| blob | 8dc7dd4ca264b888d7b24455023f7959b8ad2d31 |
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 * parallel_tuple_cost Cost of CPU time to pass a tuple from worker to leader backend
15 * parallel_setup_cost Cost of setting up shared memory for parallelism
16 *
17 * We expect that the kernel will typically do some amount of read-ahead
18 * optimization; this in conjunction with seek costs means that seq_page_cost
19 * is normally considerably less than random_page_cost. (However, if the
20 * database is fully cached in RAM, it is reasonable to set them equal.)
21 *
22 * We also use a rough estimate "effective_cache_size" of the number of
23 * disk pages in Postgres + OS-level disk cache. (We can't simply use
24 * NBuffers for this purpose because that would ignore the effects of
25 * the kernel's disk cache.)
26 *
27 * Obviously, taking constants for these values is an oversimplification,
28 * but it's tough enough to get any useful estimates even at this level of
29 * detail. Note that all of these parameters are user-settable, in case
30 * the default values are drastically off for a particular platform.
31 *
32 * seq_page_cost and random_page_cost can also be overridden for an individual
33 * tablespace, in case some data is on a fast disk and other data is on a slow
34 * disk. Per-tablespace overrides never apply to temporary work files such as
35 * an external sort or a materialize node that overflows work_mem.
36 *
37 * We compute two separate costs for each path:
38 * total_cost: total estimated cost to fetch all tuples
39 * startup_cost: cost that is expended before first tuple is fetched
40 * In some scenarios, such as when there is a LIMIT or we are implementing
41 * an EXISTS(...) sub-select, it is not necessary to fetch all tuples of the
42 * path's result. A caller can estimate the cost of fetching a partial
43 * result by interpolating between startup_cost and total_cost. In detail:
44 * actual_cost = startup_cost +
45 * (total_cost - startup_cost) * tuples_to_fetch / path->rows;
46 * Note that a base relation's rows count (and, by extension, plan_rows for
47 * plan nodes below the LIMIT node) are set without regard to any LIMIT, so
48 * that this equation works properly. (Note: while path->rows is never zero
49 * for ordinary relations, it is zero for paths for provably-empty relations,
50 * so beware of division-by-zero.) The LIMIT is applied as a top-level
51 * plan node.
52 *
53 * For largely historical reasons, most of the routines in this module use
54 * the passed result Path only to store their results (rows, startup_cost and
55 * total_cost) into. All the input data they need is passed as separate
56 * parameters, even though much of it could be extracted from the Path.
57 * An exception is made for the cost_XXXjoin() routines, which expect all
58 * the other fields of the passed XXXPath to be filled in, and similarly
59 * cost_index() assumes the passed IndexPath is valid except for its output
60 * values.
61 *
62 *
63 * Portions Copyright (c) 1996-2022, PostgreSQL Global Development Group
64 * Portions Copyright (c) 1994, Regents of the University of California
65 *
66 * IDENTIFICATION
67 * src/backend/optimizer/path/costsize.c
68 *
69 *-------------------------------------------------------------------------
70 */
74 #include <math.h>
102 #define LOG2(x) (log(x) / 0.693147180559945)
104 /*
105 * Append and MergeAppend nodes are less expensive than some other operations
106 * which use cpu_tuple_cost; instead of adding a separate GUC, estimate the
107 * per-tuple cost as cpu_tuple_cost multiplied by this value.
108 */
109 #define APPEND_CPU_COST_MULTIPLIER 0.5
111 /*
112 * Maximum value for row estimates. We cap row estimates to this to help
113 * ensure that costs based on these estimates remain within the range of what
114 * double can represent. add_path() wouldn't act sanely given infinite or NaN
115 * cost values.
116 */
117 #define MAXIMUM_ROWCOUNT 1e100
155 {
157 QualCost total;
182 Relids outer_relids,
183 Relids inner_relids,
194 /*
195 * clamp_row_est
196 * Force a row-count estimate to a sane value.
197 */
198 double
200 {
201 /*
202 * Avoid infinite and NaN row estimates. Costs derived from such values
203 * are going to be useless. Also force the estimate to be at least one
204 * row, to make explain output look better and to avoid possible
205 * divide-by-zero when interpolating costs. Make it an integer, too.
206 */
211 else
215 }
218 /*
219 * cost_seqscan
220 * Determines and returns the cost of scanning a relation sequentially.
221 *
222 * 'baserel' is the relation to be scanned
223 * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
224 */
225 void
228 {
230 Cost cpu_run_cost;
231 Cost disk_run_cost;
233 QualCost qpqual_cost;
234 Cost cpu_per_tuple;
236 /* Should only be applied to base relations */
240 /* Mark the path with the correct row estimate */
243 else
249 /* fetch estimated page cost for tablespace containing table */
251 NULL,
254 /*
255 * disk costs
256 */
259 /* CPU costs */
265 /* tlist eval costs are paid per output row, not per tuple scanned */
269 /* Adjust costing for parallelism, if used. */
271 {
274 /* The CPU cost is divided among all the workers. */
277 /*
278 * It may be possible to amortize some of the I/O cost, but probably
279 * not very much, because most operating systems already do aggressive
280 * prefetching. For now, we assume that the disk run cost can't be
281 * amortized at all.
282 */
284 /*
285 * In the case of a parallel plan, the row count needs to represent
286 * the number of tuples processed per worker.
287 */
289 }
293 }
295 /*
296 * cost_samplescan
297 * Determines and returns the cost of scanning a relation using sampling.
298 *
299 * 'baserel' is the relation to be scanned
300 * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
301 */
302 void
305 {
312 spc_random_page_cost,
313 spc_page_cost;
314 QualCost qpqual_cost;
315 Cost cpu_per_tuple;
317 /* Should only be applied to base relations with tablesample clauses */
325 /* Mark the path with the correct row estimate */
328 else
331 /* fetch estimated page cost for tablespace containing table */
336 /* if NextSampleBlock is used, assume random access, else sequential */
340 /*
341 * disk costs (recall that baserel->pages has already been set to the
342 * number of pages the sampling method will visit)
343 */
346 /*
347 * CPU costs (recall that baserel->tuples has already been set to the
348 * number of tuples the sampling method will select). Note that we ignore
349 * execution cost of the TABLESAMPLE parameter expressions; they will be
350 * evaluated only once per scan, and in most usages they'll likely be
351 * simple constants anyway. We also don't charge anything for the
352 * calculations the sampling method might do internally.
353 */
359 /* tlist eval costs are paid per output row, not per tuple scanned */
365 }
367 /*
368 * cost_gather
369 * Determines and returns the cost of gather path.
370 *
371 * 'rel' is the relation to be operated upon
372 * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
373 * 'rows' may be used to point to a row estimate; if non-NULL, it overrides
374 * both 'rel' and 'param_info'. This is useful when the path doesn't exactly
375 * correspond to any particular RelOptInfo.
376 */
377 void
381 {
385 /* Mark the path with the correct row estimate */
390 else
397 /* Parallel setup and communication cost. */
403 }
405 /*
406 * cost_gather_merge
407 * Determines and returns the cost of gather merge path.
408 *
409 * GatherMerge merges several pre-sorted input streams, using a heap that at
410 * any given instant holds the next tuple from each stream. If there are N
411 * streams, we need about N*log2(N) tuple comparisons to construct the heap at
412 * startup, and then for each output tuple, about log2(N) comparisons to
413 * replace the top heap entry with the next tuple from the same stream.
414 */
415 void
420 {
423 Cost comparison_cost;
427 /* Mark the path with the correct row estimate */
432 else
438 /*
439 * Add one to the number of workers to account for the leader. This might
440 * be overgenerous since the leader will do less work than other workers
441 * in typical cases, but we'll go with it for now.
442 */
447 /* Assumed cost per tuple comparison */
450 /* Heap creation cost */
453 /* Per-tuple heap maintenance cost */
456 /* small cost for heap management, like cost_merge_append */
459 /*
460 * Parallel setup and communication cost. Since Gather Merge, unlike
461 * Gather, requires us to block until a tuple is available from every
462 * worker, we bump the IPC cost up a little bit as compared with Gather.
463 * For lack of a better idea, charge an extra 5%.
464 */
470 }
472 /*
473 * cost_index
474 * Determines and returns the cost of scanning a relation using an index.
475 *
476 * 'path' describes the indexscan under consideration, and is complete
477 * except for the fields to be set by this routine
478 * 'loop_count' is the number of repetitions of the indexscan to factor into
479 * estimates of caching behavior
480 *
481 * In addition to rows, startup_cost and total_cost, cost_index() sets the
482 * path's indextotalcost and indexselectivity fields. These values will be
483 * needed if the IndexPath is used in a BitmapIndexScan.
484 *
485 * NOTE: path->indexquals must contain only clauses usable as index
486 * restrictions. Any additional quals evaluated as qpquals may reduce the
487 * number of returned tuples, but they won't reduce the number of tuples
488 * we have to fetch from the table, so they don't reduce the scan cost.
489 */
490 void
493 {
497 amcostestimate_function amcostestimate;
502 Cost indexStartupCost;
503 Cost indexTotalCost;
504 Selectivity indexSelectivity;
506 csquared;
508 spc_random_page_cost;
509 Cost min_IO_cost,
510 max_IO_cost;
511 QualCost qpqual_cost;
512 Cost cpu_per_tuple;
518 /* Should only be applied to base relations */
524 /*
525 * Mark the path with the correct row estimate, and identify which quals
526 * will need to be enforced as qpquals. We need not check any quals that
527 * are implied by the index's predicate, so we can use indrestrictinfo not
528 * baserestrictinfo as the list of relevant restriction clauses for the
529 * rel.
530 */
532 {
534 /* qpquals come from the rel's restriction clauses and ppi_clauses */
539 }
540 else
541 {
543 /* qpquals come from just the rel's restriction clauses */
546 }
550 /* we don't need to check enable_indexonlyscan; indxpath.c does that */
552 /*
553 * Call index-access-method-specific code to estimate the processing cost
554 * for scanning the index, as well as the selectivity of the index (ie,
555 * the fraction of main-table tuples we will have to retrieve) and its
556 * correlation to the main-table tuple order. We need a cast here because
557 * pathnodes.h uses a weak function type to avoid including amapi.h.
