PostgreSQL Source Code  git master
ts_typanalyze.c File Reference
#include "postgres.h"
#include "catalog/pg_collation.h"
#include "catalog/pg_operator.h"
#include "commands/vacuum.h"
#include "common/hashfn.h"
#include "tsearch/ts_type.h"
#include "utils/builtins.h"
#include "varatt.h"
Include dependency graph for ts_typanalyze.c:

Go to the source code of this file.

Data Structures

struct  LexemeHashKey
 
struct  TrackItem
 

Functions

static void compute_tsvector_stats (VacAttrStats *stats, AnalyzeAttrFetchFunc fetchfunc, int samplerows, double totalrows)
 
static void prune_lexemes_hashtable (HTAB *lexemes_tab, int b_current)
 
static uint32 lexeme_hash (const void *key, Size keysize)
 
static int lexeme_match (const void *key1, const void *key2, Size keysize)
 
static int lexeme_compare (const void *key1, const void *key2)
 
static int trackitem_compare_frequencies_desc (const void *e1, const void *e2, void *arg)
 
static int trackitem_compare_lexemes (const void *e1, const void *e2, void *arg)
 
Datum ts_typanalyze (PG_FUNCTION_ARGS)
 

Function Documentation

◆ compute_tsvector_stats()

static void compute_tsvector_stats ( VacAttrStats stats,
AnalyzeAttrFetchFunc  fetchfunc,
int  samplerows,
double  totalrows 
)
static

Definition at line 141 of file ts_typanalyze.c.

