Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * ts_typanalyze.c
4 : * functions for gathering statistics from tsvector columns
5 : *
6 : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
7 : *
8 : *
9 : * IDENTIFICATION
10 : * src/backend/tsearch/ts_typanalyze.c
11 : *
12 : *-------------------------------------------------------------------------
13 : */
14 : #include "postgres.h"
15 :
16 : #include "catalog/pg_collation.h"
17 : #include "catalog/pg_operator.h"
18 : #include "commands/vacuum.h"
19 : #include "common/hashfn.h"
20 : #include "tsearch/ts_type.h"
21 : #include "utils/builtins.h"
22 : #include "varatt.h"
23 :
24 :
25 : /* A hash key for lexemes */
26 : typedef struct
27 : {
28 : char *lexeme; /* lexeme (not NULL terminated!) */
29 : int length; /* its length in bytes */
30 : } LexemeHashKey;
31 :
32 : /* A hash table entry for the Lossy Counting algorithm */
33 : typedef struct
34 : {
35 : LexemeHashKey key; /* This is 'e' from the LC algorithm. */
36 : int frequency; /* This is 'f'. */
37 : int delta; /* And this is 'delta'. */
38 : } TrackItem;
39 :
40 : static void compute_tsvector_stats(VacAttrStats *stats,
41 : AnalyzeAttrFetchFunc fetchfunc,
42 : int samplerows,
43 : double totalrows);
44 : static void prune_lexemes_hashtable(HTAB *lexemes_tab, int b_current);
45 : static uint32 lexeme_hash(const void *key, Size keysize);
46 : static int lexeme_match(const void *key1, const void *key2, Size keysize);
47 : static int lexeme_compare(const void *key1, const void *key2);
48 : static int trackitem_compare_frequencies_desc(const void *e1, const void *e2,
49 : void *arg);
50 : static int trackitem_compare_lexemes(const void *e1, const void *e2,
51 : void *arg);
52 :
53 :
54 : /*
55 : * ts_typanalyze -- a custom typanalyze function for tsvector columns
56 : */
57 : Datum
58 6 : ts_typanalyze(PG_FUNCTION_ARGS)
59 : {
60 6 : VacAttrStats *stats = (VacAttrStats *) PG_GETARG_POINTER(0);
61 :
62 : /* If the attstattarget column is negative, use the default value */
63 6 : if (stats->attstattarget < 0)
64 6 : stats->attstattarget = default_statistics_target;
65 :
66 6 : stats->compute_stats = compute_tsvector_stats;
67 : /* see comment about the choice of minrows in commands/analyze.c */
68 6 : stats->minrows = 300 * stats->attstattarget;
69 :
70 6 : PG_RETURN_BOOL(true);
71 : }
72 :
73 : /*
74 : * compute_tsvector_stats() -- compute statistics for a tsvector column
75 : *
76 : * This functions computes statistics that are useful for determining @@
77 : * operations' selectivity, along with the fraction of non-null rows and
78 : * average width.
79 : *
80 : * Instead of finding the most common values, as we do for most datatypes,
81 : * we're looking for the most common lexemes. This is more useful, because
82 : * there most probably won't be any two rows with the same tsvector and thus
83 : * the notion of a MCV is a bit bogus with this datatype. With a list of the
84 : * most common lexemes we can do a better job at figuring out @@ selectivity.
85 : *
86 : * For the same reasons we assume that tsvector columns are unique when
87 : * determining the number of distinct values.
88 : *
89 : * The algorithm used is Lossy Counting, as proposed in the paper "Approximate
90 : * frequency counts over data streams" by G. S. Manku and R. Motwani, in
91 : * Proceedings of the 28th International Conference on Very Large Data Bases,
92 : * Hong Kong, China, August 2002, section 4.2. The paper is available at
93 : * http://www.vldb.org/conf/2002/S10P03.pdf
94 : *
95 : * The Lossy Counting (aka LC) algorithm goes like this:
96 : * Let s be the threshold frequency for an item (the minimum frequency we
97 : * are interested in) and epsilon the error margin for the frequency. Let D
98 : * be a set of triples (e, f, delta), where e is an element value, f is that
99 : * element's frequency (actually, its current occurrence count) and delta is
100 : * the maximum error in f. We start with D empty and process the elements in
101 : * batches of size w. (The batch size is also known as "bucket size" and is
102 : * equal to 1/epsilon.) Let the current batch number be b_current, starting
103 : * with 1. For each element e we either increment its f count, if it's
104 : * already in D, or insert a new triple into D with values (e, 1, b_current
105 : * - 1). After processing each batch we prune D, by removing from it all
106 : * elements with f + delta <= b_current. After the algorithm finishes we
107 : * suppress all elements from D that do not satisfy f >= (s - epsilon) * N,
108 : * where N is the total number of elements in the input. We emit the
109 : * remaining elements with estimated frequency f/N. The LC paper proves
110 : * that this algorithm finds all elements with true frequency at least s,
111 : * and that no frequency is overestimated or is underestimated by more than
112 : * epsilon. Furthermore, given reasonable assumptions about the input
113 : * distribution, the required table size is no more than about 7 times w.