558 */
565 /*
566 * Save amcostestimate's results for possible use in bitmap scan planning.
567 * We don't bother to save indexStartupCost or indexCorrelation, because a
568 * bitmap scan doesn't care about either.
569 */
573 /* all costs for touching index itself included here */
577 /* estimate number of main-table tuples fetched */
580 /* fetch estimated page costs for tablespace containing table */
585 /*----------
586 * Estimate number of main-table pages fetched, and compute I/O cost.
587 *
588 * When the index ordering is uncorrelated with the table ordering,
589 * we use an approximation proposed by Mackert and Lohman (see
590 * index_pages_fetched() for details) to compute the number of pages
591 * fetched, and then charge spc_random_page_cost per page fetched.
592 *
593 * When the index ordering is exactly correlated with the table ordering
594 * (just after a CLUSTER, for example), the number of pages fetched should
595 * be exactly selectivity * table_size. What's more, all but the first
596 * will be sequential fetches, not the random fetches that occur in the
597 * uncorrelated case. So if the number of pages is more than 1, we
598 * ought to charge
599 * spc_random_page_cost + (pages_fetched - 1) * spc_seq_page_cost
600 * For partially-correlated indexes, we ought to charge somewhere between
601 * these two estimates. We currently interpolate linearly between the
602 * estimates based on the correlation squared (XXX is that appropriate?).
603 *
604 * If it's an index-only scan, then we will not need to fetch any heap
605 * pages for which the visibility map shows all tuples are visible.
606 * Hence, reduce the estimated number of heap fetches accordingly.
607 * We use the measured fraction of the entire heap that is all-visible,
608 * which might not be particularly relevant to the subset of the heap
609 * that this query will fetch; but it's not clear how to do better.
610 *----------
611 */
613 {
614 /*
615 * For repeated indexscans, the appropriate estimate for the
616 * uncorrelated case is to scale up the number of tuples fetched in
617 * the Mackert and Lohman formula by the number of scans, so that we
618 * estimate the number of pages fetched by all the scans; then
619 * pro-rate the costs for one scan. In this case we assume all the
620 * fetches are random accesses.
621 */
625 root);
634 /*
635 * In the perfectly correlated case, the number of pages touched by
636 * each scan is selectivity * table_size, and we can use the Mackert
637 * and Lohman formula at the page level to estimate how much work is
638 * saved by caching across scans. We still assume all the fetches are
639 * random, though, which is an overestimate that's hard to correct for
640 * without double-counting the cache effects. (But in most cases
641 * where such a plan is actually interesting, only one page would get
642 * fetched per scan anyway, so it shouldn't matter much.)
643 */
649 root);
655 }
656 else
657 {
658 /*
659 * Normal case: apply the Mackert and Lohman formula, and then
660 * interpolate between that and the correlation-derived result.
661 */
665 root);
672 /* max_IO_cost is for the perfectly uncorrelated case (csquared=0) */
675 /* min_IO_cost is for the perfectly correlated case (csquared=1) */
682 {
686 }
687 else
689 }
692 {
693 /*
694 * For index only scans compute workers based on number of index pages
695 * fetched; the number of heap pages we fetch might be so small as to
696 * effectively rule out parallelism, which we don't want to do.
697 */
701 /*
702 * Estimate the number of parallel workers required to scan index. Use
703 * the number of heap pages computed considering heap fetches won't be
704 * sequential as for parallel scans the pages are accessed in random
705 * order.
706 */
708 rand_heap_pages,
709 index_pages,
710 max_parallel_workers_per_gather);
712 /*
713 * Fall out if workers can't be assigned for parallel scan, because in
714 * such a case this path will be rejected. So there is no benefit in
715 * doing extra computation.
716 */
721 }
723 /*
724 * Now interpolate based on estimated index order correlation to get total
725 * disk I/O cost for main table accesses.
726 */
731 /*
732 * Estimate CPU costs per tuple.
733 *
734 * What we want here is cpu_tuple_cost plus the evaluation costs of any
735 * qual clauses that we have to evaluate as qpquals.
736 */
744 /* tlist eval costs are paid per output row, not per tuple scanned */
748 /* Adjust costing for parallelism, if used. */
750 {
755 /* The CPU cost is divided among all the workers. */
757 }
763 }
765 /*
766 * extract_nonindex_conditions
767 *
768 * Given a list of quals to be enforced in an indexscan, extract the ones that
769 * will have to be applied as qpquals (ie, the index machinery won't handle
770 * them). Here we detect only whether a qual clause is directly redundant
771 * with some indexclause. If the index path is chosen for use, createplan.c
772 * will try a bit harder to get rid of redundant qual conditions; specifically
773 * it will see if quals can be proven to be implied by the indexquals. But
774 * it does not seem worth the cycles to try to factor that in at this stage,
775 * since we're only trying to estimate qual eval costs. Otherwise this must
776 * match the logic in create_indexscan_plan().
777 *
778 * qual_clauses, and the result, are lists of RestrictInfos.
779 * indexclauses is a list of IndexClauses.
780 */
783 {
788 {
795 /* ... skip the predicate proof attempt createplan.c will try ... */
797 }
799 }
801 /*
802 * index_pages_fetched
803 * Estimate the number of pages actually fetched after accounting for
804 * cache effects.
805 *
806 * We use an approximation proposed by Mackert and Lohman, "Index Scans
807 * Using a Finite LRU Buffer: A Validated I/O Model", ACM Transactions
808 * on Database Systems, Vol. 14, No. 3, September 1989, Pages 401-424.
809 * The Mackert and Lohman approximation is that the number of pages
810 * fetched is
811 * PF =
812 * min(2TNs/(2T+Ns), T) when T <= b
813 * 2TNs/(2T+Ns) when T > b and Ns <= 2Tb/(2T-b)
814 * b + (Ns - 2Tb/(2T-b))*(T-b)/T when T > b and Ns > 2Tb/(2T-b)
815 * where
816 * T = # pages in table
817 * N = # tuples in table
818 * s = selectivity = fraction of table to be scanned
819 * b = # buffer pages available (we include kernel space here)
820 *
821 * We assume that effective_cache_size is the total number of buffer pages
822 * available for the whole query, and pro-rate that space across all the
823 * tables in the query and the index currently under consideration. (This
824 * ignores space needed for other indexes used by the query, but since we
825 * don't know which indexes will get used, we can't estimate that very well;
826 * and in any case counting all the tables may well be an overestimate, since
827 * depending on the join plan not all the tables may be scanned concurrently.)
828 *
829 * The product Ns is the number of tuples fetched; we pass in that
830 * product rather than calculating it here. "pages" is the number of pages
831 * in the object under consideration (either an index or a table).
832 * "index_pages" is the amount to add to the total table space, which was
833 * computed for us by make_one_rel.
834 *
835 * Caller is expected to have ensured that tuples_fetched is greater than zero
836 * and rounded to integer (see clamp_row_est). The result will likewise be
837 * greater than zero and integral.
838 */
839 double
842 {
846 b;
848 /* T is # pages in table, but don't allow it to be zero */
851 /* Compute number of pages assumed to be competing for cache space */
856 /* b is pro-rated share of effective_cache_size */
859 /* force it positive and integral */
862 else
865 /* This part is the Mackert and Lohman formula */
867 {
868 pages_fetched =
872 else
874 }
875 else
876 {
881 {
882 pages_fetched =
884 }
885 else
886 {
887 pages_fetched =
889 }
891 }
893 }
895 /*
896 * get_indexpath_pages
897 * Determine the total size of the indexes used in a bitmap index path.
898 *
899 * Note: if the same index is used more than once in a bitmap tree, we will
900 * count it multiple times, which perhaps is the wrong thing ... but it's
901 * not completely clear, and detecting duplicates is difficult, so ignore it
902 * for now.
903 */
904 static double
906 {
911 {
915 {
917 }
918 }
920 {
924 {
926 }
927 }
929 {
933 }
934 else
938 }
940 /*
941 * cost_bitmap_heap_scan
942 * Determines and returns the cost of scanning a relation using a bitmap
943 * index-then-heap plan.
944 *
945 * 'baserel' is the relation to be scanned
946 * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
947 * 'bitmapqual' is a tree of IndexPaths, BitmapAndPaths, and BitmapOrPaths
948 * 'loop_count' is the number of repetitions of the indexscan to factor into
949 * estimates of caching behavior
950 *
951 * Note: the component IndexPaths in bitmapqual should have been costed
952 * using the same loop_count.
953 */
954 void
958 {
961 Cost indexTotalCost;
962 QualCost qpqual_cost;
963 Cost cpu_per_tuple;
964 Cost cost_per_page;
965 Cost cpu_run_cost;
969 spc_random_page_cost;
972 /* Should only be applied to base relations */
977 /* Mark the path with the correct row estimate */
980 else
993 /* Fetch estimated page costs for tablespace containing table. */
998 /*
999 * For small numbers of pages we should charge spc_random_page_cost
1000 * apiece, while if nearly all the table's pages are being read, it's more
1001 * appropriate to charge spc_seq_page_cost apiece. The effect is
1002 * nonlinear, too. For lack of a better idea, interpolate like this to
1003 * determine the cost per page.
1004 */
1009 else
1014 /*
1015 * Estimate CPU costs per tuple.
1016 *
1017 * Often the indexquals don't need to be rechecked at each tuple ... but
1018 * not always, especially not if there are enough tuples involved that the
1019 * bitmaps become lossy. For the moment, just assume they will be
1020 * rechecked always. This means we charge the full freight for all the
1021 * scan clauses.
1022 */
1029 /* Adjust costing for parallelism, if used. */
1031 {
1034 /* The CPU cost is divided among all the workers. */
1038 }
1043 /* tlist eval costs are paid per output row, not per tuple scanned */
1049 }
1051 /*
1052 * cost_bitmap_tree_node
1053 * Extract cost and selectivity from a bitmap tree node (index/and/or)
1054 */
1055 void
1057 {
1059 {
1063 /*
1064 * Charge a small amount per retrieved tuple to reflect the costs of
1065 * manipulating the bitmap. This is mostly to make sure that a bitmap
1066 * scan doesn't look to be the same cost as an indexscan to retrieve a
1067 * single tuple.