145 {
146  int num_mcelem;
147  int null_cnt = 0;
148  double total_width = 0;
149 
150  /* This is D from the LC algorithm. */
151  HTAB *lexemes_tab;
152  HASHCTL hash_ctl;
153  HASH_SEQ_STATUS scan_status;
154 
155  /* This is the current bucket number from the LC algorithm */
156  int b_current;
157 
158  /* This is 'w' from the LC algorithm */
159  int bucket_width;
160  int vector_no,
161  lexeme_no;
163 
164  /*
165  * We want statistics_target * 10 lexemes in the MCELEM array. This
166  * multiplier is pretty arbitrary, but is meant to reflect the fact that
167  * the number of individual lexeme values tracked in pg_statistic ought to
168  * be more than the number of values for a simple scalar column.
169  */
170  num_mcelem = stats->attstattarget * 10;
171 
172  /*
173  * We set bucket width equal to (num_mcelem + 10) / 0.007 as per the
174  * comment above.
175  */
176  bucket_width = (num_mcelem + 10) * 1000 / 7;
177 
178  /*
179  * Create the hashtable. It will be in local memory, so we don't need to
180  * worry about overflowing the initial size. Also we don't need to pay any
181  * attention to locking and memory management.
182  */
183  hash_ctl.keysize = sizeof(LexemeHashKey);
184  hash_ctl.entrysize = sizeof(TrackItem);
185  hash_ctl.hash = lexeme_hash;
186  hash_ctl.match = lexeme_match;
187  hash_ctl.hcxt = CurrentMemoryContext;
188  lexemes_tab = hash_create("Analyzed lexemes table",
189  num_mcelem,
190  &hash_ctl,
192 
193  /* Initialize counters. */
194  b_current = 1;
195  lexeme_no = 0;
196 
197  /* Loop over the tsvectors. */
198  for (vector_no = 0; vector_no < samplerows; vector_no++)
199  {
200  Datum value;
201  bool isnull;
202  TSVector vector;
203  WordEntry *curentryptr;
204  char *lexemesptr;
205  int j;
206 
208 
209  value = fetchfunc(stats, vector_no, &isnull);
210 
211  /*
212  * Check for null/nonnull.
213  */
214  if (isnull)
215  {
216  null_cnt++;
217  continue;
218  }
219 
220  /*
221  * Add up widths for average-width calculation. Since it's a
222  * tsvector, we know it's varlena. As in the regular
223  * compute_minimal_stats function, we use the toasted width for this
224  * calculation.
225  */
226  total_width += VARSIZE_ANY(DatumGetPointer(value));
227 
228  /*
229  * Now detoast the tsvector if needed.
230  */
231  vector = DatumGetTSVector(value);
232 
233  /*
234  * We loop through the lexemes in the tsvector and add them to our
235  * tracking hashtable.
236  */
237  lexemesptr = STRPTR(vector);
238  curentryptr = ARRPTR(vector);
239  for (j = 0; j < vector->size; j++)
240  {
241  TrackItem *item;
242  bool found;
243 
244  /*
245  * Construct a hash key. The key points into the (detoasted)
246  * tsvector value at this point, but if a new entry is created, we
247  * make a copy of it. This way we can free the tsvector value
248  * once we've processed all its lexemes.
249  */
250  hash_key.lexeme = lexemesptr + curentryptr->pos;
251  hash_key.length = curentryptr->len;
252 
253  /* Lookup current lexeme in hashtable, adding it if new */
254  item = (TrackItem *) hash_search(lexemes_tab,
255  &hash_key,
256  HASH_ENTER, &found);
257 
258  if (found)
259  {
260  /* The lexeme is already on the tracking list */
261  item->frequency++;
262  }
263  else
264  {
265  /* Initialize new tracking list element */
266  item->frequency = 1;
267  item->delta = b_current - 1;
268 
269  item->key.lexeme = palloc(hash_key.length);
270  memcpy(item->key.lexeme, hash_key.lexeme, hash_key.length);
271  }
272 
273  /* lexeme_no is the number of elements processed (ie N) */
274  lexeme_no++;
275 
276  /* We prune the D structure after processing each bucket */
277  if (lexeme_no % bucket_width == 0)
278  {
279  prune_lexemes_hashtable(lexemes_tab, b_current);
280  b_current++;
281  }
282 
283  /* Advance to the next WordEntry in the tsvector */
284  curentryptr++;
285  }
286 
287  /* If the vector was toasted, free the detoasted copy. */
288  if (TSVectorGetDatum(vector) != value)
289  pfree(vector);
290  }
291 
292  /* We can only compute real stats if we found some non-null values. */
293  if (null_cnt < samplerows)
294  {
295  int nonnull_cnt = samplerows - null_cnt;
296  int i;
297  TrackItem **sort_table;
298  TrackItem *item;
299  int track_len;
300  int cutoff_freq;
301  int minfreq,
302  maxfreq;
303 
304  stats->stats_valid = true;
305  /* Do the simple null-frac and average width stats */
306  stats->stanullfrac = (double) null_cnt / (double) samplerows;
307  stats->stawidth = total_width / (double) nonnull_cnt;
308 
309  /* Assume it's a unique column (see notes above) */
310  stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
311 
312  /*
313  * Construct an array of the interesting hashtable items, that is,
314  * those meeting the cutoff frequency (s - epsilon)*N. Also identify
315  * the minimum and maximum frequencies among these items.
316  *
317  * Since epsilon = s/10 and bucket_width = 1/epsilon, the cutoff
318  * frequency is 9*N / bucket_width.
319  */
320  cutoff_freq = 9 * lexeme_no / bucket_width;
321 
322  i = hash_get_num_entries(lexemes_tab); /* surely enough space */
323  sort_table = (TrackItem **) palloc(sizeof(TrackItem *) * i);
324 
325  hash_seq_init(&scan_status, lexemes_tab);
326  track_len = 0;
327  minfreq = lexeme_no;
328  maxfreq = 0;
329  while ((item = (TrackItem *) hash_seq_search(&scan_status)) != NULL)
330  {
331  if (item->frequency > cutoff_freq)
332  {
333  sort_table[track_len++] = item;
334  minfreq = Min(minfreq, item->frequency);
335  maxfreq = Max(maxfreq, item->frequency);
336  }
337  }
338  Assert(track_len <= i);
339 
340  /* emit some statistics for debug purposes */
341  elog(DEBUG3, "tsvector_stats: target # mces = %d, bucket width = %d, "
342  "# lexemes = %d, hashtable size = %d, usable entries = %d",
343  num_mcelem, bucket_width, lexeme_no, i, track_len);
344 
345  /*
346  * If we obtained more lexemes than we really want, get rid of those
347  * with least frequencies. The easiest way is to qsort the array into
348  * descending frequency order and truncate the array.
349  */
350  if (num_mcelem < track_len)
351  {
352  qsort_interruptible(sort_table, track_len, sizeof(TrackItem *),
354  /* reset minfreq to the smallest frequency we're keeping */
355  minfreq = sort_table[num_mcelem - 1]->frequency;
356  }
357  else
358  num_mcelem = track_len;
359 
360  /* Generate MCELEM slot entry */
361  if (num_mcelem > 0)
362  {
363  MemoryContext old_context;
364  Datum *mcelem_values;
365  float4 *mcelem_freqs;
366 
367  /*
368  * We want to store statistics sorted on the lexeme value using
