| blob | ab5a2671b221e75bcd5d2bf86a2a400dbc5c8c64 |
1 /*-------------------------------------------------------------------------
2 *
3 * costsize.c
4 * Routines to compute (and set) relation sizes and path costs
5 *
6 * Path costs are measured in arbitrary units established by these basic
7 * parameters:
8 *
9 * seq_page_cost Cost of a sequential page fetch
10 * random_page_cost Cost of a non-sequential page fetch
11 * cpu_tuple_cost Cost of typical CPU time to process a tuple
12 * cpu_index_tuple_cost Cost of typical CPU time to process an index tuple
13 * cpu_operator_cost Cost of CPU time to execute an operator or function
14 *
15 * We expect that the kernel will typically do some amount of read-ahead
16 * optimization; this in conjunction with seek costs means that seq_page_cost
17 * is normally considerably less than random_page_cost. (However, if the
18 * database is fully cached in RAM, it is reasonable to set them equal.)
19 *
20 * We also use a rough estimate "effective_cache_size" of the number of
21 * disk pages in Postgres + OS-level disk cache. (We can't simply use
22 * NBuffers for this purpose because that would ignore the effects of
23 * the kernel's disk cache.)
24 *
25 * Obviously, taking constants for these values is an oversimplification,
26 * but it's tough enough to get any useful estimates even at this level of
27 * detail. Note that all of these parameters are user-settable, in case
28 * the default values are drastically off for a particular platform.
29 *
30 * We compute two separate costs for each path:
31 * total_cost: total estimated cost to fetch all tuples
32 * startup_cost: cost that is expended before first tuple is fetched
33 * In some scenarios, such as when there is a LIMIT or we are implementing
34 * an EXISTS(...) sub-select, it is not necessary to fetch all tuples of the
35 * path's result. A caller can estimate the cost of fetching a partial
36 * result by interpolating between startup_cost and total_cost. In detail:
37 * actual_cost = startup_cost +
38 * (total_cost - startup_cost) * tuples_to_fetch / path->parent->rows;
39 * Note that a base relation's rows count (and, by extension, plan_rows for
40 * plan nodes below the LIMIT node) are set without regard to any LIMIT, so
41 * that this equation works properly. (Also, these routines guarantee not to
42 * set the rows count to zero, so there will be no zero divide.) The LIMIT is
43 * applied as a top-level plan node.
44 *
45 * For largely historical reasons, most of the routines in this module use
46 * the passed result Path only to store their startup_cost and total_cost
47 * results into. All the input data they need is passed as separate
48 * parameters, even though much of it could be extracted from the Path.
49 * An exception is made for the cost_XXXjoin() routines, which expect all
50 * the non-cost fields of the passed XXXPath to be filled in.
51 *
52 *
53 * Portions Copyright (c) 1996-2009, PostgreSQL Global Development Group
54 * Portions Copyright (c) 1994, Regents of the University of California
55 *
56 * IDENTIFICATION
57 * $PostgreSQL$
58 *
59 *-------------------------------------------------------------------------
60 */
64 #include <math.h>
81 #define LOG2(x) (log(x) / 0.693147180559945)
83 /*
84 * Some Paths return less than the nominal number of rows of their parent
85 * relations; join nodes need to do this to get the correct input count:
86 */
87 #define PATH_ROWS(path) \
88 (IsA(path, UniquePath) ? \
89 ((UniquePath *) (path))->rows : \
90 (path)->parent->rows)
114 {
116 QualCost total;
135 /*
136 * clamp_row_est
137 * Force a row-count estimate to a sane value.
138 */
139 double
141 {
142 /*
143 * Force estimate to be at least one row, to make explain output look
144 * better and to avoid possible divide-by-zero when interpolating costs.
145 * Make it an integer, too.
146 */
149 else
153 }
156 /*
157 * cost_seqscan
158 * Determines and returns the cost of scanning a relation sequentially.
159 */
160 void
163 {
166 Cost cpu_per_tuple;
168 /* Should only be applied to base relations */
175 /*
176 * disk costs
177 */
180 /* CPU costs */
187 }
189 /*
190 * cost_index
191 * Determines and returns the cost of scanning a relation using an index.
192 *
193 * 'index' is the index to be used
194 * 'indexQuals' is the list of applicable qual clauses (implicit AND semantics)
195 * 'outer_rel' is the outer relation when we are considering using the index
196 * scan as the inside of a nestloop join (hence, some of the indexQuals
197 * are join clauses, and we should expect repeated scans of the index);
198 * NULL for a plain index scan
199 *
200 * cost_index() takes an IndexPath not just a Path, because it sets a few
201 * additional fields of the IndexPath besides startup_cost and total_cost.
202 * These fields are needed if the IndexPath is used in a BitmapIndexScan.
203 *
204 * NOTE: 'indexQuals' must contain only clauses usable as index restrictions.
205 * Any additional quals evaluated as qpquals may reduce the number of returned
206 * tuples, but they won't reduce the number of tuples we have to fetch from
207 * the table, so they don't reduce the scan cost.
208 *
209 * NOTE: as of 8.0, indexQuals is a list of RestrictInfo nodes, where formerly
210 * it was a list of bare clause expressions.
211 */
212 void
217 {
221 Cost indexStartupCost;
222 Cost indexTotalCost;
223 Selectivity indexSelectivity;
225 csquared;
226 Cost min_IO_cost,
227 max_IO_cost;
228 Cost cpu_per_tuple;
232 /* Should only be applied to base relations */
241 /*
242 * Call index-access-method-specific code to estimate the processing cost
243 * for scanning the index, as well as the selectivity of the index (ie,
244 * the fraction of main-table tuples we will have to retrieve) and its
245 * correlation to the main-table tuple order.
246 */
257 /*
258 * Save amcostestimate's results for possible use in bitmap scan planning.
259 * We don't bother to save indexStartupCost or indexCorrelation, because a
260 * bitmap scan doesn't care about either.
261 */
265 /* all costs for touching index itself included here */
269 /* estimate number of main-table tuples fetched */
272 /*----------
273 * Estimate number of main-table pages fetched, and compute I/O cost.
274 *
275 * When the index ordering is uncorrelated with the table ordering,
276 * we use an approximation proposed by Mackert and Lohman (see
277 * index_pages_fetched() for details) to compute the number of pages
278 * fetched, and then charge random_page_cost per page fetched.
279 *
280 * When the index ordering is exactly correlated with the table ordering
281 * (just after a CLUSTER, for example), the number of pages fetched should
282 * be exactly selectivity * table_size. What's more, all but the first
283 * will be sequential fetches, not the random fetches that occur in the
284 * uncorrelated case. So if the number of pages is more than 1, we
285 * ought to charge
286 * random_page_cost + (pages_fetched - 1) * seq_page_cost
287 * For partially-correlated indexes, we ought to charge somewhere between
288 * these two estimates. We currently interpolate linearly between the
289 * estimates based on the correlation squared (XXX is that appropriate?).
290 *----------
291 */
293 {
294 /*
295 * For repeated indexscans, the appropriate estimate for the
296 * uncorrelated case is to scale up the number of tuples fetched in
297 * the Mackert and Lohman formula by the number of scans, so that we
298 * estimate the number of pages fetched by all the scans; then
299 * pro-rate the costs for one scan. In this case we assume all the
300 * fetches are random accesses.
301 */
307 root);
311 /*
312 * In the perfectly correlated case, the number of pages touched by
313 * each scan is selectivity * table_size, and we can use the Mackert
314 * and Lohman formula at the page level to estimate how much work is
315 * saved by caching across scans. We still assume all the fetches are
316 * random, though, which is an overestimate that's hard to correct for
317 * without double-counting the cache effects. (But in most cases
318 * where such a plan is actually interesting, only one page would get
319 * fetched per scan anyway, so it shouldn't matter much.)
