyze(onerel);
1188  slot = table_slot_create(onerel, NULL);
1189 
1190 #ifdef USE_PREFETCH
1191 
1192  /*
1193  * If we are doing prefetching, then go ahead and tell the kernel about
1194  * the first set of pages we are going to want. This also moves our
1195  * iterator out ahead of the main one being used, where we will keep it so
1196  * that we're always pre-fetching out prefetch_maximum number of blocks
1197  * ahead.
1198  */
1199  if (prefetch_maximum)
1200  {
1201  for (int i = 0; i < prefetch_maximum; i++)
1202  {
1203  BlockNumber prefetch_block;
1204 
1205  if (!BlockSampler_HasMore(&prefetch_bs))
1206  break;
1207 
1208  prefetch_block = BlockSampler_Next(&prefetch_bs);
1209  PrefetchBuffer(scan->rs_rd, MAIN_FORKNUM, prefetch_block);
1210  }
1211  }
1212 #endif
1213 
1214  /* Outer loop over blocks to sample */
1215  while (BlockSampler_HasMore(&bs))
1216  {
1217  bool block_accepted;
1218  BlockNumber targblock = BlockSampler_Next(&bs);
1219 #ifdef USE_PREFETCH
1220  BlockNumber prefetch_targblock = InvalidBlockNumber;
1221 
1222  /*
1223  * Make sure that every time the main BlockSampler is moved forward
1224  * that our prefetch BlockSampler also gets moved forward, so that we
1225  * always stay out ahead.
1226  */
1227  if (prefetch_maximum && BlockSampler_HasMore(&prefetch_bs))
1228  prefetch_targblock = BlockSampler_Next(&prefetch_bs);
1229 #endif
1230 
1232 
1233  block_accepted = table_scan_analyze_next_block(scan, targblock, vac_strategy);
1234 
1235 #ifdef USE_PREFETCH
1236 
1237  /*
1238  * When pre-fetching, after we get a block, tell the kernel about the
1239  * next one we will want, if there's any left.
1240  *
1241  * We want to do this even if the table_scan_analyze_next_block() call
1242  * above decides against analyzing the block it picked.
1243  */
1244  if (prefetch_maximum && prefetch_targblock != InvalidBlockNumber)
1245  PrefetchBuffer(scan->rs_rd, MAIN_FORKNUM, prefetch_targblock);
1246 #endif
1247 
1248  /*
1249  * Don't analyze if table_scan_analyze_next_block() indicated this
1250  * block is unsuitable for analyzing.
1251  */
1252  if (!block_accepted)
1253  continue;
1254 
1255  while (table_scan_analyze_next_tuple(scan, OldestXmin, &liverows, &deadrows, slot))
1256  {
1257  /*
1258  * The first targrows sample rows are simply copied into the
1259  * reservoir. Then we start replacing tuples in the sample until
1260  * we reach the end of the relation. This algorithm is from Jeff
1261  * Vitter's paper (see full citation in utils/misc/sampling.c). It
1262  * works by repeatedly computing the number of tuples to skip
1263  * before selecting a tuple, which replaces a randomly chosen
1264  * element of the reservoir (current set of tuples). At all times
1265  * the reservoir is a true random sample of the tuples we've
1266  * passed over so far, so when we fall off the end of the relation
1267  * we're done.
1268  */
1269  if (numrows < targrows)
1270  rows[numrows++] = ExecCopySlotHeapTuple(slot);
1271  else
1272  {
1273  /*
1274  * t in Vitter's paper is the number of records already
1275  * processed. If we need to compute a new S value, we must
1276  * use the not-yet-incremented value of samplerows as t.
1277  */
1278  if (rowstoskip < 0)
1279  rowstoskip = reservoir_get_next_S(&rstate, samplerows, targrows);
1280 
1281  if (rowstoskip <= 0)
1282  {
1283  /*
1284  * Found a suitable tuple, so save it, replacing one old
1285  * tuple at random
1286  */
1287  int k = (int) (targrows * sampler_random_fract(&rstate.randstate));
1288 
1289  Assert(k >= 0 && k < targrows);
1290  heap_freetuple(rows[k]);
1291  rows[k] = ExecCopySlotHeapTuple(slot);
1292  }
1293 
1294  rowstoskip -= 1;
1295  }
1296 
1297  samplerows += 1;
1298  }
1299 
1301  ++blksdone);
1302  }
1303 
1305  table_endscan(scan);
1306 
1307  /*
1308  * If we didn't find as many tuples as we wanted then we're done. No sort
1309  * is needed, since they're already in order.
1310  *
1311  * Otherwise we need to sort the collected tuples by position
1312  * (itempointer). It's not worth worrying about corner cases where the
1313  * tuples are already sorted.
1314  */
1315  if (numrows == targrows)
1316  qsort_interruptible(rows, numrows, sizeof(HeapTuple),
1317  compare_rows, NULL);
1318 
1319  /*
1320  * Estimate total numbers of live and dead rows in relation, extrapolating
1321  * on the assumption that the average tuple density in pages we didn't
1322  * scan is the same as in the pages we did scan. Since what we scanned is
1323  * a random sample of the pages in the relation, this should be a good
1324  * assumption.
1325  */
1326  if (bs.m > 0)
1327  {
1328  *totalrows = floor((liverows / bs.m) * totalblocks + 0.5);
1329  *totaldeadrows = floor((deadrows / bs.m) * totalblocks + 0.5);
1330  }
1331  else
1332  {
1333  *totalrows = 0.0;
1334  *totaldeadrows = 0.0;
1335  }
1336 
1337  /*
1338  * Emit some interesting relation info
1339  */
1340  ereport(elevel,
1341  (errmsg("\"%s\": scanned %d of %u pages, "
1342  "containing %.0f live rows and %.0f dead rows; "
1343  "%d rows in sample, %.0f estimated total rows",
1344  RelationGetRelationName(onerel),
1345  bs.m, totalblocks,
1346  liverows, deadrows,
1347  numrows, *totalrows)));
1348 
1349  return numrows;
1350 }
#define InvalidBlockNumber
Definition: block.h:33
PrefetchBufferResult PrefetchBuffer(Relation reln, ForkNumber forkNum, BlockNumber blockNum)
Definition: bufmgr.c:627
unsigned int uint32
Definition: c.h:493
uint32 TransactionId
Definition: c.h:639
static BufferAccessStrategy vac_strategy
Definition: analyze.c:77
static int compare_rows(const void *a, const void *b, void *arg)
Definition: analyze.c:1356
void ExecDropSingleTupleTableSlot(TupleTableSlot *slot)
Definition: execTuples.c:1341
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
void qsort_interruptible(void *base, size_t nel, size_t elsize, qsort_arg_comparator cmp, void *arg)
TransactionId GetOldestNonRemovableTransactionId(Relation rel)
Definition: procarray.c:1993
@ 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
int get_tablespace_maintenance_io_concurrency(Oid spcid)
Definition: spccache.c:229
pg_prng_state randstate
Definition: sampling.h:49
Relation rs_rd
Definition: relscan.h:34
TupleTableSlot * table_slot_create(Relation relation, List **reglist)
Definition: tableam.c:91
static bool table_scan_analyze_next_block(TableScanDesc scan, BlockNumber blockno, BufferAccessStrategy bstrategy)
Definition: tableam.h:1758
static void table_endscan(TableScanDesc scan)
Definition: tableam.h:1028
static bool table_scan_analyze_next_tuple(TableScanDesc scan, TransactionId OldestXmin, double *liverows, double *deadrows, TupleTableSlot *slot)
Definition: tableam.h:1776
static TableScanDesc table_beginscan_analyze(Relation rel)
Definition: tableam.h:1017
static HeapTuple ExecCopySlotHeapTuple(TupleTableSlot *slot)
Definition: tuptable.h:481
void vacuum_delay_point(void)
Definition: vacuum.c:2337

References Assert(), BlockSampler_HasMore(), BlockSampler_Init(), BlockSampler_Next(), compare_rows(), ereport, errmsg(), ExecCopySlotHeapTuple(), ExecDropSingleTupleTableSlot(), get_tablespace_maintenance_io_concurrency(), GetOldestNonRemovableTransactionId(), heap_freetuple(), i, InvalidBlockNumber, BlockSamplerData::m, MAIN_FORKNUM, pg_global_prng_state, pg_prng_uint32(), pgstat_progress_update_param(), PrefetchBuffer(), PROGRESS_ANALYZE_BLOCKS_DONE, PROGRESS_ANALYZE_BLOCKS_TOTAL, qsort_interruptible(), ReservoirStateData::randstate, RelationData::rd_rel, RelationGetNumberOfBlocks, RelationGetRelationName, reservoir_get_next_S(), reservoir_init_selection_state(), TableScanDescData::rs_rd, sampler_random_fract(), table_beginscan_analyze(), table_endscan(), table_scan_analyze_next_block(), table_scan_analyze_next_tuple(), table_slot_create(), vac_strategy, and vacuum_delay_point().

Referenced by acquire_inherited_sample_rows(), and analyze_rel().

◆ analyze_mcv_list()

static int analyze_mcv_list ( int *  mcv_counts,
int  num_mcv,
double  stadistinct,
double  stanullfrac,
int  samplerows,
double  totalrows 
)
static

Definition at line 2975 of file analyze.c.

2981 {
2982  double ndistinct_table;
2983  double sumcount;
2984  int i;
2985 
2986  /*
2987  * If the entire table was sampled, keep the whole list. This also
2988  * protects us against division by zero in the code below.
2989  */
2990  if (samplerows == totalrows || totalrows <= 1.0)
2991  return num_mcv;
2992 
2993  /* Re-extract the estimated number of distinct nonnull values in table */
2994  ndistinct_table = stadistinct;
2995  if (ndistinct_table < 0)
2996  ndistinct_table = -ndistinct_table * totalrows;
2997 
2998  /*
2999  * Exclude the least common values from the MCV list, if they are not
3000  * significantly more common than the estimated selectivity they would
3001  * have if they weren't in the list. All non-MCV values are assumed to be
3002  * equally common, after taking into account the frequencies of all the
3003  * values in the MCV list and the number of nulls (c.f. eqsel()).
3004  *
3005  * Here sumcount tracks the total count of all but the last (least common)
3006  * value in the MCV list, allowing us to determine the effect of excluding
3007  * that value from the list.
3008  *
3009  * Note that we deliberately do this by removing values from the full
3010  * list, rather than starting with an empty list and adding values,
3011  * because the latter approach can fail to add any values if all the most
3012  * common values have around the same frequency and make up the majority
3013  * of the table, so that the overall average frequency of all values is
3014  * roughly the same as that of the common values. This would lead to any
3015  * uncommon values being significantly overestimated.
3016  */
3017  sumcount = 0.0;
3018  for (i = 0; i < num_mcv - 1; i++)
3019  sumcount += mcv_counts[i];
3020 
3021  while (num_mcv > 0)
3022  {
3023  double selec,
3024  otherdistinct,
3025  N,
3026  n,
3027  K,
3028  variance,
3029  stddev;
3030 
3031  /*
3032  * Estimated selectivity the least common value would have if it
3033  * wasn't in the MCV list (c.f. eqsel()).
3034  */
3035  selec = 1.0 - sumcount / samplerows - stanullfrac;
3036  if (selec < 0.0)
3037  selec = 0.0;
3038  if (selec > 1.0)
3039  selec = 1.0;
3040  otherdistinct = ndistinct_table - (num_mcv - 1);
3041  if (otherdistinct > 1)
3042  selec /= otherdistinct;
3043 
3044  /*
3045  * If the value is kept in the MCV list, its population frequency is
3046  * assumed to equal its sample frequency. We use the lower end of a
3047  * textbook continuity-corrected Wald-type confidence interval to
3048  * determine if that is significantly more common than the non-MCV
3049  * frequency --- specifically we assume the population frequency is
3050  * highly likely to be within around 2 standard errors of the sample
3051  * frequency, which equates to an interval of 2 standard deviations
3052  * either side of the sample count, plus an additional 0.5 for the
3053  * continuity correction. Since we are sampling without replacement,
3054  * this is a hypergeometric distribution.
3055  *
3056  * XXX: Empirically, this approach seems to work quite well, but it
3057  * may be worth considering more advanced techniques for estimating
3058  * the confidence interval of the hypergeometric distribution.
3059  */
3060  N = totalrows;
3061  n = samplerows;
3062  K = N * mcv_counts[num_mcv - 1] / n;
3063  variance = n * K * (N - K) * (N - n) / (N * N * (N - 1));
3064  stddev = sqrt(variance);
3065 
3066  if (mcv_counts[num_mcv - 1] > selec * samplerows + 2 * stddev + 0.5)
3067  {
3068  /*
3069  * The value is significantly more common than the non-MCV
3070  * selectivity would suggest. Keep it, and all the other more
3071  * common values in the list.
3072  */
3073  break;
3074  }
3075  else
3076  {
3077  /* Discard this value and consider the next least common value */
3078  num_mcv--;
3079  if (num_mcv == 0)
3080  break;
3081  sumcount -= mcv_counts[num_mcv - 1];
3082  }
3083  }
3084  return num_mcv;
3085 }
#define K(t)
Definition: sha1.c:66

References i, and K.

