| blob | b31147c30a1115e868504e0f4cd63e8558b34a6a |
1 /*-------------------------------------------------------------------------
2 *
3 * analyze.c
4 * the Postgres statistics generator
5 *
6 * Portions Copyright (c) 1996-2008, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
8 *
9 *
10 * IDENTIFICATION
11 * $PostgreSQL$
12 *
13 *-------------------------------------------------------------------------
14 */
17 #include <math.h>
49 /* Data structure for Algorithm S from Knuth 3.4.2 */
51 {
59 /* Per-index data for ANALYZE */
61 {
69 /* Default statistics target (GUC parameter) */
72 /* A few variables that don't seem worth passing around as parameters */
87 MemoryContext col_context);
102 /*
103 * analyze_rel() -- analyze one relation
104 *
105 * If update_reltuples is true, we update reltuples and relpages columns
106 * in pg_class. Caller should pass false if we're part of VACUUM ANALYZE,
107 * and the VACUUM didn't skip any pages. We only have an approximate count,
108 * so we don't want to overwrite the accurate values already inserted by the
109 * VACUUM in that case. VACUUM always scans all indexes, however, so the
110 * pg_class entries for indexes are never updated if we're part of VACUUM
111 * ANALYZE.
112 */
113 void
116 {
117 Relation onerel;
119 tcnt,
120 i,
121 ind;
129 numrows;
131 totaldeadrows;
133 PGRUsage ru0;
135 Oid save_userid;
140 else
145 /*
146 * Use the current context for storing analysis info. vacuum.c ensures
147 * that this context will be cleared when I return, thus releasing the
148 * memory allocated here.
149 */
152 /*
153 * Check for user-requested abort. Note we want this to be inside a
154 * transaction, so xact.c doesn't issue useless WARNING.
155 */
158 /*
159 * Open the relation, getting ShareUpdateExclusiveLock to ensure that two
160 * ANALYZEs don't run on it concurrently. (This also locks out a
161 * concurrent VACUUM, which doesn't matter much at the moment but might
162 * matter if we ever try to accumulate stats on dead tuples.) If the rel
163 * has been dropped since we last saw it, we don't need to process it.
164 */
169 /*
170 * Check permissions --- this should match vacuum's check!
171 */
174 {
175 /* No need for a WARNING if we already complained during VACUUM */
177 {
186 else
190 }
193 }
195 /*
196 * Check that it's a plain table; we used to do this in get_rel_oids() but
197 * seems safer to check after we've locked the relation.
198 */
200 {
201 /* No need for a WARNING if we already complained during VACUUM */
208 }
210 /*
211 * Silently ignore tables that are temp tables of other backends ---
212 * trying to analyze these is rather pointless, since their contents are
213 * probably not up-to-date on disk. (We don't throw a warning here; it
214 * would just lead to chatter during a database-wide ANALYZE.)
215 */
217 {
220 }
222 /*
223 * We can ANALYZE any table except pg_statistic. See update_attstats
224 */
226 {
229 }
236 /*
237 * Switch to the table owner's userid, so that any index functions are
238 * run as that user.
239 */
243 /* let others know what I'm doing */
248 /* measure elapsed time iff autovacuum logging requires it */
250 {
254 }
256 /*
257 * Determine which columns to analyze
258 *
259 * Note that system attributes are never analyzed.
260 */
262 {
269 {
280 tcnt++;
281 }
283 }
284 else
285 {
291 {
294 tcnt++;
295 }
297 }
299 /*
300 * Open all indexes of the relation, and see if there are any analyzable
301 * columns in the indexes. We do not analyze index columns if there was
302 * an explicit column list in the ANALYZE command, however.
303 */
309 {
312 {
319 {
326 {
330 {
331 /* Found an index expression */
339 /*
340 * Can't analyze if the opclass uses a storage type
341 * different from the expression result type. We'd get
342 * confused because the type shown in pg_attribute for
343 * the index column doesn't match what we are getting
344 * from the expression. Perhaps this can be fixed
345 * someday, but for now, punt.
346 */
354 {
355 tcnt++;
357 }
358 }
359 }
361 }
362 }
363 }
365 /*
366 * Quit if no analyzable columns
367 */
369 {
370 /*
371 * We report that the table is empty; this is just so that the
372 * autovacuum code doesn't go nuts trying to get stats about a
373 * zero-column table.
374 */
378 }
380 /*
381 * Determine how many rows we need to sample, using the worst case from
382 * all analyzable columns. We use a lower bound of 100 rows to avoid
383 * possible overflow in Vitter's algorithm.
384 */
387 {
390 }
392 {
396 {
399 }
400 }
402 /*
403 * Acquire the sample rows
404 */
409 /*
410 * Compute the statistics. Temporary results during the calculations for
411 * each column are stored in a child context. The calc routines are
412 * responsible to make sure that whatever they store into the VacAttrStats
413 * structure is allocated in anl_context.
