1 /*-------------------------------------------------------------------------
4 * functions for gathering statistics from tsvector columns
6 * Portions Copyright (c) 1996-2011, PostgreSQL Global Development Group
10 * src/backend/tsearch/ts_typanalyze.c
12 *-------------------------------------------------------------------------
16 #include "access/hash.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/hsearch.h"
24 /* A hash key for lexemes */
27 char *lexeme; /* lexeme (not NULL terminated!) */
28 int length; /* its length in bytes */
31 /* A hash table entry for the Lossy Counting algorithm */
34 LexemeHashKey key; /* This is 'e' from the LC algorithm. */
35 int frequency; /* This is 'f'. */
36 int delta; /* And this is 'delta'. */
39 static void compute_tsvector_stats(VacAttrStats *stats,
40 AnalyzeAttrFetchFunc fetchfunc,
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);
52 * ts_typanalyze -- a custom typanalyze function for tsvector columns
55 ts_typanalyze(PG_FUNCTION_ARGS)
57 VacAttrStats *stats = (VacAttrStats *) PG_GETARG_POINTER(0);
58 Form_pg_attribute attr = stats->attr;
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;
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;
73 * compute_tsvector_stats() -- compute statistics for a tsvector column
75 * This functions computes statistics that are useful for determining @@
76 * operations' selectivity, along with the fraction of non-null rows and
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.
85 * For the same reasons we assume that tsvector columns are unique when
86 * determining the number of distinct values.
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
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.
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.
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).
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
140 compute_tsvector_stats(VacAttrStats *stats,
141 AnalyzeAttrFetchFunc fetchfunc,
147 double total_width = 0;
149 /* This is D from the LC algorithm. */
152 HASH_SEQ_STATUS scan_status;
154 /* This is the current bucket number from the LC algorithm */
157 /* This is 'w' from the LC algorithm */
161 LexemeHashKey hash_key;
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.
170 num_mcelem = stats->attr->attstattarget * 10;
173 * We set bucket width equal to (num_mcelem + 10) / 0.007 as per the
176 bucket_width = (num_mcelem + 10) * 1000 / 7;
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.
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",
192 HASH_ELEM | HASH_FUNCTION | HASH_COMPARE | HASH_CONTEXT);
194 /* Initialize counters. */
198 /* Loop over the tsvectors. */
199 for (vector_no = 0; vector_no < samplerows; vector_no++)
204 WordEntry *curentryptr;
208 vacuum_delay_point();
210 value = fetchfunc(stats, vector_no, &isnull);
213 * Check for null/nonnull.
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
227 total_width += VARSIZE_ANY(DatumGetPointer(value));
230 * Now detoast the tsvector if needed.
232 vector = DatumGetTSVector(value);
235 * We loop through the lexemes in the tsvector and add them to our
236 * tracking hashtable. Note: the hashtable entries will point into
237 * the (detoasted) tsvector value, therefore we cannot free that
238 * storage until we're done.
240 lexemesptr = STRPTR(vector);
241 curentryptr = ARRPTR(vector);
242 for (j = 0; j < vector->size; j++)
246 /* Construct a hash key */
247 hash_key.lexeme = lexemesptr + curentryptr->pos;
248 hash_key.length = curentryptr->len;
250 /* Lookup current lexeme in hashtable, adding it if new */
251 item = (TrackItem *) hash_search(lexemes_tab,
252 (const void *) &hash_key,
257 /* The lexeme is already on the tracking list */
262 /* Initialize new tracking list element */
264 item->delta = b_current - 1;
267 /* lexeme_no is the number of elements processed (ie N) */
270 /* We prune the D structure after processing each bucket */
271 if (lexeme_no % bucket_width == 0)
273 prune_lexemes_hashtable(lexemes_tab, b_current);
277 /* Advance to the next WordEntry in the tsvector */
282 /* We can only compute real stats if we found some non-null values. */
283 if (null_cnt < samplerows)
285 int nonnull_cnt = samplerows - null_cnt;
287 TrackItem **sort_table;
293 stats->stats_valid = true;
294 /* Do the simple null-frac and average width stats */
295 stats->stanullfrac = (double) null_cnt / (double) samplerows;
296 stats->stawidth = total_width / (double) nonnull_cnt;
298 /* Assume it's a unique column (see notes above) */
299 stats->stadistinct = -1.0;
302 * Construct an array of the interesting hashtable items, that is,
303 * those meeting the cutoff frequency (s - epsilon)*N. Also identify
304 * the minimum and maximum frequencies among these items.
306 * Since epsilon = s/10 and bucket_width = 1/epsilon, the cutoff
307 * frequency is 9*N / bucket_width.
309 cutoff_freq = 9 * lexeme_no / bucket_width;
311 i = hash_get_num_entries(lexemes_tab); /* surely enough space */
312 sort_table = (TrackItem **) palloc(sizeof(TrackItem *) * i);
314 hash_seq_init(&scan_status, lexemes_tab);
318 while ((item = (TrackItem *) hash_seq_search(&scan_status)) != NULL)
320 if (item->frequency > cutoff_freq)
322 sort_table[track_len++] = item;
323 minfreq = Min(minfreq, item->frequency);
324 maxfreq = Max(maxfreq, item->frequency);
327 Assert(track_len <= i);
329 /* emit some statistics for debug purposes */
330 elog(DEBUG3, "tsvector_stats: target # mces = %d, bucket width = %d, "
331 "# lexemes = %d, hashtable size = %d, usable entries = %d",
332 num_mcelem, bucket_width, lexeme_no, i, track_len);
335 * If we obtained more lexemes than we really want, get rid of those
336 * with least frequencies. The easiest way is to qsort the array into
337 * descending frequency order and truncate the array.