369  * first length, then byte-for-byte comparison. The reason for
370  * doing length comparison first is that we don't care about the
371  * ordering so long as it's consistent, and comparing lengths
372  * first gives us a chance to avoid a strncmp() call.
373  *
374  * This is different from what we do with scalar statistics --
375  * they get sorted on frequencies. The rationale is that we
376  * usually search through most common elements looking for a
377  * specific value, so we can grab its frequency. When values are
378  * presorted we can employ binary search for that. See
379  * ts_selfuncs.c for a real usage scenario.
380  */
381  qsort_interruptible(sort_table, num_mcelem, sizeof(TrackItem *),
383 
384  /* Must copy the target values into anl_context */
385  old_context = MemoryContextSwitchTo(stats->anl_context);
386 
387  /*
388  * We sorted statistics on the lexeme value, but we want to be
389  * able to find out the minimal and maximal frequency without
390  * going through all the values. We keep those two extra
391  * frequencies in two extra cells in mcelem_freqs.
392  *
393  * (Note: the MCELEM statistics slot definition allows for a third
394  * extra number containing the frequency of nulls, but we don't
395  * create that for a tsvector column, since null elements aren't
396  * possible.)
397  */
398  mcelem_values = (Datum *) palloc(num_mcelem * sizeof(Datum));
399  mcelem_freqs = (float4 *) palloc((num_mcelem + 2) * sizeof(float4));
400 
401  /*
402  * See comments above about use of nonnull_cnt as the divisor for
403  * the final frequency estimates.
404  */
405  for (i = 0; i < num_mcelem; i++)
406  {
407  TrackItem *titem = sort_table[i];
408 
409  mcelem_values[i] =
411  titem->key.length));
412  mcelem_freqs[i] = (double) titem->frequency / (double) nonnull_cnt;
413  }
414  mcelem_freqs[i++] = (double) minfreq / (double) nonnull_cnt;
415  mcelem_freqs[i] = (double) maxfreq / (double) nonnull_cnt;
416  MemoryContextSwitchTo(old_context);
417 
418  stats->stakind[0] = STATISTIC_KIND_MCELEM;
419  stats->staop[0] = TextEqualOperator;
420  stats->stacoll[0] = DEFAULT_COLLATION_OID;
421  stats->stanumbers[0] = mcelem_freqs;
422  /* See above comment about two extra frequency fields */
423  stats->numnumbers[0] = num_mcelem + 2;
424  stats->stavalues[0] = mcelem_values;
425  stats->numvalues[0] = num_mcelem;
426  /* We are storing text values */
427  stats->statypid[0] = TEXTOID;
428  stats->statyplen[0] = -1; /* typlen, -1 for varlena */
429  stats->statypbyval[0] = false;
430  stats->statypalign[0] = 'i';
431  }
432  }
433  else
434  {
435  /* We found only nulls; assume the column is entirely null */
436  stats->stats_valid = true;
437  stats->stanullfrac = 1.0;
438  stats->stawidth = 0; /* "unknown" */
439  stats->stadistinct = 0.0; /* "unknown" */
440  }
441 
442  /*
443  * We don't need to bother cleaning up any of our temporary palloc's. The
444  * hashtable should also go away, as it used a child memory context.
445  */
446 }
#define Min(x, y)
Definition: c.h:991
#define Max(x, y)
Definition: c.h:985
float float4
Definition: c.h:616
#define ARRPTR(x)
Definition: cube.c:25
static dshash_hash hash_key(dshash_table *hash_table, const void *key)
Definition: dshash.c:1061
void * hash_search(HTAB *hashp, const void *keyPtr, HASHACTION action, bool *foundPtr)
Definition: dynahash.c:955
HTAB * hash_create(const char *tabname, long nelem, const HASHCTL *info, int flags)
Definition: dynahash.c:352
long hash_get_num_entries(HTAB *hashp)
Definition: dynahash.c:1341
void * hash_seq_search(HASH_SEQ_STATUS *status)
Definition: dynahash.c:1395
void hash_seq_init(HASH_SEQ_STATUS *status, HTAB *hashp)
Definition: dynahash.c:1385
#define DEBUG3
Definition: elog.h:28
#define elog(elevel,...)
Definition: elog.h:224
@ HASH_ENTER
Definition: hsearch.h:114
#define HASH_CONTEXT
Definition: hsearch.h:102
#define HASH_ELEM
Definition: hsearch.h:95
#define HASH_COMPARE
Definition: hsearch.h:99
#define HASH_FUNCTION
Definition: hsearch.h:98
#define STRPTR(x)
Definition: hstore.h:76
static struct @150 value
int j
Definition: isn.c:74
int i
Definition: isn.c:73
Assert(fmt[strlen(fmt) - 1] !='\n')
void pfree(void *pointer)
Definition: mcxt.c:1508
MemoryContext CurrentMemoryContext
Definition: mcxt.c:131
void * palloc(Size size)
Definition: mcxt.c:1304
static MemoryContext MemoryContextSwitchTo(MemoryContext context)
Definition: palloc.h:124
void qsort_interruptible(void *base, size_t nel, size_t elsize, qsort_arg_comparator cmp, void *arg)
static Datum PointerGetDatum(const void *X)
Definition: postgres.h:322
uintptr_t Datum
Definition: postgres.h:64
static Pointer DatumGetPointer(Datum X)
Definition: postgres.h:312
Size keysize
Definition: hsearch.h:75
HashValueFunc hash
Definition: hsearch.h:78
Size entrysize
Definition: hsearch.h:76
HashCompareFunc match
Definition: hsearch.h:80
MemoryContext hcxt
Definition: hsearch.h:86
Definition: dynahash.c:220
int32 size
Definition: ts_type.h:93
LexemeHashKey key
Definition: ts_typanalyze.c:35
bool stats_valid
Definition: vacuum.h:143
float4 stanullfrac
Definition: vacuum.h:144
int16 stakind[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:147
MemoryContext anl_context
Definition: vacuum.h:129
Oid statypid[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:161
Oid staop[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:148
Oid stacoll[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:149
char statypalign[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:164
float4 * stanumbers[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:151
int attstattarget
Definition: vacuum.h:124
int32 stawidth
Definition: vacuum.h:145
bool statypbyval[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:163
int16 statyplen[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:162
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
uint32 pos
Definition: ts_type.h:46
uint32 len
Definition: ts_type.h:45
static int trackitem_compare_frequencies_desc(const void *e1, const void *e2, void *arg)
static void prune_lexemes_hashtable(HTAB *lexemes_tab, int b_current)
static int trackitem_compare_lexemes(const void *e1, const void *e2, void *arg)
static int lexeme_match(const void *key1, const void *key2, Size keysize)
static uint32 lexeme_hash(const void *key, Size keysize)
static TSVector DatumGetTSVector(Datum X)
Definition: ts_type.h:118
static Datum TSVectorGetDatum(const TSVectorData *X)
Definition: ts_type.h:130
void vacuum_delay_point(void)
Definition: vacuum.c:2337
#define VARSIZE_ANY(PTR)
Definition: varatt.h:311
text * cstring_to_text_with_len(const char *s, int len)
Definition: varlena.c:196