320 */
326 root);
329 }
330 else
331 {
332 /*
333 * Normal case: apply the Mackert and Lohman formula, and then
334 * interpolate between that and the correlation-derived result.
335 */
339 root);
341 /* max_IO_cost is for the perfectly uncorrelated case (csquared=0) */
344 /* min_IO_cost is for the perfectly correlated case (csquared=1) */
349 }
351 /*
352 * Now interpolate based on estimated index order correlation to get total
353 * disk I/O cost for main table accesses.
354 */
359 /*
360 * Estimate CPU costs per tuple.
361 *
362 * Normally the indexquals will be removed from the list of restriction
363 * clauses that we have to evaluate as qpquals, so we should subtract
364 * their costs from baserestrictcost. But if we are doing a join then
365 * some of the indexquals are join clauses and shouldn't be subtracted.
366 * Rather than work out exactly how much to subtract, we don't subtract
367 * anything.
368 */
373 {
374 QualCost index_qual_cost;
377 /* any startup cost still has to be paid ... */
379 }
385 }
387 /*
388 * index_pages_fetched
389 * Estimate the number of pages actually fetched after accounting for
390 * cache effects.
391 *
392 * We use an approximation proposed by Mackert and Lohman, "Index Scans
393 * Using a Finite LRU Buffer: A Validated I/O Model", ACM Transactions
394 * on Database Systems, Vol. 14, No. 3, September 1989, Pages 401-424.
395 * The Mackert and Lohman approximation is that the number of pages
396 * fetched is
397 * PF =
398 * min(2TNs/(2T+Ns), T) when T <= b
399 * 2TNs/(2T+Ns) when T > b and Ns <= 2Tb/(2T-b)
400 * b + (Ns - 2Tb/(2T-b))*(T-b)/T when T > b and Ns > 2Tb/(2T-b)
401 * where
402 * T = # pages in table
403 * N = # tuples in table
404 * s = selectivity = fraction of table to be scanned
405 * b = # buffer pages available (we include kernel space here)
406 *
407 * We assume that effective_cache_size is the total number of buffer pages
408 * available for the whole query, and pro-rate that space across all the
409 * tables in the query and the index currently under consideration. (This
410 * ignores space needed for other indexes used by the query, but since we
411 * don't know which indexes will get used, we can't estimate that very well;
412 * and in any case counting all the tables may well be an overestimate, since
413 * depending on the join plan not all the tables may be scanned concurrently.)
414 *
415 * The product Ns is the number of tuples fetched; we pass in that
416 * product rather than calculating it here. "pages" is the number of pages
417 * in the object under consideration (either an index or a table).
418 * "index_pages" is the amount to add to the total table space, which was
419 * computed for us by query_planner.
420 *
421 * Caller is expected to have ensured that tuples_fetched is greater than zero
422 * and rounded to integer (see clamp_row_est). The result will likewise be
423 * greater than zero and integral.
424 */
425 double
428 {
432 b;
434 /* T is # pages in table, but don't allow it to be zero */
437 /* Compute number of pages assumed to be competing for cache space */
442 /* b is pro-rated share of effective_cache_size */
445 /* force it positive and integral */
448 else
451 /* This part is the Mackert and Lohman formula */
453 {
454 pages_fetched =
458 else
460 }
461 else
462 {
467 {
468 pages_fetched =
470 }
471 else
472 {
473 pages_fetched =
475 }
477 }
479 }
481 /*
482 * get_indexpath_pages
483 * Determine the total size of the indexes used in a bitmap index path.
484 *
485 * Note: if the same index is used more than once in a bitmap tree, we will
486 * count it multiple times, which perhaps is the wrong thing ... but it's
487 * not completely clear, and detecting duplicates is difficult, so ignore it
488 * for now.
489 */
490 static double
492 {
497 {
501 {
503 }
504 }
506 {
510 {
512 }
513 }
515 {
519 }
520 else
524 }
526 /*
527 * cost_bitmap_heap_scan
528 * Determines and returns the cost of scanning a relation using a bitmap
529 * index-then-heap plan.
530 *
531 * 'baserel' is the relation to be scanned
532 * 'bitmapqual' is a tree of IndexPaths, BitmapAndPaths, and BitmapOrPaths
533 * 'outer_rel' is the outer relation when we are considering using the bitmap
534 * scan as the inside of a nestloop join (hence, some of the indexQuals
535 * are join clauses, and we should expect repeated scans of the table);
536 * NULL for a plain bitmap scan
537 *
538 * Note: if this is a join inner path, the component IndexPaths in bitmapqual
539 * should have been costed accordingly.
540 */
541 void
544 {
547 Cost indexTotalCost;
548 Selectivity indexSelectivity;
549 Cost cpu_per_tuple;
550 Cost cost_per_page;
555 /* Should only be applied to base relations */
563 /*
564 * Fetch total cost of obtaining the bitmap, as well as its total
565 * selectivity.
566 */
571 /*
572 * Estimate number of main-table pages fetched.
573 */
579 {
580 /*
581 * For repeated bitmap scans, scale up the number of tuples fetched in
582 * the Mackert and Lohman formula by the number of scans, so that we
583 * estimate the number of pages fetched by all the scans. Then
584 * pro-rate for one scan.
585 */
591 root);
593 }
594 else
595 {
596 /*
597 * For a single scan, the number of heap pages that need to be fetched
598 * is the same as the Mackert and Lohman formula for the case T <= b
599 * (ie, no re-reads needed).
600 */
602 }
605 else
608 /*
609 * For small numbers of pages we should charge random_page_cost apiece,
610 * while if nearly all the table's pages are being read, it's more
611 * appropriate to charge seq_page_cost apiece. The effect is nonlinear,
612 * too. For lack of a better idea, interpolate like this to determine the
613 * cost per page.
614 */
618 else
623 /*
624 * Estimate CPU costs per tuple.
625 *
626 * Often the indexquals don't need to be rechecked at each tuple ... but
627 * not always, especially not if there are enough tuples involved that the
628 * bitmaps become lossy. For the moment, just assume they will be
629 * rechecked always.
630 */
638 }
640 /*
641 * cost_bitmap_tree_node
642 * Extract cost and selectivity from a bitmap tree node (index/and/or)
643 */
644 void
646 {
648 {
652 /*
653 * Charge a small amount per retrieved tuple to reflect the costs of
654 * manipulating the bitmap. This is mostly to make sure that a bitmap
655 * scan doesn't look to be the same cost as an indexscan to retrieve a
656 * single tuple.
657 */
659 }
661 {
664 }
666 {
669 }
670 else
671 {
674 }
675 }
677 /*
678 * cost_bitmap_and_node
679 * Estimate the cost of a BitmapAnd node
680 *
681 * Note that this considers only the costs of index scanning and bitmap
682 * creation, not the eventual heap access. In that sense the object isn't
683 * truly a Path, but it has enough path-like properties (costs in particular)
684 * to warrant treating it as one.
685 */
686 void
688 {
689 Cost totalCost;
690 Selectivity selec;
693 /*
694 * We estimate AND selectivity on the assumption that the inputs are
695 * independent. This is probably often wrong, but we don't have the info
696 * to do better.
697 *
698 * The runtime cost of the BitmapAnd itself is estimated at 100x
699 * cpu_operator_cost for each tbm_intersect needed. Probably too small,
700 * definitely too simplistic?
701 */
705 {
707 Cost subCost;
708 Selectivity subselec;
717 }
721 }
723 /*
724 * cost_bitmap_or_node
725 * Estimate the cost of a BitmapOr node
726 *
727 * See comments for cost_bitmap_and_node.