114 : *
115 : * We set s to be the estimated frequency of the K'th word in a natural
116 : * language's frequency table, where K is the target number of entries in
117 : * the MCELEM array plus an arbitrary constant, meant to reflect the fact
118 : * that the most common words in any language would usually be stopwords
119 : * so we will not actually see them in the input. We assume that the
120 : * distribution of word frequencies (including the stopwords) follows Zipf's
121 : * law with an exponent of 1.
122 : *
123 : * Assuming Zipfian distribution, the frequency of the K'th word is equal
124 : * to 1/(K * H(W)) where H(n) is 1/2 + 1/3 + ... + 1/n and W is the number of
125 : * words in the language. Putting W as one million, we get roughly 0.07/K.
126 : * Assuming top 10 words are stopwords gives s = 0.07/(K + 10). We set
127 : * epsilon = s/10, which gives bucket width w = (K + 10)/0.007 and
128 : * maximum expected hashtable size of about 1000 * (K + 10).
129 : *
130 : * Note: in the above discussion, s, epsilon, and f/N are in terms of a
131 : * lexeme's frequency as a fraction of all lexemes seen in the input.
132 : * However, what we actually want to store in the finished pg_statistic
133 : * entry is each lexeme's frequency as a fraction of all rows that it occurs
134 : * in. Assuming that the input tsvectors are correctly constructed, no
135 : * lexeme occurs more than once per tsvector, so the final count f is a
136 : * correct estimate of the number of input tsvectors it occurs in, and we
137 : * need only change the divisor from N to nonnull_cnt to get the number we
138 : * want.
139 : */
140 : static void
141 6 : compute_tsvector_stats(VacAttrStats *stats,
142 : AnalyzeAttrFetchFunc fetchfunc,
143 : int samplerows,
144 : double totalrows)
145 : {
146 : int num_mcelem;
147 6 : int null_cnt = 0;
148 6 : 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;
162 : LexemeHashKey hash_key;
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 6 : 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 6 : 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 6 : hash_ctl.keysize = sizeof(LexemeHashKey);
184 6 : hash_ctl.entrysize = sizeof(TrackItem);
185 6 : hash_ctl.hash = lexeme_hash;
186 6 : hash_ctl.match = lexeme_match;
187 6 : hash_ctl.hcxt = CurrentMemoryContext;
188 6 : lexemes_tab = hash_create("Analyzed lexemes table",
189 : num_mcelem,
190 : &hash_ctl,
191 : HASH_ELEM | HASH_FUNCTION | HASH_COMPARE | HASH_CONTEXT);
192 :
193 : /* Initialize counters. */
194 6 : b_current = 1;
195 6 : lexeme_no = 0;
196 :
197 : /* Loop over the tsvectors. */
198 3054 : 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 :
207 3048 : vacuum_delay_point();
208 :
209 3048 : value = fetchfunc(stats, vector_no, &isnull);
210 :
211 : /*
212 : * Check for null/nonnull.
213 : */
214 3048 : if (isnull)
215 : {
216 0 : null_cnt++;
217 0 : 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 3048 : total_width += VARSIZE_ANY(DatumGetPointer(value));
227 :
228 : /*
229 : * Now detoast the tsvector if needed.
230 : */
231 3048 : vector = DatumGetTSVector(value);
232 :
233 : /*
234 : * We loop through the lexemes in the tsvector and add them to our
235 : * tracking hashtable.