414 */
416 {
417 MemoryContext col_context,
418 old_context;
422 ALLOCSET_DEFAULT_MINSIZE,
423 ALLOCSET_DEFAULT_INITSIZE,
424 ALLOCSET_DEFAULT_MAXSIZE);
428 {
434 std_fetch_func,
435 numrows,
436 totalrows);
438 }
444 col_context);
449 /*
450 * Emit the completed stats rows into pg_statistic, replacing any
451 * previous statistics for the target columns. (If there are stats in
452 * pg_statistic for columns we didn't process, we leave them alone.)
453 */
457 {
462 }
463 }
465 /*
466 * Update pages/tuples stats in pg_class.
467 */
469 {
473 /* report results to the stats collector, too */
475 }
477 /*
478 * Same for indexes. Vacuum always scans all indexes, so if we're part of
479 * VACUUM ANALYZE, don't overwrite the accurate count already inserted by
480 * VACUUM.
481 */
483 {
485 {
493 }
494 }
496 /* We skip to here if there were no analyzable columns */
497 cleanup:
499 /* Done with indexes */
502 /*
503 * Close source relation now, but keep lock so that no one deletes it
504 * before we commit. (If someone did, they'd fail to clean up the entries
505 * we made in pg_statistic. Also, releasing the lock before commit would
506 * expose us to concurrent-update failures in update_attstats.)
507 */
510 /* Log the action if appropriate */
512 {
515 Log_autovacuum_min_duration))
522 }
524 /*
525 * Reset my PGPROC flag. Note: we need this here, and not in vacuum_rel,
526 * because the vacuum flag is cleared by the end-of-xact code.
527 */
532 /* Restore userid */
534 }
536 /*
537 * Compute statistics about indexes of a relation
538 */
539 static void
543 MemoryContext col_context)
544 {
545 MemoryContext ind_context,
546 old_context;
550 i;
554 ALLOCSET_DEFAULT_MINSIZE,
555 ALLOCSET_DEFAULT_INITSIZE,
556 ALLOCSET_DEFAULT_MAXSIZE);
560 {
571 tcnt,
572 rowno;
575 /* Ignore index if no columns to analyze and not partial */
579 /*
580 * Need an EState for evaluation of index expressions and
581 * partial-index predicates. Create it in the per-index context to be
582 * sure it gets cleaned up at the bottom of the loop.
583 */
586 /* Need a slot to hold the current heap tuple, too */
589 /* Arrange for econtext's scan tuple to be the tuple under test */
592 /* Set up execution state for predicate. */
595 estate);
597 /* Compute and save index expression values */
603 {
606 /* Set up for predicate or expression evaluation */
609 /* If index is partial, check predicate */
611 {
614 }
615 numindexrows++;
618 {
619 /*
620 * Evaluate the index row to compute expression values. We
621 * could do this by hand, but FormIndexDatum is convenient.
622 */
624 slot,
625 estate,
626 values,
627 isnull);
629 /*
630 * Save just the columns we care about.
631 */
633 {
639 tcnt++;
640 }
641 }
642 }
644 /*
645 * Having counted the number of rows that pass the predicate in the
646 * sample, we can estimate the total number of rows in the index.
647 */
651 /*
652 * Now we can compute the statistics for the expression columns.
653 */
655 {
658 {
665 ind_fetch_func,
666 numindexrows,
667 totalindexrows);
669 }
670 }
672 /* And clean up */
678 }
682 }
684 /*
685 * examine_attribute -- pre-analysis of a single column
686 *
687 * Determine whether the column is analyzable; if so, create and initialize
688 * a VacAttrStats struct for it. If not, return NULL.
689 */
692 {
694 HeapTuple typtuple;
699 /* Never analyze dropped columns */
703 /* Don't analyze column if user has specified not to */
707 /*
708 * Create the VacAttrStats struct.
709 */
724 /*
725 * The fields describing the stats->stavalues[n] element types default
726 * to the type of the field being analyzed, but the type-specific
727 * typanalyze function can change them if it wants to store something
728 * else.
729 */
731 {
736 }
738 /*
739 * Call the type-specific typanalyze function. If none is specified, use
740 * std_typanalyze().
741 */
745 else
749 {
754 }
757 }
759 /*
760 * BlockSampler_Init -- prepare for random sampling of blocknumbers
761 *
762 * BlockSampler is used for stage one of our new two-stage tuple
763 * sampling mechanism as discussed on pgsql-hackers 2004-04-02 (subject
764 * "Large DB"). It selects a random sample of samplesize blocks out of
765 * the nblocks blocks in the table. If the table has less than
766 * samplesize blocks, all blocks are selected.
