1 | /*------------------------------------------------------------------------- |
2 | * |
3 | * ts_typanalyze.c |
4 | * functions for gathering statistics from tsvector columns |
5 | * |
6 | * Portions Copyright (c) 1996-2019, 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 "tsearch/ts_type.h" |
20 | #include "utils/builtins.h" |
21 | #include "utils/hashutils.h" |
22 | |
23 | |
24 | /* A hash key for lexemes */ |
25 | typedef struct |
26 | { |
27 | char *lexeme; /* lexeme (not NULL terminated!) */ |
28 | int length; /* its length in bytes */ |
29 | } LexemeHashKey; |
30 | |
31 | /* A hash table entry for the Lossy Counting algorithm */ |
32 | typedef struct |
33 | { |
34 | LexemeHashKey key; /* This is 'e' from the LC algorithm. */ |
35 | int frequency; /* This is 'f'. */ |
36 | int delta; /* And this is 'delta'. */ |
37 | } TrackItem; |
38 | |
39 | static void compute_tsvector_stats(VacAttrStats *stats, |
40 | AnalyzeAttrFetchFunc fetchfunc, |
41 | int samplerows, |
42 | double totalrows); |
43 | static void prune_lexemes_hashtable(HTAB *lexemes_tab, int b_current); |
44 | static uint32 lexeme_hash(const void *key, Size keysize); |
45 | static int lexeme_match(const void *key1, const void *key2, Size keysize); |
46 | static int lexeme_compare(const void *key1, const void *key2); |
47 | static int trackitem_compare_frequencies_desc(const void *e1, const void *e2); |
48 | static int trackitem_compare_lexemes(const void *e1, const void *e2); |
49 | |
50 | |
51 | /* |
52 | * ts_typanalyze -- a custom typanalyze function for tsvector columns |
53 | */ |
54 | Datum |
55 | ts_typanalyze(PG_FUNCTION_ARGS) |
56 | { |
57 | VacAttrStats *stats = (VacAttrStats *) PG_GETARG_POINTER(0); |
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 | |
65 | stats->compute_stats = compute_tsvector_stats; |
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 | } |
71 | |
72 | /* |
73 | * compute_tsvector_stats() -- compute statistics for a tsvector column |
74 | * |
75 | * This functions computes statistics that are useful for determining @@ |
76 | * operations' selectivity, along with the fraction of non-null rows and |
77 | * average width. |
78 | * |
79 | * Instead of finding the most common values, as we do for most datatypes, |
80 | * we're looking for the most common lexemes. This is more useful, because |
81 | * there most probably won't be any two rows with the same tsvector and thus |
82 | * the notion of a MCV is a bit bogus with this datatype. With a list of the |
83 | * most common lexemes we can do a better job at figuring out @@ selectivity. |
84 | * |
85 | * For the same reasons we assume that tsvector columns are unique when |
86 | * determining the number of distinct values. |
87 | * |
88 | * The algorithm used is Lossy Counting, as proposed in the paper "Approximate |
89 | * frequency counts over data streams" by G. S. Manku and R. Motwani, in |
90 | * Proceedings of the 28th International Conference on Very Large Data Bases, |
91 | * Hong Kong, China, August 2002, section 4.2. The paper is available at |
92 | * http://www.vldb.org/conf/2002/S10P03.pdf |
93 | * |
94 | * The Lossy Counting (aka LC) algorithm goes like this: |
95 | * Let s be the threshold frequency for an item (the minimum frequency we |
96 | * are interested in) and epsilon the error margin for the frequency. Let D |
97 | * be a set of triples (e, f, delta), where e is an element value, f is that |
98 | * element's frequency (actually, its current occurrence count) and delta is |
99 | * the maximum error in f. We start with D empty and process the elements in |
100 | * batches of size w. (The batch size is also known as "bucket size" and is |
101 | * equal to 1/epsilon.) Let the current batch number be b_current, starting |
102 | * with 1. For each element e we either increment its f count, if it's |
103 | * already in D, or insert a new triple into D with values (e, 1, b_current |
104 | * - 1). After processing each batch we prune D, by removing from it all |
105 | * elements with f + delta <= b_current. After the algorithm finishes we |
106 | * suppress all elements from D that do not satisfy f >= (s - epsilon) * N, |
107 | * where N is the total number of elements in the input. We emit the |
108 | * remaining elements with estimated frequency f/N. The LC paper proves |