236 : */
237 3048 : lexemesptr = STRPTR(vector);
238 3048 : curentryptr = ARRPTR(vector);
239 175848 : 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 172800 : hash_key.lexeme = lexemesptr + curentryptr->pos;
251 172800 : hash_key.length = curentryptr->len;
252 :
253 : /* Lookup current lexeme in hashtable, adding it if new */
254 172800 : item = (TrackItem *) hash_search(lexemes_tab,
255 : &hash_key,
256 : HASH_ENTER, &found);
257 :
258 172800 : if (found)
259 : {
260 : /* The lexeme is already on the tracking list */
261 165954 : item->frequency++;
262 : }
263 : else
264 : {
265 : /* Initialize new tracking list element */
266 6846 : item->frequency = 1;
267 6846 : item->delta = b_current - 1;
268 :
269 6846 : item->key.lexeme = palloc(hash_key.length);
270 6846 : memcpy(item->key.lexeme, hash_key.lexeme, hash_key.length);
271 : }
272 :
273 : /* lexeme_no is the number of elements processed (ie N) */
274 172800 : lexeme_no++;
275 :
276 : /* We prune the D structure after processing each bucket */
277 172800 : if (lexeme_no % bucket_width == 0)
278 : {
279 0 : prune_lexemes_hashtable(lexemes_tab, b_current);
280 0 : b_current++;
281 : }
282 :
283 : /* Advance to the next WordEntry in the tsvector */
284 172800 : curentryptr++;
285 : }
286 :
287 : /* If the vector was toasted, free the detoasted copy. */
288 3048 : if (TSVectorGetDatum(vector) != value)
289 384 : pfree(vector);
290 : }
291 :
292 : /* We can only compute real stats if we found some non-null values. */
293 6 : if (null_cnt < samplerows)
294 : {
295 6 : 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 6 : stats->stats_valid = true;
305 : /* Do the simple null-frac and average width stats */
306 6 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
307 6 : stats->stawidth = total_width / (double) nonnull_cnt;
308 :
309 : /* Assume it's a unique column (see notes above) */
310 6 : 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 6 : cutoff_freq = 9 * lexeme_no / bucket_width;
321 :
322 6 : i = hash_get_num_entries(lexemes_tab); /* surely enough space */
323 6 : sort_table = (TrackItem **) palloc(sizeof(TrackItem *) * i);
324 :
325 6 : hash_seq_init(&scan_status, lexemes_tab);
326 6 : track_len = 0;
327 6 : minfreq = lexeme_no;
328 6 : maxfreq = 0;
329 6852 : while ((item = (TrackItem *) hash_seq_search(&scan_status)) != NULL)
330 : {
331 6846 : if (item->frequency > cutoff_freq)
332 : {
333 6318 : sort_table[track_len++] = item;
334 6318 : minfreq = Min(minfreq, item->frequency);
335 6318 : maxfreq = Max(maxfreq, item->frequency);
336 : }
337 : }
338 : Assert(track_len <= i);
339 :
340 : /* emit some statistics for debug purposes */
341 6 : 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 6 : if (num_mcelem < track_len)
351 : {
352 6 : qsort_interruptible(sort_table, track_len, sizeof(TrackItem *),
353 : trackitem_compare_frequencies_desc, NULL);
354 : /* reset minfreq to the smallest frequency we're keeping */
355 6 : minfreq = sort_table[num_mcelem - 1]->frequency;
356 : }
357 : else
358 0 : num_mcelem = track_len;
359 :
360 : /* Generate MCELEM slot entry */
361 6 : 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 6 : qsort_interruptible(sort_table, num_mcelem, sizeof(TrackItem *),
382 : trackitem_compare_lexemes, NULL);
383 :
384 : /* Must copy the target values into anl_context */
385 6 : 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 6 : mcelem_values = (Datum *) palloc(num_mcelem * sizeof(Datum));
399 6 : 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 6006 : for (i = 0; i < num_mcelem; i++)
406 : {
407 6000 : TrackItem *titem = sort_table[i];
408 :
409 12000 : mcelem_values[i] =
410 6000 : PointerGetDatum(cstring_to_text_with_len(titem->key.lexeme,
411 : titem->key.length));
412 6000 : mcelem_freqs[i] = (double) titem->frequency / (double) nonnull_cnt;
413 : }
414 6 : mcelem_freqs[i++] = (double) minfreq / (double) nonnull_cnt;
415 6 : mcelem_freqs[i] = (double) maxfreq / (double) nonnull_cnt;
416 6 : MemoryContextSwitchTo(old_context);
417 :
418 6 : stats->stakind[0] = STATISTIC_KIND_MCELEM;
419 6 : stats->staop[0] = TextEqualOperator;
420 6 : stats->stacoll[0] = DEFAULT_COLLATION_OID;
421 6 : stats->stanumbers[0] = mcelem_freqs;
422 : /* See above comment about two extra frequency fields */
423 6 : stats->numnumbers[0] = num_mcelem + 2;
424 6 : stats->stavalues[0] = mcelem_values;
425 6 : stats->numvalues[0] = num_mcelem;
426 : /* We are storing text values */
427 6 : stats->statypid[0] = TEXTOID;
428 6 : stats->statyplen[0] = -1; /* typlen, -1 for varlena */
429 6 : stats->statypbyval[0] = false;
430 6 : stats->statypalign[0] = 'i';
431 : }
432 : }
433 : else
434 : {
435 : /* We found only nulls; assume the column is entirely null */
436 0 : stats->stats_valid = true;
437 0 : stats->stanullfrac = 1.0;
438 0 : stats->stawidth = 0; /* "unknown" */
439 0 : 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 6 : }
447 :
448 : /*
449 : * A function to prune the D structure from the Lossy Counting algorithm.
