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"
23 /* A hash key for lexemes */
26 char *lexeme; /* lexeme (not NULL terminated!) */
27 int length; /* its length in bytes */
30 /* A hash table entry for the Lossy Counting algorithm */
33 LexemeHashKey key; /* This is 'e' from the LC algorithm. */
34 int frequency; /* This is 'f'. */
35 int delta; /* And this is 'delta'. */
38 static void compute_tsvector_stats(VacAttrStats *stats,
39 AnalyzeAttrFetchFunc fetchfunc,
42 static void prune_lexemes_hashtable(HTAB *lexemes_tab, int b_current);
43 static uint32 lexeme_hash(const void *key, Size keysize);
44 static int lexeme_match(const void *key1, const void *key2, Size keysize);
45 static int lexeme_compare(const void *key1, const void *key2);
46 static int trackitem_compare_frequencies_desc(const void *e1, const void *e2);
47 static int trackitem_compare_lexemes(const void *e1, const void *e2);
51 * ts_typanalyze -- a custom typanalyze function for tsvector columns
54 ts_typanalyze(PG_FUNCTION_ARGS)
56 VacAttrStats *stats = (VacAttrStats *) PG_GETARG_POINTER(0);
57 Form_pg_attribute attr = stats->attr;
59 /* If the attstattarget column is negative, use the default value */
60 /* NB: it is okay to scribble on stats->attr since it's a copy */
61 if (attr->attstattarget < 0)
62 attr->attstattarget = default_statistics_target;
64 stats->compute_stats = compute_tsvector_stats;
65 /* see comment about the choice of minrows in commands/analyze.c */
66 stats->minrows = 300 * attr->attstattarget;
72 * compute_tsvector_stats() -- compute statistics for a tsvector column
74 * This functions computes statistics that are useful for determining @@
75 * operations' selectivity, along with the fraction of non-null rows and
78 * Instead of finding the most common values, as we do for most datatypes,
79 * we're looking for the most common lexemes. This is more useful, because
80 * there most probably won't be any two rows with the same tsvector and thus
81 * the notion of a MCV is a bit bogus with this datatype. With a list of the
82 * most common lexemes we can do a better job at figuring out @@ selectivity.
84 * For the same reasons we assume that tsvector columns are unique when
85 * determining the number of distinct values.
87 * The algorithm used is Lossy Counting, as proposed in the paper "Approximate
88 * frequency counts over data streams" by G. S. Manku and R. Motwani, in
89 * Proceedings of the 28th International Conference on Very Large Data Bases,
90 * Hong Kong, China, August 2002, section 4.2. The paper is available at
91 * http://www.vldb.org/conf/2002/S10P03.pdf
93 * The Lossy Counting (aka LC) algorithm goes like this:
94 * Let s be the threshold frequency for an item (the minimum frequency we
95 * are interested in) and epsilon the error margin for the frequency. Let D
96 * be a set of triples (e, f, delta), where e is an element value, f is that
97 * element's frequency (actually, its current occurrence count) and delta is
98 * the maximum error in f. We start with D empty and process the elements in
99 * batches of size w. (The batch size is also known as "bucket size" and is
100 * equal to 1/epsilon.) Let the current batch number be b_current, starting
101 * with 1. For each element e we either increment its f count, if it's
102 * already in D, or insert a new triple into D with values (e, 1, b_current
103 * - 1). After processing each batch we prune D, by removing from it all
104 * elements with f + delta <= b_current. After the algorithm finishes we
105 * suppress all elements from D that do not satisfy f >= (s - epsilon) * N,
106 * where N is the total number of elements in the input. We emit the
107 * remaining elements with estimated frequency f/N. The LC paper proves
108 * that this algorithm finds all elements with true frequency at least s,
109 * and that no frequency is overestimated or is underestimated by more than
110 * epsilon. Furthermore, given reasonable assumptions about the input
111 * distribution, the required table size is no more than about 7 times w.