1068 */
1070 }
1072 {
1075 }
1077 {
1080 }
1081 else
1082 {
1085 }
1086 }
1088 /*
1089 * cost_bitmap_and_node
1090 * Estimate the cost of a BitmapAnd node
1091 *
1092 * Note that this considers only the costs of index scanning and bitmap
1093 * creation, not the eventual heap access. In that sense the object isn't
1094 * truly a Path, but it has enough path-like properties (costs in particular)
1095 * to warrant treating it as one. We don't bother to set the path rows field,
1096 * however.
1097 */
1098 void
1100 {
1101 Cost totalCost;
1102 Selectivity selec;
1105 /*
1106 * We estimate AND selectivity on the assumption that the inputs are
1107 * independent. This is probably often wrong, but we don't have the info
1108 * to do better.
1109 *
1110 * The runtime cost of the BitmapAnd itself is estimated at 100x
1111 * cpu_operator_cost for each tbm_intersect needed. Probably too small,
1112 * definitely too simplistic?
1113 */
1117 {
1119 Cost subCost;
1120 Selectivity subselec;
1129 }
1134 }
1136 /*
1137 * cost_bitmap_or_node
1138 * Estimate the cost of a BitmapOr node
1139 *
1140 * See comments for cost_bitmap_and_node.
1141 */
1142 void
1144 {
1145 Cost totalCost;
1146 Selectivity selec;
1149 /*
1150 * We estimate OR selectivity on the assumption that the inputs are
1151 * non-overlapping, since that's often the case in "x IN (list)" type
1152 * situations. Of course, we clamp to 1.0 at the end.
1153 *
1154 * The runtime cost of the BitmapOr itself is estimated at 100x
1155 * cpu_operator_cost for each tbm_union needed. Probably too small,
1156 * definitely too simplistic? We are aware that the tbm_unions are
1157 * optimized out when the inputs are BitmapIndexScans.
1158 */
1162 {
1164 Cost subCost;
1165 Selectivity subselec;
1175 }
1180 }
1182 /*
1183 * cost_tidscan
1184 * Determines and returns the cost of scanning a relation using TIDs.
1185 *
1186 * 'baserel' is the relation to be scanned
1187 * 'tidquals' is the list of TID-checkable quals
1188 * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1189 */
1190 void
1193 {
1197 QualCost qpqual_cost;
1198 Cost cpu_per_tuple;
1199 QualCost tid_qual_cost;
1204 /* Should only be applied to base relations */
1208 /* Mark the path with the correct row estimate */
1211 else
1214 /* Count how many tuples we expect to retrieve */
1217 {
1222 {
1223 /* Each element of the array yields 1 tuple */
1228 }
1230 {
1231 /* CURRENT OF yields 1 tuple */
1233 ntuples++;
1234 }
1235 else
1236 {
1237 /* It's just CTID = something, count 1 tuple */
1238 ntuples++;
1239 }
1240 }
1242 /*
1243 * We must force TID scan for WHERE CURRENT OF, because only nodeTidscan.c
1244 * understands how to do it correctly. Therefore, honor enable_tidscan
1245 * only when CURRENT OF isn't present. Also note that cost_qual_eval
1246 * counts a CurrentOfExpr as having startup cost disable_cost, which we
1247 * subtract off here; that's to prevent other plan types such as seqscan
1248 * from winning.
1249 */
1251 {
1254 }
1258 /*
1259 * The TID qual expressions will be computed once, any other baserestrict
1260 * quals once per retrieved tuple.
1261 */
1264 /* fetch estimated page cost for tablespace containing table */
1267 NULL);
1269 /* disk costs --- assume each tuple on a different page */
1272 /* Add scanning CPU costs */
1275 /* XXX currently we assume TID quals are a subset of qpquals */
1281 /* tlist eval costs are paid per output row, not per tuple scanned */
1287 }
1289 /*
1290 * cost_tidrangescan
1291 * Determines and sets the costs of scanning a relation using a range of
1292 * TIDs for 'path'
1293 *
1294 * 'baserel' is the relation to be scanned
1295 * 'tidrangequals' is the list of TID-checkable range quals
1296 * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1297 */
1298 void
1302 {
1303 Selectivity selectivity;
1307 QualCost qpqual_cost;
1308 Cost cpu_per_tuple;
1309 QualCost tid_qual_cost;
1315 /* Should only be applied to base relations */
1319 /* Mark the path with the correct row estimate */
1322 else
1325 /* Count how many tuples and pages we expect to scan */
1333 /*
1334 * The first page in a range requires a random seek, but each subsequent
1335 * page is just a normal sequential page read. NOTE: it's desirable for
1336 * TID Range Scans to cost more than the equivalent Sequential Scans,
1337 * because Seq Scans have some performance advantages such as scan
1338 * synchronization and parallelizability, and we'd prefer one of them to
1339 * be picked unless a TID Range Scan really is better.
1340 */
1347 /*
1348 * The TID qual expressions will be computed once, any other baserestrict
1349 * quals once per retrieved tuple.
1350 */
1353 /* fetch estimated page cost for tablespace containing table */
1358 /* disk costs; 1 random page and the remainder as seq pages */
1361 /* Add scanning CPU costs */
1364 /*
1365 * XXX currently we assume TID quals are a subset of qpquals at this
1366 * point; they will be removed (if possible) when we create the plan, so
1367 * we subtract their cost from the total qpqual cost. (If the TID quals
1368 * can't be removed, this is a mistake and we're going to underestimate
1369 * the CPU cost a bit.)
1370 */
1376 /* tlist eval costs are paid per output row, not per tuple scanned */
1382 }
1384 /*
1385 * cost_subqueryscan
1386 * Determines and returns the cost of scanning a subquery RTE.
1387 *
1388 * 'baserel' is the relation to be scanned
1389 * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1390 */
1391 void
1394 {
1395 Cost startup_cost;
1396 Cost run_cost;
1397 QualCost qpqual_cost;
1398 Cost cpu_per_tuple;
1400 /* Should only be applied to base relations that are subqueries */
1404 /* Mark the path with the correct row estimate */
1407 else
1410 /*
1411 * Cost of path is cost of evaluating the subplan, plus cost of evaluating
1412 * any restriction clauses and tlist that will be attached to the
1413 * SubqueryScan node, plus cpu_tuple_cost to account for selection and
1414 * projection overhead.
1415 */
1425 /* tlist eval costs are paid per output row, not per tuple scanned */
1431 }
1433 /*
1434 * cost_functionscan
1435 * Determines and returns the cost of scanning a function RTE.
1436 *
1437 * 'baserel' is the relation to be scanned
1438 * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1439 */
1440 void
1443 {
1446 QualCost qpqual_cost;
1447 Cost cpu_per_tuple;
1449 QualCost exprcost;
1451 /* Should only be applied to base relations that are functions */
1456 /* Mark the path with the correct row estimate */
1459 else
1462 /*
1463 * Estimate costs of executing the function expression(s).
1464 *
1465 * Currently, nodeFunctionscan.c always executes the functions to
1466 * completion before returning any rows, and caches the results in a
1467 * tuplestore. So the function eval cost is all startup cost, and per-row
1468 * costs are minimal.
1469 *
1470 * XXX in principle we ought to charge tuplestore spill costs if the
1471 * number of rows is large. However, given how phony our rowcount
1472 * estimates for functions tend to be, there's not a lot of point in that
1473 * refinement right now.
1474 */
1479 /* Add scanning CPU costs */
1486 /* tlist eval costs are paid per output row, not per tuple scanned */
1492 }
1494 /*
1495 * cost_tablefuncscan
1496 * Determines and returns the cost of scanning a table function.
1497 *
1498 * 'baserel' is the relation to be scanned
1499 * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1500 */
1501 void
1504 {
1507 QualCost qpqual_cost;
1508 Cost cpu_per_tuple;
1510 QualCost exprcost;
1512 /* Should only be applied to base relations that are functions */
1517 /* Mark the path with the correct row estimate */
1520 else
1523 /*
1524 * Estimate costs of executing the table func expression(s).
1525 *
1526 * XXX in principle we ought to charge tuplestore spill costs if the
1527 * number of rows is large. However, given how phony our rowcount
1528 * estimates for tablefuncs tend to be, there's not a lot of point in that
1529 * refinement right now.
1530 */
1535 /* Add scanning CPU costs */
1542 /* tlist eval costs are paid per output row, not per tuple scanned */
1548 }
1550 /*
1551 * cost_valuesscan
1552 * Determines and returns the cost of scanning a VALUES RTE.
1553 *
1554 * 'baserel' is the relation to be scanned
1555 * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1556 */
1557 void
1560 {
1563 QualCost qpqual_cost;
1564 Cost cpu_per_tuple;
1566 /* Should only be applied to base relations that are values lists */
1570 /* Mark the path with the correct row estimate */
1573 else
1576 /*
1577 * For now, estimate list evaluation cost at one operator eval per list
1578 * (probably pretty bogus, but is it worth being smarter?)
1579 */
1582 /* Add scanning CPU costs */
1589 /* tlist eval costs are paid per output row, not per tuple scanned */
1595 }
1597 /*
1598 * cost_ctescan
1599 * Determines and returns the cost of scanning a CTE RTE.
1600 *
1601 * Note: this is used for both self-reference and regular CTEs; the
1602 * possible cost differences are below the threshold of what we could
1603 * estimate accurately anyway. Note that the costs of evaluating the
1604 * referenced CTE query are added into the final plan as initplan costs,
1605 * and should NOT be counted here.
1606 */
1607 void
1610 {
1613 QualCost qpqual_cost;
1614 Cost cpu_per_tuple;
1616 /* Should only be applied to base relations that are CTEs */
1620 /* Mark the path with the correct row estimate */
1623 else
1626 /* Charge one CPU tuple cost per row for tuplestore manipulation */
1629 /* Add scanning CPU costs */
1636 /* tlist eval costs are paid per output row, not per tuple scanned */
1642 }
1644 /*
1645 * cost_namedtuplestorescan
1646 * Determines and returns the cost of scanning a named tuplestore.