728 */
729 void
731 {
732 Cost totalCost;
733 Selectivity selec;
736 /*
737 * We estimate OR selectivity on the assumption that the inputs are
738 * non-overlapping, since that's often the case in "x IN (list)" type
739 * situations. Of course, we clamp to 1.0 at the end.
740 *
741 * The runtime cost of the BitmapOr itself is estimated at 100x
742 * cpu_operator_cost for each tbm_union needed. Probably too small,
743 * definitely too simplistic? We are aware that the tbm_unions are
744 * optimized out when the inputs are BitmapIndexScans.
745 */
749 {
751 Cost subCost;
752 Selectivity subselec;
762 }
766 }
768 /*
769 * cost_tidscan
770 * Determines and returns the cost of scanning a relation using TIDs.
771 */
772 void
775 {
779 Cost cpu_per_tuple;
780 QualCost tid_qual_cost;
784 /* Should only be applied to base relations */
788 /* Count how many tuples we expect to retrieve */
791 {
793 {
794 /* Each element of the array yields 1 tuple */
799 }
801 {
802 /* CURRENT OF yields 1 tuple */
804 ntuples++;
805 }
806 else
807 {
808 /* It's just CTID = something, count 1 tuple */
809 ntuples++;
810 }
811 }
813 /*
814 * We must force TID scan for WHERE CURRENT OF, because only nodeTidscan.c
815 * understands how to do it correctly. Therefore, honor enable_tidscan
816 * only when CURRENT OF isn't present. Also note that cost_qual_eval
817 * counts a CurrentOfExpr as having startup cost disable_cost, which we
818 * subtract off here; that's to prevent other plan types such as seqscan
819 * from winning.
820 */
822 {
825 }
829 /*
830 * The TID qual expressions will be computed once, any other baserestrict
831 * quals once per retrived tuple.
832 */
835 /* disk costs --- assume each tuple on a different page */
838 /* CPU costs */
847 }
849 /*
850 * cost_subqueryscan
851 * Determines and returns the cost of scanning a subquery RTE.
852 */
853 void
855 {
856 Cost startup_cost;
857 Cost run_cost;
858 Cost cpu_per_tuple;
860 /* Should only be applied to base relations that are subqueries */
864 /*
865 * Cost of path is cost of evaluating the subplan, plus cost of evaluating
866 * any restriction clauses that will be attached to the SubqueryScan node,
867 * plus cpu_tuple_cost to account for selection and projection overhead.
868 */
878 }
880 /*
881 * cost_functionscan
882 * Determines and returns the cost of scanning a function RTE.
883 */
884 void
886 {
889 Cost cpu_per_tuple;
891 QualCost exprcost;
893 /* Should only be applied to base relations that are functions */
898 /* Estimate costs of executing the function expression */
904 /* Add scanning CPU costs */
911 }
913 /*
914 * cost_valuesscan
915 * Determines and returns the cost of scanning a VALUES RTE.
916 */
917 void
919 {
922 Cost cpu_per_tuple;
924 /* Should only be applied to base relations that are values lists */
928 /*
929 * For now, estimate list evaluation cost at one operator eval per list
930 * (probably pretty bogus, but is it worth being smarter?)
931 */
934 /* Add scanning CPU costs */
941 }
943 /*
944 * cost_ctescan
945 * Determines and returns the cost of scanning a CTE RTE.
946 *
947 * Note: this is used for both self-reference and regular CTEs; the
948 * possible cost differences are below the threshold of what we could
949 * estimate accurately anyway. Note that the costs of evaluating the
950 * referenced CTE query are added into the final plan as initplan costs,
951 * and should NOT be counted here.
952 */
953 void
955 {
958 Cost cpu_per_tuple;
960 /* Should only be applied to base relations that are CTEs */
964 /* Charge one CPU tuple cost per row for tuplestore manipulation */
967 /* Add scanning CPU costs */
974 }
976 /*
977 * cost_recursive_union
978 * Determines and returns the cost of performing a recursive union,
979 * and also the estimated output size.
980 *
981 * We are given Plans for the nonrecursive and recursive terms.
982 *
983 * Note that the arguments and output are Plans, not Paths as in most of
984 * the rest of this module. That's because we don't bother setting up a
985 * Path representation for recursive union --- we have only one way to do it.
986 */
987 void
989 {
990 Cost startup_cost;
991 Cost total_cost;
994 /* We probably have decent estimates for the non-recursive term */
999 /*
1000 * We arbitrarily assume that about 10 recursive iterations will be
1001 * needed, and that we've managed to get a good fix on the cost and output
1002 * size of each one of them. These are mighty shaky assumptions but it's
1003 * hard to see how to do better.
1004 */
1008 /*
1009 * Also charge cpu_tuple_cost per row to account for the costs of
1010 * manipulating the tuplestores. (We don't worry about possible
1011 * spill-to-disk costs.)
1012 */
1019 }
1021 /*
1022 * cost_sort
1023 * Determines and returns the cost of sorting a relation, including
1024 * the cost of reading the input data.
1025 *
1026 * If the total volume of data to sort is less than work_mem, we will do
1027 * an in-memory sort, which requires no I/O and about t*log2(t) tuple
1028 * comparisons for t tuples.
1029 *
1030 * If the total volume exceeds work_mem, we switch to a tape-style merge
1031 * algorithm. There will still be about t*log2(t) tuple comparisons in
1032 * total, but we will also need to write and read each tuple once per
1033 * merge pass. We expect about ceil(logM(r)) merge passes where r is the
1034 * number of initial runs formed and M is the merge order used by tuplesort.c.
1035 * Since the average initial run should be about twice work_mem, we have
1036 * disk traffic = 2 * relsize * ceil(logM(p / (2*work_mem)))
1037 * cpu = comparison_cost * t * log2(t)
1038 *
1039 * If the sort is bounded (i.e., only the first k result tuples are needed)
1040 * and k tuples can fit into work_mem, we use a heap method that keeps only
1041 * k tuples in the heap; this will require about t*log2(k) tuple comparisons.
1042 *
1043 * The disk traffic is assumed to be 3/4ths sequential and 1/4th random
1044 * accesses (XXX can't we refine that guess?)
1045 *
1046 * We charge two operator evals per tuple comparison, which should be in
1047 * the right ballpark in most cases.
1048 *
1049 * 'pathkeys' is a list of sort keys
1050 * 'input_cost' is the total cost for reading the input data
1051 * 'tuples' is the number of tuples in the relation
1052 * 'width' is the average tuple width in bytes
1053 * 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
1054 *
1055 * NOTE: some callers currently pass NIL for pathkeys because they
1056 * can't conveniently supply the sort keys. Since this routine doesn't
1057 * currently do anything with pathkeys anyway, that doesn't matter...
1058 * but if it ever does, it should react gracefully to lack of key data.
1059 * (Actually, the thing we'd most likely be interested in is just the number
1060 * of sort keys, which all callers *could* supply.)
1061 */
1062 void
1066 {
1077 /*
1078 * We want to be sure the cost of a sort is never estimated as zero, even
1079 * if passed-in tuple count is zero. Besides, mustn't do log(0)...
