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 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)
 
static int trackitem_compare_lexemes (const void *e1, const void *e2)
 
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 140 of file ts_typanalyze.c.

References VacAttrStats::anl_context, ARRPTR, Assert, VacAttrStats::attr, 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, 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, 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().

144 {
145  int num_mcelem;
146  int null_cnt = 0;
147  double total_width = 0;
148 
149  /* This is D from the LC algorithm. */
150  HTAB *lexemes_tab;
151  HASHCTL hash_ctl;
152  HASH_SEQ_STATUS scan_status;
153 
154  /* This is the current bucket number from the LC algorithm */
155  int b_current;
156 
157  /* This is 'w' from the LC algorithm */
158  int bucket_width;
159  int vector_no,
160  lexeme_no;
162  TrackItem *item;
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->attr->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  bool found;
242 
243  /*
244  * Construct a hash key. The key points into the (detoasted)
245  * tsvector value at this point, but if a new entry is created, we
246  * make a copy of it. This way we can free the tsvector value
247  * once we've processed all its lexemes.
248  */
249  hash_key.lexeme = lexemesptr + curentryptr->pos;
250  hash_key.length = curentryptr->len;
251 
252  /* Lookup current lexeme in hashtable, adding it if new */
253  item = (TrackItem *) hash_search(lexemes_tab,
254  (const void *) &hash_key,
255  HASH_ENTER, &found);
256 
257  if (found)
258  {
259  /* The lexeme is already on the tracking list */
260  item->frequency++;
261  }
262  else
263  {
264  /* Initialize new tracking list element */
265  item->frequency = 1;
266  item->delta = b_current - 1;
267 
268  item->key.lexeme = palloc(hash_key.length);
269  memcpy(item->key.lexeme, hash_key.lexeme, hash_key.length);
270  }
271 
272  /* lexeme_no is the number of elements processed (ie N) */
273  lexeme_no++;
274 
275  /* We prune the D structure after processing each bucket */
276  if (lexeme_no % bucket_width == 0)
277  {
278  prune_lexemes_hashtable(lexemes_tab, b_current);
279  b_current++;
280  }
281 
282  /* Advance to the next WordEntry in the tsvector */
283  curentryptr++;
284  }
285 
286  /* If the vector was toasted, free the detoasted copy. */
287  if (TSVectorGetDatum(vector) != value)
288  pfree(vector);
289  }
290 
291  /* We can only compute real stats if we found some non-null values. */
292  if (null_cnt < samplerows)
293  {
294  int nonnull_cnt = samplerows - null_cnt;
295  int i;
296  TrackItem **sort_table;
297  int track_len;
298  int cutoff_freq;
299  int minfreq,
300  maxfreq;
301 
302  stats->stats_valid = true;
303  /* Do the simple null-frac and average width stats */
304  stats->stanullfrac = (double) null_cnt / (double) samplerows;
305  stats->stawidth = total_width / (double) nonnull_cnt;
306 
307  /* Assume it's a unique column (see notes above) */
308  stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
309 
310  /*
311  * Construct an array of the interesting hashtable items, that is,
312  * those meeting the cutoff frequency (s - epsilon)*N. Also identify
313  * the minimum and maximum frequencies among these items.
314  *
315  * Since epsilon = s/10 and bucket_width = 1/epsilon, the cutoff
316  * frequency is 9*N / bucket_width.
317  */
318  cutoff_freq = 9 * lexeme_no / bucket_width;
319 
320  i = hash_get_num_entries(lexemes_tab); /* surely enough space */
321  sort_table = (TrackItem **) palloc(sizeof(TrackItem *) * i);
322 
323  hash_seq_init(&scan_status, lexemes_tab);
324  track_len = 0;
325  minfreq = lexeme_no;
326  maxfreq = 0;
327  while ((item = (TrackItem *) hash_seq_search(&scan_status)) != NULL)
328  {
329  if (item->frequency > cutoff_freq)
330  {