Referenced by compute_distinct_stats(), and compute_scalar_stats().

◆ analyze_rel()

void analyze_rel ( Oid  relid,
RangeVar relation,
VacuumParams params,
List va_cols,
bool  in_outer_xact,
BufferAccessStrategy  bstrategy 
)

Definition at line 111 of file analyze.c.

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 }
void pgstat_progress_start_command(ProgressCommandType cmdtype, Oid relid)
void pgstat_progress_end_command(void)
@ PROGRESS_COMMAND_ANALYZE
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
#define WARNING
Definition: elog.h:36
#define DEBUG2
Definition: elog.h:29
#define INFO
Definition: elog.h:34
#define ShareUpdateExclusiveLock
Definition: lockdefs.h:39
#define CHECK_FOR_INTERRUPTS()
Definition: miscadmin.h:122
void relation_close(Relation relation, LOCKMODE lockmode)
Definition: relation.c:205
bits32 options
Definition: vacuum.h:218
int log_min_duration
Definition: vacuum.h:226
Relation vacuum_open_relation(Oid relid, RangeVar *relation, bits32 options, bool verbose, LOCKMODE lmode)
Definition: vacuum.c:759
bool vacuum_is_permitted_for_relation(Oid relid, Form_pg_class reltuple, bits32 options)
Definition: vacuum.c:707
#define VACOPT_VACUUM
Definition: vacuum.h:179
#define VACOPT_VERBOSE
Definition: vacuum.h:181

References acquire_sample_rows(), FdwRoutine::AnalyzeForeignTable, CHECK_FOR_INTERRUPTS, DEBUG2, do_analyze_rel(), ereport, errmsg(), GetFdwRoutineForRelation(), INFO, VacuumParams::log_min_duration, NoLock, VacuumParams::options, pgstat_progress_end_command(), pgstat_progress_start_command(), PROGRESS_COMMAND_ANALYZE, RelationData::rd_rel, relation_close(), RELATION_IS_OTHER_TEMP, RelationGetNumberOfBlocks, RelationGetRelationName, RelationGetRelid, ShareUpdateExclusiveLock, vac_strategy, VACOPT_VACUUM, VACOPT_VERBOSE, vacuum_is_permitted_for_relation(), vacuum_open_relation(), and WARNING.

Referenced by vacuum().

◆ compare_mcvs()

static int compare_mcvs ( const void *  a,
const void *  b,
void *  arg 
)
static

Definition at line 2957 of file analyze.c.

2958 {
2959  int da = ((const ScalarMCVItem *) a)->first;
2960  int db = ((const ScalarMCVItem *) b)->first;
2961 
2962  return da - db;
2963 }
int b
Definition: isn.c:70
int a
Definition: isn.c:69

References a, and b.

Referenced by compute_scalar_stats().

◆ compare_rows()

static int compare_rows ( const void *  a,
const void *  b,
void *  arg 
)
static

Definition at line 1356 of file analyze.c.

1357 {
1358  HeapTuple ha = *(const HeapTuple *) a;
1359  HeapTuple hb = *(const HeapTuple *) b;
1364 
1365  if (ba < bb)
1366  return -1;
1367  if (ba > bb)
1368  return 1;
1369  if (oa < ob)
1370  return -1;
1371  if (oa > ob)
1372  return 1;
1373  return 0;
1374 }
static OffsetNumber ItemPointerGetOffsetNumber(const ItemPointerData *pointer)
Definition: itemptr.h:124
static BlockNumber ItemPointerGetBlockNumber(const ItemPointerData *pointer)
Definition: itemptr.h:103
uint16 OffsetNumber
Definition: off.h:24
ItemPointerData t_self
Definition: htup.h:65

References a, b, ItemPointerGetBlockNumber(), ItemPointerGetOffsetNumber(), and HeapTupleData::t_self.

Referenced by acquire_sample_rows().

◆ compare_scalars()

static int compare_scalars ( const void *  a,
const void *  b,
void *  arg 
)
static

Definition at line 2926 of file analyze.c.

2927 {
2928  Datum da = ((const ScalarItem *) a)->value;
2929  int ta = ((const ScalarItem *) a)->tupno;
2930  Datum db = ((const ScalarItem *) b)->value;
2931  int tb = ((const ScalarItem *) b)->tupno;
2933  int compare;
2934 
2935  compare = ApplySortComparator(da, false, db, false, cxt->ssup);
2936  if (compare != 0)
2937  return compare;
2938 
2939  /*
2940  * The two datums are equal, so update cxt->tupnoLink[].
2941  */
2942  if (cxt->tupnoLink[ta] < tb)
2943  cxt->tupnoLink[ta] = tb;
2944  if (cxt->tupnoLink[tb] < ta)
2945  cxt->tupnoLink[tb] = ta;
2946 
2947  /*
2948  * For equal datums, sort by tupno
2949  */
2950  return ta - tb;
2951 }
static int compare(const void *arg1, const void *arg2)
Definition: geqo_pool.c:145
void * arg
uintptr_t Datum
Definition: postgres.h:64
static int ApplySortComparator(Datum datum1, bool isNull1, Datum datum2, bool isNull2, SortSupport ssup)
Definition: sortsupport.h:200
SortSupport ssup
Definition: analyze.c:1855

References a, ApplySortComparator(), arg, b, compare(), CompareScalarsContext::ssup, and CompareScalarsContext::tupnoLink.

Referenced by compute_scalar_stats().

◆ compute_distinct_stats()

static void compute_distinct_stats ( VacAttrStatsP  stats,
AnalyzeAttrFetchFunc  fetchfunc,
int  samplerows,
double  totalrows 
)
static

Definition at line 2054 of file analyze.c.