767 *
768 * Since we know the total number of blocks in advance, we can use the
769 * straightforward Algorithm S from Knuth 3.4.2, rather than Vitter's
770 * algorithm.
771 */
772 static void
774 {
777 /*
778 * If we decide to reduce samplesize for tables that have less or not much
779 * more than samplesize blocks, here is the place to do it.
780 */
784 }
786 static bool
788 {
790 }
792 static BlockNumber
794 {
803 {
804 /* need all the rest */
807 }
809 /*----------
810 * It is not obvious that this code matches Knuth's Algorithm S.
811 * Knuth says to skip the current block with probability 1 - k/K.
812 * If we are to skip, we should advance t (hence decrease K), and
813 * repeat the same probabilistic test for the next block. The naive
814 * implementation thus requires a random_fract() call for each block
815 * number. But we can reduce this to one random_fract() call per
816 * selected block, by noting that each time the while-test succeeds,
817 * we can reinterpret V as a uniform random number in the range 0 to p.
818 * Therefore, instead of choosing a new V, we just adjust p to be
819 * the appropriate fraction of its former value, and our next loop
820 * makes the appropriate probabilistic test.
821 *
822 * We have initially K > k > 0. If the loop reduces K to equal k,
823 * the next while-test must fail since p will become exactly zero
824 * (we assume there will not be roundoff error in the division).
825 * (Note: Knuth suggests a "<=" loop condition, but we use "<" just
826 * to be doubly sure about roundoff error.) Therefore K cannot become
827 * less than k, which means that we cannot fail to select enough blocks.
828 *----------
829 */
833 {
834 /* skip */
838 /* adjust p to be new cutoff point in reduced range */
840 }
842 /* select */
845 }
847 /*
848 * acquire_sample_rows -- acquire a random sample of rows from the table
849 *
850 * As of May 2004 we use a new two-stage method: Stage one selects up
851 * to targrows random blocks (or all blocks, if there aren't so many).
852 * Stage two scans these blocks and uses the Vitter algorithm to create
853 * a random sample of targrows rows (or less, if there are less in the
854 * sample of blocks). The two stages are executed simultaneously: each
855 * block is processed as soon as stage one returns its number and while
856 * the rows are read stage two controls which ones are to be inserted
857 * into the sample.
858 *
859 * Although every row has an equal chance of ending up in the final
860 * sample, this sampling method is not perfect: not every possible
861 * sample has an equal chance of being selected. For large relations
862 * the number of different blocks represented by the sample tends to be
863 * too small. We can live with that for now. Improvements are welcome.
864 *
865 * We also estimate the total numbers of live and dead rows in the table,
866 * and return them into *totalrows and *totaldeadrows, respectively.
867 *
868 * An important property of this sampling method is that because we do
869 * look at a statistically unbiased set of blocks, we should get
870 * unbiased estimates of the average numbers of live and dead rows per
871 * block. The previous sampling method put too much credence in the row
872 * density near the start of the table.
873 *
874 * The returned list of tuples is in order by physical position in the table.
875 * (We will rely on this later to derive correlation estimates.)
876 */
877 static int
880 {
886 BlockNumber totalblocks;
887 TransactionId OldestXmin;
888 BlockSamplerData bs;
895 /* Need a cutoff xmin for HeapTupleSatisfiesVacuum */
898 /* Prepare for sampling block numbers */
900 /* Prepare for sampling rows */
903 /* Outer loop over blocks to sample */
905 {
907 Buffer targbuffer;
908 Page targpage;
909 OffsetNumber targoffset,
910 maxoffset;
914 /*
915 * We must maintain a pin on the target page's buffer to ensure that
916 * the maxoffset value stays good (else concurrent VACUUM might delete
917 * tuples out from under us). Hence, pin the page until we are done
918 * looking at it. We also choose to hold sharelock on the buffer
919 * throughout --- we could release and re-acquire sharelock for
920 * each tuple, but since we aren't doing much work per tuple, the
921 * extra lock traffic is probably better avoided.
922 */
929 /* Inner loop over all tuples on the selected page */
931 {
932 ItemId itemid;
933 HeapTupleData targtuple;
938 /*
939 * We ignore unused and redirect line pointers. DEAD line
940 * pointers should be counted as dead, because we need vacuum
941 * to run to get rid of them. Note that this rule agrees with
942 * the way that heap_page_prune() counts things.
943 */
945 {
949 }
957 OldestXmin,
958 targbuffer))
959 {
967 /* Count dead and recently-dead rows */
972 /*
973 * Insert-in-progress rows are not counted. We assume
974 * that when the inserting transaction commits or aborts,
975 * it will send a stats message to increment the proper
976 * count. This works right only if that transaction ends
977 * after we finish analyzing the table; if things happen
978 * in the other order, its stats update will be
979 * overwritten by ours. However, the error will be
980 * large only if the other transaction runs long enough
981 * to insert many tuples, so assuming it will finish
982 * after us is the safer option.