331  sort_table[track_len++] = item;
332  minfreq = Min(minfreq, item->frequency);
333  maxfreq = Max(maxfreq, item->frequency);
334  }
335  }
336  Assert(track_len <= i);
337 
338  /* emit some statistics for debug purposes */
339  elog(DEBUG3, "tsvector_stats: target # mces = %d, bucket width = %d, "
340  "# lexemes = %d, hashtable size = %d, usable entries = %d",
341  num_mcelem, bucket_width, lexeme_no, i, track_len);
342 
343  /*
344  * If we obtained more lexemes than we really want, get rid of those
345  * with least frequencies. The easiest way is to qsort the array into
346  * descending frequency order and truncate the array.
347  */
348  if (num_mcelem < track_len)
349  {
350  qsort(sort_table, track_len, sizeof(TrackItem *),
352  /* reset minfreq to the smallest frequency we're keeping */
353  minfreq = sort_table[num_mcelem - 1]->frequency;
354  }
355  else
356  num_mcelem = track_len;
357 
358  /* Generate MCELEM slot entry */
359  if (num_mcelem > 0)
360  {
361  MemoryContext old_context;
362  Datum *mcelem_values;
363  float4 *mcelem_freqs;
364 
365  /*
366  * We want to store statistics sorted on the lexeme value using
367  * first length, then byte-for-byte comparison. The reason for
368  * doing length comparison first is that we don't care about the
369  * ordering so long as it's consistent, and comparing lengths
370  * first gives us a chance to avoid a strncmp() call.
371  *
372  * This is different from what we do with scalar statistics --
373  * they get sorted on frequencies. The rationale is that we
374  * usually search through most common elements looking for a
375  * specific value, so we can grab its frequency. When values are
376  * presorted we can employ binary search for that. See
377  * ts_selfuncs.c for a real usage scenario.
378  */
379  qsort(sort_table, num_mcelem, sizeof(TrackItem *),
381 
382  /* Must copy the target values into anl_context */
383  old_context = MemoryContextSwitchTo(stats->anl_context);
384 
385  /*
386  * We sorted statistics on the lexeme value, but we want to be
387  * able to find out the minimal and maximal frequency without
388  * going through all the values. We keep those two extra
389  * frequencies in two extra cells in mcelem_freqs.
390  *
391  * (Note: the MCELEM statistics slot definition allows for a third
392  * extra number containing the frequency of nulls, but we don't
393  * create that for a tsvector column, since null elements aren't
394  * possible.)
395  */
396  mcelem_values = (Datum *) palloc(num_mcelem * sizeof(Datum));
397  mcelem_freqs = (float4 *) palloc((num_mcelem + 2) * sizeof(float4));
398 
399  /*
400  * See comments above about use of nonnull_cnt as the divisor for
401  * the final frequency estimates.
402  */
403  for (i = 0; i < num_mcelem; i++)
404  {
405  TrackItem *item = sort_table[i];
406 
407  mcelem_values[i] =
409  item->key.length));
410  mcelem_freqs[i] = (double) item->frequency / (double) nonnull_cnt;
411  }
412  mcelem_freqs[i++] = (double) minfreq / (double) nonnull_cnt;
413  mcelem_freqs[i] = (double) maxfreq / (double) nonnull_cnt;
414  MemoryContextSwitchTo(old_context);
415 
416  stats->stakind[0] = STATISTIC_KIND_MCELEM;
417  stats->staop[0] = TextEqualOperator;
418  stats->stacoll[0] = DEFAULT_COLLATION_OID;
419  stats->stanumbers[0] = mcelem_freqs;
420  /* See above comment about two extra frequency fields */
421  stats->numnumbers[0] = num_mcelem + 2;
422  stats->stavalues[0] = mcelem_values;
423  stats->numvalues[0] = num_mcelem;
424  /* We are storing text values */
425  stats->statypid[0] = TEXTOID;
426  stats->statyplen[0] = -1; /* typlen, -1 for varlena */
427  stats->statypbyval[0] = false;
428  stats->statypalign[0] = 'i';
429  }
430  }
431  else
432  {
433  /* We found only nulls; assume the column is entirely null */
434  stats->stats_valid = true;
435  stats->stanullfrac = 1.0;