339 if (num_mcelem < track_len)
341 qsort(sort_table, track_len, sizeof(TrackItem *),
342 trackitem_compare_frequencies_desc);
343 /* reset minfreq to the smallest frequency we're keeping */
344 minfreq = sort_table[num_mcelem - 1]->frequency;
347 num_mcelem = track_len;
349 /* Generate MCELEM slot entry */
352 MemoryContext old_context;
353 Datum *mcelem_values;
354 float4 *mcelem_freqs;
357 * We want to store statistics sorted on the lexeme value using
358 * first length, then byte-for-byte comparison. The reason for
359 * doing length comparison first is that we don't care about the
360 * ordering so long as it's consistent, and comparing lengths
361 * first gives us a chance to avoid a strncmp() call.
363 * This is different from what we do with scalar statistics --
364 * they get sorted on frequencies. The rationale is that we
365 * usually search through most common elements looking for a
366 * specific value, so we can grab its frequency. When values are
367 * presorted we can employ binary search for that. See
368 * ts_selfuncs.c for a real usage scenario.
370 qsort(sort_table, num_mcelem, sizeof(TrackItem *),
371 trackitem_compare_lexemes);
373 /* Must copy the target values into anl_context */
374 old_context = MemoryContextSwitchTo(stats->anl_context);
377 * We sorted statistics on the lexeme value, but we want to be
378 * able to find out the minimal and maximal frequency without
379 * going through all the values. We keep those two extra
380 * frequencies in two extra cells in mcelem_freqs.
382 mcelem_values = (Datum *) palloc(num_mcelem * sizeof(Datum));
383 mcelem_freqs = (float4 *) palloc((num_mcelem + 2) * sizeof(float4));
386 * See comments above about use of nonnull_cnt as the divisor for
387 * the final frequency estimates.
389 for (i = 0; i < num_mcelem; i++)
391 TrackItem *item = sort_table[i];
394 PointerGetDatum(cstring_to_text_with_len(item->key.lexeme,
396 mcelem_freqs[i] = (double) item->frequency / (double) nonnull_cnt;
398 mcelem_freqs[i++] = (double) minfreq / (double) nonnull_cnt;
399 mcelem_freqs[i] = (double) maxfreq / (double) nonnull_cnt;
400 MemoryContextSwitchTo(old_context);
402 stats->stakind[0] = STATISTIC_KIND_MCELEM;
403 stats->staop[0] = TextEqualOperator;
404 stats->stanumbers[0] = mcelem_freqs;
405 /* See above comment about two extra frequency fields */
406 stats->numnumbers[0] = num_mcelem + 2;
407 stats->stavalues[0] = mcelem_values;
408 stats->numvalues[0] = num_mcelem;
409 /* We are storing text values */
410 stats->statypid[0] = TEXTOID;
411 stats->statyplen[0] = -1; /* typlen, -1 for varlena */
412 stats->statypbyval[0] = false;
413 stats->statypalign[0] = 'i';
418 /* We found only nulls; assume the column is entirely null */
419 stats->stats_valid = true;
420 stats->stanullfrac = 1.0;
421 stats->stawidth = 0; /* "unknown" */
422 stats->stadistinct = 0.0; /* "unknown" */
426 * We don't need to bother cleaning up any of our temporary palloc's. The
427 * hashtable should also go away, as it used a child memory context.
432 * A function to prune the D structure from the Lossy Counting algorithm.
433 * Consult compute_tsvector_stats() for wider explanation.
436 prune_lexemes_hashtable(HTAB *lexemes_tab, int b_current)
438 HASH_SEQ_STATUS scan_status;
441 hash_seq_init(&scan_status, lexemes_tab);
442 while ((item = (TrackItem *) hash_seq_search(&scan_status)) != NULL)
444 if (item->frequency + item->delta <= b_current)
446 if (hash_search(lexemes_tab, (const void *) &item->key,
447 HASH_REMOVE, NULL) == NULL)
448 elog(ERROR, "hash table corrupted");
454 * Hash functions for lexemes. They are strings, but not NULL terminated,
455 * so we need a special hash function.
458 lexeme_hash(const void *key, Size keysize)
460 const LexemeHashKey *l = (const LexemeHashKey *) key;
462 return DatumGetUInt32(hash_any((const unsigned char *) l->lexeme,
467 * Matching function for lexemes, to be used in hashtable lookups.
470 lexeme_match(const void *key1, const void *key2, Size keysize)
472 /* The keysize parameter is superfluous, the keys store their lengths */
473 return lexeme_compare(key1, key2);
477 * Comparison function for lexemes.
480 lexeme_compare(const void *key1, const void *key2)
482 const LexemeHashKey *d1 = (const LexemeHashKey *) key1;
483 const LexemeHashKey *d2 = (const LexemeHashKey *) key2;
485 /* First, compare by length */
486 if (d1->length > d2->length)
488 else if (d1->length < d2->length)
490 /* Lengths are equal, do a byte-by-byte comparison */
491 return strncmp(d1->lexeme, d2->lexeme, d1->length);
495 * qsort() comparator for sorting TrackItems on frequencies (descending sort)
498 trackitem_compare_frequencies_desc(const void *e1, const void *e2)
500 const TrackItem *const * t1 = (const TrackItem *const *) e1;
501 const TrackItem *const * t2 = (const TrackItem *const *) e2;
503 return (*t2)->frequency - (*t1)->frequency;
507 * qsort() comparator for sorting TrackItems on lexemes
510 trackitem_compare_lexemes(const void *e1, const void *e2)
512 const TrackItem *const * t1 = (const TrackItem *const *) e1;
513 const TrackItem *const * t2 = (const TrackItem *const *) e2;
515 return lexeme_compare(&(*t1)->key, &(*t2)->key);