2058 {
2059  int i;
2060  int null_cnt = 0;
2061  int nonnull_cnt = 0;
2062  int toowide_cnt = 0;
2063  double total_width = 0;
2064  bool is_varlena = (!stats->attrtype->typbyval &&
2065  stats->attrtype->typlen == -1);
2066  bool is_varwidth = (!stats->attrtype->typbyval &&
2067  stats->attrtype->typlen < 0);
2068  FmgrInfo f_cmpeq;
2069  typedef struct
2070  {
2071  Datum value;
2072  int count;
2073  } TrackItem;
2074  TrackItem *track;
2075  int track_cnt,
2076  track_max;
2077  int num_mcv = stats->attstattarget;
2078  StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2079 
2080  /*
2081  * We track up to 2*n values for an n-element MCV list; but at least 10
2082  */
2083  track_max = 2 * num_mcv;
2084  if (track_max < 10)
2085  track_max = 10;
2086  track = (TrackItem *) palloc(track_max * sizeof(TrackItem));
2087  track_cnt = 0;
2088 
2089  fmgr_info(mystats->eqfunc, &f_cmpeq);
2090 
2091  for (i = 0; i < samplerows; i++)
2092  {
2093  Datum value;
2094  bool isnull;
2095  bool match;
2096  int firstcount1,
2097  j;
2098 
2100 
2101  value = fetchfunc(stats, i, &isnull);
2102 
2103  /* Check for null/nonnull */
2104  if (isnull)
2105  {
2106  null_cnt++;
2107  continue;
2108  }
2109  nonnull_cnt++;
2110 
2111  /*
2112  * If it's a variable-width field, add up widths for average width
2113  * calculation. Note that if the value is toasted, we use the toasted
2114  * width. We don't bother with this calculation if it's a fixed-width
2115  * type.
2116  */
2117  if (is_varlena)
2118  {
2119  total_width += VARSIZE_ANY(DatumGetPointer(value));
2120 
2121  /*
2122  * If the value is toasted, we want to detoast it just once to
2123  * avoid repeated detoastings and resultant excess memory usage
2124  * during the comparisons. Also, check to see if the value is
2125  * excessively wide, and if so don't detoast at all --- just
2126  * ignore the value.
2127  */
2129  {
2130  toowide_cnt++;
2131  continue;
2132  }
2134  }
2135  else if (is_varwidth)
2136  {
2137  /* must be cstring */
2138  total_width += strlen(DatumGetCString(value)) + 1;
2139  }
2140 
2141  /*
2142  * See if the value matches anything we're already tracking.
2143  */
2144  match = false;
2145  firstcount1 = track_cnt;
2146  for (j = 0; j < track_cnt; j++)
2147  {
2148  if (DatumGetBool(FunctionCall2Coll(&f_cmpeq,
2149  stats->attrcollid,
2150  value, track[j].value)))
2151  {
2152  match = true;
2153  break;
2154  }
2155  if (j < firstcount1 && track[j].count == 1)
2156  firstcount1 = j;
2157  }
2158 
2159  if (match)
2160  {
2161  /* Found a match */
2162  track[j].count++;
2163  /* This value may now need to "bubble up" in the track list */
2164  while (j > 0 && track[j].count > track[j - 1].count)
2165  {
2166  swapDatum(track[j].value, track[j - 1].value);
2167  swapInt(track[j].count, track[j - 1].count);
2168  j--;
2169  }
2170  }
2171  else
2172  {
2173  /* No match. Insert at head of count-1 list */
2174  if (track_cnt < track_max)
2175  track_cnt++;
2176  for (j = track_cnt - 1; j > firstcount1; j--)
2177  {
2178  track[j].value = track[j - 1].value;
2179  track[j].count = track[j - 1].count;
2180  }
2181  if (firstcount1 < track_cnt)
2182  {
2183  track[firstcount1].value = value;
2184  track[firstcount1].count = 1;
2185  }
2186  }
2187  }
2188 
2189  /* We can only compute real stats if we found some non-null values. */
2190  if (nonnull_cnt > 0)
2191  {
2192  int nmultiple,
2193  summultiple;
2194 
2195  stats->stats_valid = true;
2196  /* Do the simple null-frac and width stats */
2197  stats->stanullfrac = (double) null_cnt / (double) samplerows;
2198  if (is_varwidth)
2199  stats->stawidth = total_width / (double) nonnull_cnt;
2200  else
2201  stats->stawidth = stats->attrtype->typlen;
2202 
2203  /* Count the number of values we found multiple times */
2204  summultiple = 0;
2205  for (nmultiple = 0; nmultiple < track_cnt; nmultiple++)
2206  {
2207  if (track[nmultiple].count == 1)
2208  break;
2209  summultiple += track[nmultiple].count;
2210  }
2211 
2212  if (nmultiple == 0)
2213  {
2214  /*
2215  * If we found no repeated non-null values, assume it's a unique
2216  * column; but be sure to discount for any nulls we found.
2217  */
2218  stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2219  }
2220  else if (track_cnt < track_max && toowide_cnt == 0 &&
2221  nmultiple == track_cnt)
2222  {
2223  /*
2224  * Our track list includes every value in the sample, and every
2225  * value appeared more than once. Assume the column has just
2226  * these values. (This case is meant to address columns with
2227  * small, fixed sets of possible values, such as boolean or enum
2228  * columns. If there are any values that appear just once in the
2229  * sample, including too-wide values, we should assume that that's
2230  * not what we're dealing with.)
2231  */
2232  stats->stadistinct = track_cnt;
2233  }
2234  else
2235  {
2236  /*----------
2237  * Estimate the number of distinct values using the estimator
2238  * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2239  * n*d / (n - f1 + f1*n/N)
2240  * where f1 is the number of distinct values that occurred
2241  * exactly once in our sample of n rows (from a total of N),
2242  * and d is the total number of distinct values in the sample.
2243  * This is their Duj1 estimator; the other estimators they
2244  * recommend are considerably more complex, and are numerically
2245  * very unstable when n is much smaller than N.
2246  *
2247  * In this calculation, we consider only non-nulls. We used to
2248  * include rows with null values in the n and N counts, but that
2249  * leads to inaccurate answers in columns with many nulls, and
2250  * it's intuitively bogus anyway considering the desired result is
2251  * the number of distinct non-null values.
2252  *
2253  * We assume (not very reliably!) that all the multiply-occurring
2254  * values are reflected in the final track[] list, and the other
2255  * nonnull values all appeared but once. (XXX this usually
2256  * results in a drastic overestimate of ndistinct. Can we do
2257  * any better?)
2258  *----------
2259  */
2260  int f1 = nonnull_cnt - summultiple;
2261  int d = f1 + nmultiple;
2262  double n = samplerows - null_cnt;
2263  double N = totalrows * (1.0 - stats->stanullfrac);
2264  double stadistinct;
2265 
2266  /* N == 0 shouldn't happen, but just in case ... */
2267  if (N > 0)
2268  stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2269  else
2270  stadistinct = 0;
2271 
2272  /* Clamp to sane range in case of roundoff error */
2273  if (stadistinct < d)
2274  stadistinct = d;
2275  if (stadistinct > N)
2276  stadistinct = N;
2277  /* And round to integer */
2278  stats->stadistinct = floor(stadistinct + 0.5);
2279  }
2280 
2281  /*
2282  * If we estimated the number of distinct values at more than 10% of
2283  * the total row count (a very arbitrary limit), then assume that
2284  * stadistinct should scale with the row count rather than be a fixed
2285  * value.
2286  */
2287  if (stats->stadistinct > 0.1 * totalrows)
2288  stats->stadistinct = -(stats->stadistinct / totalrows);
2289 
2290  /*
2291  * Decide how many values are worth storing as most-common values. If
2292  * we are able to generate a complete MCV list (all the values in the
2293  * sample will fit, and we think these are all the ones in the table),
2294  * then do so. Otherwise, store only those values that are
2295  * significantly more common than the values not in the list.
2296  *
2297  * Note: the first of these cases is meant to address columns with
2298  * small, fixed sets of possible values, such as boolean or enum
2299  * columns. If we can *completely* represent the column population by
2300  * an MCV list that will fit into the stats target, then we should do
2301  * so and thus provide the planner with complete information. But if
2302  * the MCV list is not complete, it's generally worth being more
2303  * selective, and not just filling it all the way up to the stats
2304  * target.
2305  */
2306  if (track_cnt < track_max && toowide_cnt == 0 &&
2307  stats->stadistinct > 0 &&
2308  track_cnt <= num_mcv)
2309  {
2310  /* Track list includes all values seen, and all will fit */
2311  num_mcv = track_cnt;
2312  }
2313  else
2314  {
2315  int *mcv_counts;
2316 
2317  /* Incomplete list; decide how many values are worth keeping */
2318  if (num_mcv > track_cnt)
2319  num_mcv = track_cnt;
2320 
2321  if (num_mcv > 0)
2322  {
2323  mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2324  for (i = 0; i < num_mcv; i++)
2325  mcv_counts[i] = track[i].count;
2326 
2327  num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2328  stats->stadistinct,
2329  stats->stanullfrac,
2330  samplerows, totalrows);
2331  }
2332  }
2333 
2334  /* Generate MCV slot entry */
2335  if (num_mcv > 0)
2336  {
2337  MemoryContext old_context;
2338  Datum *mcv_values;
2339  float4 *mcv_freqs;
2340 
2341  /* Must copy the target values into anl_context */
2342  old_context = MemoryContextSwitchTo(stats->anl_context);
2343  mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2344  mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2345  for (i = 0; i < num_mcv; i++)
2346  {
2347  mcv_values[i] = datumCopy(track[i].value,
2348  stats->attrtype->typbyval,
2349  stats->attrtype->typlen);
2350  mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2351  }
2352  MemoryContextSwitchTo(old_context);
2353 
2354  stats->stakind[0] = STATISTIC_KIND_MCV;
2355  stats->staop[0] = mystats->eqopr;
2356  stats->stacoll[0] = stats->attrcollid;
2357  stats->stanumbers[0] = mcv_freqs;
2358  stats->numnumbers[0] = num_mcv;
2359  stats->stavalues[0] = mcv_values;
2360  stats->numvalues[0] = num_mcv;
2361 
2362  /*
2363  * Accept the defaults for stats->statypid and others. They have
2364  * been set before we were called (see vacuum.h)
2365  */
2366  }
2367  }
2368  else if (null_cnt > 0)
2369  {
2370  /* We found only nulls; assume the column is entirely null */
2371  stats->stats_valid = true;
2372  stats->stanullfrac = 1.0;
2373  if (is_varwidth)
2374  stats->stawidth = 0; /* "unknown" */
2375  else
2376  stats->stawidth = stats->attrtype->typlen;
2377  stats->stadistinct = 0.0; /* "unknown" */
2378  }
2379 
2380  /* We don't need to bother cleaning up any of our temporary palloc's */
2381 }
float float4
Definition: c.h:616
#define swapInt(a, b)
Definition: analyze.c:1841
#define swapDatum(a, b)
Definition: analyze.c:1842
#define WIDTH_THRESHOLD
Definition: analyze.c:1839
static int analyze_mcv_list(int *mcv_counts, int num_mcv, double stadistinct, double stanullfrac, int samplerows, double totalrows)
Definition: analyze.c:2975
Datum datumCopy(Datum value, bool typByVal, int typLen)
Definition: datum.c:132
Size toast_raw_datum_size(Datum value)
Definition: detoast.c:545
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 PG_DETOAST_DATUM(datum)
Definition: fmgr.h:240
static struct @154 value
if(TABLE==NULL||TABLE_index==NULL)
Definition: isn.c:77
static bool DatumGetBool(Datum X)
Definition: postgres.h:90
static Datum PointerGetDatum(const void *X)
Definition: postgres.h:322
static char * DatumGetCString(Datum X)
Definition: postgres.h:335
static Pointer DatumGetPointer(Datum X)
Definition: postgres.h:312
MemoryContextSwitchTo(old_ctx)
int f1[ARRAY_SIZE]
Definition: sql-declare.c:113
Definition: fmgr.h:57
bool stats_valid
Definition: vacuum.h:143
float4 stanullfrac
Definition: vacuum.h:144
Form_pg_type attrtype
Definition: vacuum.h:127
int16 stakind[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:147
MemoryContext anl_context
Definition: vacuum.h:129
Oid staop[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:148
Oid stacoll[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:149
float4 * stanumbers[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:151
int attstattarget
Definition: vacuum.h:124
int32 stawidth
Definition: vacuum.h:145
void * extra_data
Definition: vacuum.h:137
int numvalues[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:152
Datum * stavalues[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:153
float4 stadistinct
Definition: vacuum.h:146
int numnumbers[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:150
Oid attrcollid
Definition: vacuum.h:128
#define VARSIZE_ANY(PTR)
Definition: varatt.h:311

References analyze_mcv_list(), VacAttrStats::anl_context, VacAttrStats::attrcollid, VacAttrStats::attrtype, VacAttrStats::attstattarget, datumCopy(), DatumGetBool(), DatumGetCString(), DatumGetPointer(), StdAnalyzeData::eqfunc, StdAnalyzeData::eqopr, VacAttrStats::extra_data, f1, fmgr_info(), FunctionCall2Coll(), i, if(), j, MemoryContextSwitchTo(), VacAttrStats::numnumbers, VacAttrStats::numvalues, palloc(), PG_DETOAST_DATUM, PointerGetDatum(), VacAttrStats::stacoll, VacAttrStats::stadistinct, VacAttrStats::stakind, VacAttrStats::stanullfrac, VacAttrStats::stanumbers, VacAttrStats::staop, VacAttrStats::stats_valid, VacAttrStats::stavalues, VacAttrStats::stawidth, swapDatum, swapInt, toast_raw_datum_size(), vacuum_delay_point(), value, VARSIZE_ANY, and WIDTH_THRESHOLD.

Referenced by std_typanalyze().

◆ compute_index_stats()

static void compute_index_stats ( Relation  onerel,
double  totalrows,
AnlIndexData indexdata,
int  nindexes,
HeapTuple rows,
int  numrows,
MemoryContext  col_context 
)
static

Definition at line 823 of file analyze.c.

827 {
828  MemoryContext ind_context,
829  old_context;
831  bool isnull[INDEX_MAX_KEYS];
832  int ind,
833  i;
834 
835  ind_context = AllocSetContextCreate(anl_context,
836  "Analyze Index",
838  old_context = MemoryContextSwitchTo(ind_context);
839 
840  for (ind = 0; ind < nindexes; ind++)
841  {
842  AnlIndexData *thisdata = &indexdata[ind];
843  IndexInfo *indexInfo = thisdata->indexInfo;
844  int attr_cnt = thisdata->attr_cnt;
845  TupleTableSlot *slot;
846  EState *estate;
847  ExprContext *econtext;
848  ExprState *predicate;
849  Datum *exprvals;
850  bool *exprnulls;
851  int numindexrows,
852  tcnt,
853  rowno;
854  double totalindexrows;
855 
856  /* Ignore index if no columns to analyze and not partial */
857  if (attr_cnt == 0 && indexInfo->ii_Predicate == NIL)
858  continue;
859 
860  /*
861  * Need an EState for evaluation of index expressions and
862  * partial-index predicates. Create it in the per-index context to be
863  * sure it gets cleaned up at the bottom of the loop.
864  */
865  estate = CreateExecutorState();
866  econtext = GetPerTupleExprContext(estate);
867  /* Need a slot to hold the current heap tuple, too */
869  &TTSOpsHeapTuple);
870 
871  /* Arrange for econtext's scan tuple to be the tuple under test */
872  econtext->ecxt_scantuple = slot;
873 
874  /* Set up execution state for predicate. */
875  predicate = ExecPrepareQual(indexInfo->ii_Predicate, estate);
876 
877  /* Compute and save index expression values */
878  exprvals = (Datum *) palloc(numrows * attr_cnt * sizeof(Datum));
879  exprnulls = (bool *) palloc(numrows * attr_cnt * sizeof(bool));
880  numindexrows = 0;
881  tcnt = 0;
882  for (rowno = 0; rowno < numrows; rowno++)
883  {
884  HeapTuple heapTuple = rows[rowno];
885 
887 
888  /*
889  * Reset the per-tuple context each time, to reclaim any cruft
890  * left behind by evaluating the predicate or index expressions.
891  */
892  ResetExprContext(econtext);
893 
894  /* Set up for predicate or expression evaluation */
895  ExecStoreHeapTuple(heapTuple, slot, false);
896 
897  /* If index is partial, check predicate */
898  if (predicate != NULL)
899  {
900  if (!ExecQual(predicate, econtext))
901  continue;
902  }
903  numindexrows++;
904 
905  if (attr_cnt > 0)
906  {
907  /*
908  * Evaluate the index row to compute expression values. We
909  * could do this by hand, but FormIndexDatum is convenient.
910  */
911  FormIndexDatum(indexInfo,
912  slot,
913  estate,
914  values,
915  isnull);
916 
917  /*
918  * Save just the columns we care about. We copy the values
919  * into ind_context from the estate's per-tuple context.
920  */
921  for (i = 0; i < attr_cnt; i++)
922  {
923  VacAttrStats *stats = thisdata->vacattrstats[i];
924  int attnum = stats->tupattnum;
925 
926  if (isnull[attnum - 1])
927  {
928  exprvals[tcnt] = (Datum) 0;
929  exprnulls[tcnt] = true;
930  }
931  else
932  {
933  exprvals[tcnt] = datumCopy(values[attnum - 1],
934  stats->attrtype->typbyval,
935  stats->attrtype->typlen);
936  exprnulls[tcnt] = false;
937  }
938  tcnt++;
939  }
940  }
941  }
942 
943  /*
944  * Having counted the number of rows that pass the predicate in the
945  * sample, we can estimate the total number of rows in the index.
946  */
947  thisdata->tupleFract = (double) numindexrows / (double) numrows;
948  totalindexrows = ceil(thisdata->tupleFract * totalrows);
949 
950  /*
951  * Now we can compute the statistics for the expression columns.
952  */
953  if (numindexrows > 0)
954  {
955  MemoryContextSwitchTo(col_context);
956  for (i = 0; i < attr_cnt; i++)
957  {
958  VacAttrStats *stats = thisdata->vacattrstats[i];
959 
960  stats->exprvals = exprvals + i;
961  stats->exprnulls = exprnulls + i;
962  stats->rowstride = attr_cnt;
963  stats->compute_stats(stats,
965  numindexrows,
966  totalindexrows);
967 
968  MemoryContextReset(col_context);
969  }
970  }
971 
972  /* And clean up */
973  MemoryContextSwitchTo(ind_context);
974 
976  FreeExecutorState(estate);
977  MemoryContextReset(ind_context);
978  }
979 
980  MemoryContextSwitchTo(old_context);
981  MemoryContextDelete(ind_context);
982 }
static Datum values[MAXATTR]
Definition: bootstrap.c:152
static MemoryContext anl_context
Definition: analyze.c:76
static Datum ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
Definition: analyze.c:1809
ExprState * ExecPrepareQual(List *qual, EState *estate)
Definition: execExpr.c:767
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
void FormIndexDatum(IndexInfo *indexInfo, TupleTableSlot *slot, EState *estate, Datum *values, bool *isnull)
Definition: index.c:2705
void MemoryContextReset(MemoryContext context)
Definition: mcxt.c:371
void MemoryContextDelete(MemoryContext context)
Definition: mcxt.c:442
#define AllocSetContextCreate
Definition: memutils.h:129
#define ALLOCSET_DEFAULT_SIZES
Definition: memutils.h:153
int16 attnum
Definition: pg_attribute.h:74
#define INDEX_MAX_KEYS
#define NIL
Definition: pg_list.h:68
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
TupleTableSlot * ecxt_scantuple
Definition: execnodes.h:255
List * ii_Predicate
Definition: execnodes.h:189
int tupattnum
Definition: vacuum.h:170
int rowstride
Definition: vacuum.h:175
bool * exprnulls
Definition: vacuum.h:174
Datum * exprvals
Definition: vacuum.h:173
AnalyzeAttrComputeStatsFunc compute_stats
Definition: vacuum.h:135