References VacAttrStats::anl_context, ARRPTR, Assert(), VacAttrStats::attstattarget, cstring_to_text_with_len(), CurrentMemoryContext, DatumGetPointer(), DatumGetTSVector(), DEBUG3, TrackItem::delta, elog, HASHCTL::entrysize, TrackItem::frequency, HASHCTL::hash, HASH_COMPARE, HASH_CONTEXT, hash_create(), HASH_ELEM, HASH_ENTER, HASH_FUNCTION, hash_get_num_entries(), hash_key(), hash_search(), hash_seq_init(), hash_seq_search(), HASHCTL::hcxt, i, j, TrackItem::key, HASHCTL::keysize, WordEntry::len, LexemeHashKey::length, LexemeHashKey::lexeme, lexeme_hash(), lexeme_match(), HASHCTL::match, Max, MemoryContextSwitchTo(), Min, VacAttrStats::numnumbers, VacAttrStats::numvalues, palloc(), pfree(), PointerGetDatum(), WordEntry::pos, prune_lexemes_hashtable(), qsort_interruptible(), TSVectorData::size, VacAttrStats::stacoll, VacAttrStats::stadistinct, VacAttrStats::stakind, VacAttrStats::stanullfrac, VacAttrStats::stanumbers, VacAttrStats::staop, VacAttrStats::stats_valid, VacAttrStats::statypalign, VacAttrStats::statypbyval, VacAttrStats::statypid, VacAttrStats::statyplen, VacAttrStats::stavalues, VacAttrStats::stawidth, STRPTR, trackitem_compare_frequencies_desc(), trackitem_compare_lexemes(), TSVectorGetDatum(), vacuum_delay_point(), value, and VARSIZE_ANY.