983 *
984 * A special case is that the inserting transaction might
985 * be our own. In this case we should count and sample
986 * the row, to accommodate users who load a table and
987 * analyze it in one transaction. (pgstat_report_analyze
988 * has to adjust the numbers we send to the stats collector
989 * to make this come out right.)
990 */
992 {
995 }
999 /*
1000 * We count delete-in-progress rows as still live, using
1001 * the same reasoning given above; but we don't bother to
1002 * include them in the sample.
1003 *
1004 * If the delete was done by our own transaction, however,
1005 * we must count the row as dead to make
1006 * pgstat_report_analyze's stats adjustments come out
1007 * right. (Note: this works out properly when the row
1008 * was both inserted and deleted in our xact.)
1009 */
1012 else
1019 }
1022 {
1023 /*
1024 * The first targrows sample rows are simply copied into the
1025 * reservoir. Then we start replacing tuples in the sample
1026 * until we reach the end of the relation. This algorithm is
1027 * from Jeff Vitter's paper (see full citation below). It
1028 * works by repeatedly computing the number of tuples to skip
1029 * before selecting a tuple, which replaces a randomly chosen
1030 * element of the reservoir (current set of tuples). At all
1031 * times the reservoir is a true random sample of the tuples
1032 * we've passed over so far, so when we fall off the end of
1033 * the relation we're done.
1034 */
1037 else
1038 {
1039 /*
1040 * t in Vitter's paper is the number of records already
1041 * processed. If we need to compute a new S value, we
1042 * must use the not-yet-incremented value of samplerows
1043 * as t.
1044 */
1049 {
1050 /*
1051 * Found a suitable tuple, so save it, replacing one
1052 * old tuple at random
1053 */
1059 }
1062 }
1065 }
1066 }
1068 /* Now release the lock and pin on the page */
1070 }
1072 /*
1073 * If we didn't find as many tuples as we wanted then we're done. No sort
1074 * is needed, since they're already in order.
1075 *
1076 * Otherwise we need to sort the collected tuples by position
1077 * (itempointer). It's not worth worrying about corner cases where the
1078 * tuples are already sorted.
1079 */
1083 /*
1084 * Estimate total numbers of rows in relation.
1085 */
1087 {
1090 }
1091 else
1092 {
1095 }
1097 /*
1098 * Emit some interesting relation info
1099 */
1102 "containing %.0f live rows and %.0f dead rows; "
1110 }
1112 /* Select a random value R uniformly distributed in (0 - 1) */
1113 static double
1115 {
1117 }
1119 /*
1120 * These two routines embody Algorithm Z from "Random sampling with a
1121 * reservoir" by Jeffrey S. Vitter, in ACM Trans. Math. Softw. 11, 1
1122 * (Mar. 1985), Pages 37-57. Vitter describes his algorithm in terms
1123 * of the count S of records to skip before processing another record.
1124 * It is computed primarily based on t, the number of records already read.
1125 * The only extra state needed between calls is W, a random state variable.
1126 *
1127 * init_selection_state computes the initial W value.
1128 *
1129 * Given that we've already read t records (t >= n), get_next_S
1130 * determines the number of records to skip before the next record is
1131 * processed.
1132 */
1133 static double
1135 {
1136 /* Initial value of W (for use when Algorithm Z is first applied) */
1138 }
1140 static double
1142 {
1145 /* The magic constant here is T from Vitter's paper */
1147 {
1148 /* Process records using Algorithm X until t is large enough */
1150 quot;
1155 /* Note: "num" in Vitter's code is always equal to t - n */
1157 /* Find min S satisfying (4.1) */
1159 {
1163 }
1164 }
1165 else
1166 {
1167 /* Now apply Algorithm Z */
1172 {
1174 numer_lim,
1175 denom;
1177 X,
1178 lhs,
1179 rhs,
1180 y,
1181 tmp;
1183 /* Generate U and X */
1187 /* Test if U <= h(S)/cg(X) in the manner of (6.3) */
1192 {
1195 }
1196 /* Test if U <= f(S)/cg(X) */
1199 {
1202 }
1203 else
1204 {
1207 }
1209 {
1212 }
1216 }
1218 }
1220 }
1222 /*
1223 * qsort comparator for sorting rows[] array
1224 */
1225 static int
1227 {
1244 }
1247 /*
1248 * update_attstats() -- update attribute statistics for one relation
1249 *
1250 * Statistics are stored in several places: the pg_class row for the
1251 * relation has stats about the whole relation, and there is a
1252 * pg_statistic row for each (non-system) attribute that has ever
1253 * been analyzed. The pg_class values are updated by VACUUM, not here.