1647 */
1648 void
1651 {
1654 QualCost qpqual_cost;
1655 Cost cpu_per_tuple;
1657 /* Should only be applied to base relations that are Tuplestores */
1661 /* Mark the path with the correct row estimate */
1664 else
1667 /* Charge one CPU tuple cost per row for tuplestore manipulation */
1670 /* Add scanning CPU costs */
1679 }
1681 /*
1682 * cost_resultscan
1683 * Determines and returns the cost of scanning an RTE_RESULT relation.
1684 */
1685 void
1688 {
1691 QualCost qpqual_cost;
1692 Cost cpu_per_tuple;
1694 /* Should only be applied to RTE_RESULT base relations */
1698 /* Mark the path with the correct row estimate */
1701 else
1704 /* We charge qual cost plus cpu_tuple_cost */
1713 }
1715 /*
1716 * cost_recursive_union
1717 * Determines and returns the cost of performing a recursive union,
1718 * and also the estimated output size.
1719 *
1720 * We are given Paths for the nonrecursive and recursive terms.
1721 */
1722 void
1724 {
1725 Cost startup_cost;
1726 Cost total_cost;
1729 /* We probably have decent estimates for the non-recursive term */
1734 /*
1735 * We arbitrarily assume that about 10 recursive iterations will be
1736 * needed, and that we've managed to get a good fix on the cost and output
1737 * size of each one of them. These are mighty shaky assumptions but it's
1738 * hard to see how to do better.
1739 */
1743 /*
1744 * Also charge cpu_tuple_cost per row to account for the costs of
1745 * manipulating the tuplestores. (We don't worry about possible
1746 * spill-to-disk costs.)
1747 */
1755 }
1757 /*
1758 * cost_tuplesort
1759 * Determines and returns the cost of sorting a relation using tuplesort,
1760 * not including the cost of reading the input data.
1761 *
1762 * If the total volume of data to sort is less than sort_mem, we will do
1763 * an in-memory sort, which requires no I/O and about t*log2(t) tuple
1764 * comparisons for t tuples.
1765 *
1766 * If the total volume exceeds sort_mem, we switch to a tape-style merge
1767 * algorithm. There will still be about t*log2(t) tuple comparisons in
1768 * total, but we will also need to write and read each tuple once per
1769 * merge pass. We expect about ceil(logM(r)) merge passes where r is the
1770 * number of initial runs formed and M is the merge order used by tuplesort.c.
1771 * Since the average initial run should be about sort_mem, we have
1772 * disk traffic = 2 * relsize * ceil(logM(p / sort_mem))
1773 * cpu = comparison_cost * t * log2(t)
1774 *
1775 * If the sort is bounded (i.e., only the first k result tuples are needed)
1776 * and k tuples can fit into sort_mem, we use a heap method that keeps only
1777 * k tuples in the heap; this will require about t*log2(k) tuple comparisons.
1778 *
1779 * The disk traffic is assumed to be 3/4ths sequential and 1/4th random
1780 * accesses (XXX can't we refine that guess?)
1781 *
1782 * By default, we charge two operator evals per tuple comparison, which should
1783 * be in the right ballpark in most cases. The caller can tweak this by
1784 * specifying nonzero comparison_cost; typically that's used for any extra
1785 * work that has to be done to prepare the inputs to the comparison operators.
1786 *
1787 * 'tuples' is the number of tuples in the relation
1788 * 'width' is the average tuple width in bytes
1789 * 'comparison_cost' is the extra cost per comparison, if any
1790 * 'sort_mem' is the number of kilobytes of work memory allowed for the sort
1791 * 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
1792 */
1793 static void
1798 {
1804 /*
1805 * We want to be sure the cost of a sort is never estimated as zero, even
1806 * if passed-in tuple count is zero. Besides, mustn't do log(0)...
1807 */
1811 /* Include the default cost-per-comparison */
1814 /* Do we have a useful LIMIT? */
1816 {
1819 }
1820 else
1821 {
1824 }
1827 {
1828 /*
1829 * We'll have to use a disk-based sort of all the tuples
1830 */
1837 /*
1838 * CPU costs
1839 *
1840 * Assume about N log2 N comparisons
1841 */
1844 /* Disk costs */
1846 /* Compute logM(r) as log(r) / log(M) */
1849 else
1852 /* Assume 3/4ths of accesses are sequential, 1/4th are not */
1855 }
1857 {
1858 /*
1859 * We'll use a bounded heap-sort keeping just K tuples in memory, for
1860 * a total number of tuple comparisons of N log2 K; but the constant
1861 * factor is a bit higher than for quicksort. Tweak it so that the
1862 * cost curve is continuous at the crossover point.
1863 */
1865 }
1866 else
1867 {
1868 /* We'll use plain quicksort on all the input tuples */
1870 }
1872 /*
1873 * Also charge a small amount (arbitrarily set equal to operator cost) per
1874 * extracted tuple. We don't charge cpu_tuple_cost because a Sort node
1875 * doesn't do qual-checking or projection, so it has less overhead than
1876 * most plan nodes. Note it's correct to use tuples not output_tuples
1877 * here --- the upper LIMIT will pro-rate the run cost so we'd be double
1878 * counting the LIMIT otherwise.
1879 */
1881 }
1883 /*
1884 * cost_incremental_sort
1885 * Determines and returns the cost of sorting a relation incrementally, when
1886 * the input path is presorted by a prefix of the pathkeys.
1887 *
1888 * 'presorted_keys' is the number of leading pathkeys by which the input path
1889 * is sorted.
1890 *
1891 * We estimate the number of groups into which the relation is divided by the
1892 * leading pathkeys, and then calculate the cost of sorting a single group
1893 * with tuplesort using cost_tuplesort().
1894 */
1895 void
1901 {
1906 input_groups;
1907 Cost group_startup_cost,
1908 group_run_cost,
1909 group_input_run_cost;
1917 /*
1918 * We want to be sure the cost of a sort is never estimated as zero, even
1919 * if passed-in tuple count is zero. Besides, mustn't do log(0)...
1920 */
1924 /* Default estimate of number of groups, capped to one group per row. */
1927 /*
1928 * Extract presorted keys as list of expressions.
1929 *
1930 * We need to be careful about Vars containing "varno 0" which might have
1931 * been introduced by generate_append_tlist, which would confuse
1932 * estimate_num_groups (in fact it'd fail for such expressions). See
1933 * recurse_set_operations which has to deal with the same issue.
1934 *
1935 * Unlike recurse_set_operations we can't access the original target list
1936 * here, and even if we could it's not very clear how useful would that be
1937 * for a set operation combining multiple tables. So we simply detect if
1938 * there are any expressions with "varno 0" and use the default
1939 * DEFAULT_NUM_DISTINCT in that case.
1940 *
1941 * We might also use either 1.0 (a single group) or input_tuples (each row
1942 * being a separate group), pretty much the worst and best case for
1943 * incremental sort. But those are extreme cases and using something in
1944 * between seems reasonable. Furthermore, generate_append_tlist is used
1945 * for set operations, which are likely to produce mostly unique output
1946 * anyway - from that standpoint the DEFAULT_NUM_DISTINCT is defensive
1947 * while maintaining lower startup cost.
1948 */
1950 {
1955 /*
1956 * Check if the expression contains Var with "varno 0" so that we
1957 * don't call estimate_num_groups in that case.
1958 */
1960 {
1963 }
1965 /* expression not containing any Vars with "varno 0" */
1968 i++;
1971 }
1973 /* Estimate number of groups with equal presorted keys. */
1981 /*
1982 * Estimate average cost of sorting of one group where presorted keys are
1983 * equal. Incremental sort is sensitive to distribution of tuples to the
1984 * groups, where we're relying on quite rough assumptions. Thus, we're
1985 * pessimistic about incremental sort performance and increase its average
1986 * group size by half.
1987 */
1990 limit_tuples);
1992 /*
1993 * Startup cost of incremental sort is the startup cost of its first group
1994 * plus the cost of its input.
1995 */
1996 startup_cost += group_startup_cost
1999 /*
2000 * After we started producing tuples from the first group, the cost of
2001 * producing all the tuples is given by the cost to finish processing this
2002 * group, plus the total cost to process the remaining groups, plus the
2003 * remaining cost of input.
2004 */
2005 run_cost += group_run_cost
2009 /*
2010 * Incremental sort adds some overhead by itself. Firstly, it has to
2011 * detect the sort groups. This is roughly equal to one extra copy and
2012 * comparison per tuple. Secondly, it has to reset the tuplesort context
2013 * for every group.
2014 */
2021 }
2023 /*
2024 * cost_sort
2025 * Determines and returns the cost of sorting a relation, including
2026 * the cost of reading the input data.
2027 *
2028 * NOTE: some callers currently pass NIL for pathkeys because they
2029 * can't conveniently supply the sort keys. Since this routine doesn't
2030 * currently do anything with pathkeys anyway, that doesn't matter...
2031 * but if it ever does, it should react gracefully to lack of key data.
2032 * (Actually, the thing we'd most likely be interested in is just the number
2033 * of sort keys, which all callers *could* supply.)
2034 */
2035 void
2041 {
2042 Cost startup_cost;
2043 Cost run_cost;
2048 limit_tuples);
2058 }
2060 /*
2061 * append_nonpartial_cost
2062 * Estimate the cost of the non-partial paths in a Parallel Append.
2063 * The non-partial paths are assumed to be the first "numpaths" paths
2064 * from the subpaths list, and to be in order of decreasing cost.
2065 */
2066 static Cost
2068 {
2081 /*
2082 * Array length is number of workers or number of relevant paths,
2083 * whichever is less.
2084 */
2088 /* The first few paths will each be claimed by a different worker. */
2091 {
2097 }
2099 /*
2100 * Since subpaths are sorted by decreasing cost, the last one will have
2101 * the minimum cost.
2102 */
2105 /*
2106 * For each of the remaining subpaths, add its cost to the array element
2107 * with minimum cost.
2108 */
2110 {
2114 /* Consider only the non-partial paths */
2120 /* Update the new min cost array index */
2122 {
2125 }
2126 }
2128 /* Return the highest cost from the array */
2130 {
2133 }
2136 }
2138 /*
2139 * cost_append
2140 * Determines and returns the cost of an Append node.