113 * We set s to be the estimated frequency of the K'th word in a natural
114 * language's frequency table, where K is the target number of entries in
115 * the MCELEM array plus an arbitrary constant, meant to reflect the fact
116 * that the most common words in any language would usually be stopwords
117 * so we will not actually see them in the input. We assume that the
118 * distribution of word frequencies (including the stopwords) follows Zipf's
119 * law with an exponent of 1.
121 * Assuming Zipfian distribution, the frequency of the K'th word is equal
122 * to 1/(K * H(W)) where H(n) is 1/2 + 1/3 + ... + 1/n and W is the number of
123 * words in the language. Putting W as one million, we get roughly 0.07/K.
124 * Assuming top 10 words are stopwords gives s = 0.07/(K + 10). We set
125 * epsilon = s/10, which gives bucket width w = (K + 10)/0.007 and
126 * maximum expected hashtable size of about 1000 * (K + 10).
128 * Note: in the above discussion, s, epsilon, and f/N are in terms of a
129 * lexeme's frequency as a fraction of all lexemes seen in the input.
130 * However, what we actually want to store in the finished pg_statistic
131 * entry is each lexeme's frequency as a fraction of all rows that it occurs
132 * in. Assuming that the input tsvectors are correctly constructed, no
133 * lexeme occurs more than once per tsvector, so the final count f is a
134 * correct estimate of the number of input tsvectors it occurs in, and we
135 * need only change the divisor from N to nonnull_cnt to get the number we
139 compute_tsvector_stats(VacAttrStats *stats,
140 AnalyzeAttrFetchFunc fetchfunc,
146 double total_width = 0;
148 /* This is D from the LC algorithm. */
151 HASH_SEQ_STATUS scan_status;
153 /* This is the current bucket number from the LC algorithm */
156 /* This is 'w' from the LC algorithm */
160 LexemeHashKey hash_key;
164 * We want statistics_target * 10 lexemes in the MCELEM array. This
165 * multiplier is pretty arbitrary, but is meant to reflect the fact that
166 * the number of individual lexeme values tracked in pg_statistic ought to
167 * be more than the number of values for a simple scalar column.
169 num_mcelem = stats->attr->attstattarget * 10;
172 * We set bucket width equal to (num_mcelem + 10) / 0.007 as per the
175 bucket_width = (num_mcelem + 10) * 1000 / 7;
178 * Create the hashtable. It will be in local memory, so we don't need to
179 * worry about overflowing the initial size. Also we don't need to pay any
180 * attention to locking and memory management.
182 MemSet(&hash_ctl, 0, sizeof(hash_ctl));
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",
191 HASH_ELEM | HASH_FUNCTION | HASH_COMPARE | HASH_CONTEXT);
193 /* Initialize counters. */
197 /* Loop over the tsvectors. */
198 for (vector_no = 0; vector_no < samplerows; vector_no++)
203 WordEntry *curentryptr;
207 vacuum_delay_point();
209 value = fetchfunc(stats, vector_no, &isnull);
212 * Check for null/nonnull.
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
226 total_width += VARSIZE_ANY(DatumGetPointer(value));
229 * Now detoast the tsvector if needed.
231 vector = DatumGetTSVector(value);
234 * We loop through the lexemes in the tsvector and add them to our
235 * tracking hashtable. Note: the hashtable entries will point into
236 * the (detoasted) tsvector value, therefore we cannot free that
237 * storage until we're done.