1254 *
1255 * pg_statistic rows are just added or updated normally. This means
1256 * that pg_statistic will probably contain some deleted rows at the
1257 * completion of a vacuum cycle, unless it happens to get vacuumed last.
1258 *
1259 * To keep things simple, we punt for pg_statistic, and don't try
1260 * to compute or store rows for pg_statistic itself in pg_statistic.
1261 * This could possibly be made to work, but it's not worth the trouble.
1262 * Note analyze_rel() has seen to it that we won't come here when
1263 * vacuuming pg_statistic itself.
1264 *
1265 * Note: there would be a race condition here if two backends could
1266 * ANALYZE the same table concurrently. Presently, we lock that out
1267 * by taking a self-exclusive lock on the relation in analyze_rel().
1268 */
1269 static void
1271 {
1272 Relation sd;
1281 {
1283 HeapTuple stup,
1284 oldtup;
1286 k,
1287 n;
1292 /* Ignore attr if we weren't able to collect stats */
1296 /*
1297 * Construct a new pg_statistic tuple
1298 */
1300 {
1303 }
1312 {
1314 }
1316 {
1318 }
1320 {
1324 {
1330 /* XXX knows more than it should about type float4: */
1332 FLOAT4OID,
1335 }
1336 else
1337 {
1340 }
1341 }
1343 {
1345 {
1355 }
1356 else
1357 {
1360 }
1361 }
1363 /* Is there already a pg_statistic tuple for this attribute? */
1370 {
1371 /* Yes, replace it */
1374 values,
1375 nulls,
1376 replaces);
1379 }
1380 else
1381 {
1382 /* No, insert new tuple */
1385 }
1387 /* update indexes too */
1391 }
1394 }
1396 /*
1397 * Standard fetch function for use by compute_stats subroutines.
1398 *
1399 * This exists to provide some insulation between compute_stats routines
1400 * and the actual storage of the sample data.
1401 */
1402 static Datum
1404 {
1410 }
1412 /*
1413 * Fetch function for analyzing index expressions.
1414 *
1415 * We have not bothered to construct index tuples, instead the data is
1416 * just in Datum arrays.
1417 */
1418 static Datum
1420 {
1423 /* exprvals and exprnulls are already offset for proper column */
1427 }
1430 /*==========================================================================
1431 *
1432 * Code below this point represents the "standard" type-specific statistics
1433 * analysis algorithms. This code can be replaced on a per-data-type basis
1434 * by setting a nonzero value in pg_type.typanalyze.
1435 *
1436 *==========================================================================
1437 */
1440 /*
1441 * To avoid consuming too much memory during analysis and/or too much space
1442 * in the resulting pg_statistic rows, we ignore varlena datums that are wider
1443 * than WIDTH_THRESHOLD (after detoasting!). This is legitimate for MCV
1444 * and distinct-value calculations since a wide value is unlikely to be
1445 * duplicated at all, much less be a most-common value. For the same reason,
1446 * ignoring wide values will not affect our estimates of histogram bin
1447 * boundaries very much.
1448 */
1449 #define WIDTH_THRESHOLD 1024
1451 #define swapInt(a,b) do {int _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1452 #define swapDatum(a,b) do {Datum _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1454 /*
1455 * Extra information used by the default analysis routines
1456 */
1458 {
1465 {
1471 {
1477 {
1485 AnalyzeAttrFetchFunc fetchfunc,
1489 AnalyzeAttrFetchFunc fetchfunc,
1496 /*
1497 * std_typanalyze -- the default type-specific typanalyze function
1498 */
1499 static bool
1501 {
1503 Oid ltopr;
1504 Oid eqopr;
1507 /* If the attstattarget column is negative, use the default value */
1508 /* NB: it is okay to scribble on stats->attr since it's a copy */
1512 /* Look for default "<" and "=" operators for column's type */
1517 /* If column has no "=" operator, we can't do much of anything */
1521 /* Save the operator info for compute_stats routines */
1528 /*
1529 * Determine which standard statistics algorithm to use
1530 */
1532 {
1533 /* Seems to be a scalar datatype */
1535 /*--------------------
1536 * The following choice of minrows is based on the paper
1537 * "Random sampling for histogram construction: how much is enough?"
1538 * by Surajit Chaudhuri, Rajeev Motwani and Vivek Narasayya, in
1539 * Proceedings of ACM SIGMOD International Conference on Management
1540 * of Data, 1998, Pages 436-447. Their Corollary 1 to Theorem 5
1541 * says that for table size n, histogram size k, maximum relative
1542 * error in bin size f, and error probability gamma, the minimum
1543 * random sample size is
1544 * r = 4 * k * ln(2*n/gamma) / f^2
1545 * Taking f = 0.5, gamma = 0.01, n = 10^6 rows, we obtain
1546 * r = 305.82 * k
1547 * Note that because of the log function, the dependence on n is
1548 * quite weak; even at n = 10^12, a 300*k sample gives <= 0.66
1549 * bin size error with probability 0.99. So there's no real need to
1550 * scale for n, which is a good thing because we don't necessarily
1551 * know it at this point.