1080 */
1084 /* Do we have a useful LIMIT? */
1086 {
1089 }
1090 else
1091 {
1094 }
1097 {
1098 /*
1099 * We'll have to use a disk-based sort of all the tuples
1100 */
1107 /*
1108 * CPU costs
1109 *
1110 * Assume about two operator evals per tuple comparison and N log2 N
1111 * comparisons
1112 */
1115 /* Disk costs */
1117 /* Compute logM(r) as log(r) / log(M) */
1120 else
1123 /* Assume 3/4ths of accesses are sequential, 1/4th are not */
1126 }
1128 {
1129 /*
1130 * We'll use a bounded heap-sort keeping just K tuples in memory, for
1131 * a total number of tuple comparisons of N log2 K; but the constant
1132 * factor is a bit higher than for quicksort. Tweak it so that the
1133 * cost curve is continuous at the crossover point.
1134 */
1136 }
1137 else
1138 {
1139 /* We'll use plain quicksort on all the input tuples */
1141 }
1143 /*
1144 * Also charge a small amount (arbitrarily set equal to operator cost) per
1145 * extracted tuple. Note it's correct to use tuples not output_tuples
1146 * here --- the upper LIMIT will pro-rate the run cost so we'd be double
1147 * counting the LIMIT otherwise.
1148 */
1153 }
1155 /*
1156 * sort_exceeds_work_mem
1157 * Given a finished Sort plan node, detect whether it is expected to
1158 * spill to disk (ie, will need more than work_mem workspace)
1159 *
1160 * This assumes there will be no available LIMIT.
1161 */
1162 bool
1164 {
1170 }
1172 /*
1173 * cost_material
1174 * Determines and returns the cost of materializing a relation, including
1175 * the cost of reading the input data.
1176 *
1177 * If the total volume of data to materialize exceeds work_mem, we will need
1178 * to write it to disk, so the cost is much higher in that case.
1179 */
1180 void
1183 {
1189 /* disk costs */
1191 {
1194 /* We'll write during startup and read during retrieval */
1197 }
1199 /*
1200 * Charge a very small amount per inserted tuple, to reflect bookkeeping
1201 * costs. We use cpu_tuple_cost/10 for this. This is needed to break the
1202 * tie that would otherwise exist between nestloop with A outer,
1203 * materialized B inner and nestloop with B outer, materialized A inner.
1204 * The extra cost ensures we'll prefer materializing the smaller rel.
1205 */
1208 /*
1209 * Also charge a small amount per extracted tuple. We use cpu_tuple_cost
1210 * so that it doesn't appear worthwhile to materialize a bare seqscan.
1211 */
1216 }
1218 /*
1219 * cost_agg
1220 * Determines and returns the cost of performing an Agg plan node,
1221 * including the cost of its input.
1222 *
1223 * Note: when aggstrategy == AGG_SORTED, caller must ensure that input costs
1224 * are for appropriately-sorted input.
1225 */
1226 void
1232 {
1233 Cost startup_cost;
1234 Cost total_cost;
1236 /*
1237 * We charge one cpu_operator_cost per aggregate function per input tuple,
1238 * and another one per output tuple (corresponding to transfn and finalfn
1239 * calls respectively). If we are grouping, we charge an additional
1240 * cpu_operator_cost per grouping column per input tuple for grouping
1241 * comparisons.
1242 *
1243 * We will produce a single output tuple if not grouping, and a tuple per
1244 * group otherwise. We charge cpu_tuple_cost for each output tuple.
1245 *
1246 * Note: in this cost model, AGG_SORTED and AGG_HASHED have exactly the
1247 * same total CPU cost, but AGG_SORTED has lower startup cost. If the
1248 * input path is already sorted appropriately, AGG_SORTED should be
1249 * preferred (since it has no risk of memory overflow). This will happen
1250 * as long as the computed total costs are indeed exactly equal --- but if
1251 * there's roundoff error we might do the wrong thing. So be sure that
1252 * the computations below form the same intermediate values in the same
1253 * order.
1254 *
1255 * Note: ideally we should use the pg_proc.procost costs of each
1256 * aggregate's component functions, but for now that seems like an
1257 * excessive amount of work.
1258 */
1260 {
1263 /* we aren't grouping */
1265 }
1267 {
1268 /* Here we are able to deliver output on-the-fly */
1271 /* calcs phrased this way to match HASHED case, see note above */
1276 }
1277 else
1278 {
1279 /* must be AGG_HASHED */
1286 }
1290 }
1292 /*
1293 * cost_windowagg
1294 * Determines and returns the cost of performing a WindowAgg plan node,
1295 * including the cost of its input.
1296 *
1297 * Input is assumed already properly sorted.
1298 */
1299 void
1304 {
1305 Cost startup_cost;
1306 Cost total_cost;
1311 /*
1312 * We charge one cpu_operator_cost per window function per tuple (often a
1313 * drastic underestimate, but without a way to gauge how many tuples the
1314 * window function will fetch, it's hard to do better). We also charge
1315 * cpu_operator_cost per grouping column per tuple for grouping
1316 * comparisons, plus cpu_tuple_cost per tuple for general overhead.
1317 */
1324 }
1326 /*
1327 * cost_group
1328 * Determines and returns the cost of performing a Group plan node,
1329 * including the cost of its input.
1330 *
1331 * Note: caller must ensure that input costs are for appropriately-sorted
1332 * input.
1333 */
1334 void
1339 {
1340 Cost startup_cost;
1341 Cost total_cost;
1346 /*
1347 * Charge one cpu_operator_cost per comparison per input tuple. We assume
1348 * all columns get compared at most of the tuples.
1349 */
1354 }
1356 /*
1357 * If a nestloop's inner path is an indexscan, be sure to use its estimated
1358 * output row count, which may be lower than the restriction-clause-only row
1359 * count of its parent. (We don't include this case in the PATH_ROWS macro
1360 * because it applies *only* to a nestloop's inner relation.) We have to
1361 * be prepared to recurse through Append nodes in case of an appendrel.
1362 */
1363 static double
1365 {
1373 {
1378 {
1380 }
1381 }
1382 else
1386 }
1388 /*
1389 * cost_nestloop
1390 * Determines and returns the cost of joining two relations using the
1391 * nested loop algorithm.
1392 *
1393 * 'path' is already filled in except for the cost fields
1394 * 'sjinfo' is extra info about the join for selectivity estimation
1395 */
1396 void
1398 {
1403 Cost inner_run_cost;
1404 Cost cpu_per_tuple;
1405 QualCost restrict_qual_cost;
1409 Selectivity outer_match_frac;
1410 Selectivity match_count;
1416 /* cost of source data */
1418 /*
1419 * NOTE: clearly, we must pay both outer and inner paths' startup_cost
1420 * before we can start returning tuples, so the join's startup cost is
1421 * their sum. What's not so clear is whether the inner path's
1422 * startup_cost must be paid again on each rescan of the inner path. This
1423 * is not true if the inner path is materialized or is a hashjoin, but
1424 * probably is true otherwise.
1425 */
1430 {
1431 /* charge only run cost for each iteration of inner path */
1432 }
1433 else
1434 {
1435 /*
1436 * charge startup cost for each iteration of inner path, except we
1437 * already charged the first startup_cost in our own startup
1438 */
1440 }
1447 {
1449 Selectivity inner_scan_frac;
1451 /*
1452 * SEMI or ANTI join: executor will stop after first match.
1453 *
1454 * For an outer-rel row that has at least one match, we can expect the
1455 * inner scan to stop after a fraction 1/(match_count+1) of the inner
1456 * rows, if the matches are evenly distributed. Since they probably
1457 * aren't quite evenly distributed, we apply a fuzz factor of 2.0 to
1458 * that fraction. (If we used a larger fuzz factor, we'd have to
1459 * clamp inner_scan_frac to at most 1.0; but since match_count is at
1460 * least 1, no such clamp is needed now.)