Referenced by ts_typanalyze().

◆ lexeme_compare()

static int lexeme_compare ( const void *  key1,
const void *  key2 
)
static

Definition at line 500 of file ts_typanalyze.c.

501 {
502  const LexemeHashKey *d1 = (const LexemeHashKey *) key1;
503  const LexemeHashKey *d2 = (const LexemeHashKey *) key2;
504 
505  /* First, compare by length */
506  if (d1->length > d2->length)
507  return 1;
508  else if (d1->length < d2->length)
509  return -1;
510  /* Lengths are equal, do a byte-by-byte comparison */
511  return strncmp(d1->lexeme, d2->lexeme, d1->length);
512 }

References LexemeHashKey::length, and LexemeHashKey::lexeme.

Referenced by lexeme_match(), and trackitem_compare_lexemes().

◆ lexeme_hash()

static uint32 lexeme_hash ( const void *  key,
Size  keysize 
)
static

Definition at line 478 of file ts_typanalyze.c.

479 {
480  const LexemeHashKey *l = (const LexemeHashKey *) key;
481 
482  return DatumGetUInt32(hash_any((const unsigned char *) l->lexeme,
483  l->length));
484 }
static Datum hash_any(const unsigned char *k, int keylen)
Definition: hashfn.h:31
static uint32 DatumGetUInt32(Datum X)
Definition: postgres.h:222

References DatumGetUInt32(), hash_any(), sort-test::key, LexemeHashKey::length, and LexemeHashKey::lexeme.

Referenced by compute_tsvector_stats().