1552 *--------------------
1553 */
1555 }
1556 else
1557 {
1558 /* Can't do much but the minimal stuff */
1560 /* Might as well use the same minrows as above */
1562 }
1565 }
1567 /*
1568 * compute_minimal_stats() -- compute minimal column statistics
1569 *
1570 * We use this when we can find only an "=" operator for the datatype.
1571 *
1572 * We determine the fraction of non-null rows, the average width, the
1573 * most common values, and the (estimated) number of distinct values.
1574 *
1575 * The most common values are determined by brute force: we keep a list
1576 * of previously seen values, ordered by number of times seen, as we scan
1577 * the samples. A newly seen value is inserted just after the last
1578 * multiply-seen value, causing the bottommost (oldest) singly-seen value
1579 * to drop off the list. The accuracy of this method, and also its cost,
1580 * depend mainly on the length of the list we are willing to keep.
1581 */
1582 static void
1584 AnalyzeAttrFetchFunc fetchfunc,
1587 {
1597 FmgrInfo f_cmpeq;
1599 {
1600 Datum value;
1605 track_max;
1609 /*
1610 * We track up to 2*n values for an n-element MCV list; but at least 10
1611 */
1621 {
1622 Datum value;
1626 j;
1632 /* Check for null/nonnull */
1634 {
1635 null_cnt++;
1637 }
1638 nonnull_cnt++;
1640 /*
1641 * If it's a variable-width field, add up widths for average width
1642 * calculation. Note that if the value is toasted, we use the toasted
1643 * width. We don't bother with this calculation if it's a fixed-width
1644 * type.
1645 */
1647 {
1650 /*
1651 * If the value is toasted, we want to detoast it just once to
1652 * avoid repeated detoastings and resultant excess memory usage
1653 * during the comparisons. Also, check to see if the value is
1654 * excessively wide, and if so don't detoast at all --- just
1655 * ignore the value.
1656 */
1658 {
1659 toowide_cnt++;
1661 }
1663 }
1665 {
1666 /* must be cstring */
1668 }
1670 /*
1671 * See if the value matches anything we're already tracking.
1672 */
1676 {
1678 {
1681 }
1684 }
1687 {
1688 /* Found a match */
1690 /* This value may now need to "bubble up" in the track list */
1692 {
1695 j--;
1696 }
1697 }
1698 else
1699 {
1700 /* No match. Insert at head of count-1 list */
1702 track_cnt++;
1704 {
1707 }
1709 {
1712 }
1713 }
1714 }
1716 /* We can only compute real stats if we found some non-null values. */
1718 {
1720 summultiple;
1723 /* Do the simple null-frac and width stats */
1727 else
1730 /* Count the number of values we found multiple times */
1733 {
1737 }
1740 {
1741 /* If we found no repeated values, assume it's a unique column */
1743 }
1746 {
1747 /*
1748 * Our track list includes every value in the sample, and every
1749 * value appeared more than once. Assume the column has just
1750 * these values.
1751 */
1753 }
1754 else
1755 {
1756 /*----------
1757 * Estimate the number of distinct values using the estimator
1758 * proposed by Haas and Stokes in IBM Research Report RJ 10025:
1759 * n*d / (n - f1 + f1*n/N)
1760 * where f1 is the number of distinct values that occurred
1761 * exactly once in our sample of n rows (from a total of N),
1762 * and d is the total number of distinct values in the sample.
1763 * This is their Duj1 estimator; the other estimators they
1764 * recommend are considerably more complex, and are numerically
1765 * very unstable when n is much smaller than N.
1766 *
1767 * We assume (not very reliably!) that all the multiply-occurring
1768 * values are reflected in the final track[] list, and the other
1769 * nonnull values all appeared but once. (XXX this usually
1770 * results in a drastic overestimate of ndistinct. Can we do
1771 * any better?)
1772 *----------
1773 */
1777 denom,
1778 stadistinct;
1786 /* Clamp to sane range in case of roundoff error */
1792 }
1794 /*
1795 * If we estimated the number of distinct values at more than 10% of
1796 * the total row count (a very arbitrary limit), then assume that
1797 * stadistinct should scale with the row count rather than be a fixed
1798 * value.