References ALLOCSET_DEFAULT_SIZES, AllocSetContextCreate, anl_context, attnum, AnlIndexData::attr_cnt, VacAttrStats::attrtype, VacAttrStats::compute_stats, CreateExecutorState(), datumCopy(), ExprContext::ecxt_scantuple, ExecDropSingleTupleTableSlot(), ExecPrepareQual(), ExecQual(), ExecStoreHeapTuple(), VacAttrStats::exprnulls, VacAttrStats::exprvals, FormIndexDatum(), FreeExecutorState(), GetPerTupleExprContext, i, IndexInfo::ii_Predicate, ind_fetch_func(), INDEX_MAX_KEYS, AnlIndexData::indexInfo, MakeSingleTupleTableSlot(), MemoryContextDelete(), MemoryContextReset(), MemoryContextSwitchTo(), NIL, palloc(), RelationGetDescr, ResetExprContext, VacAttrStats::rowstride, TTSOpsHeapTuple, VacAttrStats::tupattnum, AnlIndexData::tupleFract, AnlIndexData::vacattrstats, vacuum_delay_point(), and values.

Referenced by do_analyze_rel().

◆ compute_scalar_stats()

static void compute_scalar_stats ( VacAttrStatsP  stats,
AnalyzeAttrFetchFunc  fetchfunc,
int  samplerows,
double  totalrows 
)
static

Definition at line 2397 of file analyze.c.