2141 */
2142 void
2144 {
2155 {
2159 {
2162 /*
2163 * For an unordered, non-parallel-aware Append we take the startup
2164 * cost as the startup cost of the first subpath.
2165 */
2168 /* Compute rows and costs as sums of subplan rows and costs. */
2170 {
2175 }
2176 }
2177 else
2178 {
2179 /*
2180 * For an ordered, non-parallel-aware Append we take the startup
2181 * cost as the sum of the subpath startup costs. This ensures
2182 * that we don't underestimate the startup cost when a query's
2183 * LIMIT is such that several of the children have to be run to
2184 * satisfy it. This might be overkill --- another plausible hack
2185 * would be to take the Append's startup cost as the maximum of
2186 * the child startup costs. But we don't want to risk believing
2187 * that an ORDER BY LIMIT query can be satisfied at small cost
2188 * when the first child has small startup cost but later ones
2189 * don't. (If we had the ability to deal with nonlinear cost
2190 * interpolation for partial retrievals, we would not need to be
2191 * so conservative about this.)
2192 *
2193 * This case is also different from the above in that we have to
2194 * account for possibly injecting sorts into subpaths that aren't
2195 * natively ordered.
2196 */
2198 {
2203 {
2204 /*
2205 * We'll need to insert a Sort node, so include costs for
2206 * that. We can use the parent's LIMIT if any, since we
2207 * certainly won't pull more than that many tuples from
2208 * any child.
2209 */
2212 pathkeys,
2217 work_mem,
2220 }
2225 }
2226 }
2227 }
2229 {
2233 /* Parallel-aware Append never produces ordered output. */
2236 /* Calculate startup cost. */
2238 {
2241 /*
2242 * Append will start returning tuples when the child node having
2243 * lowest startup cost is done setting up. We consider only the
2244 * first few subplans that immediately get a worker assigned.
2245 */
2252 /*
2253 * Apply parallel divisor to subpaths. Scale the number of rows
2254 * for each partial subpath based on the ratio of the parallel
2255 * divisor originally used for the subpath to the one we adopted.
2256 * Also add the cost of partial paths to the total cost, but
2257 * ignore non-partial paths for now.
2258 */
2261 else
2262 {
2267 parallel_divisor);
2269 }
2273 i++;
2274 }
2276 /* Add cost for non-partial subpaths. */
2281 }
2283 /*
2284 * Although Append does not do any selection or projection, it's not free;
2285 * add a small per-tuple overhead.
2286 */
2289 }
2291 /*
2292 * cost_merge_append
2293 * Determines and returns the cost of a MergeAppend node.
2294 *
2295 * MergeAppend merges several pre-sorted input streams, using a heap that
2296 * at any given instant holds the next tuple from each stream. If there
2297 * are N streams, we need about N*log2(N) tuple comparisons to construct
2298 * the heap at startup, and then for each output tuple, about log2(N)
2299 * comparisons to replace the top entry.
2300 *
2301 * (The effective value of N will drop once some of the input streams are
2302 * exhausted, but it seems unlikely to be worth trying to account for that.)
2303 *
2304 * The heap is never spilled to disk, since we assume N is not very large.
2305 * So this is much simpler than cost_sort.
2306 *
2307 * As in cost_sort, we charge two operator evals per tuple comparison.
2308 *
2309 * 'pathkeys' is a list of sort keys
2310 * 'n_streams' is the number of input streams
2311 * 'input_startup_cost' is the sum of the input streams' startup costs
2312 * 'input_total_cost' is the sum of the input streams' total costs
2313 * 'tuples' is the number of tuples in all the streams
2314 */
2315 void
2320 {
2323 Cost comparison_cost;
2327 /*
2328 * Avoid log(0)...
2329 */
2333 /* Assumed cost per tuple comparison */
2336 /* Heap creation cost */
2339 /* Per-tuple heap maintenance cost */
2342 /*
2343 * Although MergeAppend does not do any selection or projection, it's not
2344 * free; add a small per-tuple overhead.
2345 */
2350 }
2352 /*
2353 * cost_material
2354 * Determines and returns the cost of materializing a relation, including
2355 * the cost of reading the input data.
2356 *
2357 * If the total volume of data to materialize exceeds work_mem, we will need
2358 * to write it to disk, so the cost is much higher in that case.
2359 *
2360 * Note that here we are estimating the costs for the first scan of the
2361 * relation, so the materialization is all overhead --- any savings will
2362 * occur only on rescan, which is estimated in cost_rescan.
2363 */
2364 void
2368 {
2376 /*
2377 * Whether spilling or not, charge 2x cpu_operator_cost per tuple to
2378 * reflect bookkeeping overhead. (This rate must be more than what
2379 * cost_rescan charges for materialize, ie, cpu_operator_cost per tuple;
2380 * if it is exactly the same then there will be a cost tie between
2381 * nestloop with A outer, materialized B inner and nestloop with B outer,
2382 * materialized A inner. The extra cost ensures we'll prefer
2383 * materializing the smaller rel.) Note that this is normally a good deal
2384 * less than cpu_tuple_cost; which is OK because a Material plan node
2385 * doesn't do qual-checking or projection, so it's got less overhead than
2386 * most plan nodes.
2387 */
2390 /*
2391 * If we will spill to disk, charge at the rate of seq_page_cost per page.
2392 * This cost is assumed to be evenly spread through the plan run phase,
2393 * which isn't exactly accurate but our cost model doesn't allow for
2394 * nonuniform costs within the run phase.
2395 */
2397 {
2401 }
2405 }
2407 /*
2408 * cost_memoize_rescan
2409 * Determines the estimated cost of rescanning a Memoize node.
2410 *
2411 * In order to estimate this, we must gain knowledge of how often we expect to
2412 * be called and how many distinct sets of parameters we are likely to be
2413 * called with. If we expect a good cache hit ratio, then we can set our
2414 * costs to account for that hit ratio, plus a little bit of cost for the
2415 * caching itself. Caching will not work out well if we expect to be called
2416 * with too many distinct parameter values. The worst-case here is that we
2417 * never see any parameter value twice, in which case we'd never get a cache
2418 * hit and caching would be a complete waste of effort.
2419 */
2420 static void
2423 {
2424 EstimationInfo estinfo;
2437 Cost startup_cost;
2438 Cost total_cost;
2440 /* available cache space */
2443 /*
2444 * Set the number of bytes each cache entry should consume in the cache.
2445 * To provide us with better estimations on how many cache entries we can
2446 * store at once, we make a call to the executor here to ask it what
2447 * memory overheads there are for a single cache entry.
2448 *
2449 * XXX we also store the cache key, but that's not accounted for here.
2450 */
2454 /* estimate on the upper limit of cache entries we can hold at once */
2457 /* estimate on the distinct number of parameter values */
2461 /*
2462 * When the estimation fell back on using a default value, it's a bit too
2463 * risky to assume that it's ok to use a Memoize node. The use of a
2464 * default could cause us to use a Memoize node when it's really
2465 * inappropriate to do so. If we see that this has been done, then we'll
2466 * assume that every call will have unique parameters, which will almost
2467 * certainly mean a MemoizePath will never survive add_path().
2468 */
2472 /*
2473 * Since we've already estimated the maximum number of entries we can
2474 * store at once and know the estimated number of distinct values we'll be
2475 * called with, we'll take this opportunity to set the path's est_entries.
2476 * This will ultimately determine the hash table size that the executor
2477 * will use. If we leave this at zero, the executor will just choose the
2478 * size itself. Really this is not the right place to do this, but it's
2479 * convenient since everything is already calculated.
2480 */
2482 PG_UINT32_MAX);
2484 /*
2485 * When the number of distinct parameter values is above the amount we can
2486 * store in the cache, then we'll have to evict some entries from the
2487 * cache. This is not free. Here we estimate how often we'll incur the
2488 * cost of that eviction.
2489 */
2492 /*
2493 * In order to estimate how costly a single scan will be, we need to
2494 * attempt to estimate what the cache hit ratio will be. To do that we
2495 * must look at how many scans are estimated in total for this node and
2496 * how many of those scans we expect to get a cache hit.
2497 */
2501 /* Ensure we don't go negative */
2504 /*
2505 * Set the total_cost accounting for the expected cache hit ratio. We
2506 * also add on a cpu_operator_cost to account for a cache lookup. This
2507 * will happen regardless of whether it's a cache hit or not.
2508 */
2511 /* Now adjust the total cost to account for cache evictions */
2513 /* Charge a cpu_tuple_cost for evicting the actual cache entry */
2516 /*
2517 * Charge a 10th of cpu_operator_cost to evict every tuple in that entry.
2518 * The per-tuple eviction is really just a pfree, so charging a whole
2519 * cpu_operator_cost seems a little excessive.
2520 */
2523 /*
2524 * Now adjust for storing things in the cache, since that's not free
2525 * either. Everything must go in the cache. We don't proportion this
2526 * over any ratio, just apply it once for the scan. We charge a
2527 * cpu_tuple_cost for the creation of the cache entry and also a
2528 * cpu_operator_cost for each tuple we expect to cache.
2529 */
2532 /*
2533 * Getting the first row must be also be proportioned according to the
2534 * expected cache hit ratio.
2535 */
2538 /*
2539 * Additionally we charge a cpu_tuple_cost to account for cache lookups,
2540 * which we'll do regardless of whether it was a cache hit or not.
2541 */
2546 }
2548 /*
2549 * cost_agg
2550 * Determines and returns the cost of performing an Agg plan node,
2551 * including the cost of its input.
2552 *
2553 * aggcosts can be NULL when there are no actual aggregate functions (i.e.,
2554 * we are using a hashed Agg node just to do grouping).
2555 *
2556 * Note: when aggstrategy == AGG_SORTED, caller must ensure that input costs
2557 * are for appropriately-sorted input.