1461 */
1465 /* Add inner run cost for outer tuples having matches */
1468 /* Compute number of tuples processed (not number emitted!) */
1471 /*
1472 * For unmatched outer-rel rows, there are two cases. If the inner
1473 * path is an indexscan using all the joinquals as indexquals, then an
1474 * unmatched row results in an indexscan returning no rows, which is
1475 * probably quite cheap. We estimate this case as the same cost to
1476 * return the first tuple of a nonempty scan. Otherwise, the executor
1477 * will have to scan the whole inner rel; not so cheap.
1478 */
1480 {
1483 /* We won't be evaluating any quals at all for these rows */
1484 }
1485 else
1486 {
1488 inner_run_cost;
1490 inner_path_rows;
1491 }
1492 }
1493 else
1494 {
1495 /* Normal case; we'll scan whole input rel for each outer row */
1498 /* Compute number of tuples processed (not number emitted!) */
1500 }
1502 /* CPU costs */
1510 }
1512 /*
1513 * cost_mergejoin
1514 * Determines and returns the cost of joining two relations using the
1515 * merge join algorithm.
1516 *
1517 * 'path' is already filled in except for the cost fields
1518 * 'sjinfo' is extra info about the join for selectivity estimation
1519 *
1520 * Notes: path's mergeclauses should be a subset of the joinrestrictinfo list;
1521 * outersortkeys and innersortkeys are lists of the keys to be used
1522 * to sort the outer and inner relations, or NIL if no explicit
1523 * sort is needed because the source path is already ordered.
1524 */
1525 void
1527 {
1535 Cost cpu_per_tuple;
1536 QualCost merge_qual_cost;
1537 QualCost qp_qual_cost;
1541 inner_rows,
1542 outer_skip_rows,
1543 inner_skip_rows;
1545 rescannedtuples;
1547 Selectivity outerstartsel,
1548 outerendsel,
1549 innerstartsel,
1550 innerendsel;
1553 /* Protect some assumptions below that rowcounts aren't zero */
1562 /*
1563 * Compute cost of the mergequals and qpquals (other restriction clauses)
1564 * separately.
1565 */
1571 /*
1572 * Get approx # tuples passing the mergequals. We use approx_tuple_count
1573 * here because we need an estimate done with JOIN_INNER semantics.
1574 */
1577 /*
1578 * When there are equal merge keys in the outer relation, the mergejoin
1579 * must rescan any matching tuples in the inner relation. This means
1580 * re-fetching inner tuples. Our cost model for this is that a re-fetch
1581 * costs the same as an original fetch, which is probably an overestimate;
1582 * but on the other hand we ignore the bookkeeping costs of mark/restore.
1583 * Not clear if it's worth developing a more refined model.
1584 *
1585 * For regular inner and outer joins, the number of re-fetches can be
1586 * estimated approximately as size of merge join output minus size of
1587 * inner relation. Assume that the distinct key values are 1, 2, ..., and
1588 * denote the number of values of each key in the outer relation as m1,
1589 * m2, ...; in the inner relation, n1, n2, ... Then we have
1590 *
1591 * size of join = m1 * n1 + m2 * n2 + ...
1592 *
1593 * number of rescanned tuples = (m1 - 1) * n1 + (m2 - 1) * n2 + ... = m1 *
1594 * n1 + m2 * n2 + ... - (n1 + n2 + ...) = size of join - size of inner
1595 * relation
1596 *
1597 * This equation works correctly for outer tuples having no inner match
1598 * (nk = 0), but not for inner tuples having no outer match (mk = 0); we
1599 * are effectively subtracting those from the number of rescanned tuples,
1600 * when we should not. Can we do better without expensive selectivity
1601 * computations?
1602 *
1603 * The whole issue is moot if we are working from a unique-ified outer
1604 * input.
1605 */
1608 else
1609 {
1611 /* Must clamp because of possible underestimate */
1614 }
1615 /* We'll inflate inner run cost this much to account for rescanning */
1618 /*
1619 * A merge join will stop as soon as it exhausts either input stream
1620 * (unless it's an outer join, in which case the outer side has to be
1621 * scanned all the way anyway). Estimate fraction of the left and right
1622 * inputs that will actually need to be scanned. Likewise, we can
1623 * estimate the number of rows that will be skipped before the first join
1624 * pair is found, which should be factored into startup cost. We use only
1625 * the first (most significant) merge clause for this purpose. Since
1626 * mergejoinscansel() is a fairly expensive computation, we cache the
1627 * results in the merge clause RestrictInfo.
1628 */
1630 {
1638 /* Get the input pathkeys to determine the sort-order details */
1645 /* debugging check */
1651 /* Get the selectivity with caching */
1656 {
1657 /* left side of clause is outer */
1662 }
1663 else
1664 {
1665 /* left side of clause is inner */
1670 }
1673 {
1676 }
1678 {
1681 }
1682 }
1683 else
1684 {
1685 /* cope with clauseless or full mergejoin */
1688 }
1690 /*
1691 * Convert selectivities to row counts. We force outer_rows and
1692 * inner_rows to be at least 1, but the skip_rows estimates can be zero.
1693 */
1702 /*
1703 * Readjust scan selectivities to account for above rounding. This is
1704 * normally an insignificant effect, but when there are only a few rows in
1705 * the inputs, failing to do this makes for a large percentage error.
1706 */
1715 /* cost of source data */
1718 {
1720 root,
1721 outersortkeys,
1723 outer_path_rows,
1731 }
1732 else
1733 {
1739 }
1742 {
1744 root,
1745 innersortkeys,
1747 inner_path_rows,
1756 /*
1757 * If the inner sort is expected to spill to disk, we want to add a
1758 * materialize node to shield it from the need to handle mark/restore.
1759 * This will allow it to perform the last merge pass on-the-fly, while
1760 * in most cases not requiring the materialize to spill to disk.
1761 * Charge an extra cpu_tuple_cost per tuple to account for the
1762 * materialize node. (Keep this estimate in sync with similar ones in
1763 * create_mergejoin_path and create_mergejoin_plan.)
1764 */
1768 }
1769 else
1770 {
1776 }
1778 /* CPU costs */
1780 /*
1781 * The number of tuple comparisons needed is approximately number of outer
1782 * rows plus number of inner rows plus number of rescanned tuples (can we
1783 * refine this?). At each one, we need to evaluate the mergejoin quals.
1784 */
1792 /*
1793 * For each tuple that gets through the mergejoin proper, we charge
1794 * cpu_tuple_cost plus the cost of evaluating additional restriction
1795 * clauses that are to be applied at the join. (This is pessimistic since
1796 * not all of the quals may get evaluated at each tuple.)
1797 *
1798 * Note: we could adjust for SEMI/ANTI joins skipping some qual
1799 * evaluations here, but it's probably not worth the trouble.
1800 */
1807 }
1809 /*
1810 * run mergejoinscansel() with caching
1811 */
1814 {
1817 Selectivity leftstartsel,
1818 leftendsel,
1819 rightstartsel,
1820 rightendsel;
1821 MemoryContext oldcontext;
1823 /* Do we have this result already? */
1825 {
1831 }
1833 /* Nope, do the computation */
1844 /* Cache the result in suitably long-lived workspace */
1861 }
1863 /*
1864 * cost_hashjoin
1865 * Determines and returns the cost of joining two relations using the
1866 * hash join algorithm.
1867 *
1868 * 'path' is already filled in except for the cost fields
1869 * 'sjinfo' is extra info about the join for selectivity estimation
1870 *
1871 * Note: path's hashclauses should be a subset of the joinrestrictinfo list
1872 */
1873 void
1875 {
1881 Cost cpu_per_tuple;
1882 QualCost hash_qual_cost;
1883 QualCost qp_qual_cost;
1892 Selectivity innerbucketsize;
1893 Selectivity outer_match_frac;
1894 Selectivity match_count;
1900 /*
1901 * Compute cost of the hashquals and qpquals (other restriction clauses)
1902 * separately.