2401 {
2402  int i;
2403  int null_cnt = 0;
2404  int nonnull_cnt = 0;
2405  int toowide_cnt = 0;
2406  double total_width = 0;
2407  bool is_varlena = (!stats->attrtype->typbyval &&
2408  stats->attrtype->typlen == -1);
2409  bool is_varwidth = (!stats->attrtype->typbyval &&
2410  stats->attrtype->typlen < 0);
2411  double corr_xysum;
2412  SortSupportData ssup;
2413  ScalarItem *values;
2414  int values_cnt = 0;
2415  int *tupnoLink;
2416  ScalarMCVItem *track;
2417  int track_cnt = 0;
2418  int num_mcv = stats->attstattarget;
2419  int num_bins = stats->attstattarget;
2420  StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2421 
2422  values = (ScalarItem *) palloc(samplerows * sizeof(ScalarItem));
2423  tupnoLink = (int *) palloc(samplerows * sizeof(int));
2424  track = (ScalarMCVItem *) palloc(num_mcv * sizeof(ScalarMCVItem));
2425 
2426  memset(&ssup, 0, sizeof(ssup));
2428  ssup.ssup_collation = stats->attrcollid;
2429  ssup.ssup_nulls_first = false;
2430 
2431  /*
2432  * For now, don't perform abbreviated key conversion, because full values
2433  * are required for MCV slot generation. Supporting that optimization
2434  * would necessitate teaching compare_scalars() to call a tie-breaker.
2435  */
2436  ssup.abbreviate = false;
2437 
2438  PrepareSortSupportFromOrderingOp(mystats->ltopr, &ssup);
2439 
2440  /* Initial scan to find sortable values */
2441  for (i = 0; i < samplerows; i++)
2442  {
2443  Datum value;
2444  bool isnull;
2445 
2447 
2448  value = fetchfunc(stats, i, &isnull);
2449 
2450  /* Check for null/nonnull */
2451  if (isnull)
2452  {
2453  null_cnt++;
2454  continue;
2455  }
2456  nonnull_cnt++;
2457 
2458  /*
2459  * If it's a variable-width field, add up widths for average width
2460  * calculation. Note that if the value is toasted, we use the toasted
2461  * width. We don't bother with this calculation if it's a fixed-width
2462  * type.
2463  */
2464  if (is_varlena)
2465  {
2466  total_width += VARSIZE_ANY(DatumGetPointer(value));
2467 
2468  /*
2469  * If the value is toasted, we want to detoast it just once to
2470  * avoid repeated detoastings and resultant excess memory usage
2471  * during the comparisons. Also, check to see if the value is
2472  * excessively wide, and if so don't detoast at all --- just
2473  * ignore the value.
2474  */
2476  {
2477  toowide_cnt++;
2478  continue;
2479  }
2481  }
2482  else if (is_varwidth)
2483  {
2484  /* must be cstring */
2485  total_width += strlen(DatumGetCString(value)) + 1;
2486  }
2487 
2488  /* Add it to the list to be sorted */
2489  values[values_cnt].value = value;
2490  values[values_cnt].tupno = values_cnt;
2491  tupnoLink[values_cnt] = values_cnt;
2492  values_cnt++;
2493  }
2494 
2495  /* We can only compute real stats if we found some sortable values. */
2496  if (values_cnt > 0)
2497  {
2498  int ndistinct, /* # distinct values in sample */
2499  nmultiple, /* # that appear multiple times */
2500  num_hist,
2501  dups_cnt;
2502  int slot_idx = 0;
2504 
2505  /* Sort the collected values */
2506  cxt.ssup = &ssup;
2507  cxt.tupnoLink = tupnoLink;
2508  qsort_interruptible(values, values_cnt, sizeof(ScalarItem),
2509  compare_scalars, &cxt);
2510 
2511  /*
2512  * Now scan the values in order, find the most common ones, and also
2513  * accumulate ordering-correlation statistics.
2514  *
2515  * To determine which are most common, we first have to count the
2516  * number of duplicates of each value. The duplicates are adjacent in
2517  * the sorted list, so a brute-force approach is to compare successive
2518  * datum values until we find two that are not equal. However, that
2519  * requires N-1 invocations of the datum comparison routine, which are
2520  * completely redundant with work that was done during the sort. (The
2521  * sort algorithm must at some point have compared each pair of items
2522  * that are adjacent in the sorted order; otherwise it could not know
2523  * that it's ordered the pair correctly.) We exploit this by having
2524  * compare_scalars remember the highest tupno index that each
2525  * ScalarItem has been found equal to. At the end of the sort, a
2526  * ScalarItem's tupnoLink will still point to itself if and only if it
2527  * is the last item of its group of duplicates (since the group will
2528  * be ordered by tupno).
2529  */
2530  corr_xysum = 0;
2531  ndistinct = 0;
2532  nmultiple = 0;
2533  dups_cnt = 0;
2534  for (i = 0; i < values_cnt; i++)
2535  {
2536  int tupno = values[i].tupno;
2537 
2538  corr_xysum += ((double) i) * ((double) tupno);
2539  dups_cnt++;
2540  if (tupnoLink[tupno] == tupno)
2541  {
2542  /* Reached end of duplicates of this value */
2543  ndistinct++;
2544  if (dups_cnt > 1)
2545  {
2546  nmultiple++;
2547  if (track_cnt < num_mcv ||
2548  dups_cnt > track[track_cnt - 1].count)
2549  {
2550  /*
2551  * Found a new item for the mcv list; find its
2552  * position, bubbling down old items if needed. Loop
2553  * invariant is that j points at an empty/ replaceable
2554  * slot.
2555  */
2556  int j;
2557 
2558  if (track_cnt < num_mcv)
2559  track_cnt++;
2560  for (j = track_cnt - 1; j > 0; j--)
2561  {
2562  if (dups_cnt <= track[j - 1].count)
2563  break;
2564  track[j].count = track[j - 1].count;
2565  track[j].first = track[j - 1].first;
2566  }
2567  track[j].count = dups_cnt;
2568  track[j].first = i + 1 - dups_cnt;
2569  }
2570  }
2571  dups_cnt = 0;
2572  }
2573  }
2574 
2575  stats->stats_valid = true;
2576  /* Do the simple null-frac and width stats */
2577  stats->stanullfrac = (double) null_cnt / (double) samplerows;
2578  if (is_varwidth)
2579  stats->stawidth = total_width / (double) nonnull_cnt;
2580  else
2581  stats->stawidth = stats->attrtype->typlen;
2582 
2583  if (nmultiple == 0)
2584  {
2585  /*
2586  * If we found no repeated non-null values, assume it's a unique
2587  * column; but be sure to discount for any nulls we found.
2588  */
2589  stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2590  }
2591  else if (toowide_cnt == 0 && nmultiple == ndistinct)
2592  {
2593  /*
2594  * Every value in the sample appeared more than once. Assume the
2595  * column has just these values. (This case is meant to address
2596  * columns with small, fixed sets of possible values, such as
2597  * boolean or enum columns. If there are any values that appear
2598  * just once in the sample, including too-wide values, we should
2599  * assume that that's not what we're dealing with.)
2600  */
2601  stats->stadistinct = ndistinct;
2602  }
2603  else
2604  {
2605  /*----------
2606  * Estimate the number of distinct values using the estimator
2607  * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2608  * n*d / (n - f1 + f1*n/N)
2609  * where f1 is the number of distinct values that occurred
2610  * exactly once in our sample of n rows (from a total of N),
2611  * and d is the total number of distinct values in the sample.
2612  * This is their Duj1 estimator; the other estimators they
2613  * recommend are considerably more complex, and are numerically
2614  * very unstable when n is much smaller than N.
2615  *
2616  * In this calculation, we consider only non-nulls. We used to
2617  * include rows with null values in the n and N counts, but that
2618  * leads to inaccurate answers in columns with many nulls, and
2619  * it's intuitively bogus anyway considering the desired result is
2620  * the number of distinct non-null values.
2621  *
2622  * Overwidth values are assumed to have been distinct.
2623  *----------
2624  */
2625  int f1 = ndistinct - nmultiple + toowide_cnt;
2626  int d = f1 + nmultiple;
2627  double n = samplerows - null_cnt;
2628  double N = totalrows * (1.0 - stats->stanullfrac);
2629  double stadistinct;
2630 
2631  /* N == 0 shouldn't happen, but just in case ... */
2632  if (N > 0)
2633  stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2634  else
2635  stadistinct = 0;
2636 
2637  /* Clamp to sane range in case of roundoff error */
2638  if (stadistinct < d)
2639  stadistinct = d;
2640  if (stadistinct > N)
2641  stadistinct = N;
2642  /* And round to integer */
2643  stats->stadistinct = floor(stadistinct + 0.5);
2644  }
2645 
2646  /*
2647  * If we estimated the number of distinct values at more than 10% of
2648  * the total row count (a very arbitrary limit), then assume that
2649  * stadistinct should scale with the row count rather than be a fixed
2650  * value.
2651  */
2652  if (stats->stadistinct > 0.1 * totalrows)
2653  stats->stadistinct = -(stats->stadistinct / totalrows);
2654 
2655  /*
2656  * Decide how many values are worth storing as most-common values. If
2657  * we are able to generate a complete MCV list (all the values in the
2658  * sample will fit, and we think these are all the ones in the table),
2659  * then do so. Otherwise, store only those values that are
2660  * significantly more common than the values not in the list.
2661  *
2662  * Note: the first of these cases is meant to address columns with
2663  * small, fixed sets of possible values, such as boolean or enum
2664  * columns. If we can *completely* represent the column population by
2665  * an MCV list that will fit into the stats target, then we should do
2666  * so and thus provide the planner with complete information. But if
2667  * the MCV list is not complete, it's generally worth being more
2668  * selective, and not just filling it all the way up to the stats
2669  * target.
2670  */
2671  if (track_cnt == ndistinct && toowide_cnt == 0 &&
2672  stats->stadistinct > 0 &&
2673  track_cnt <= num_mcv)
2674  {
2675  /* Track list includes all values seen, and all will fit */
2676  num_mcv = track_cnt;
2677  }
2678  else
2679  {
2680  int *mcv_counts;
2681 
2682  /* Incomplete list; decide how many values are worth keeping */
2683  if (num_mcv > track_cnt)
2684  num_mcv = track_cnt;
2685 
2686  if (num_mcv > 0)
2687  {
2688  mcv_counts = (int *) palloc(num_mcv * sizeof(int));
2689  for (i = 0; i < num_mcv; i++)
2690  mcv_counts[i] = track[i].count;
2691 
2692  num_mcv = analyze_mcv_list(mcv_counts, num_mcv,
2693  stats->stadistinct,
2694  stats->stanullfrac,
2695  samplerows, totalrows);
2696  }
2697  }
2698 
2699  /* Generate MCV slot entry */
2700  if (num_mcv > 0)
2701  {
2702  MemoryContext old_context;
2703  Datum *mcv_values;
2704  float4 *mcv_freqs;
2705 
2706  /* Must copy the target values into anl_context */
2707  old_context = MemoryContextSwitchTo(stats->anl_context);
2708  mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2709  mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2710  for (i = 0; i < num_mcv; i++)
2711  {
2712  mcv_values[i] = datumCopy(values[track[i].first].value,
2713  stats->attrtype->typbyval,
2714  stats->attrtype->typlen);
2715  mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2716  }
2717  MemoryContextSwitchTo(old_context);
2718 
2719  stats->stakind[slot_idx] = STATISTIC_KIND_MCV;
2720  stats->staop[slot_idx] = mystats->eqopr;
2721  stats->stacoll[slot_idx] = stats->attrcollid;
2722  stats->stanumbers[slot_idx] = mcv_freqs;
2723  stats->numnumbers[slot_idx] = num_mcv;
2724  stats->stavalues[slot_idx] = mcv_values;
2725  stats->numvalues[slot_idx] = num_mcv;
2726 
2727  /*
2728  * Accept the defaults for stats->statypid and others. They have
2729  * been set before we were called (see vacuum.h)
2730  */
2731  slot_idx++;
2732  }
2733 
2734  /*
2735  * Generate a histogram slot entry if there are at least two distinct
2736  * values not accounted for in the MCV list. (This ensures the
2737  * histogram won't collapse to empty or a singleton.)
2738  */
2739  num_hist = ndistinct - num_mcv;
2740  if (num_hist > num_bins)
2741  num_hist = num_bins + 1;
2742  if (num_hist >= 2)
2743  {
2744  MemoryContext old_context;
2745  Datum *hist_values;
2746  int nvals;
2747  int pos,
2748  posfrac,
2749  delta,
2750  deltafrac;
2751 
2752  /* Sort the MCV items into position order to speed next loop */
2753  qsort_interruptible(track, num_mcv, sizeof(ScalarMCVItem),
2754  compare_mcvs, NULL);
2755 
2756  /*
2757  * Collapse out the MCV items from the values[] array.
2758  *
2759  * Note we destroy the values[] array here... but we don't need it
2760  * for anything more. We do, however, still need values_cnt.
2761  * nvals will be the number of remaining entries in values[].
2762  */
2763  if (num_mcv > 0)
2764  {
2765  int src,
2766  dest;
2767  int j;
2768 
2769  src = dest = 0;
2770  j = 0; /* index of next interesting MCV item */
2771  while (src < values_cnt)
2772  {
2773  int ncopy;
2774 
2775  if (j < num_mcv)
2776  {
2777  int first = track[j].first;
2778 
2779  if (src >= first)
2780  {
2781  /* advance past this MCV item */
2782  src = first + track[j].count;
2783  j++;
2784  continue;
2785  }
2786  ncopy = first - src;
2787  }
2788  else
2789  ncopy = values_cnt - src;
2790  memmove(&values[dest], &values[src],
2791  ncopy * sizeof(ScalarItem));
2792  src += ncopy;
2793  dest += ncopy;
2794  }
2795  nvals = dest;
2796  }
2797  else
2798  nvals = values_cnt;
2799  Assert(nvals >= num_hist);
2800 
2801  /* Must copy the target values into anl_context */
2802  old_context = MemoryContextSwitchTo(stats->anl_context);
2803  hist_values = (Datum *) palloc(num_hist * sizeof(Datum));
2804 
2805  /*
2806  * The object of this loop is to copy the first and last values[]
2807  * entries along with evenly-spaced values in between. So the
2808  * i'th value is values[(i * (nvals - 1)) / (num_hist - 1)]. But
2809  * computing that subscript directly risks integer overflow when
2810  * the stats target is more than a couple thousand. Instead we
2811  * add (nvals - 1) / (num_hist - 1) to pos at each step, tracking
2812  * the integral and fractional parts of the sum separately.
2813  */
2814  delta = (nvals - 1) / (num_hist - 1);
2815  deltafrac = (nvals - 1) % (num_hist - 1);
2816  pos = posfrac = 0;
2817 
2818  for (i = 0; i < num_hist; i++)
2819  {
2820  hist_values[i] = datumCopy(values[pos].value,
2821  stats->attrtype->typbyval,
2822  stats->attrtype->typlen);
2823  pos += delta;
2824  posfrac += deltafrac;
2825  if (posfrac >= (num_hist - 1))
2826  {
2827  /* fractional part exceeds 1, carry to integer part */
2828  pos++;
2829  posfrac -= (num_hist - 1);
2830  }
2831  }
2832 
2833  MemoryContextSwitchTo(old_context);
2834 
2835  stats->stakind[slot_idx] = STATISTIC_KIND_HISTOGRAM;
2836  stats->staop[slot_idx] = mystats->ltopr;
2837  stats->stacoll[slot_idx] = stats->attrcollid;
2838  stats->stavalues[slot_idx] = hist_values;
2839  stats->numvalues[slot_idx] = num_hist;
2840 
2841  /*
2842  * Accept the defaults for stats->statypid and others. They have
2843  * been set before we were called (see vacuum.h)
2844  */
2845  slot_idx++;
2846  }
2847 
2848  /* Generate a correlation entry if there are multiple values */
2849  if (values_cnt > 1)
2850  {
2851  MemoryContext old_context;
2852  float4 *corrs;
2853  double corr_xsum,
2854  corr_x2sum;
2855 
2856  /* Must copy the target values into anl_context */
2857  old_context = MemoryContextSwitchTo(stats->anl_context);
2858  corrs = (float4 *) palloc(sizeof(float4));
2859  MemoryContextSwitchTo(old_context);
2860 
2861  /*----------
2862  * Since we know the x and y value sets are both
2863  * 0, 1, ..., values_cnt-1
2864  * we have sum(x) = sum(y) =
2865  * (values_cnt-1)*values_cnt / 2
2866  * and sum(x^2) = sum(y^2) =
2867  * (values_cnt-1)*values_cnt*(2*values_cnt-1) / 6.
2868  *----------
2869  */
2870  corr_xsum = ((double) (values_cnt - 1)) *
2871  ((double) values_cnt) / 2.0;
2872  corr_x2sum = ((double) (values_cnt - 1)) *
2873  ((double) values_cnt) * (double) (2 * values_cnt - 1) / 6.0;
2874 
2875  /* And the correlation coefficient reduces to */
2876  corrs[0] = (values_cnt * corr_xysum - corr_xsum * corr_xsum) /
2877  (values_cnt * corr_x2sum - corr_xsum * corr_xsum);
2878 
2879  stats->stakind[slot_idx] = STATISTIC_KIND_CORRELATION;
2880  stats->staop[slot_idx] = mystats->ltopr;
2881  stats->stacoll[slot_idx] = stats->attrcollid;
2882  stats->stanumbers[slot_idx] = corrs;
2883  stats->numnumbers[slot_idx] = 1;
2884  slot_idx++;
2885  }
2886  }
2887  else if (nonnull_cnt > 0)
2888  {
2889  /* We found some non-null values, but they were all too wide */
2890  Assert(nonnull_cnt == toowide_cnt);
2891  stats->stats_valid = true;
2892  /* Do the simple null-frac and width stats */
2893  stats->stanullfrac = (double) null_cnt / (double) samplerows;
2894  if (is_varwidth)
2895  stats->stawidth = total_width / (double) nonnull_cnt;
2896  else
2897  stats->stawidth = stats->attrtype->typlen;
2898  /* Assume all too-wide values are distinct, so it's a unique column */
2899  stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
2900  }
2901  else if (null_cnt > 0)
2902  {
2903  /* We found only nulls; assume the column is entirely null */
2904  stats->stats_valid = true;
2905  stats->stanullfrac = 1.0;
2906  if (is_varwidth)
2907  stats->stawidth = 0; /* "unknown" */
2908  else
2909  stats->stawidth = stats->attrtype->typlen;
2910  stats->stadistinct = 0.0; /* "unknown" */
2911  }
2912 
2913  /* We don't need to bother cleaning up any of our temporary palloc's */
2914 }
static int compare_mcvs(const void *a, const void *b, void *arg)
Definition: analyze.c:2957
static int compare_scalars(const void *a, const void *b, void *arg)
Definition: analyze.c:2926
MemoryContext CurrentMemoryContext
Definition: mcxt.c:131
void PrepareSortSupportFromOrderingOp(Oid orderingOp, SortSupport ssup)
Definition: sortsupport.c:134
bool ssup_nulls_first
Definition: sortsupport.h:75
MemoryContext ssup_cxt
Definition: sortsupport.h:66

References SortSupportData::abbreviate, analyze_mcv_list(), VacAttrStats::anl_context, Assert(), VacAttrStats::attrcollid, VacAttrStats::attrtype, VacAttrStats::attstattarget, compare_mcvs(), compare_scalars(), ScalarMCVItem::count, CurrentMemoryContext, datumCopy(), DatumGetCString(), DatumGetPointer(), generate_unaccent_rules::dest, StdAnalyzeData::eqopr, VacAttrStats::extra_data, f1, ScalarMCVItem::first, i, j, StdAnalyzeData::ltopr, MemoryContextSwitchTo(), VacAttrStats::numnumbers, VacAttrStats::numvalues, palloc(), PG_DETOAST_DATUM, PointerGetDatum(), PrepareSortSupportFromOrderingOp(), qsort_interruptible(), CompareScalarsContext::ssup, SortSupportData::ssup_collation, SortSupportData::ssup_cxt, SortSupportData::ssup_nulls_first, VacAttrStats::stacoll, VacAttrStats::stadistinct, VacAttrStats::stakind, VacAttrStats::stanullfrac, VacAttrStats::stanumbers, VacAttrStats::staop, VacAttrStats::stats_valid, VacAttrStats::stavalues, VacAttrStats::stawidth, toast_raw_datum_size(), CompareScalarsContext::tupnoLink, vacuum_delay_point(), value, values, VARSIZE_ANY, and WIDTH_THRESHOLD.

Referenced by std_typanalyze().

◆ compute_trivial_stats()

static void compute_trivial_stats ( VacAttrStatsP  stats,
AnalyzeAttrFetchFunc  fetchfunc,
int  samplerows,
double  totalrows 
)
static

Definition at line 1964 of file analyze.c.