109 | * that this algorithm finds all elements with true frequency at least s, |
110 | * and that no frequency is overestimated or is underestimated by more than |
111 | * epsilon. Furthermore, given reasonable assumptions about the input |
112 | * distribution, the required table size is no more than about 7 times w. |
113 | * |
114 | * We set s to be the estimated frequency of the K'th word in a natural |
115 | * language's frequency table, where K is the target number of entries in |
116 | * the MCELEM array plus an arbitrary constant, meant to reflect the fact |
117 | * that the most common words in any language would usually be stopwords |
118 | * so we will not actually see them in the input. We assume that the |
119 | * distribution of word frequencies (including the stopwords) follows Zipf's |
120 | * law with an exponent of 1. |
121 | * |
122 | * Assuming Zipfian distribution, the frequency of the K'th word is equal |
123 | * to 1/(K * H(W)) where H(n) is 1/2 + 1/3 + ... + 1/n and W is the number of |
124 | * words in the language. Putting W as one million, we get roughly 0.07/K. |
125 | * Assuming top 10 words are stopwords gives s = 0.07/(K + 10). We set |
126 | * epsilon = s/10, which gives bucket width w = (K + 10)/0.007 and |
127 | * maximum expected hashtable size of about 1000 * (K + 10). |
128 | * |
129 | * Note: in the above discussion, s, epsilon, and f/N are in terms of a |
130 | * lexeme's frequency as a fraction of all lexemes seen in the input. |
131 | * However, what we actually want to store in the finished pg_statistic |
132 | * entry is each lexeme's frequency as a fraction of all rows that it occurs |
133 | * in. Assuming that the input tsvectors are correctly constructed, no |
134 | * lexeme occurs more than once per tsvector, so the final count f is a |
135 | * correct estimate of the number of input tsvectors it occurs in, and we |
136 | * need only change the divisor from N to nonnull_cnt to get the number we |
137 | * want. |
138 | */ |
139 | static void |
140 | compute_tsvector_stats(VacAttrStats *stats, |
141 | AnalyzeAttrFetchFunc fetchfunc, |
142 | int samplerows, |
143 | double totalrows) |
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; |
161 | LexemeHashKey hash_key; |
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 | MemSet(&hash_ctl, 0, sizeof(hash_ctl)); |
184 | hash_ctl.keysize = sizeof(LexemeHashKey); |
185 | hash_ctl.entrysize = sizeof(TrackItem); |
186 | hash_ctl.hash = lexeme_hash; |
187 | hash_ctl.match = lexeme_match; |
188 | hash_ctl.hcxt = CurrentMemoryContext; |
189 | lexemes_tab = hash_create("Analyzed lexemes table" , |
190 | num_mcelem, |
191 | &hash_ctl, |
192 | HASH_ELEM | HASH_FUNCTION | HASH_COMPARE | HASH_CONTEXT); |
193 | |
194 | /* Initialize counters. */ |
195 | b_current = 1; |
196 | lexeme_no = 0; |
197 | |
198 | /* Loop over the tsvectors. */ |
199 | for (vector_no = 0; vector_no < samplerows; vector_no++) |
200 | { |
201 | Datum value; |
202 | bool isnull; |
203 | TSVector vector; |
204 | WordEntry *curentryptr; |
205 | char *lexemesptr; |
206 | int j; |
207 | |
208 | vacuum_delay_point(); |
209 | |
210 | value = fetchfunc(stats, vector_no, &isnull); |
211 | |
212 | /* |
213 | * Check for null/nonnull. |
214 | */ |
215 | if (isnull) |
216 | { |
217 | null_cnt++; |
218 | continue; |
219 | } |
220 | |
221 | /* |
222 | * Add up widths for average-width calculation. Since it's a |
223 | * tsvector, we know it's varlena. As in the regular |
224 | * compute_minimal_stats function, we use the toasted width for this |
225 | * calculation. |
226 | */ |
227 | total_width += VARSIZE_ANY(DatumGetPointer(value)); |
228 | |
229 | /* |
230 | * Now detoast the tsvector if needed. |
231 | */ |
232 | vector = DatumGetTSVector(value); |
233 | |
234 | /* |
235 | * We loop through the lexemes in the tsvector and add them to our |
236 | * tracking hashtable. |
237 | */ |
238 | lexemesptr = STRPTR(vector); |
239 | curentryptr = ARRPTR(vector); |
240 | for (j = 0; j < vector->size; j++) |
241 | { |
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 | (const void *) &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 | int track_len; |
299 | int cutoff_freq; |
300 | int minfreq, |
301 | maxfreq; |
302 | |
303 | stats->stats_valid = true; |
304 | /* Do the simple null-frac and average width stats */ |