1903 */
1909 /* cost of source data */
1914 /*
1915 * Cost of computing hash function: must do it once per input tuple. We
1916 * charge one cpu_operator_cost for each column's hash function. Also,
1917 * tack on one cpu_tuple_cost per inner row, to model the costs of
1918 * inserting the row into the hashtable.
1919 *
1920 * XXX when a hashclause is more complex than a single operator, we really
1921 * should charge the extra eval costs of the left or right side, as
1922 * appropriate, here. This seems more work than it's worth at the moment.
1923 */
1928 /*
1929 * Get hash table size that executor would use for inner relation.
1930 *
1931 * XXX for the moment, always assume that skew optimization will be
1932 * performed. As long as SKEW_WORK_MEM_PERCENT is small, it's not worth
1933 * trying to determine that for sure.
1934 *
1935 * XXX at some point it might be interesting to try to account for skew
1936 * optimization in the cost estimate, but for now, we don't.
1937 */
1946 /* mark the path with estimated # of batches */
1949 /*
1950 * Determine bucketsize fraction for inner relation. We use the smallest
1951 * bucketsize estimated for any individual hashclause; this is undoubtedly
1952 * conservative.
1953 *
1954 * BUT: if inner relation has been unique-ified, we can assume it's good
1955 * for hashing. This is important both because it's the right answer, and
1956 * because we avoid contaminating the cache with a value that's wrong for
1957 * non-unique-ified paths.
1958 */
1961 else
1962 {
1965 {
1967 Selectivity thisbucketsize;
1971 /*
1972 * First we have to figure out which side of the hashjoin clause
1973 * is the inner side.
1974 *
1975 * Since we tend to visit the same clauses over and over when
1976 * planning a large query, we cache the bucketsize estimate in the
1977 * RestrictInfo node to avoid repeated lookups of statistics.
1978 */
1981 {
1982 /* righthand side is inner */
1985 {
1986 /* not cached yet */
1987 thisbucketsize =
1990 virtualbuckets);
1992 }
1993 }
1994 else
1995 {
1998 /* lefthand side is inner */
2001 {
2002 /* not cached yet */
2003 thisbucketsize =
2006 virtualbuckets);
2008 }
2009 }
2013 }
2014 }
2016 /*
2017 * If inner relation is too big then we will need to "batch" the join,
2018 * which implies writing and reading most of the tuples to disk an extra
2019 * time. Charge seq_page_cost per page, since the I/O should be nice and
2020 * sequential. Writing the inner rel counts as startup cost, all the rest
2021 * as run cost.
2022 */
2024 {
2032 }
2034 /* CPU costs */
2039 NULL))
2040 {
2042 Selectivity inner_scan_frac;
2044 /*
2045 * SEMI or ANTI join: executor will stop after first match.
2046 *
2047 * For an outer-rel row that has at least one match, we can expect the
2048 * bucket scan to stop after a fraction 1/(match_count+1) of the
2049 * bucket's rows, if the matches are evenly distributed. Since they
2050 * probably aren't quite evenly distributed, we apply a fuzz factor of
2051 * 2.0 to that fraction. (If we used a larger fuzz factor, we'd have
2052 * to clamp inner_scan_frac to at most 1.0; but since match_count is
2053 * at least 1, no such clamp is needed now.)
2054 */
2062 /*
2063 * For unmatched outer-rel rows, the picture is quite a lot different.
2064 * In the first place, there is no reason to assume that these rows
2065 * preferentially hit heavily-populated buckets; instead assume they
2066 * are uncorrelated with the inner distribution and so they see an
2067 * average bucket size of inner_path_rows / virtualbuckets. In the
2068 * second place, it seems likely that they will have few if any exact
2069 * hash-code matches and so very few of the tuples in the bucket will
2070 * actually require eval of the hash quals. We don't have any good
2071 * way to estimate how many will, but for the moment assume that the
2072 * effective cost per bucket entry is one-tenth what it is for
2073 * matchable tuples.
2074 */
2079 /* Get # of tuples that will pass the basic join */
2082 else
2084 }
2085 else
2086 {
2087 /*
2088 * The number of tuple comparisons needed is the number of outer
2089 * tuples times the typical number of tuples in a hash bucket, which
2090 * is the inner relation size times its bucketsize fraction. At each
2091 * one, we need to evaluate the hashjoin quals. But actually,
2092 * charging the full qual eval cost at each tuple is pessimistic,
2093 * since we don't evaluate the quals unless the hash values match
2094 * exactly. For lack of a better idea, halve the cost estimate to
2095 * allow for that.
2096 */
2101 /*
2102 * Get approx # tuples passing the hashquals. We use
2103 * approx_tuple_count here because we need an estimate done with
2104 * JOIN_INNER semantics.
2105 */
2107 }
2109 /*
2110 * For each tuple that gets through the hashjoin proper, we charge
2111 * cpu_tuple_cost plus the cost of evaluating additional restriction
2112 * clauses that are to be applied at the join. (This is pessimistic since
2113 * not all of the quals may get evaluated at each tuple.)
2114 */
2121 }
2124 /*
2125 * cost_subplan
2126 * Figure the costs for a SubPlan (or initplan).
2127 *
2128 * Note: we could dig the subplan's Plan out of the root list, but in practice
2129 * all callers have it handy already, so we make them pass it.
2130 */
2131 void
2133 {
2134 QualCost sp_cost;
2136 /* Figure any cost for evaluating the testexpr */
2139 root);
2142 {
2143 /*
2144 * If we are using a hash table for the subquery outputs, then the
2145 * cost of evaluating the query is a one-time cost. We charge one
2146 * cpu_operator_cost per tuple for the work of loading the hashtable,
2147 * too.
2148 */
2152 /*
2153 * The per-tuple costs include the cost of evaluating the lefthand
2154 * expressions, plus the cost of probing the hashtable. We already
2155 * accounted for the lefthand expressions as part of the testexpr, and
2156 * will also have counted one cpu_operator_cost for each comparison
2157 * operator. That is probably too low for the probing cost, but it's
2158 * hard to make a better estimate, so live with it for now.
2159 */
2160 }
2161 else
2162 {
2163 /*
2164 * Otherwise we will be rescanning the subplan output on each
2165 * evaluation. We need to estimate how much of the output we will
2166 * actually need to scan. NOTE: this logic should agree with the
2167 * tuple_fraction estimates used by make_subplan() in
2168 * plan/subselect.c.
2169 */
2173 {
2174 /* we only need to fetch 1 tuple */
2176 }
2179 {
2180 /* assume we need 50% of the tuples */
2182 /* also charge a cpu_operator_cost per row examined */
2184 }
2185 else
2186 {
2187 /* assume we need all tuples */
2189 }
2191 /*
2192 * Also account for subplan's startup cost. If the subplan is
2193 * uncorrelated or undirect correlated, AND its topmost node is a Sort
2194 * or Material node, assume that we'll only need to pay its startup
2195 * cost once; otherwise assume we pay the startup cost every time.
2196 */
2201 else
2203 }
2207 }
2210 /*
2211 * cost_qual_eval
2212 * Estimate the CPU costs of evaluating a WHERE clause.
2213 * The input can be either an implicitly-ANDed list of boolean
2214 * expressions, or a list of RestrictInfo nodes. (The latter is
2215 * preferred since it allows caching of the results.)
2216 * The result includes both a one-time (startup) component,
2217 * and a per-evaluation component.