1968 {
1969  int i;
1970  int null_cnt = 0;
1971  int nonnull_cnt = 0;
1972  double total_width = 0;
1973  bool is_varlena = (!stats->attrtype->typbyval &&
1974  stats->attrtype->typlen == -1);
1975  bool is_varwidth = (!stats->attrtype->typbyval &&
1976  stats->attrtype->typlen < 0);
1977 
1978  for (i = 0; i < samplerows; i++)
1979  {
1980  Datum value;
1981  bool isnull;
1982 
1984 
1985  value = fetchfunc(stats, i, &isnull);
1986 
1987  /* Check for null/nonnull */
1988  if (isnull)
1989  {
1990  null_cnt++;
1991  continue;
1992  }
1993  nonnull_cnt++;
1994 
1995  /*
1996  * If it's a variable-width field, add up widths for average width
1997  * calculation. Note that if the value is toasted, we use the toasted
1998  * width. We don't bother with this calculation if it's a fixed-width
1999  * type.
2000  */
2001  if (is_varlena)
2002  {
2003  total_width += VARSIZE_ANY(DatumGetPointer(value));
2004  }
2005  else if (is_varwidth)
2006  {
2007  /* must be cstring */
2008  total_width += strlen(DatumGetCString(value)) + 1;
2009  }
2010  }
2011 
2012  /* We can only compute average width if we found some non-null values. */
2013  if (nonnull_cnt > 0)
2014  {
2015  stats->stats_valid = true;
2016  /* Do the simple null-frac and width stats */
2017  stats->stanullfrac = (double) null_cnt / (double) samplerows;
2018  if (is_varwidth)
2019  stats->stawidth = total_width / (double) nonnull_cnt;
2020  else
2021  stats->stawidth = stats->attrtype->typlen;
2022  stats->stadistinct = 0.0; /* "unknown" */
2023  }
2024  else if (null_cnt > 0)
2025  {
2026  /* We found only nulls; assume the column is entirely null */
2027  stats->stats_valid = true;
2028  stats->stanullfrac = 1.0;
2029  if (is_varwidth)
2030  stats->stawidth = 0; /* "unknown" */
2031  else
2032  stats->stawidth = stats->attrtype->typlen;
2033  stats->stadistinct = 0.0; /* "unknown" */
2034  }
2035 }

References VacAttrStats::attrtype, DatumGetCString(), DatumGetPointer(), i, VacAttrStats::stadistinct, VacAttrStats::stanullfrac, VacAttrStats::stats_valid, VacAttrStats::stawidth, vacuum_delay_point(), value, and VARSIZE_ANY.

Referenced by std_typanalyze().

◆ do_analyze_rel()

static void do_analyze_rel ( Relation  onerel,
VacuumParams params,
List va_cols,
AcquireSampleRowsFunc  acquirefunc,
BlockNumber  relpages,
bool  inh,
bool  in_outer_xact,
int  elevel 
)
static

Definition at line 280 of file analyze.c.

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  /* Update pg_class for table relation */
633  vac_update_relstats(onerel,
634  relpages,
635  totalrows,
636  relallvisible,
637  hasindex,
640  NULL, NULL,
641  in_outer_xact);
642 
643  /* Same for indexes */
644  for (ind = 0; ind < nindexes; ind++)
645  {
646  AnlIndexData *thisdata = &indexdata[ind];
647  double totalindexrows;
648 
649  totalindexrows = ceil(thisdata->tupleFract * totalrows);
650  vac_update_relstats(Irel[ind],
652  totalindexrows,
653  0,
654  false,
657  NULL, NULL,
658  in_outer_xact);
659  }
660  }
661  else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
662  {
663  /*
664  * Partitioned tables don't have storage, so we don't set any fields
665  * in their pg_class entries except for reltuples and relhasindex.
666  */
667  vac_update_relstats(onerel, -1, totalrows,
668  0, hasindex, InvalidTransactionId,
670  NULL, NULL,
671  in_outer_xact);
672  }
673 
674  /*
675  * Now report ANALYZE to the cumulative stats system. For regular tables,
676  * we do it only if not doing inherited stats. For partitioned tables, we
677  * only do it for inherited stats. (We're never called for not-inherited
678  * stats on partitioned tables anyway.)
679  *
680  * Reset the changes_since_analyze counter only if we analyzed all
681  * columns; otherwise, there is still work for auto-analyze to do.
682  */
683  if (!inh)
684  pgstat_report_analyze(onerel, totalrows, totaldeadrows,
685  (va_cols == NIL));
686  else if (onerel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
687  pgstat_report_analyze(onerel, 0, 0, (va_cols == NIL));
688 
689  /*
690  * If this isn't part of VACUUM ANALYZE, let index AMs do cleanup.
691  *
692  * Note that most index AMs perform a no-op as a matter of policy for
693  * amvacuumcleanup() when called in ANALYZE-only mode. The only exception
694  * among core index AMs is GIN/ginvacuumcleanup().
695  */
696  if (!(params->options & VACOPT_VACUUM))
697  {
698  for (ind = 0; ind < nindexes; ind++)
699  {
700  IndexBulkDeleteResult *stats;
701  IndexVacuumInfo ivinfo;
702 
703  ivinfo.index = Irel[ind];
704  ivinfo.heaprel = onerel;
705  ivinfo.analyze_only = true;
706  ivinfo.estimated_count = true;
707  ivinfo.message_level = elevel;
708  ivinfo.num_heap_tuples = onerel->rd_rel->reltuples;
709  ivinfo.strategy = vac_strategy;
710 
711  stats = index_vacuum_cleanup(&ivinfo, NULL);
712 
713  if (stats)
714  pfree(stats);
715  }
716  }
717 
718  /* Done with indexes */
719  vac_close_indexes(nindexes, Irel, NoLock);
720 
721  /* Log the action if appropriate */
722  if (AmAutoVacuumWorkerProcess() && params->log_min_duration >= 0)
723  {
724  TimestampTz endtime = GetCurrentTimestamp();
725 
726  if (params->log_min_duration == 0 ||
727  TimestampDifferenceExceeds(starttime, endtime,
728  params->log_min_duration))
729  {
730  long delay_in_ms;
731  double read_rate = 0;
732  double write_rate = 0;
734 
735  /*
736  * Calculate the difference in the Page Hit/Miss/Dirty that
737  * happened as part of the analyze by subtracting out the
738  * pre-analyze values which we saved above.
739  */
740  AnalyzePageHit = VacuumPageHit - AnalyzePageHit;
741  AnalyzePageMiss = VacuumPageMiss - AnalyzePageMiss;
742  AnalyzePageDirty = VacuumPageDirty - AnalyzePageDirty;
743 
744  /*
745  * We do not expect an analyze to take > 25 days and it simplifies
746  * things a bit to use TimestampDifferenceMilliseconds.
747  */
748  delay_in_ms = TimestampDifferenceMilliseconds(starttime, endtime);
749 
750  /*
751  * Note that we are reporting these read/write rates in the same
752  * manner as VACUUM does, which means that while the 'average read
753  * rate' here actually corresponds to page misses and resulting
754  * reads which are also picked up by track_io_timing, if enabled,
755  * the 'average write rate' is actually talking about the rate of
756  * pages being dirtied, not being written out, so it's typical to
757  * have a non-zero 'avg write rate' while I/O timings only reports
758  * reads.
759  *
760  * It's not clear that an ANALYZE will ever result in
761  * FlushBuffer() being called, but we track and support reporting
762  * on I/O write time in case that changes as it's practically free
763  * to do so anyway.
764  */
765 
766  if (delay_in_ms > 0)
767  {
768  read_rate = (double) BLCKSZ * AnalyzePageMiss / (1024 * 1024) /
769  (delay_in_ms / 1000.0);
770  write_rate = (double) BLCKSZ * AnalyzePageDirty / (1024 * 1024) /
771  (delay_in_ms / 1000.0);
772  }
773 
774  /*
775  * We split this up so we don't emit empty I/O timing values when
776  * track_io_timing isn't enabled.
777  */
778 
780  appendStringInfo(&buf, _("automatic analyze of table \"%s.%s.%s\"\n"),
783  RelationGetRelationName(onerel));
784  if (track_io_timing)
785  {
786  double read_ms = (double) (pgStatBlockReadTime - startreadtime) / 1000;
787  double write_ms = (double) (pgStatBlockWriteTime - startwritetime) / 1000;
788 
789  appendStringInfo(&buf, _("I/O timings: read: %.3f ms, write: %.3f ms\n"),
790  read_ms, write_ms);
791  }
792  appendStringInfo(&buf, _("avg read rate: %.3f MB/s, avg write rate: %.3f MB/s\n"),
793  read_rate, write_rate);
794  appendStringInfo(&buf, _("buffer usage: %lld hits, %lld misses, %lld dirtied\n"),
795  (long long) AnalyzePageHit,
796  (long long) AnalyzePageMiss,
797  (long long) AnalyzePageDirty);
798  appendStringInfo(&buf, _("system usage: %s"), pg_rusage_show(&ru0));
799 
800  ereport(LOG,
801  (errmsg_internal("%s", buf.data)));
802 
803  pfree(buf.data);
804  }
805  }
806 
807  /* Roll back any GUC changes executed by index functions */
808  AtEOXact_GUC(false, save_nestlevel);
809 
810  /* Restore userid and security context */
811  SetUserIdAndSecContext(save_userid, save_sec_context);
812 
813  /* Restore current context and release memory */
814  MemoryContextSwitchTo(caller_context);
816  anl_context = NULL;
817 }
#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:1766
bool TimestampDifferenceExceeds(TimestampTz start_time, TimestampTz stop_time, int msec)
Definition: timestamp.c:1790
TimestampTz GetCurrentTimestamp(void)
Definition: timestamp.c:1654
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
bool track_io_timing
Definition: bufmgr.c:138
double float8
Definition: c.h:617
static Datum std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
Definition: analyze.c:1793
static void update_attstats(Oid relid, bool inh, int natts, VacAttrStats **vacattrstats)
Definition: analyze.c:1650
static int acquire_inherited_sample_rows(Relation onerel, int elevel, HeapTuple *rows, int targrows, double *totalrows, double *totaldeadrows)
Definition: analyze.c:1386
static VacAttrStats * examine_attribute(Relation onerel, int attnum, Node *index_expr)
Definition: analyze.c:994
static void compute_index_stats(Relation onerel, double totalrows, AnlIndexData *indexdata, int nindexes, HeapTuple *rows, int numrows, MemoryContext col_context)
Definition: analyze.c:823
int64 TimestampTz
Definition: timestamp.h:39
char * get_database_name(Oid dbid)
Definition: dbcommands.c:3153
int errmsg_internal(const char *fmt,...)
Definition: elog.c:1159
int errcode(int sqlerrcode)
Definition: elog.c:859
#define _(x)
Definition: elog.c:90
#define LOG
Definition: elog.h:31
#define ERROR
Definition: elog.h:39
#define elog(elevel,...)
Definition: elog.h:224
int ComputeExtStatisticsRows(Relation onerel, int natts, VacAttrStats **vacattrstats)
void BuildRelationExtStatistics(Relation onerel, bool inh, double totalrows, int numrows, HeapTuple *rows, int natts, VacAttrStats **vacattrstats)
int64 VacuumPageHit
Definition: globals.c:154
int64 VacuumPageMiss
Definition: globals.c:155
int64 VacuumPageDirty
Definition: globals.c:156
Oid MyDatabaseId
Definition: globals.c:91
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
IndexInfo * BuildIndexInfo(Relation index)
Definition: index.c:2407
IndexBulkDeleteResult * index_vacuum_cleanup(IndexVacuumInfo *info, IndexBulkDeleteResult *istat)
Definition: indexam.c:772
void list_free(List *list)
Definition: list.c:1546
void pfree(void *pointer)
Definition: mcxt.c:1508
void * palloc0(Size size)
Definition: mcxt.c:1334
#define AmAutoVacuumWorkerProcess()
Definition: miscadmin.h:375
#define SECURITY_RESTRICTED_OPERATION
Definition: miscadmin.h:315
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
int attnameAttNum(Relation rd, const char *attname, bool sysColOK)
#define lfirst(lc)
Definition: pg_list.h:172
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
const char * pg_rusage_show(const PGRUsage *ru0)
Definition: pg_rusage.c:40
void pg_rusage_init(PGRUsage *ru0)
Definition: pg_rusage.c:27
static char * buf
Definition: pg_test_fsync.c:73
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)
#define PROGRESS_ANALYZE_PHASE_FINALIZE_ANALYZE
Definition: progress.h:54
#define PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS_INH
Definition: progress.h:51
#define PROGRESS_ANALYZE_PHASE
Definition: progress.h:40
#define PROGRESS_ANALYZE_PHASE_COMPUTE_STATS
Definition: progress.h:52
#define PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS
Definition: progress.h:50
List * RelationGetIndexList(Relation relation)
Definition: relcache.c:4751
void appendStringInfo(StringInfo str, const char *fmt,...)
Definition: stringinfo.c:97
void initStringInfo(StringInfo str)
Definition: stringinfo.c:59
float8 n_distinct
Definition: attoptcache.h:22
float8 n_distinct_inherited
Definition: attoptcache.h:23
int ii_NumIndexAttrs
Definition: execnodes.h:184
List * ii_Expressions
Definition: execnodes.h:187
AttrNumber ii_IndexAttrNumbers[INDEX_MAX_KEYS]
Definition: execnodes.h:186
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: nodes.h:129
TupleDesc rd_att
Definition: rel.h:112
HeapTuple * rows
Definition: vacuum.h:171
int minrows
Definition: vacuum.h:136
TupleDesc tupDesc
Definition: vacuum.h:172
#define InvalidTransactionId
Definition: transam.h:31
void vac_open_indexes(Relation relation, LOCKMODE lockmode, int *nindexes, Relation **Irel)
Definition: vacuum.c:2273
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:1399
void vac_close_indexes(int nindexes, Relation *Irel, LOCKMODE lockmode)
Definition: vacuum.c:2316
#define strVal(v)
Definition: value.h:82
void visibilitymap_count(Relation rel, BlockNumber *all_visible, BlockNumber *all_frozen)