305 | stats->stanullfrac = (double) null_cnt / (double) samplerows; |
306 | stats->stawidth = total_width / (double) nonnull_cnt; |
307 | |
308 | /* Assume it's a unique column (see notes above) */ |
309 | stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac); |
310 | |
311 | /* |
312 | * Construct an array of the interesting hashtable items, that is, |
313 | * those meeting the cutoff frequency (s - epsilon)*N. Also identify |
314 | * the minimum and maximum frequencies among these items. |
315 | * |
316 | * Since epsilon = s/10 and bucket_width = 1/epsilon, the cutoff |
317 | * frequency is 9*N / bucket_width. |
318 | */ |
319 | cutoff_freq = 9 * lexeme_no / bucket_width; |
320 | |
321 | i = hash_get_num_entries(lexemes_tab); /* surely enough space */ |
322 | sort_table = (TrackItem **) palloc(sizeof(TrackItem *) * i); |
323 | |
324 | hash_seq_init(&scan_status, lexemes_tab); |
325 | track_len = 0; |
326 | minfreq = lexeme_no; |
327 | maxfreq = 0; |
328 | while ((item = (TrackItem *) hash_seq_search(&scan_status)) != NULL) |
329 | { |
330 | if (item->frequency > cutoff_freq) |
331 | { |
332 | sort_table[track_len++] = item; |
333 | minfreq = Min(minfreq, item->frequency); |
334 | maxfreq = Max(maxfreq, item->frequency); |
335 | } |
336 | } |
337 | Assert(track_len <= i); |
338 | |
339 | /* emit some statistics for debug purposes */ |
340 | elog(DEBUG3, "tsvector_stats: target # mces = %d, bucket width = %d, " |
341 | "# lexemes = %d, hashtable size = %d, usable entries = %d" , |
342 | num_mcelem, bucket_width, lexeme_no, i, track_len); |
343 | |
344 | /* |
345 | * If we obtained more lexemes than we really want, get rid of those |
346 | * with least frequencies. The easiest way is to qsort the array into |
347 | * descending frequency order and truncate the array. |
348 | */ |
349 | if (num_mcelem < track_len) |
350 | { |
351 | qsort(sort_table, track_len, sizeof(TrackItem *), |
352 | trackitem_compare_frequencies_desc); |
353 | /* reset minfreq to the smallest frequency we're keeping */ |
354 | minfreq = sort_table[num_mcelem - 1]->frequency; |
355 | } |
356 | else |
357 | num_mcelem = track_len; |
358 | |
359 | /* Generate MCELEM slot entry */ |
360 | if (num_mcelem > 0) |
361 | { |
362 | MemoryContext old_context; |
363 | Datum *mcelem_values; |
364 | float4 *mcelem_freqs; |
365 | |
366 | /* |
367 | * We want to store statistics sorted on the lexeme value using |
368 | * first length, then byte-for-byte comparison. The reason for |
369 | * doing length comparison first is that we don't care about the |
370 | * ordering so long as it's consistent, and comparing lengths |
371 | * first gives us a chance to avoid a strncmp() call. |
372 | * |
373 | * This is different from what we do with scalar statistics -- |
374 | * they get sorted on frequencies. The rationale is that we |
375 | * usually search through most common elements looking for a |
376 | * specific value, so we can grab its frequency. When values are |
377 | * presorted we can employ binary search for that. See |
378 | * ts_selfuncs.c for a real usage scenario. |
379 | */ |
380 | qsort(sort_table, num_mcelem, sizeof(TrackItem *), |
381 | trackitem_compare_lexemes); |
382 | |
383 | /* Must copy the target values into anl_context */ |
384 | old_context = MemoryContextSwitchTo(stats->anl_context); |
385 | |
386 | /* |
387 | * We sorted statistics on the lexeme value, but we want to be |
388 | * able to find out the minimal and maximal frequency without |
389 | * going through all the values. We keep those two extra |
390 | * frequencies in two extra cells in mcelem_freqs. |
391 | * |
392 | * (Note: the MCELEM statistics slot definition allows for a third |
393 | * extra number containing the frequency of nulls, but we don't |
394 | * create that for a tsvector column, since null elements aren't |
395 | * possible.) |
396 | */ |
397 | mcelem_values = (Datum *) palloc(num_mcelem * sizeof(Datum)); |
398 | mcelem_freqs = (float4 *) palloc((num_mcelem + 2) * sizeof(float4)); |
399 | |
400 | /* |
401 | * See comments above about use of nonnull_cnt as the divisor for |
402 | * the final frequency estimates. |
403 | */ |
404 | for (i = 0; i < num_mcelem; i++) |
405 | { |
406 | TrackItem *item = sort_table[i]; |
407 | |
408 | mcelem_values[i] = |
409 | PointerGetDatum(cstring_to_text_with_len(item->key.lexeme, |