2558 */
2559 void
2566 {
2568 Cost startup_cost;
2569 Cost total_cost;
2570 AggClauseCosts dummy_aggcosts;
2572 /* Use all-zero per-aggregate costs if NULL is passed */
2574 {
2578 }
2580 /*
2581 * The transCost.per_tuple component of aggcosts should be charged once
2582 * per input tuple, corresponding to the costs of evaluating the aggregate
2583 * transfns and their input expressions. The finalCost.per_tuple component
2584 * is charged once per output tuple, corresponding to the costs of
2585 * evaluating the finalfns. Startup costs are of course charged but once.
2586 *
2587 * If we are grouping, we charge an additional cpu_operator_cost per
2588 * grouping column per input tuple for grouping comparisons.
2589 *
2590 * We will produce a single output tuple if not grouping, and a tuple per
2591 * group otherwise. We charge cpu_tuple_cost for each output tuple.
2592 *
2593 * Note: in this cost model, AGG_SORTED and AGG_HASHED have exactly the
2594 * same total CPU cost, but AGG_SORTED has lower startup cost. If the
2595 * input path is already sorted appropriately, AGG_SORTED should be
2596 * preferred (since it has no risk of memory overflow). This will happen
2597 * as long as the computed total costs are indeed exactly equal --- but if
2598 * there's roundoff error we might do the wrong thing. So be sure that
2599 * the computations below form the same intermediate values in the same
2600 * order.
2601 */
2603 {
2609 /* we aren't grouping */
2612 }
2614 {
2615 /* Here we are able to deliver output on-the-fly */
2619 {
2622 }
2623 /* calcs phrased this way to match HASHED case, see note above */
2631 }
2632 else
2633 {
2634 /* must be AGG_HASHED */
2640 /* cost of computing hash value */
2646 /* cost of retrieving from hash table */
2649 }
2651 /*
2652 * Add the disk costs of hash aggregation that spills to disk.
2653 *
2654 * Groups that go into the hash table stay in memory until finalized, so
2655 * spilling and reprocessing tuples doesn't incur additional invocations
2656 * of transCost or finalCost. Furthermore, the computed hash value is
2657 * stored with the spilled tuples, so we don't incur extra invocations of
2658 * the hash function.
2659 *
2660 * Hash Agg begins returning tuples after the first batch is complete.
2661 * Accrue writes (spilled tuples) to startup_cost and to total_cost;
2662 * accrue reads only to total_cost.
2663 */
2665 {
2672 Size mem_limit;
2673 uint64 ngroups_limit;
2677 /*
2678 * Estimate number of batches based on the computed limits. If less
2679 * than or equal to one, all groups are expected to fit in memory;
2680 * otherwise we expect to spill.
2681 */
2683 input_width,
2694 /*
2695 * The number of partitions can change at different levels of
2696 * recursion; but for the purposes of this calculation assume it stays
2697 * constant.
2698 */
2701 /*
2702 * Estimate number of pages read and written. For each level of
2703 * recursion, a tuple must be written and then later read.
2704 */
2708 /*
2709 * HashAgg has somewhat worse IO behavior than Sort on typical
2710 * hardware/OS combinations. Account for this with a generic penalty.
2711 */
2719 /* account for CPU cost of spilling a tuple and reading it back */
2723 }
2725 /*
2726 * If there are quals (HAVING quals), account for their cost and
2727 * selectivity.
2728 */
2730 {
2731 QualCost qual_cost;
2739 quals,
2741 JOIN_INNER,
2742 NULL));
2743 }
2748 }
2750 /*
2751 * cost_windowagg
2752 * Determines and returns the cost of performing a WindowAgg plan node,
2753 * including the cost of its input.
2754 *
2755 * Input is assumed already properly sorted.
2756 */
2757 void
2762 {
2763 Cost startup_cost;
2764 Cost total_cost;
2770 /*
2771 * Window functions are assumed to cost their stated execution cost, plus
2772 * the cost of evaluating their input expressions, per tuple. Since they
2773 * may in fact evaluate their inputs at multiple rows during each cycle,
2774 * this could be a drastic underestimate; but without a way to know how
2775 * many rows the window function will fetch, it's hard to do better. In
2776 * any case, it's a good estimate for all the built-in window functions,
2777 * so we'll just do this for now.
2778 */
2780 {
2782 Cost wfunccost;
2783 QualCost argcosts;
2791 /* also add the input expressions' cost to per-input-row costs */
2796 /*
2797 * Add the filter's cost to per-input-row costs. XXX We should reduce
2798 * input expression costs according to filter selectivity.
2799 */
2805 }
2807 /*
2808 * We also charge cpu_operator_cost per grouping column per tuple for
2809 * grouping comparisons, plus cpu_tuple_cost per tuple for general
2810 * overhead.
2811 *
2812 * XXX this neglects costs of spooling the data to disk when it overflows
2813 * work_mem. Sooner or later that should get accounted for.
2814 */
2821 }
2823 /*
2824 * cost_group
2825 * Determines and returns the cost of performing a Group plan node,
2826 * including the cost of its input.
2827 *
2828 * Note: caller must ensure that input costs are for appropriately-sorted
2829 * input.
2830 */
2831 void
2837 {
2839 Cost startup_cost;
2840 Cost total_cost;
2846 /*
2847 * Charge one cpu_operator_cost per comparison per input tuple. We assume
2848 * all columns get compared at most of the tuples.
2849 */
2852 /*
2853 * If there are quals (HAVING quals), account for their cost and
2854 * selectivity.
2855 */
2857 {
2858 QualCost qual_cost;
2866 quals,
2868 JOIN_INNER,
2869 NULL));
2870 }
2875 }
2877 /*
2878 * initial_cost_nestloop
2879 * Preliminary estimate of the cost of a nestloop join path.
2880 *
2881 * This must quickly produce lower-bound estimates of the path's startup and
2882 * total costs. If we are unable to eliminate the proposed path from
2883 * consideration using the lower bounds, final_cost_nestloop will be called
2884 * to obtain the final estimates.
2885 *
2886 * The exact division of labor between this function and final_cost_nestloop
2887 * is private to them, and represents a tradeoff between speed of the initial
2888 * estimate and getting a tight lower bound. We choose to not examine the
2889 * join quals here, since that's by far the most expensive part of the
2890 * calculations. The end result is that CPU-cost considerations must be
2891 * left for the second phase; and for SEMI/ANTI joins, we must also postpone
2892 * incorporation of the inner path's run cost.
2893 *
2894 * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
2895 * other data to be used by final_cost_nestloop
2896 * 'jointype' is the type of join to be performed
2897 * 'outer_path' is the outer input to the join
2898 * 'inner_path' is the inner input to the join
2899 * 'extra' contains miscellaneous information about the join
2900 */
2901 void
2903 JoinType jointype,
2906 {
2910 Cost inner_rescan_start_cost;
2911 Cost inner_rescan_total_cost;
2912 Cost inner_run_cost;
2913 Cost inner_rescan_run_cost;
2915 /* estimate costs to rescan the inner relation */
2920 /* cost of source data */
2922 /*
2923 * NOTE: clearly, we must pay both outer and inner paths' startup_cost
2924 * before we can start returning tuples, so the join's startup cost is
2925 * their sum. We'll also pay the inner path's rescan startup cost
2926 * multiple times.
2927 */
2938 {
2939 /*
2940 * With a SEMI or ANTI join, or if the innerrel is known unique, the
2941 * executor will stop after the first match.
2942 *
2943 * Getting decent estimates requires inspection of the join quals,
2944 * which we choose to postpone to final_cost_nestloop.
2945 */
2947 /* Save private data for final_cost_nestloop */
2950 }
2951 else
2952 {
2953 /* Normal case; we'll scan whole input rel for each outer row */
2957 }
2959 /* CPU costs left for later */
2961 /* Public result fields */
2964 /* Save private data for final_cost_nestloop */
2966 }
2968 /*
2969 * final_cost_nestloop
2970 * Final estimate of the cost and result size of a nestloop join path.
2971 *
2972 * 'path' is already filled in except for the rows and cost fields
2973 * 'workspace' is the result from initial_cost_nestloop
2974 * 'extra' contains miscellaneous information about the join
2975 */
2976 void
2980 {
2987 Cost cpu_per_tuple;
2988 QualCost restrict_qual_cost;
2991 /* Protect some assumptions below that rowcounts aren't zero */
2996 /* Mark the path with the correct row estimate */
2999 else
3002 /* For partial paths, scale row estimate. */
3004 {
3009 }
3011 /*
3012 * We could include disable_cost in the preliminary estimate, but that
3013 * would amount to optimizing for the case where the join method is
3014 * disabled, which doesn't seem like the way to bet.
3015 */
3019 /* cost of inner-relation source data (we already dealt with outer rel) */
3023 {
3024 /*
3025 * With a SEMI or ANTI join, or if the innerrel is known unique, the
3026 * executor will stop after the first match.
3027 */
3032 Selectivity inner_scan_frac;
3034 /*
3035 * For an outer-rel row that has at least one match, we can expect the
3036 * inner scan to stop after a fraction 1/(match_count+1) of the inner
3037 * rows, if the matches are evenly distributed. Since they probably
3038 * aren't quite evenly distributed, we apply a fuzz factor of 2.0 to
3039 * that fraction. (If we used a larger fuzz factor, we'd have to
3040 * clamp inner_scan_frac to at most 1.0; but since match_count is at
3041 * least 1, no such clamp is needed now.)
3042 */
3047 /*
3048 * Compute number of tuples processed (not number emitted!). First,
3049 * account for successfully-matched outer rows.
3050 */
3053 /*
3054 * Now we need to estimate the actual costs of scanning the inner
3055 * relation, which may be quite a bit less than N times inner_run_cost
3056 * due to early scan stops. We consider two cases. If the inner path
3057 * is an indexscan using all the joinquals as indexquals, then an
3058 * unmatched outer row results in an indexscan returning no rows,
3059 * which is probably quite cheap. Otherwise, the executor will have
3060 * to scan the whole inner rel for an unmatched row; not so cheap.