436  stats->stawidth = 0; /* "unknown" */
437  stats->stadistinct = 0.0; /* "unknown" */
438  }
439 
440  /*
441  * We don't need to bother cleaning up any of our temporary palloc's. The
442  * hashtable should also go away, as it used a child memory context.
443  */
444 }
#define HASH_CONTEXT
Definition: hsearch.h:102
#define HASH_ELEM
Definition: hsearch.h:95
static struct @142 value
MemoryContext hcxt
Definition: hsearch.h:86
#define DEBUG3
Definition: elog.h:23
#define PointerGetDatum(X)
Definition: postgres.h:600
#define Min(x, y)
Definition: c.h:986
Datum * stavalues[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:152
static MemoryContext MemoryContextSwitchTo(MemoryContext context)
Definition: palloc.h:109
Size entrysize
Definition: hsearch.h:76
uint32 len
Definition: ts_type.h:44
long hash_get_num_entries(HTAB *hashp)
Definition: dynahash.c:1382
bool statypbyval[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:162
void * hash_search(HTAB *hashp, const void *keyPtr, HASHACTION action, bool *foundPtr)
Definition: dynahash.c:954
char statypalign[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:163
Form_pg_attribute attr
Definition: vacuum.h:123
Definition: dynahash.c:219
void pfree(void *pointer)
Definition: mcxt.c:1169
int32 stawidth
Definition: vacuum.h:144
Oid stacoll[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:148
text * cstring_to_text_with_len(const char *s, int len)
Definition: varlena.c:202
int numnumbers[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:149
LexemeHashKey key
Definition: ts_typanalyze.c:34
int32 size
Definition: ts_type.h:93
static int lexeme_match(const void *key1, const void *key2, Size keysize)
HTAB * hash_create(const char *tabname, long nelem, const HASHCTL *info, int flags)
Definition: dynahash.c:349
static void prune_lexemes_hashtable(HTAB *lexemes_tab, int b_current)
#define TSVectorGetDatum(X)
Definition: ts_type.h:119
MemoryContext CurrentMemoryContext
Definition: mcxt.c:42
#define DatumGetTSVector(X)
Definition: ts_type.h:117
float4 stanullfrac
Definition: vacuum.h:143
Oid staop[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:147
bool stats_valid
Definition: vacuum.h:142
float float4
Definition: c.h:564
uintptr_t Datum
Definition: postgres.h:411
int16 stakind[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:146
Size keysize
Definition: hsearch.h:75
HashCompareFunc match
Definition: hsearch.h:80
#define VARSIZE_ANY(PTR)
Definition: postgres.h:348
Oid statypid[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:160
static uint32 lexeme_hash(const void *key, Size keysize)
#define Max(x, y)
Definition: c.h:980
#define Assert(condition)
Definition: c.h:804
static int trackitem_compare_frequencies_desc(const void *e1, const void *e2)
float4 * stanumbers[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:150
#define HASH_COMPARE
Definition: hsearch.h:99
void * hash_seq_search(HASH_SEQ_STATUS *status)
Definition: dynahash.c:1436
uint32 pos
Definition: ts_type.h:44
void hash_seq_init(HASH_SEQ_STATUS *status, HTAB *hashp)
Definition: dynahash.c:1426
MemoryContext anl_context
Definition: vacuum.h:128
static int trackitem_compare_lexemes(const void *e1, const void *e2)
#define DatumGetPointer(X)
Definition: postgres.h:593
static dshash_hash hash_key(dshash_table *hash_table, const void *key)
Definition: dshash.c:883
int numvalues[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:151
void * palloc(Size size)
Definition: mcxt.c:1062
#define STRPTR(x)
Definition: hstore.h:76
int16 statyplen[STATISTIC_NUM_SLOTS]
Definition: vacuum.h:161
#define elog(elevel,...)
Definition: elog.h:232
int i
#define ARRPTR(x)
Definition: cube.c:25
#define qsort(a, b, c, d)
Definition: port.h:504
void vacuum_delay_point(void)
Definition: vacuum.c:2116
HashValueFunc hash
Definition: hsearch.h:78
#define HASH_FUNCTION
Definition: hsearch.h:98
float4 stadistinct
Definition: vacuum.h:145