410 | item->key.length)); |
411 | mcelem_freqs[i] = (double) item->frequency / (double) nonnull_cnt; |
412 | } |
413 | mcelem_freqs[i++] = (double) minfreq / (double) nonnull_cnt; |
414 | mcelem_freqs[i] = (double) maxfreq / (double) nonnull_cnt; |
415 | MemoryContextSwitchTo(old_context); |
416 | |
417 | stats->stakind[0] = STATISTIC_KIND_MCELEM; |
418 | stats->staop[0] = TextEqualOperator; |
419 | stats->stacoll[0] = DEFAULT_COLLATION_OID; |
420 | stats->stanumbers[0] = mcelem_freqs; |
421 | /* See above comment about two extra frequency fields */ |
422 | stats->numnumbers[0] = num_mcelem + 2; |
423 | stats->stavalues[0] = mcelem_values; |
424 | stats->numvalues[0] = num_mcelem; |
425 | /* We are storing text values */ |
426 | stats->statypid[0] = TEXTOID; |
427 | stats->statyplen[0] = -1; /* typlen, -1 for varlena */ |
428 | stats->statypbyval[0] = false; |
429 | stats->statypalign[0] = 'i'; |
430 | } |
431 | } |
432 | else |
433 | { |
434 | /* We found only nulls; assume the column is entirely null */ |
435 | stats->stats_valid = true; |
436 | stats->stanullfrac = 1.0; |
437 | stats->stawidth = 0; /* "unknown" */ |
438 | stats->stadistinct = 0.0; /* "unknown" */ |
439 | } |
440 | |
441 | /* |
442 | * We don't need to bother cleaning up any of our temporary palloc's. The |
443 | * hashtable should also go away, as it used a child memory context. |
444 | */ |
445 | } |
446 | |
447 | /* |
448 | * A function to prune the D structure from the Lossy Counting algorithm. |
449 | * Consult compute_tsvector_stats() for wider explanation. |
450 | */ |
451 | static void |
452 | prune_lexemes_hashtable(HTAB *lexemes_tab, int b_current) |
453 | { |
454 | HASH_SEQ_STATUS scan_status; |
455 | TrackItem *item; |
456 | |
457 | hash_seq_init(&scan_status, lexemes_tab); |
458 | while ((item = (TrackItem *) hash_seq_search(&scan_status)) != NULL) |
459 | { |
460 | if (item->frequency + item->delta <= b_current) |
461 | { |
462 | char *lexeme = item->key.lexeme; |
463 | |
464 | if (hash_search(lexemes_tab, (const void *) &item->key, |
465 | HASH_REMOVE, NULL) == NULL) |
466 | elog(ERROR, "hash table corrupted" ); |
467 | pfree(lexeme); |
468 | } |
469 | } |
470 | } |
471 | |
472 | /* |
473 | * Hash functions for lexemes. They are strings, but not NULL terminated, |
474 | * so we need a special hash function. |
475 | */ |
476 | static uint32 |
477 | lexeme_hash(const void *key, Size keysize) |
478 | { |
479 | const LexemeHashKey *l = (const LexemeHashKey *) key; |
480 | |
481 | return DatumGetUInt32(hash_any((const unsigned char *) l->lexeme, |
482 | l->length)); |
483 | } |
484 | |
485 | /* |
486 | * Matching function for lexemes, to be used in hashtable lookups. |
487 | */ |
488 | static int |
489 | lexeme_match(const void *key1, const void *key2, Size keysize) |
490 | { |
491 | /* The keysize parameter is superfluous, the keys store their lengths */ |
492 | return lexeme_compare(key1, key2); |
493 | } |
494 | |
495 | /* |
496 | * Comparison function for lexemes. |
497 | */ |
498 | static int |
499 | lexeme_compare(const void *key1, const void *key2) |
500 | { |
501 | const LexemeHashKey *d1 = (const LexemeHashKey *) key1; |
502 | const LexemeHashKey *d2 = (const LexemeHashKey *) key2; |
503 | |
504 | /* First, compare by length */ |
505 | if (d1->length > d2->length) |
506 | return 1; |
507 | else if (d1->length < d2->length) |
508 | return -1; |
509 | /* Lengths are equal, do a byte-by-byte comparison */ |
510 | return strncmp(d1->lexeme, d2->lexeme, d1->length); |
511 | } |
512 | |
513 | /* |
514 | * qsort() comparator for sorting TrackItems on frequencies (descending sort) |
515 | */ |
516 | static int |
517 | trackitem_compare_frequencies_desc(const void *e1, const void *e2) |
518 | { |
519 | const TrackItem *const *t1 = (const TrackItem *const *) e1; |
520 | const TrackItem *const *t2 = (const TrackItem *const *) e2; |
521 | |
522 | return (*t2)->frequency - (*t1)->frequency; |
523 | } |
524 | |
525 | /* |
526 | * qsort() comparator for sorting TrackItems on lexemes |
527 | */ |
528 | static int |
529 | trackitem_compare_lexemes(const void *e1, const void *e2) |
530 | { |
531 | const TrackItem *const *t1 = (const TrackItem *const *) e1; |
532 | const TrackItem *const *t2 = (const TrackItem *const *) e2; |
533 | |
534 | return lexeme_compare(&(*t1)->key, &(*t2)->key); |
535 | } |
536 | |