450 : * Consult compute_tsvector_stats() for wider explanation.
451 : */
452 : static void
453 0 : prune_lexemes_hashtable(HTAB *lexemes_tab, int b_current)
454 : {
455 : HASH_SEQ_STATUS scan_status;
456 : TrackItem *item;
457 :
458 0 : hash_seq_init(&scan_status, lexemes_tab);
459 0 : while ((item = (TrackItem *) hash_seq_search(&scan_status)) != NULL)
460 : {
461 0 : if (item->frequency + item->delta <= b_current)
462 : {
463 0 : char *lexeme = item->key.lexeme;
464 :
465 0 : if (hash_search(lexemes_tab, &item->key,
466 : HASH_REMOVE, NULL) == NULL)
467 0 : elog(ERROR, "hash table corrupted");
468 0 : pfree(lexeme);
469 : }
470 : }
471 0 : }
472 :
473 : /*
474 : * Hash functions for lexemes. They are strings, but not NULL terminated,
475 : * so we need a special hash function.
476 : */
477 : static uint32
478 172800 : lexeme_hash(const void *key, Size keysize)
479 : {
480 172800 : const LexemeHashKey *l = (const LexemeHashKey *) key;
481 :
482 172800 : return DatumGetUInt32(hash_any((const unsigned char *) l->lexeme,
483 : l->length));
484 : }
485 :
486 : /*
487 : * Matching function for lexemes, to be used in hashtable lookups.
488 : */
489 : static int
490 165954 : lexeme_match(const void *key1, const void *key2, Size keysize)
491 : {
492 : /* The keysize parameter is superfluous, the keys store their lengths */
493 165954 : return lexeme_compare(key1, key2);
494 : }
495 :
496 : /*
497 : * Comparison function for lexemes.
498 : */
499 : static int
500 227028 : lexeme_compare(const void *key1, const void *key2)
501 : {
502 227028 : const LexemeHashKey *d1 = (const LexemeHashKey *) key1;
503 227028 : const LexemeHashKey *d2 = (const LexemeHashKey *) key2;
504 :
505 : /* First, compare by length */
506 227028 : if (d1->length > d2->length)
507 0 : return 1;
508 227028 : else if (d1->length < d2->length)
509 0 : return -1;
510 : /* Lengths are equal, do a byte-by-byte comparison */
511 227028 : return strncmp(d1->lexeme, d2->lexeme, d1->length);
512 : }
513 :
514 : /*
515 : * Comparator for sorting TrackItems on frequencies (descending sort)
516 : */
517 : static int
518 38460 : trackitem_compare_frequencies_desc(const void *e1, const void *e2, void *arg)
519 : {
520 38460 : const TrackItem *const *t1 = (const TrackItem *const *) e1;
521 38460 : const TrackItem *const *t2 = (const TrackItem *const *) e2;
522 :
523 38460 : return (*t2)->frequency - (*t1)->frequency;
524 : }
525 :
526 : /*
527 : * Comparator for sorting TrackItems on lexemes
528 : */
529 : static int
530 61074 : trackitem_compare_lexemes(const void *e1, const void *e2, void *arg)
531 : {
532 61074 : const TrackItem *const *t1 = (const TrackItem *const *) e1;
533 61074 : const TrackItem *const *t2 = (const TrackItem *const *) e2;
534 :
535 61074 : return lexeme_compare(&(*t1)->key, &(*t2)->key);
536 : }
|