2218 */
2219 void
2221 {
2222 cost_qual_eval_context context;
2229 /* We don't charge any cost for the implicit ANDing at top level ... */
2232 {
2236 }
2239 }
2241 /*
2242 * cost_qual_eval_node
2243 * As above, for a single RestrictInfo or expression.
2244 */
2245 void
2247 {
2248 cost_qual_eval_context context;
2257 }
2259 static bool
2261 {
2265 /*
2266 * RestrictInfo nodes contain an eval_cost field reserved for this
2267 * routine's use, so that it's not necessary to evaluate the qual clause's
2268 * cost more than once. If the clause's cost hasn't been computed yet,
2269 * the field's startup value will contain -1.
2270 */
2272 {
2276 {
2277 cost_qual_eval_context locContext;
2283 /*
2284 * For an OR clause, recurse into the marked-up tree so that we
2285 * set the eval_cost for contained RestrictInfos too.
2286 */
2289 else
2292 /*
2293 * If the RestrictInfo is marked pseudoconstant, it will be tested
2294 * only once, so treat its cost as all startup cost.
2295 */
2297 {
2298 /* count one execution during startup */
2301 }
2303 }
2306 /* do NOT recurse into children */
2308 }
2310 /*
2311 * For each operator or function node in the given tree, we charge the
2312 * estimated execution cost given by pg_proc.procost (remember to multiply
2313 * this by cpu_operator_cost).
2314 *
2315 * Vars and Consts are charged zero, and so are boolean operators (AND,
2316 * OR, NOT). Simplistic, but a lot better than no model at all.
2317 *
2318 * Note that Aggref and WindowFunc nodes are (and should be) treated like
2319 * Vars --- whatever execution cost they have is absorbed into
2320 * plan-node-specific costing. As far as expression evaluation is
2321 * concerned they're just like Vars.
2322 *
2323 * Should we try to account for the possibility of short-circuit
2324 * evaluation of AND/OR? Probably *not*, because that would make the
2325 * results depend on the clause ordering, and we are not in any position
2326 * to expect that the current ordering of the clauses is the one that's
2327 * going to end up being used. (Is it worth applying order_qual_clauses
2328 * much earlier in the planning process to fix this?)
2329 */
2331 {
2334 }
2338 {
2339 /* rely on struct equivalence to treat these all alike */
2343 }
2345 {
2346 /*
2347 * Estimate that the operator will be applied to about half of the
2348 * array elements before the answer is determined.
2349 */
2356 }
2358 {
2360 Oid iofunc;
2361 Oid typioparam;
2364 /* check the result type's input function */
2368 /* check the input type's output function */
2372 }
2374 {
2381 }
2383 {
2384 /* Conservatively assume we will check all the columns */
2389 {
2393 cpu_operator_cost;
2394 }
2395 }
2397 {
2398 /* Report high cost to prevent selection of anything but TID scan */
2400 }
2402 {
2403 /* This routine should not be applied to un-planned expressions */
2405 }
2407 {
2408 /*
2409 * A subplan node in an expression typically indicates that the
2410 * subplan will be executed on each evaluation, so charge accordingly.
2411 * (Sub-selects that can be executed as InitPlans have already been
2412 * removed from the expression.)
2413 */
2419 /*
2420 * We don't want to recurse into the testexpr, because it was already
2421 * counted in the SubPlan node's costs. So we're done.
2422 */
2424 }
2426 {
2427 /*
2428 * Arbitrarily use the first alternative plan for costing. (We should
2429 * certainly only include one alternative, and we don't yet have
2430 * enough information to know which one the executor is most likely to
2431 * use.)
2432 */
2436 context);
2437 }
2439 /* recurse into children */
2442 }
2445 /*
2446 * adjust_semi_join
2447 * Estimate how much of the inner input a SEMI or ANTI join
2448 * can be expected to scan.
2449 *
2450 * In a hash or nestloop SEMI/ANTI join, the executor will stop scanning
2451 * inner rows as soon as it finds a match to the current outer row.
2452 * We should therefore adjust some of the cost components for this effect.
2453 * This function computes some estimates needed for these adjustments.
2454 *
2455 * 'path' is already filled in except for the cost fields
2456 * 'sjinfo' is extra info about the join for selectivity estimation
2457 *
2458 * Returns TRUE if this is a SEMI or ANTI join, FALSE if not.
2459 *
2460 * Output parameters (set only in TRUE-result case):
2461 * *outer_match_frac is set to the fraction of the outer tuples that are
2462 * expected to have at least one match.
2463 * *match_count is set to the average number of matches expected for
2464 * outer tuples that have at least one match.
2465 * *indexed_join_quals is set to TRUE if all the joinquals are used as
2466 * inner index quals, FALSE if not.
2467 *
2468 * indexed_join_quals can be passed as NULL if that information is not
2469 * relevant (it is only useful for the nestloop case).
2470 */
2471 static bool
2476 {
2478 Selectivity jselec;
2479 Selectivity nselec;
2480 Selectivity avgmatch;
2481 SpecialJoinInfo norm_sjinfo;
2485 /* Fall out if it's not JOIN_SEMI or JOIN_ANTI */
2489 /*
2490 * Note: it's annoying to repeat this selectivity estimation on each call,
2491 * when the joinclause list will be the same for all path pairs
2492 * implementing a given join. clausesel.c will save us from the worst
2493 * effects of this by caching at the RestrictInfo level; but perhaps it'd
2494 * be worth finding a way to cache the results at a higher level.
2495 */
2497 /*
2498 * In an ANTI join, we must ignore clauses that are "pushed down", since
2499 * those won't affect the match logic. In a SEMI join, we do not
2500 * distinguish joinquals from "pushed down" quals, so just use the whole
2501 * restrictinfo list.
2502 */
2504 {
2507 {
2513 }
2514 }
2515 else
2518 /*
2519 * Get the JOIN_SEMI or JOIN_ANTI selectivity of the join clauses.
2520 */
2522 joinquals,
2524 jointype,
2525 sjinfo);
2527 /*
2528 * Also get the normal inner-join selectivity of the join clauses.
2529 */
2536 /* we don't bother trying to make the remaining fields valid */
2542 joinquals,
2544 JOIN_INNER,
2547 /* Avoid leaking a lot of ListCells */
2551 /*
2552 * jselec can be interpreted as the fraction of outer-rel rows that have
2553 * any matches (this is true for both SEMI and ANTI cases). And nselec is
2554 * the fraction of the Cartesian product that matches. So, the average
2555 * number of matches for each outer-rel row that has at least one match is
2556 * nselec * inner_rows / jselec.
2557 *
2558 * Note: it is correct to use the inner rel's "rows" count here, not
2559 * PATH_ROWS(), even if the inner path under consideration is an inner
2560 * indexscan. This is because we have included all the join clauses in
2561 * the selectivity estimate, even ones used in an inner indexscan.
2562 */
2564 {
2566 /* Clamp to sane range */
2568 }
2569 else
2575 /*
2576 * If requested, check whether the inner path uses all the joinquals as
2577 * indexquals. (If that's true, we can assume that an unmatched outer
2578 * tuple is cheap to process, whereas otherwise it's probably expensive.)
2579 */
2581 {
2588 }
2591 }
2594 /*
2595 * approx_tuple_count
2596 * Quick-and-dirty estimation of the number of join rows passing
2597 * a set of qual conditions.
2598 *
2599 * The quals can be either an implicitly-ANDed list of boolean expressions,
2600 * or a list of RestrictInfo nodes (typically the latter).
2601 *
2602 * We intentionally compute the selectivity under JOIN_INNER rules, even
2603 * if it's some type of outer join. This is appropriate because we are
2604 * trying to figure out how many tuples pass the initial merge or hash
2605 * join step.