239 lexemesptr = STRPTR(vector);
240 curentryptr = ARRPTR(vector);
241 for (j = 0; j < vector->size; j++)
245 /* Construct a hash key */
246 hash_key.lexeme = lexemesptr + curentryptr->pos;
247 hash_key.length = curentryptr->len;
249 /* Lookup current lexeme in hashtable, adding it if new */
250 item = (TrackItem *) hash_search(lexemes_tab,
251 (const void *) &hash_key,
256 /* The lexeme is already on the tracking list */
261 /* Initialize new tracking list element */
263 item->delta = b_current - 1;
266 /* lexeme_no is the number of elements processed (ie N) */
269 /* We prune the D structure after processing each bucket */
270 if (lexeme_no % bucket_width == 0)
272 prune_lexemes_hashtable(lexemes_tab, b_current);
276 /* Advance to the next WordEntry in the tsvector */
281 /* We can only compute real stats if we found some non-null values. */
282 if (null_cnt < samplerows)
284 int nonnull_cnt = samplerows - null_cnt;
286 TrackItem **sort_table;
292 stats->stats_valid = true;
293 /* Do the simple null-frac and average width stats */
294 stats->stanullfrac = (double) null_cnt / (double) samplerows;
295 stats->stawidth = total_width / (double) nonnull_cnt;
297 /* Assume it's a unique column (see notes above) */
298 stats->stadistinct = -1.0;
301 * Construct an array of the interesting hashtable items, that is,
302 * those meeting the cutoff frequency (s - epsilon)*N. Also identify
303 * the minimum and maximum frequencies among these items.
305 * Since epsilon = s/10 and bucket_width = 1/epsilon, the cutoff
306 * frequency is 9*N / bucket_width.
308 cutoff_freq = 9 * lexeme_no / bucket_width;
310 i = hash_get_num_entries(lexemes_tab); /* surely enough space */
311 sort_table = (TrackItem **) palloc(sizeof(TrackItem *) * i);
313 hash_seq_init(&scan_status, lexemes_tab);
317 while ((item = (TrackItem *) hash_seq_search(&scan_status)) != NULL)
319 if (item->frequency > cutoff_freq)
321 sort_table[track_len++] = item;
322 minfreq = Min(minfreq, item->frequency);
323 maxfreq = Max(maxfreq, item->frequency);
326 Assert(track_len <= i);
328 /* emit some statistics for debug purposes */
329 elog(DEBUG3, "tsvector_stats: target # mces = %d, bucket width = %d, "
330 "# lexemes = %d, hashtable size = %d, usable entries = %d",
331 num_mcelem, bucket_width, lexeme_no, i, track_len);
334 * If we obtained more lexemes than we really want, get rid of those
335 * with least frequencies. The easiest way is to qsort the array into
336 * descending frequency order and truncate the array.
338 if (num_mcelem < track_len)
340 qsort(sort_table, track_len, sizeof(TrackItem *),
341 trackitem_compare_frequencies_desc);
342 /* reset minfreq to the smallest frequency we're keeping */
343 minfreq = sort_table[num_mcelem - 1]->frequency;
346 num_mcelem = track_len;
348 /* Generate MCELEM slot entry */
351 MemoryContext old_context;
352 Datum *mcelem_values;
353 float4 *mcelem_freqs;
356 * We want to store statistics sorted on the lexeme value using
357 * first length, then byte-for-byte comparison. The reason for
358 * doing length comparison first is that we don't care about the
359 * ordering so long as it's consistent, and comparing lengths
360 * first gives us a chance to avoid a strncmp() call.
362 * This is different from what we do with scalar statistics --
363 * they get sorted on frequencies. The rationale is that we
364 * usually search through most common elements looking for a
365 * specific value, so we can grab its frequency. When values are
366 * presorted we can employ binary search for that. See
367 * ts_selfuncs.c for a real usage scenario.
369 qsort(sort_table, num_mcelem, sizeof(TrackItem *),
370 trackitem_compare_lexemes);
372 /* Must copy the target values into anl_context */
373 old_context = MemoryContextSwitchTo(stats->anl_context);
376 * We sorted statistics on the lexeme value, but we want to be
377 * able to find out the minimal and maximal frequency without
378 * going through all the values. We keep those two extra
379 * frequencies in two extra cells in mcelem_freqs.
381 mcelem_values = (Datum *) palloc(num_mcelem * sizeof(Datum));
382 mcelem_freqs = (float4 *) palloc((num_mcelem + 2) * sizeof(float4));
385 * See comments above about use of nonnull_cnt as the divisor for
386 * the final frequency estimates.