◆ lexeme_compare()

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

Definition at line 498 of file ts_typanalyze.c.

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

Referenced by lexeme_match(), and trackitem_compare_lexemes().

499 {
500  const LexemeHashKey *d1 = (const LexemeHashKey *) key1;
501  const LexemeHashKey *d2 = (const LexemeHashKey *) key2;
502 
503  /* First, compare by length */
504  if (d1->length > d2->length)
505  return 1;
506  else if (d1->length < d2->length)
507  return -1;
508  /* Lengths are equal, do a byte-by-byte comparison */
509  return strncmp(d1->lexeme, d2->lexeme, d1->length);
510 }

◆ lexeme_hash()

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

Definition at line 476 of file ts_typanalyze.c.

References DatumGetUInt32, hash_any(), LexemeHashKey::length, and LexemeHashKey::lexeme.

Referenced by compute_tsvector_stats().

477 {
478  const LexemeHashKey *l = (const LexemeHashKey *) key;
479 
480  return DatumGetUInt32(hash_any((const unsigned char *) l->lexeme,
481  l->length));
482 }
#define DatumGetUInt32(X)
Definition: postgres.h:530
static Datum hash_any(const unsigned char *k, int keylen)
Definition: hashfn.h:31

◆ lexeme_match()

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

Definition at line 488 of file ts_typanalyze.c.

References lexeme_compare().

Referenced by compute_tsvector_stats().

489 {
490  /* The keysize parameter is superfluous, the keys store their lengths */
491  return lexeme_compare(key1, key2);
492 }
static int lexeme_compare(const void *key1, const void *key2)

◆ prune_lexemes_hashtable()

static void prune_lexemes_hashtable ( HTAB lexemes_tab,
int  b_current 
)
static

Definition at line 451 of file ts_typanalyze.c.

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().

452 {
453  HASH_SEQ_STATUS scan_status;
454  TrackItem *item;
455 
456  hash_seq_init(&scan_status, lexemes_tab);
457  while ((item = (TrackItem *) hash_seq_search(&scan_status)) != NULL)
458  {
459  if (item->frequency + item->delta <= b_current)
460  {
461  char *lexeme = item->key.lexeme;
462 
463  if (hash_search(lexemes_tab, (const void *) &item->key,
464  HASH_REMOVE, NULL) == NULL)
465  elog(ERROR, "hash table corrupted");
466  pfree(lexeme);
467  }
468  }
469 }
void * hash_search(HTAB *hashp, const void *keyPtr, HASHACTION action, bool *foundPtr)
Definition: dynahash.c:954
void pfree(void *pointer)
Definition: mcxt.c:1169
#define ERROR
Definition: elog.h:46
LexemeHashKey key
Definition: ts_typanalyze.c:34
void * hash_seq_search(HASH_SEQ_STATUS *status)
Definition: dynahash.c:1436
void hash_seq_init(HASH_SEQ_STATUS *status, HTAB *hashp)
Definition: dynahash.c:1426
#define elog(elevel,...)
Definition: elog.h:232

◆ trackitem_compare_frequencies_desc()

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

Definition at line 516 of file ts_typanalyze.c.

References TrackItem::frequency.

Referenced by compute_tsvector_stats().

517 {
518  const TrackItem *const *t1 = (const TrackItem *const *) e1;
519  const TrackItem *const *t2 = (const TrackItem *const *) e2;
520 
521  return (*t2)->frequency - (*t1)->frequency;
522 }

◆ trackitem_compare_lexemes()

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

Definition at line 528 of file ts_typanalyze.c.

References lexeme_compare().

Referenced by compute_tsvector_stats().

529 {
530  const TrackItem *const *t1 = (const TrackItem *const *) e1;
531  const TrackItem *const *t2 = (const TrackItem *const *) e2;
532 
533  return lexeme_compare(&(*t1)->key, &(*t2)->key);
534 }
static int lexeme_compare(const void *key1, const void *key2)

◆ ts_typanalyze()

Datum ts_typanalyze ( PG_FUNCTION_ARGS  )

Definition at line 55 of file ts_typanalyze.c.

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

56 {
58  Form_pg_attribute attr = stats->attr;
59 
60  /* If the attstattarget column is negative, use the default value */
61  /* NB: it is okay to scribble on stats->attr since it's a copy */
62  if (attr->attstattarget < 0)
63  attr->attstattarget = default_statistics_target;
64 
66  /* see comment about the choice of minrows in commands/analyze.c */
67  stats->minrows = 300 * attr->attstattarget;
68 
69  PG_RETURN_BOOL(true);
70 }
int minrows
Definition: vacuum.h:135
#define PG_GETARG_POINTER(n)
Definition: fmgr.h:276
Form_pg_attribute attr
Definition: vacuum.h:123
static void compute_tsvector_stats(VacAttrStats *stats, AnalyzeAttrFetchFunc fetchfunc, int samplerows, double totalrows)
FormData_pg_attribute * Form_pg_attribute
Definition: pg_attribute.h:203
#define PG_RETURN_BOOL(x)
Definition: fmgr.h:359
AnalyzeAttrComputeStatsFunc compute_stats
Definition: vacuum.h:134
int default_statistics_target
Definition: analyze.c:82