2606 *
2607 * This is quick-and-dirty because we bypass clauselist_selectivity, and
2608 * simply multiply the independent clause selectivities together. Now
2609 * clauselist_selectivity often can't do any better than that anyhow, but
2610 * for some situations (such as range constraints) it is smarter. However,
2611 * we can't effectively cache the results of clauselist_selectivity, whereas
2612 * the individual clause selectivities can be and are cached.
2613 *
2614 * Since we are only using the results to estimate how many potential
2615 * output tuples are generated and passed through qpqual checking, it
2616 * seems OK to live with the approximation.
2617 */
2618 static double
2620 {
2624 SpecialJoinInfo sjinfo;
2628 /*
2629 * Make up a SpecialJoinInfo for JOIN_INNER semantics.
2630 */
2637 /* we don't bother trying to make the remaining fields valid */
2642 /* Get the approximate selectivity */
2644 {
2647 /* Note that clause_selectivity will be able to cache its result */
2649 }
2651 /* Apply it to the input relation sizes */
2655 }
2658 /*
2659 * set_baserel_size_estimates
2660 * Set the size estimates for the given base relation.
2661 *
2662 * The rel's targetlist and restrictinfo list must have been constructed
2663 * already.
2664 *
2665 * We set the following fields of the rel node:
2666 * rows: the estimated number of output tuples (after applying
2667 * restriction clauses).
2668 * width: the estimated average output tuple width in bytes.
2669 * baserestrictcost: estimated cost of evaluating baserestrictinfo clauses.
2670 */
2671 void
2673 {
2676 /* Should only be applied to base relations */
2683 JOIN_INNER,
2684 NULL);
2691 }
2693 /*
2694 * set_joinrel_size_estimates
2695 * Set the size estimates for the given join relation.
2696 *
2697 * The rel's targetlist must have been constructed already, and a
2698 * restriction clause list that matches the given component rels must
2699 * be provided.
2700 *
2701 * Since there is more than one way to make a joinrel for more than two
2702 * base relations, the results we get here could depend on which component
2703 * rel pair is provided. In theory we should get the same answers no matter
2704 * which pair is provided; in practice, since the selectivity estimation
2705 * routines don't handle all cases equally well, we might not. But there's
2706 * not much to be done about it. (Would it make sense to repeat the
2707 * calculations for each pair of input rels that's encountered, and somehow
2708 * average the results? Probably way more trouble than it's worth.)
2709 *
2710 * We set only the rows field here. The width field was already set by
2711 * build_joinrel_tlist, and baserestrictcost is not used for join rels.
2712 */
2713 void
2719 {
2721 Selectivity jselec;
2722 Selectivity pselec;
2725 /*
2726 * Compute joinclause selectivity. Note that we are only considering
2727 * clauses that become restriction clauses at this join level; we are not
2728 * double-counting them because they were not considered in estimating the
2729 * sizes of the component rels.
2730 *
2731 * For an outer join, we have to distinguish the selectivity of the join's
2732 * own clauses (JOIN/ON conditions) from any clauses that were "pushed
2733 * down". For inner joins we just count them all as joinclauses.
2734 */
2736 {
2741 /* Grovel through the clauses to separate into two lists */
2743 {
2749 else
2751 }
2753 /* Get the separate selectivities */
2755 joinquals,
2757 jointype,
2758 sjinfo);
2760 pushedquals,
2762 jointype,
2763 sjinfo);
2765 /* Avoid leaking a lot of ListCells */
2768 }
2769 else
2770 {
2772 restrictlist,
2774 jointype,
2775 sjinfo);
2777 }
2779 /*
2780 * Basically, we multiply size of Cartesian product by selectivity.
2781 *
2782 * If we are doing an outer join, take that into account: the joinqual
2783 * selectivity has to be clamped using the knowledge that the output must
2784 * be at least as large as the non-nullable input. However, any
2785 * pushed-down quals are applied after the outer join, so their
2786 * selectivity applies fully.
2787 *
2788 * For JOIN_SEMI and JOIN_ANTI, the selectivity is defined as the fraction
2789 * of LHS rows that have matches, and we apply that straightforwardly.
2790 */
2792 {
2812 /* pselec not used */
2819 /* other values not expected here */
2823 }
2826 }
2828 /*
2829 * set_function_size_estimates
2830 * Set the size estimates for a base relation that is a function call.
2831 *
2832 * The rel's targetlist and restrictinfo list must have been constructed
2833 * already.
2834 *
2835 * We set the same fields as set_baserel_size_estimates.
2836 */
2837 void
2839 {
2842 /* Should only be applied to base relations that are functions */
2847 /* Estimate number of rows the function itself will return */
2850 /* Now estimate number of output rows, etc */
2852 }
2854 /*
2855 * set_values_size_estimates
2856 * Set the size estimates for a base relation that is a values list.
2857 *
2858 * The rel's targetlist and restrictinfo list must have been constructed
2859 * already.
2860 *
2861 * We set the same fields as set_baserel_size_estimates.
2862 */
2863 void
2865 {
2868 /* Should only be applied to base relations that are values lists */
2873 /*
2874 * Estimate number of rows the values list will return. We know this
2875 * precisely based on the list length (well, barring set-returning
2876 * functions in list items, but that's a refinement not catered for
2877 * anywhere else either).
2878 */
2881 /* Now estimate number of output rows, etc */
2883 }
2885 /*
2886 * set_cte_size_estimates
2887 * Set the size estimates for a base relation that is a CTE reference.
2888 *
2889 * The rel's targetlist and restrictinfo list must have been constructed
2890 * already, and we need the completed plan for the CTE (if a regular CTE)
2891 * or the non-recursive term (if a self-reference).
2892 *
2893 * We set the same fields as set_baserel_size_estimates.
2894 */
2895 void
2897 {
2900 /* Should only be applied to base relations that are CTE references */
2906 {
2907 /*
2908 * In a self-reference, arbitrarily assume the average worktable size
2909 * is about 10 times the nonrecursive term's size.
2910 */
2912 }
2913 else
2914 {
2915 /* Otherwise just believe the CTE plan's output estimate */
2917 }
2919 /* Now estimate number of output rows, etc */
2921 }
2924 /*
2925 * set_rel_width
2926 * Set the estimated output width of a base relation.
2927 *
2928 * NB: this works best on plain relations because it prefers to look at
2929 * real Vars. It will fail to make use of pg_statistic info when applied
2930 * to a subquery relation, even if the subquery outputs are simple vars
2931 * that we could have gotten info for. Is it worth trying to be smarter
2932 * about subqueries?
2933 *
2934 * The per-attribute width estimates are cached for possible re-use while
2935 * building join relations.
2936 */
2937 static void
2939 {
2945 {
2949 {
2952 int32 item_width;
2960 /*
2961 * The width probably hasn't been cached yet, but may as well
2962 * check
2963 */
2965 {
2968 }
2970 /* Try to get column width from statistics */
2972 {
2975 {
2979 }
2980 }
2982 /*
2983 * Not a plain relation, or can't find statistics for it. Estimate
2984 * using just the type info.
2985 */
2990 }
2992 {
2997 }
2998 else
2999 {
3000 /* For now, punt on whole-row child Vars */
3002 }
3003 }
3006 }
3008 /*
3009 * relation_byte_size
3010 * Estimate the storage space in bytes for a given number of tuples
3011 * of a given width (size in bytes).
3012 */
3013 static double
3015 {
3017 }
3019 /*
3020 * page_size
3021 * Returns an estimate of the number of pages covered by a given
3022 * number of tuples of a given width (size in bytes).
3023 */
3024 static double
3026 {
3028 }