3061 */
3063 {
3064 /*
3065 * Successfully-matched outer rows will only require scanning
3066 * inner_scan_frac of the inner relation. In this case, we don't
3067 * need to charge the full inner_run_cost even when that's more
3068 * than inner_rescan_run_cost, because we can assume that none of
3069 * the inner scans ever scan the whole inner relation. So it's
3070 * okay to assume that all the inner scan executions can be
3071 * fractions of the full cost, even if materialization is reducing
3072 * the rescan cost. At this writing, it's impossible to get here
3073 * for a materialized inner scan, so inner_run_cost and
3074 * inner_rescan_run_cost will be the same anyway; but just in
3075 * case, use inner_run_cost for the first matched tuple and
3076 * inner_rescan_run_cost for additional ones.
3077 */
3082 /*
3083 * Add the cost of inner-scan executions for unmatched outer rows.
3084 * We estimate this as the same cost as returning the first tuple
3085 * of a nonempty scan. We consider that these are all rescans,
3086 * since we used inner_run_cost once already.
3087 */
3091 /*
3092 * We won't be evaluating any quals at all for unmatched rows, so
3093 * don't add them to ntuples.
3094 */
3095 }
3096 else
3097 {
3098 /*
3099 * Here, a complicating factor is that rescans may be cheaper than
3100 * first scans. If we never scan all the way to the end of the
3101 * inner rel, it might be (depending on the plan type) that we'd
3102 * never pay the whole inner first-scan run cost. However it is
3103 * difficult to estimate whether that will happen (and it could
3104 * not happen if there are any unmatched outer rows!), so be
3105 * conservative and always charge the whole first-scan cost once.
3106 * We consider this charge to correspond to the first unmatched
3107 * outer row, unless there isn't one in our estimate, in which
3108 * case blame it on the first matched row.
3109 */
3111 /* First, count all unmatched join tuples as being processed */
3114 /* Now add the forced full scan, and decrement appropriate count */
3118 else
3121 /* Add inner run cost for additional outer tuples having matches */
3125 /* Add inner run cost for additional unmatched outer tuples */
3128 }
3129 }
3130 else
3131 {
3132 /* Normal-case source costs were included in preliminary estimate */
3134 /* Compute number of tuples processed (not number emitted!) */
3136 }
3138 /* CPU costs */
3144 /* tlist eval costs are paid per output row, not per tuple scanned */
3150 }
3152 /*
3153 * initial_cost_mergejoin
3154 * Preliminary estimate of the cost of a mergejoin path.
3155 *
3156 * This must quickly produce lower-bound estimates of the path's startup and
3157 * total costs. If we are unable to eliminate the proposed path from
3158 * consideration using the lower bounds, final_cost_mergejoin will be called
3159 * to obtain the final estimates.
3160 *
3161 * The exact division of labor between this function and final_cost_mergejoin
3162 * is private to them, and represents a tradeoff between speed of the initial
3163 * estimate and getting a tight lower bound. We choose to not examine the
3164 * join quals here, except for obtaining the scan selectivity estimate which
3165 * is really essential (but fortunately, use of caching keeps the cost of
3166 * getting that down to something reasonable).
3167 * We also assume that cost_sort is cheap enough to use here.
3168 *
3169 * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3170 * other data to be used by final_cost_mergejoin
3171 * 'jointype' is the type of join to be performed
3172 * 'mergeclauses' is the list of joinclauses to be used as merge clauses
3173 * 'outer_path' is the outer input to the join
3174 * 'inner_path' is the inner input to the join
3175 * 'outersortkeys' is the list of sort keys for the outer path
3176 * 'innersortkeys' is the list of sort keys for the inner path
3177 * 'extra' contains miscellaneous information about the join
3178 *
3179 * Note: outersortkeys and innersortkeys should be NIL if no explicit
3180 * sort is needed because the respective source path is already ordered.
3181 */
3182 void
3184 JoinType jointype,
3189 {
3194 Cost inner_run_cost;
3196 inner_rows,
3197 outer_skip_rows,
3198 inner_skip_rows;
3199 Selectivity outerstartsel,
3200 outerendsel,
3201 innerstartsel,
3202 innerendsel;
3205 /* Protect some assumptions below that rowcounts aren't zero */
3211 /*
3212 * A merge join will stop as soon as it exhausts either input stream
3213 * (unless it's an outer join, in which case the outer side has to be
3214 * scanned all the way anyway). Estimate fraction of the left and right
3215 * inputs that will actually need to be scanned. Likewise, we can
3216 * estimate the number of rows that will be skipped before the first join
3217 * pair is found, which should be factored into startup cost. We use only
3218 * the first (most significant) merge clause for this purpose. Since
3219 * mergejoinscansel() is a fairly expensive computation, we cache the
3220 * results in the merge clause RestrictInfo.
3221 */
3223 {
3231 /* Get the input pathkeys to determine the sort-order details */
3238 /* debugging check */
3245 /* Get the selectivity with caching */
3250 {
3251 /* left side of clause is outer */
3256 }
3257 else
3258 {
3259 /* left side of clause is inner */
3264 }
3267 {
3270 }
3272 {
3275 }
3276 }
3277 else
3278 {
3279 /* cope with clauseless or full mergejoin */
3282 }
3284 /*
3285 * Convert selectivities to row counts. We force outer_rows and
3286 * inner_rows to be at least 1, but the skip_rows estimates can be zero.
3287 */
3296 /*
3297 * Readjust scan selectivities to account for above rounding. This is
3298 * normally an insignificant effect, but when there are only a few rows in
3299 * the inputs, failing to do this makes for a large percentage error.
3300 */
3309 /* cost of source data */
3312 {
3314 root,
3315 outersortkeys,
3317 outer_path_rows,
3320 work_mem,
3327 }
3328 else
3329 {
3335 }
3338 {
3340 root,
3341 innersortkeys,
3343 inner_path_rows,
3346 work_mem,
3353 }
3354 else
3355 {
3361 }
3363 /*
3364 * We can't yet determine whether rescanning occurs, or whether
3365 * materialization of the inner input should be done. The minimum
3366 * possible inner input cost, regardless of rescan and materialization
3367 * considerations, is inner_run_cost. We include that in
3368 * workspace->total_cost, but not yet in run_cost.
3369 */
3371 /* CPU costs left for later */
3373 /* Public result fields */
3376 /* Save private data for final_cost_mergejoin */
3383 }
3385 /*
3386 * final_cost_mergejoin
3387 * Final estimate of the cost and result size of a mergejoin path.
3388 *
3389 * Unlike other costsize functions, this routine makes two actual decisions:
3390 * whether the executor will need to do mark/restore, and whether we should
3391 * materialize the inner path. It would be logically cleaner to build
3392 * separate paths testing these alternatives, but that would require repeating
3393 * most of the cost calculations, which are not all that cheap. Since the
3394 * choice will not affect output pathkeys or startup cost, only total cost,
3395 * there is no possibility of wanting to keep more than one path. So it seems
3396 * best to make the decisions here and record them in the path's
3397 * skip_mark_restore and materialize_inner fields.
3398 *
3399 * Mark/restore overhead is usually required, but can be skipped if we know
3400 * that the executor need find only one match per outer tuple, and that the
3401 * mergeclauses are sufficient to identify a match.
3402 *
3403 * We materialize the inner path if we need mark/restore and either the inner
3404 * path can't support mark/restore, or it's cheaper to use an interposed
3405 * Material node to handle mark/restore.
3406 *
3407 * 'path' is already filled in except for the rows and cost fields and
3408 * skip_mark_restore and materialize_inner
3409 * 'workspace' is the result from initial_cost_mergejoin
3410 * 'extra' contains miscellaneous information about the join
3411 */
3412 void
3416 {
3429 Cost cpu_per_tuple,
3430 bare_inner_cost,
3431 mat_inner_cost;
3432 QualCost merge_qual_cost;
3433 QualCost qp_qual_cost;
3435 rescannedtuples;
3438 /* Protect some assumptions below that rowcounts aren't zero */
3442 /* Mark the path with the correct row estimate */
3445 else
3448 /* For partial paths, scale row estimate. */
3450 {
3455 }
3457 /*
3458 * We could include disable_cost in the preliminary estimate, but that
3459 * would amount to optimizing for the case where the join method is
3460 * disabled, which doesn't seem like the way to bet.
3461 */
3465 /*
3466 * Compute cost of the mergequals and qpquals (other restriction clauses)
3467 * separately.
3468 */
3474 /*
3475 * With a SEMI or ANTI join, or if the innerrel is known unique, the
3476 * executor will stop scanning for matches after the first match. When
3477 * all the joinclauses are merge clauses, this means we don't ever need to
3478 * back up the merge, and so we can skip mark/restore overhead.
3479 */
3486 else
3489 /*
3490 * Get approx # tuples passing the mergequals. We use approx_tuple_count
3491 * here because we need an estimate done with JOIN_INNER semantics.
3492 */
3495 /*
3496 * When there are equal merge keys in the outer relation, the mergejoin
3497 * must rescan any matching tuples in the inner relation. This means
3498 * re-fetching inner tuples; we have to estimate how often that happens.
3499 *
3500 * For regular inner and outer joins, the number of re-fetches can be
3501 * estimated approximately as size of merge join output minus size of
3502 * inner relation. Assume that the distinct key values are 1, 2, ..., and
3503 * denote the number of values of each key in the outer relation as m1,
3504 * m2, ...; in the inner relation, n1, n2, ... Then we have
3505 *
3506 * size of join = m1 * n1 + m2 * n2 + ...
3507 *
3508 * number of rescanned tuples = (m1 - 1) * n1 + (m2 - 1) * n2 + ... = m1 *
3509 * n1 + m2 * n2 + ... - (n1 + n2 + ...) = size of join - size of inner
3510 * relation
3511 *
3512 * This equation works correctly for outer tuples having no inner match
3513 * (nk = 0), but not for inner tuples having no outer match (mk = 0); we
3514 * are effectively subtracting those from the number of rescanned tuples,
3515 * when we should not. Can we do better without expensive selectivity
3516 * computations?
3517 *
3518 * The whole issue is moot if we are working from a unique-ified outer
3519 * input, or if we know we don't need to mark/restore at all.
3520 */
3523 else
3524 {
3526 /* Must clamp because of possible underestimate */
3529 }
3531 /*
3532 * We'll inflate various costs this much to account for rescanning. Note
3533 * that this is to be multiplied by something involving inner_rows, or
3534 * another number related to the portion of the inner rel we'll scan.