References _, AccessShareLock, acquire_inherited_sample_rows(), ALLOCSET_DEFAULT_SIZES, AllocSetContextCreate, AmAutoVacuumWorkerProcess, IndexVacuumInfo::analyze_only, anl_context, appendStringInfo(), AtEOXact_GUC(), attnameAttNum(), AnlIndexData::attr_cnt, bms_add_member(), bms_is_member(), buf, BuildIndexInfo(), BuildRelationExtStatistics(), compute_index_stats(), VacAttrStats::compute_stats, ComputeExtStatisticsRows(), CurrentMemoryContext, elog, ereport, errcode(), errmsg(), errmsg_internal(), ERROR, IndexVacuumInfo::estimated_count, examine_attribute(), get_attribute_options(), get_database_name(), get_namespace_name(), GetCurrentTimestamp(), GetUserIdAndSecContext(), IndexVacuumInfo::heaprel, i, IndexInfo::ii_Expressions, IndexInfo::ii_IndexAttrNumbers, IndexInfo::ii_NumIndexAttrs, IndexVacuumInfo::index, index_vacuum_cleanup(), AnlIndexData::indexInfo, initStringInfo(), InvalidAttrNumber, InvalidMultiXactId, InvalidTransactionId, lfirst, list_free(), list_head(), list_length(), lnext(), LOG, VacuumParams::log_min_duration, MemoryContextDelete(), MemoryContextReset(), MemoryContextSwitchTo(), IndexVacuumInfo::message_level, VacAttrStats::minrows, MyDatabaseId, AttributeOpts::n_distinct, AttributeOpts::n_distinct_inherited, TupleDescData::natts, NewGUCNestLevel(), NIL, NoLock, IndexVacuumInfo::num_heap_tuples, VacuumParams::options, palloc(), palloc0(), pfree(), pg_rusage_init(), pg_rusage_show(), pgstat_progress_update_param(), pgstat_report_analyze(), pgStatBlockReadTime, pgStatBlockWriteTime, PROGRESS_ANALYZE_PHASE, PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS, PROGRESS_ANALYZE_PHASE_ACQUIRE_SAMPLE_ROWS_INH, PROGRESS_ANALYZE_PHASE_COMPUTE_STATS, PROGRESS_ANALYZE_PHASE_FINALIZE_ANALYZE, RelationData::rd_att, RelationData::rd_id, RelationData::rd_rel, RelationGetIndexList(), RelationGetNamespace, RelationGetNumberOfBlocks, RelationGetRelationName, RelationGetRelid, RestrictSearchPath(), VacAttrStats::rows, SECURITY_RESTRICTED_OPERATION, SetUserIdAndSecContext(), VacAttrStats::stadistinct, std_fetch_func(), IndexVacuumInfo::strategy, strVal, TimestampDifferenceExceeds(), TimestampDifferenceMilliseconds(), track_io_timing, VacAttrStats::tupattnum, VacAttrStats::tupDesc, AnlIndexData::tupleFract, update_attstats(), vac_close_indexes(), vac_open_indexes(), vac_strategy, vac_update_relstats(), AnlIndexData::vacattrstats, VACOPT_VACUUM, VacuumPageDirty, VacuumPageHit, VacuumPageMiss, and visibilitymap_count().

Referenced by analyze_rel().

◆ examine_attribute()

static VacAttrStats * examine_attribute ( Relation  onerel,
int  attnum,
Node index_expr 
)
static

Definition at line 994 of file analyze.c.

995 {
996  Form_pg_attribute attr = TupleDescAttr(onerel->rd_att, attnum - 1);
997  int attstattarget;
998  HeapTuple atttuple;
999  Datum dat;
1000  bool isnull;
1001  HeapTuple typtuple;
1002  VacAttrStats *stats;
1003  int i;
1004  bool ok;
1005 
1006  /* Never analyze dropped columns */
1007  if (attr->attisdropped)
1008  return NULL;
1009 
1010  /*
1011  * Get attstattarget value. Set to -1 if null. (Analyze functions expect
1012  * -1 to mean use default_statistics_target; see for example
1013  * std_typanalyze.)
1014  */
1015  atttuple = SearchSysCache2(ATTNUM, ObjectIdGetDatum(RelationGetRelid(onerel)), Int16GetDatum(attnum));
1016  if (!HeapTupleIsValid(atttuple))
1017  elog(ERROR, "cache lookup failed for attribute %d of relation %u",
1018  attnum, RelationGetRelid(onerel));
1019  dat = SysCacheGetAttr(ATTNUM, atttuple, Anum_pg_attribute_attstattarget, &isnull);
1020  attstattarget = isnull ? -1 : DatumGetInt16(dat);
1021  ReleaseSysCache(atttuple);
1022 
1023  /* Don't analyze column if user has specified not to */
1024  if (attstattarget == 0)
1025  return NULL;
1026 
1027  /*
1028  * Create the VacAttrStats struct.
1029  */
1030  stats = (VacAttrStats *) palloc0(sizeof(VacAttrStats));
1031  stats->attstattarget = attstattarget;
1032 
1033  /*
1034  * When analyzing an expression index, believe the expression tree's type
1035  * not the column datatype --- the latter might be the opckeytype storage
1036  * type of the opclass, which is not interesting for our purposes. (Note:
1037  * if we did anything with non-expression index columns, we'd need to
1038  * figure out where to get the correct type info from, but for now that's
1039  * not a problem.) It's not clear whether anyone will care about the
1040  * typmod, but we store that too just in case.
1041  */
1042  if (index_expr)
1043  {
1044  stats->attrtypid = exprType(index_expr);
1045  stats->attrtypmod = exprTypmod(index_expr);
1046 
1047  /*
1048  * If a collation has been specified for the index column, use that in
1049  * preference to anything else; but if not, fall back to whatever we
1050  * can get from the expression.
1051  */
1052  if (OidIsValid(onerel->rd_indcollation[attnum - 1]))
1053  stats->attrcollid = onerel->rd_indcollation[attnum - 1];
1054  else
1055  stats->attrcollid = exprCollation(index_expr);
1056  }
1057  else
1058  {
1059  stats->attrtypid = attr->atttypid;
1060  stats->attrtypmod = attr->atttypmod;
1061  stats->attrcollid = attr->attcollation;
1062  }
1063 
1064  typtuple = SearchSysCacheCopy1(TYPEOID,
1065  ObjectIdGetDatum(stats->attrtypid));
1066  if (!HeapTupleIsValid(typtuple))
1067  elog(ERROR, "cache lookup failed for type %u", stats->attrtypid);
1068  stats->attrtype = (Form_pg_type) GETSTRUCT(typtuple);
1069  stats->anl_context = anl_context;
1070  stats->tupattnum = attnum;
1071 
1072  /*
1073  * The fields describing the stats->stavalues[n] element types default to
1074  * the type of the data being analyzed, but the type-specific typanalyze
1075  * function can change them if it wants to store something else.
1076  */
1077  for (i = 0; i < STATISTIC_NUM_SLOTS; i++)
1078  {
1079  stats->statypid[i] = stats->attrtypid;
1080  stats->statyplen[i] = stats->attrtype->typlen;
1081  stats->statypbyval[i] = stats->attrtype->typbyval;
1082  stats->statypalign[i] = stats->attrtype->typalign;
1083  }
1084 
1085  /*
1086  * Call the type-specific typanalyze function. If none is specified, use
1087  * std_typanalyze().
1088  */
1089  if (OidIsValid(stats->attrtype->typanalyze))
1090  ok = DatumGetBool(OidFunctionCall1(stats->attrtype->typanalyze,
1091  PointerGetDatum(stats)));
1092  else
1093  ok = std_typanalyze(stats);
1094 
1095  if (!ok || stats->compute_stats == NULL || stats->minrows <= 0)
1096  {
1097  heap_freetuple(typtuple);
1098  pfree(stats);
1099  return NULL;
1100  }
1101 
1102  return stats;
1103 }
#define OidIsValid(objectId)
Definition: c.h:762
bool std_typanalyze(VacAttrStats *stats)
Definition: analyze.c:1886
#define OidFunctionCall1(functionId, arg1)
Definition: fmgr.h:680
#define HeapTupleIsValid(tuple)
Definition: htup.h:78
#define GETSTRUCT(TUP)
Definition: htup_details.h:653
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
FormData_pg_attribute * Form_pg_attribute
Definition: pg_attribute.h:209
#define STATISTIC_NUM_SLOTS
Definition: pg_statistic.h:127
FormData_pg_type * Form_pg_type
Definition: pg_type.h:261
static Datum Int16GetDatum(int16 X)
Definition: postgres.h:172
static Datum ObjectIdGetDatum(Oid X)
Definition: postgres.h:252
static int16 DatumGetInt16(Datum X)
Definition: postgres.h:162
Oid * rd_indcollation
Definition: rel.h:217
int32 attrtypmod
Definition: vacuum.h:126
Oid statypid[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:161
char statypalign[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:164
Oid attrtypid
Definition: vacuum.h:125
bool statypbyval[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:163
int16 statyplen[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:162
void ReleaseSysCache(HeapTuple tuple)
Definition: syscache.c:266
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
#define TupleDescAttr(tupdesc, i)
Definition: tupdesc.h:92

References anl_context, VacAttrStats::anl_context, attnum, VacAttrStats::attrcollid, VacAttrStats::attrtype, VacAttrStats::attrtypid, VacAttrStats::attrtypmod, VacAttrStats::attstattarget, VacAttrStats::compute_stats, DatumGetBool(), DatumGetInt16(), elog, ERROR, exprCollation(), exprType(), exprTypmod(), GETSTRUCT, heap_freetuple(), HeapTupleIsValid, i, Int16GetDatum(), VacAttrStats::minrows, ObjectIdGetDatum(), OidFunctionCall1, OidIsValid, palloc0(), pfree(), PointerGetDatum(), RelationData::rd_att, RelationData::rd_indcollation, RelationGetRelid, ReleaseSysCache(), SearchSysCache2(), SearchSysCacheCopy1, STATISTIC_NUM_SLOTS, VacAttrStats::statypalign, VacAttrStats::statypbyval, VacAttrStats::statypid, VacAttrStats::statyplen, std_typanalyze(), SysCacheGetAttr(), VacAttrStats::tupattnum, and TupleDescAttr.

Referenced by do_analyze_rel().

◆ ind_fetch_func()

static Datum ind_fetch_func ( VacAttrStatsP  stats,
int  rownum,
bool isNull 
)
static

Definition at line 1809 of file analyze.c.