1799 */
1803 /*
1804 * Decide how many values are worth storing as most-common values. If
1805 * we are able to generate a complete MCV list (all the values in the
1806 * sample will fit, and we think these are all the ones in the table),
1807 * then do so. Otherwise, store only those values that are
1808 * significantly more common than the (estimated) average. We set the
1809 * threshold rather arbitrarily at 25% more than average, with at
1810 * least 2 instances in the sample.
1811 */
1815 {
1816 /* Track list includes all values seen, and all will fit */
1818 }
1819 else
1820 {
1823 mincount;
1827 /* estimate # of occurrences in sample of a typical value */
1829 /* set minimum threshold count to store a value */
1836 {
1838 {
1841 }
1842 }
1843 }
1845 /* Generate MCV slot entry */
1847 {
1848 MemoryContext old_context;
1852 /* Must copy the target values into anl_context */
1857 {
1862 }
1871 /*
1872 * Accept the defaults for stats->statypid and others.
1873 * They have been set before we were called (see vacuum.h)
1874 */
1875 }
1876 }
1878 {
1879 /* We found only nulls; assume the column is entirely null */
1884 else
1887 }
1889 /* We don't need to bother cleaning up any of our temporary palloc's */
1890 }
1893 /*
1894 * compute_scalar_stats() -- compute column statistics
1895 *
1896 * We use this when we can find "=" and "<" operators for the datatype.
1897 *
1898 * We determine the fraction of non-null rows, the average width, the
1899 * most common values, the (estimated) number of distinct values, the
1900 * distribution histogram, and the correlation of physical to logical order.
1901 *
1902 * The desired stats can be determined fairly easily after sorting the
1903 * data values into order.
1904 */
1905 static void
1907 AnalyzeAttrFetchFunc fetchfunc,
1910 {
1921 Oid cmpFn;
1923 FmgrInfo f_cmpfn;
1940 /* Initial scan to find sortable values */
1942 {
1943 Datum value;
1950 /* Check for null/nonnull */
1952 {
1953 null_cnt++;
1955 }
1956 nonnull_cnt++;
1958 /*
1959 * If it's a variable-width field, add up widths for average width
1960 * calculation. Note that if the value is toasted, we use the toasted
1961 * width. We don't bother with this calculation if it's a fixed-width
1962 * type.
1963 */
1965 {
1968 /*
1969 * If the value is toasted, we want to detoast it just once to
1970 * avoid repeated detoastings and resultant excess memory usage
1971 * during the comparisons. Also, check to see if the value is
1972 * excessively wide, and if so don't detoast at all --- just
1973 * ignore the value.
1974 */
1976 {
1977 toowide_cnt++;
1979 }
1981 }
1983 {
1984 /* must be cstring */
1986 }
1988 /* Add it to the list to be sorted */
1992 values_cnt++;
1993 }
1995 /* We can only compute real stats if we found some sortable values. */
1997 {
2000 num_hist,
2001 dups_cnt;
2003 CompareScalarsContext cxt;
2005 /* Sort the collected values */
2012 /*
2013 * Now scan the values in order, find the most common ones, and also
2014 * accumulate ordering-correlation statistics.
2015 *
2016 * To determine which are most common, we first have to count the
2017 * number of duplicates of each value. The duplicates are adjacent in
2018 * the sorted list, so a brute-force approach is to compare successive
2019 * datum values until we find two that are not equal. However, that
2020 * requires N-1 invocations of the datum comparison routine, which are
2021 * completely redundant with work that was done during the sort. (The
2022 * sort algorithm must at some point have compared each pair of items
2023 * that are adjacent in the sorted order; otherwise it could not know
2024 * that it's ordered the pair correctly.) We exploit this by having
2025 * compare_scalars remember the highest tupno index that each
2026 * ScalarItem has been found equal to. At the end of the sort, a
2027 * ScalarItem's tupnoLink will still point to itself if and only if it
2028 * is the last item of its group of duplicates (since the group will
2029 * be ordered by tupno).
2030 */
2036 {
2040 dups_cnt++;
2042 {
2043 /* Reached end of duplicates of this value */
2044 ndistinct++;
2046 {
2047 nmultiple++;
2050 {
2051 /*
2052 * Found a new item for the mcv list; find its
2053 * position, bubbling down old items if needed. Loop
2054 * invariant is that j points at an empty/ replaceable
2055 * slot.
2056 */
2060 track_cnt++;
2062 {
2067 }
2070 }
2071 }
2073 }
2074 }
2077 /* Do the simple null-frac and width stats */
2081 else
2085 {
2086 /* If we found no repeated values, assume it's a unique column */
2088 }
2090 {
2091 /*
2092 * Every value in the sample appeared more than once. Assume the
2093 * column has just these values.
2094 */
2096 }
2097 else
2098 {
2099 /*----------
2100 * Estimate the number of distinct values using the estimator
2101 * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2102 * n*d / (n - f1 + f1*n/N)
2103 * where f1 is the number of distinct values that occurred
2104 * exactly once in our sample of n rows (from a total of N),
2105 * and d is the total number of distinct values in the sample.