388 for (i = 0; i < num_mcelem; i++)
390 TrackItem *item = sort_table[i];
393 PointerGetDatum(cstring_to_text_with_len(item->key.lexeme,
395 mcelem_freqs[i] = (double) item->frequency / (double) nonnull_cnt;
397 mcelem_freqs[i++] = (double) minfreq / (double) nonnull_cnt;
398 mcelem_freqs[i] = (double) maxfreq / (double) nonnull_cnt;
399 MemoryContextSwitchTo(old_context);
401 stats->stakind[0] = STATISTIC_KIND_MCELEM;
402 stats->staop[0] = TextEqualOperator;
403 stats->stanumbers[0] = mcelem_freqs;
404 /* See above comment about two extra frequency fields */
405 stats->numnumbers[0] = num_mcelem + 2;
406 stats->stavalues[0] = mcelem_values;
407 stats->numvalues[0] = num_mcelem;
408 /* We are storing text values */
409 stats->statypid[0] = TEXTOID;
410 stats->statyplen[0] = -1; /* typlen, -1 for varlena */
411 stats->statypbyval[0] = false;
412 stats->statypalign[0] = 'i';
417 /* We found only nulls; assume the column is entirely null */
418 stats->stats_valid = true;
419 stats->stanullfrac = 1.0;
420 stats->stawidth = 0; /* "unknown" */
421 stats->stadistinct = 0.0; /* "unknown" */
425 * We don't need to bother cleaning up any of our temporary palloc's. The
426 * hashtable should also go away, as it used a child memory context.
431 * A function to prune the D structure from the Lossy Counting algorithm.
432 * Consult compute_tsvector_stats() for wider explanation.
435 prune_lexemes_hashtable(HTAB *lexemes_tab, int b_current)
437 HASH_SEQ_STATUS scan_status;
440 hash_seq_init(&scan_status, lexemes_tab);
441 while ((item = (TrackItem *) hash_seq_search(&scan_status)) != NULL)
443 if (item->frequency + item->delta <= b_current)
445 if (hash_search(lexemes_tab, (const void *) &item->key,
446 HASH_REMOVE, NULL) == NULL)
447 elog(ERROR, "hash table corrupted");
453 * Hash functions for lexemes. They are strings, but not NULL terminated,
454 * so we need a special hash function.
457 lexeme_hash(const void *key, Size keysize)
459 const LexemeHashKey *l = (const LexemeHashKey *) key;
461 return DatumGetUInt32(hash_any((const unsigned char *) l->lexeme,
466 * Matching function for lexemes, to be used in hashtable lookups.
469 lexeme_match(const void *key1, const void *key2, Size keysize)
471 /* The keysize parameter is superfluous, the keys store their lengths */
472 return lexeme_compare(key1, key2);
476 * Comparison function for lexemes.
479 lexeme_compare(const void *key1, const void *key2)
481 const LexemeHashKey *d1 = (const LexemeHashKey *) key1;
482 const LexemeHashKey *d2 = (const LexemeHashKey *) key2;
484 /* First, compare by length */
485 if (d1->length > d2->length)
487 else if (d1->length < d2->length)
489 /* Lengths are equal, do a byte-by-byte comparison */
490 return strncmp(d1->lexeme, d2->lexeme, d1->length);
494 * qsort() comparator for sorting TrackItems on frequencies (descending sort)
497 trackitem_compare_frequencies_desc(const void *e1, const void *e2)
499 const TrackItem *const * t1 = (const TrackItem *const *) e1;
500 const TrackItem *const * t2 = (const TrackItem *const *) e2;
502 return (*t2)->frequency - (*t1)->frequency;
506 * qsort() comparator for sorting TrackItems on lexemes
509 trackitem_compare_lexemes(const void *e1, const void *e2)
511 const TrackItem *const * t1 = (const TrackItem *const *) e1;
512 const TrackItem *const * t2 = (const TrackItem *const *) e2;
514 return lexeme_compare(&(*t1)->key, &(*t2)->key);