1810 {
1811  int i;
1812 
1813  /* exprvals and exprnulls are already offset for proper column */
1814  i = rownum * stats->rowstride;
1815  *isNull = stats->exprnulls[i];
1816  return stats->exprvals[i];
1817 }

References VacAttrStats::exprnulls, VacAttrStats::exprvals, i, and VacAttrStats::rowstride.

Referenced by compute_index_stats().

◆ std_fetch_func()

static Datum std_fetch_func ( VacAttrStatsP  stats,
int  rownum,
bool isNull 
)
static

Definition at line 1793 of file analyze.c.

1794 {
1795  int attnum = stats->tupattnum;
1796  HeapTuple tuple = stats->rows[rownum];
1797  TupleDesc tupDesc = stats->tupDesc;
1798 
1799  return heap_getattr(tuple, attnum, tupDesc, isNull);
1800 }
static Datum heap_getattr(HeapTuple tup, int attnum, TupleDesc tupleDesc, bool *isnull)
Definition: htup_details.h:792

References attnum, heap_getattr(), VacAttrStats::rows, VacAttrStats::tupattnum, and VacAttrStats::tupDesc.

Referenced by do_analyze_rel().

◆ std_typanalyze()

bool std_typanalyze ( VacAttrStats stats)

Definition at line 1886 of file analyze.c.

1887 {
1888  Oid ltopr;
1889  Oid eqopr;
1890  StdAnalyzeData *mystats;
1891 
1892  /* If the attstattarget column is negative, use the default value */
1893  if (stats->attstattarget < 0)
1895 
1896  /* Look for default "<" and "=" operators for column's type */
1898  false, false, false,
1899  &ltopr, &eqopr, NULL,
1900  NULL);
1901 
1902  /* Save the operator info for compute_stats routines */
1903  mystats = (StdAnalyzeData *) palloc(sizeof(StdAnalyzeData));
1904  mystats->eqopr = eqopr;
1905  mystats->eqfunc = OidIsValid(eqopr) ? get_opcode(eqopr) : InvalidOid;
1906  mystats->ltopr = ltopr;
1907  stats->extra_data = mystats;
1908 
1909  /*
1910  * Determine which standard statistics algorithm to use
1911  */
1912  if (OidIsValid(eqopr) && OidIsValid(ltopr))
1913  {
1914  /* Seems to be a scalar datatype */
1916  /*--------------------
1917  * The following choice of minrows is based on the paper
1918  * "Random sampling for histogram construction: how much is enough?"
1919  * by Surajit Chaudhuri, Rajeev Motwani and Vivek Narasayya, in
1920  * Proceedings of ACM SIGMOD International Conference on Management
1921  * of Data, 1998, Pages 436-447. Their Corollary 1 to Theorem 5
1922  * says that for table size n, histogram size k, maximum relative
1923  * error in bin size f, and error probability gamma, the minimum
1924  * random sample size is
1925  * r = 4 * k * ln(2*n/gamma) / f^2
1926  * Taking f = 0.5, gamma = 0.01, n = 10^6 rows, we obtain
1927  * r = 305.82 * k
1928  * Note that because of the log function, the dependence on n is
1929  * quite weak; even at n = 10^12, a 300*k sample gives <= 0.66
1930  * bin size error with probability 0.99. So there's no real need to
1931  * scale for n, which is a good thing because we don't necessarily
1932  * know it at this point.
1933  *--------------------
1934  */
1935  stats->minrows = 300 * stats->attstattarget;
1936  }
1937  else if (OidIsValid(eqopr))
1938  {
1939  /* We can still recognize distinct values */
1941  /* Might as well use the same minrows as above */
1942  stats->minrows = 300 * stats->attstattarget;
1943  }
1944  else
1945  {
1946  /* Can't do much but the trivial stuff */
1948  /* Might as well use the same minrows as above */
1949  stats->minrows = 300 * stats->attstattarget;
1950  }
1951 
1952  return true;
1953 }
static void compute_scalar_stats(VacAttrStatsP stats, AnalyzeAttrFetchFunc fetchfunc, int samplerows, double totalrows)
Definition: analyze.c:2397
int default_statistics_target
Definition: analyze.c:73
static void compute_distinct_stats(VacAttrStatsP stats, AnalyzeAttrFetchFunc fetchfunc, int samplerows, double totalrows)
Definition: analyze.c:2054
static void compute_trivial_stats(VacAttrStatsP stats, AnalyzeAttrFetchFunc fetchfunc, int samplerows, double totalrows)
Definition: analyze.c:1964
RegProcedure get_opcode(Oid opno)
Definition: lsyscache.c:1263
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
#define InvalidOid
Definition: postgres_ext.h:36

References VacAttrStats::attrtypid, VacAttrStats::attstattarget, compute_distinct_stats(), compute_scalar_stats(), VacAttrStats::compute_stats, compute_trivial_stats(), default_statistics_target, StdAnalyzeData::eqfunc, StdAnalyzeData::eqopr, VacAttrStats::extra_data, get_opcode(), get_sort_group_operators(), InvalidOid, StdAnalyzeData::ltopr, VacAttrStats::minrows, OidIsValid, and palloc().

Referenced by array_typanalyze(), examine_attribute(), and examine_expression().

◆ update_attstats()

static void update_attstats ( Oid  relid,
bool  inh,
int  natts,
VacAttrStats **  vacattrstats 
)
static

Definition at line 1650 of file analyze.c.

1651 {
1652  Relation sd;
1653  int attno;
1654  CatalogIndexState indstate = NULL;
1655 
1656  if (natts <= 0)
1657  return; /* nothing to do */
1658 
1659  sd = table_open(StatisticRelationId, RowExclusiveLock);
1660 
1661  for (attno = 0; attno < natts; attno++)
1662  {
1663  VacAttrStats *stats = vacattrstats[attno];
1664  HeapTuple stup,
1665  oldtup;
1666  int i,
1667  k,
1668  n;
1669  Datum values[Natts_pg_statistic];
1670  bool nulls[Natts_pg_statistic];
1671  bool replaces[Natts_pg_statistic];
1672 
1673  /* Ignore attr if we weren't able to collect stats */
1674  if (!stats->stats_valid)
1675  continue;
1676 
1677  /*
1678  * Construct a new pg_statistic tuple
1679  */
1680  for (i = 0; i < Natts_pg_statistic; ++i)
1681  {
1682  nulls[i] = false;
1683  replaces[i] = true;
1684  }
1685 
1686  values[Anum_pg_statistic_starelid - 1] = ObjectIdGetDatum(relid);
1687  values[Anum_pg_statistic_staattnum - 1] = Int16GetDatum(stats->tupattnum);
1688  values[Anum_pg_statistic_stainherit - 1] = BoolGetDatum(inh);
1689  values[Anum_pg_statistic_stanullfrac - 1] = Float4GetDatum(stats->stanullfrac);
1690  values[Anum_pg_statistic_stawidth - 1] = Int32GetDatum(stats->stawidth);
1691  values[Anum_pg_statistic_stadistinct - 1] = Float4GetDatum(stats->stadistinct);
1692  i = Anum_pg_statistic_stakind1 - 1;
1693  for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1694  {
1695  values[i++] = Int16GetDatum(stats->stakind[k]); /* stakindN */
1696  }
1697  i = Anum_pg_statistic_staop1 - 1;
1698  for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1699  {
1700  values[i++] = ObjectIdGetDatum(stats->staop[k]); /* staopN */
1701  }
1702  i = Anum_pg_statistic_stacoll1 - 1;
1703  for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1704  {
1705  values[i++] = ObjectIdGetDatum(stats->stacoll[k]); /* stacollN */
1706  }
1707  i = Anum_pg_statistic_stanumbers1 - 1;
1708  for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1709  {
1710  int nnum = stats->numnumbers[k];
1711 
1712  if (nnum > 0)
1713  {
1714  Datum *numdatums = (Datum *) palloc(nnum * sizeof(Datum));
1715  ArrayType *arry;
1716 
1717  for (n = 0; n < nnum; n++)
1718  numdatums[n] = Float4GetDatum(stats->stanumbers[k][n]);
1719  arry = construct_array_builtin(numdatums, nnum, FLOAT4OID);
1720  values[i++] = PointerGetDatum(arry); /* stanumbersN */
1721  }
1722  else
1723  {
1724  nulls[i] = true;
1725  values[i++] = (Datum) 0;
1726  }
1727  }
1728  i = Anum_pg_statistic_stavalues1 - 1;
1729  for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1730  {
1731  if (stats->numvalues[k] > 0)
1732  {
1733  ArrayType *arry;
1734 
1735  arry = construct_array(stats->stavalues[k],
1736  stats->numvalues[k],
1737  stats->statypid[k],
1738  stats->statyplen[k],
1739  stats->statypbyval[k],
1740  stats->statypalign[k]);
1741  values[i++] = PointerGetDatum(arry); /* stavaluesN */
1742  }
1743  else
1744  {
1745  nulls[i] = true;
1746  values[i++] = (Datum) 0;
1747  }
1748  }
1749 
1750  /* Is there already a pg_statistic tuple for this attribute? */
1751  oldtup = SearchSysCache3(STATRELATTINH,
1752  ObjectIdGetDatum(relid),
1753  Int16GetDatum(stats->tupattnum),
1754  BoolGetDatum(inh));
1755 
1756  /* Open index information when we know we need it */
1757  if (indstate == NULL)
1758  indstate = CatalogOpenIndexes(sd);
1759 
1760  if (HeapTupleIsValid(oldtup))
1761  {
1762  /* Yes, replace it */
1763  stup = heap_modify_tuple(oldtup,
1764  RelationGetDescr(sd),
1765  values,
1766  nulls,
1767  replaces);
1768  ReleaseSysCache(oldtup);
1769  CatalogTupleUpdateWithInfo(sd, &stup->t_self, stup, indstate);
1770  }
1771  else
1772  {
1773  /* No, insert new tuple */
1774  stup = heap_form_tuple(RelationGetDescr(sd), values, nulls);
1775  CatalogTupleInsertWithInfo(sd, stup, indstate);
1776  }
1777 
1778  heap_freetuple(stup);
1779  }
1780 
1781  if (indstate != NULL)
1782  CatalogCloseIndexes(indstate);
1784 }
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
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 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
#define RowExclusiveLock
Definition: lockdefs.h:38
static Datum Float4GetDatum(float4 X)
Definition: postgres.h:475
static Datum BoolGetDatum(bool X)
Definition: postgres.h:102
static Datum Int32GetDatum(int32 X)
Definition: postgres.h:212
HeapTuple SearchSysCache3(int cacheId, Datum key1, Datum key2, Datum key3)
Definition: syscache.c:240

References BoolGetDatum(), CatalogCloseIndexes(), CatalogOpenIndexes(), CatalogTupleInsertWithInfo(), CatalogTupleUpdateWithInfo(), construct_array(), construct_array_builtin(), Float4GetDatum(), heap_form_tuple(), heap_freetuple(), heap_modify_tuple(), HeapTupleIsValid, i, Int16GetDatum(), Int32GetDatum(), VacAttrStats::numnumbers, VacAttrStats::numvalues, ObjectIdGetDatum(), palloc(), PointerGetDatum(), RelationGetDescr, ReleaseSysCache(), RowExclusiveLock, SearchSysCache3(), VacAttrStats::stacoll, VacAttrStats::stadistinct, VacAttrStats::stakind, VacAttrStats::stanullfrac, VacAttrStats::stanumbers, VacAttrStats::staop, STATISTIC_NUM_SLOTS, VacAttrStats::stats_valid, VacAttrStats::statypalign, VacAttrStats::statypbyval, VacAttrStats::statypid, VacAttrStats::statyplen, VacAttrStats::stavalues, VacAttrStats::stawidth, HeapTupleData::t_self, table_close(), table_open(), VacAttrStats::tupattnum, and values.

Referenced by do_analyze_rel().

Variable Documentation

◆ anl_context

MemoryContext anl_context = NULL
static

Definition at line 76 of file analyze.c.

Referenced by compute_index_stats(), do_analyze_rel(), and examine_attribute().

◆ default_statistics_target

int default_statistics_target = 100

◆ vac_strategy

BufferAccessStrategy vac_strategy
static

Definition at line 77 of file analyze.c.

Referenced by acquire_sample_rows(), analyze_rel(), and do_analyze_rel().