2106 * This is their Duj1 estimator; the other estimators they
2107 * recommend are considerably more complex, and are numerically
2108 * very unstable when n is much smaller than N.
2109 *
2110 * Overwidth values are assumed to have been distinct.
2111 *----------
2112 */
2116 denom,
2117 stadistinct;
2125 /* Clamp to sane range in case of roundoff error */
2131 }
2133 /*
2134 * If we estimated the number of distinct values at more than 10% of
2135 * the total row count (a very arbitrary limit), then assume that
2136 * stadistinct should scale with the row count rather than be a fixed
2137 * value.
2138 */
2142 /*
2143 * Decide how many values are worth storing as most-common values. If
2144 * we are able to generate a complete MCV list (all the values in the
2145 * sample will fit, and we think these are all the ones in the table),
2146 * then do so. Otherwise, store only those values that are
2147 * significantly more common than the (estimated) average. We set the
2148 * threshold rather arbitrarily at 25% more than average, with at
2149 * least 2 instances in the sample. Also, we won't suppress values
2150 * that have a frequency of at least 1/K where K is the intended
2151 * number of histogram bins; such values might otherwise cause us to
2152 * emit duplicate histogram bin boundaries.
2153 */
2157 {
2158 /* Track list includes all values seen, and all will fit */
2160 }
2161 else
2162 {
2165 mincount,
2166 maxmincount;
2170 /* estimate # of occurrences in sample of a typical value */
2172 /* set minimum threshold count to store a value */
2176 /* don't let threshold exceed 1/K, however */
2183 {
2185 {
2188 }
2189 }
2190 }
2192 /* Generate MCV slot entry */
2194 {
2195 MemoryContext old_context;
2199 /* Must copy the target values into anl_context */
2204 {
2209 }
2218 /*
2219 * Accept the defaults for stats->statypid and others.
2220 * They have been set before we were called (see vacuum.h)
2221 */
2222 slot_idx++;
2223 }
2225 /*
2226 * Generate a histogram slot entry if there are at least two distinct
2227 * values not accounted for in the MCV list. (This ensures the
2228 * histogram won't collapse to empty or a singleton.)
2229 */
2234 {
2235 MemoryContext old_context;
2239 /* Sort the MCV items into position order to speed next loop */
2243 /*
2244 * Collapse out the MCV items from the values[] array.
2245 *
2246 * Note we destroy the values[] array here... but we don't need it
2247 * for anything more. We do, however, still need values_cnt.
2248 * nvals will be the number of remaining entries in values[].
2249 */
2251 {
2253 dest;
2259 {
2263 {
2267 {
2268 /* advance past this MCV item */
2270 j++;
2272 }
2274 }
2275 else
2281 }
2283 }
2284 else
2288 /* Must copy the target values into anl_context */
2292 {
2299 }
2306 /*
2307 * Accept the defaults for stats->statypid and others.
2308 * They have been set before we were called (see vacuum.h)
2309 */
2310 slot_idx++;
2311 }
2313 /* Generate a correlation entry if there are multiple values */
2315 {
2316 MemoryContext old_context;
2319 corr_x2sum;
2321 /* Must copy the target values into anl_context */
2326 /*----------
2327 * Since we know the x and y value sets are both
2328 * 0, 1, ..., values_cnt-1
2329 * we have sum(x) = sum(y) =
2330 * (values_cnt-1)*values_cnt / 2
2331 * and sum(x^2) = sum(y^2) =
2332 * (values_cnt-1)*values_cnt*(2*values_cnt-1) / 6.
2333 *----------
2334 */
2340 /* And the correlation coefficient reduces to */
2348 slot_idx++;
2349 }
2350 }
2352 {
2353 /* We found only nulls; assume the column is entirely null */
2358 else
2361 }
2363 /* We don't need to bother cleaning up any of our temporary palloc's */
2364 }
2366 /*
2367 * qsort_arg comparator for sorting ScalarItems
2368 *
2369 * Aside from sorting the items, we update the tupnoLink[] array
2370 * whenever two ScalarItems are found to contain equal datums. The array
2371 * is indexed by tupno; for each ScalarItem, it contains the highest
2372 * tupno that that item's datum has been found to be equal to. This allows
2373 * us to avoid additional comparisons in compute_scalar_stats().
2374 */
2375 static int
2377 {
2383 int32 compare;
2390 /*
2391 * The two datums are equal, so update cxt->tupnoLink[].
2392 */
2398 /*
2399 * For equal datums, sort by tupno
2400 */
2402 }
2404 /*
2405 * qsort comparator for sorting ScalarMCVItems by position
2406 */
2407 static int
2409 {
2414 }