◆ lexeme_match()

static int lexeme_match ( const void *  key1,
const void *  key2,
Size  keysize 
)
static

Definition at line 490 of file ts_typanalyze.c.

491 {
492  /* The keysize parameter is superfluous, the keys store their lengths */
493  return lexeme_compare(key1, key2);
494 }
static int lexeme_compare(const void *key1, const void *key2)

References lexeme_compare().

Referenced by compute_tsvector_stats().

◆ prune_lexemes_hashtable()

static void prune_lexemes_hashtable ( HTAB lexemes_tab,
int  b_current 
)
static

Definition at line 453 of file ts_typanalyze.c.

454 {
455  HASH_SEQ_STATUS scan_status;
456  TrackItem *item;
457 
458  hash_seq_init(&scan_status, lexemes_tab);
459  while ((item = (TrackItem *) hash_seq_search(&scan_status)) != NULL)
460  {
461  if (item->frequency + item->delta <= b_current)
462  {
463  char *lexeme = item->key.lexeme;
464 
465  if (hash_search(lexemes_tab, &item->key,
466  HASH_REMOVE, NULL) == NULL)
467  elog(ERROR, "hash table corrupted");
468  pfree(lexeme);
469  }
470  }
471 }
#define ERROR
Definition: elog.h:39
@ HASH_REMOVE
Definition: hsearch.h:115

References TrackItem::delta, elog, ERROR, TrackItem::frequency, HASH_REMOVE, hash_search(), hash_seq_init(), hash_seq_search(), TrackItem::key, LexemeHashKey::lexeme, and pfree().

Referenced by compute_tsvector_stats().

◆ trackitem_compare_frequencies_desc()

static int trackitem_compare_frequencies_desc ( const void *  e1,
const void *  e2,
void *  arg 
)
static

Definition at line 518 of file ts_typanalyze.c.

519 {
520  const TrackItem *const *t1 = (const TrackItem *const *) e1;
521  const TrackItem *const *t2 = (const TrackItem *const *) e2;
522 
523  return (*t2)->frequency - (*t1)->frequency;
524 }

References TrackItem::frequency.

Referenced by compute_tsvector_stats().

◆ trackitem_compare_lexemes()

static int trackitem_compare_lexemes ( const void *  e1,
const void *  e2,
void *  arg 
)
static

Definition at line 530 of file ts_typanalyze.c.

531 {
532  const TrackItem *const *t1 = (const TrackItem *const *) e1;
533  const TrackItem *const *t2 = (const TrackItem *const *) e2;
534 
535  return lexeme_compare(&(*t1)->key, &(*t2)->key);
536 }

References lexeme_compare().

Referenced by compute_tsvector_stats().

◆ ts_typanalyze()

Datum ts_typanalyze ( PG_FUNCTION_ARGS  )

Definition at line 58 of file ts_typanalyze.c.

59 {
61 
62  /* If the attstattarget column is negative, use the default value */
63  if (stats->attstattarget < 0)
65 
67  /* see comment about the choice of minrows in commands/analyze.c */
68  stats->minrows = 300 * stats->attstattarget;
69 
70  PG_RETURN_BOOL(true);
71 }
int default_statistics_target
Definition: analyze.c:73
#define PG_GETARG_POINTER(n)
Definition: fmgr.h:276
#define PG_RETURN_BOOL(x)
Definition: fmgr.h:359
int minrows
Definition: vacuum.h:136
AnalyzeAttrComputeStatsFunc compute_stats
Definition: vacuum.h:135
static void compute_tsvector_stats(VacAttrStats *stats, AnalyzeAttrFetchFunc fetchfunc, int samplerows, double totalrows)

References VacAttrStats::attstattarget, VacAttrStats::compute_stats, compute_tsvector_stats(), default_statistics_target, VacAttrStats::minrows, PG_GETARG_POINTER, and PG_RETURN_BOOL.