3535 */
3538 /*
3539 * Decide whether we want to materialize the inner input to shield it from
3540 * mark/restore and performing re-fetches. Our cost model for regular
3541 * re-fetches is that a re-fetch costs the same as an original fetch,
3542 * which is probably an overestimate; but on the other hand we ignore the
3543 * bookkeeping costs of mark/restore. Not clear if it's worth developing
3544 * a more refined model. So we just need to inflate the inner run cost by
3545 * rescanratio.
3546 */
3549 /*
3550 * When we interpose a Material node the re-fetch cost is assumed to be
3551 * just cpu_operator_cost per tuple, independently of the underlying
3552 * plan's cost; and we charge an extra cpu_operator_cost per original
3553 * fetch as well. Note that we're assuming the materialize node will
3554 * never spill to disk, since it only has to remember tuples back to the
3555 * last mark. (If there are a huge number of duplicates, our other cost
3556 * factors will make the path so expensive that it probably won't get
3557 * chosen anyway.) So we don't use cost_rescan here.
3558 *
3559 * Note: keep this estimate in sync with create_mergejoin_plan's labeling
3560 * of the generated Material node.
3561 */
3565 /*
3566 * If we don't need mark/restore at all, we don't need materialization.
3567 */
3571 /*
3572 * Prefer materializing if it looks cheaper, unless the user has asked to
3573 * suppress materialization.
3574 */
3578 /*
3579 * Even if materializing doesn't look cheaper, we *must* do it if the
3580 * inner path is to be used directly (without sorting) and it doesn't
3581 * support mark/restore.
3582 *
3583 * Since the inner side must be ordered, and only Sorts and IndexScans can
3584 * create order to begin with, and they both support mark/restore, you
3585 * might think there's no problem --- but you'd be wrong. Nestloop and
3586 * merge joins can *preserve* the order of their inputs, so they can be
3587 * selected as the input of a mergejoin, and they don't support
3588 * mark/restore at present.
3589 *
3590 * We don't test the value of enable_material here, because
3591 * materialization is required for correctness in this case, and turning
3592 * it off does not entitle us to deliver an invalid plan.
3593 */
3598 /*
3599 * Also, force materializing if the inner path is to be sorted and the
3600 * sort is expected to spill to disk. This is because the final merge
3601 * pass can be done on-the-fly if it doesn't have to support mark/restore.
3602 * We don't try to adjust the cost estimates for this consideration,
3603 * though.
3604 *
3605 * Since materialization is a performance optimization in this case,
3606 * rather than necessary for correctness, we skip it if enable_material is
3607 * off.
3608 */
3614 else
3617 /* Charge the right incremental cost for the chosen case */
3620 else
3623 /* CPU costs */
3625 /*
3626 * The number of tuple comparisons needed is approximately number of outer
3627 * rows plus number of inner rows plus number of rescanned tuples (can we
3628 * refine this?). At each one, we need to evaluate the mergejoin quals.
3629 */
3637 /*
3638 * For each tuple that gets through the mergejoin proper, we charge
3639 * cpu_tuple_cost plus the cost of evaluating additional restriction
3640 * clauses that are to be applied at the join. (This is pessimistic since
3641 * not all of the quals may get evaluated at each tuple.)
3642 *
3643 * Note: we could adjust for SEMI/ANTI joins skipping some qual
3644 * evaluations here, but it's probably not worth the trouble.
3645 */
3650 /* tlist eval costs are paid per output row, not per tuple scanned */
3656 }
3658 /*
3659 * run mergejoinscansel() with caching
3660 */
3663 {
3666 Selectivity leftstartsel,
3667 leftendsel,
3668 rightstartsel,
3669 rightendsel;
3670 MemoryContext oldcontext;
3672 /* Do we have this result already? */
3674 {
3681 }
3683 /* Nope, do the computation */
3694 /* Cache the result in suitably long-lived workspace */
3712 }
3714 /*
3715 * initial_cost_hashjoin
3716 * Preliminary estimate of the cost of a hashjoin path.
3717 *
3718 * This must quickly produce lower-bound estimates of the path's startup and
3719 * total costs. If we are unable to eliminate the proposed path from
3720 * consideration using the lower bounds, final_cost_hashjoin will be called
3721 * to obtain the final estimates.
3722 *
3723 * The exact division of labor between this function and final_cost_hashjoin
3724 * is private to them, and represents a tradeoff between speed of the initial
3725 * estimate and getting a tight lower bound. We choose to not examine the
3726 * join quals here (other than by counting the number of hash clauses),
3727 * so we can't do much with CPU costs. We do assume that
3728 * ExecChooseHashTableSize is cheap enough to use here.
3729 *
3730 * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3731 * other data to be used by final_cost_hashjoin
3732 * 'jointype' is the type of join to be performed
3733 * 'hashclauses' is the list of joinclauses to be used as hash clauses
3734 * 'outer_path' is the outer input to the join
3735 * 'inner_path' is the inner input to the join
3736 * 'extra' contains miscellaneous information about the join
3737 * 'parallel_hash' indicates that inner_path is partial and that a shared
3738 * hash table will be built in parallel
3739 */
3740 void
3742 JoinType jointype,
3747 {
3759 /* cost of source data */
3764 /*
3765 * Cost of computing hash function: must do it once per input tuple. We
3766 * charge one cpu_operator_cost for each column's hash function. Also,
3767 * tack on one cpu_tuple_cost per inner row, to model the costs of
3768 * inserting the row into the hashtable.
3769 *
3770 * XXX when a hashclause is more complex than a single operator, we really
3771 * should charge the extra eval costs of the left or right side, as
3772 * appropriate, here. This seems more work than it's worth at the moment.
3773 */
3778 /*
3779 * If this is a parallel hash build, then the value we have for
3780 * inner_rows_total currently refers only to the rows returned by each
3781 * participant. For shared hash table size estimation, we need the total
3782 * number, so we need to undo the division.
3783 */
3787 /*
3788 * Get hash table size that executor would use for inner relation.
3789 *
3790 * XXX for the moment, always assume that skew optimization will be
3791 * performed. As long as SKEW_HASH_MEM_PERCENT is small, it's not worth
3792 * trying to determine that for sure.
3793 *
3794 * XXX at some point it might be interesting to try to account for skew
3795 * optimization in the cost estimate, but for now, we don't.
3796 */
3807 /*
3808 * If inner relation is too big then we will need to "batch" the join,
3809 * which implies writing and reading most of the tuples to disk an extra
3810 * time. Charge seq_page_cost per page, since the I/O should be nice and
3811 * sequential. Writing the inner rel counts as startup cost, all the rest
3812 * as run cost.
3813 */
3815 {
3823 }
3825 /* CPU costs left for later */
3827 /* Public result fields */
3830 /* Save private data for final_cost_hashjoin */
3835 }
3837 /*
3838 * final_cost_hashjoin
3839 * Final estimate of the cost and result size of a hashjoin path.
3840 *
3841 * Note: the numbatches estimate is also saved into 'path' for use later
3842 *
3843 * 'path' is already filled in except for the rows and cost fields and
3844 * num_batches
3845 * 'workspace' is the result from initial_cost_hashjoin
3846 * 'extra' contains miscellaneous information about the join
3847 */
3848 void
3852 {
3863 Cost cpu_per_tuple;
3864 QualCost hash_qual_cost;
3865 QualCost qp_qual_cost;
3868 Selectivity innerbucketsize;
3869 Selectivity innermcvfreq;
3872 /* Mark the path with the correct row estimate */
3875 else
3878 /* For partial paths, scale row estimate. */
3880 {
3885 }
3887 /*
3888 * We could include disable_cost in the preliminary estimate, but that
3889 * would amount to optimizing for the case where the join method is
3890 * disabled, which doesn't seem like the way to bet.
3891 */
3895 /* mark the path with estimated # of batches */
3898 /* store the total number of tuples (sum of partial row estimates) */
3901 /* and compute the number of "virtual" buckets in the whole join */
3904 /*
3905 * Determine bucketsize fraction and MCV frequency for the inner relation.
3906 * We use the smallest bucketsize or MCV frequency estimated for any
3907 * individual hashclause; this is undoubtedly conservative.
3908 *
3909 * BUT: if inner relation has been unique-ified, we can assume it's good
3910 * for hashing. This is important both because it's the right answer, and
3911 * because we avoid contaminating the cache with a value that's wrong for
3912 * non-unique-ified paths.
3913 */
3915 {
3918 }
3919 else
3920 {
3924 {
3926 Selectivity thisbucketsize;
3927 Selectivity thismcvfreq;
3929 /*
3930 * First we have to figure out which side of the hashjoin clause
3931 * is the inner side.
3932 *
3933 * Since we tend to visit the same clauses over and over when
3934 * planning a large query, we cache the bucket stats estimates in
3935 * the RestrictInfo node to avoid repeated lookups of statistics.
3936 */
3939 {
3940 /* righthand side is inner */
3943 {
3944 /* not cached yet */
3947 virtualbuckets,
3951 }
3953 }
3954 else
3955 {
3958 /* lefthand side is inner */
3961 {
3962 /* not cached yet */
3965 virtualbuckets,
3969 }
3971 }
3977 }
3978 }
3980 /*
3981 * If the bucket holding the inner MCV would exceed hash_mem, we don't
3982 * want to hash unless there is really no other alternative, so apply
3983 * disable_cost. (The executor normally copes with excessive memory usage
3984 * by splitting batches, but obviously it cannot separate equal values
3985 * that way, so it will be unable to drive the batch size below hash_mem
3986 * when this is true.)
3987 */
3992 /*
3993 * Compute cost of the hashquals and qpquals (other restriction clauses)
3994 * separately.
3995 */
4001 /* CPU costs */
4006 {
4008 Selectivity inner_scan_frac;
4010 /*
4011 * With a SEMI or ANTI join, or if the innerrel is known unique, the
4012 * executor will stop after the first match.
4013 *
4014 * For an outer-rel row that has at least one match, we can expect the
4015 * bucket scan to stop after a fraction 1/(match_count+1) of the
4016 * bucket's rows, if the matches are evenly distributed. Since they
4017 * probably aren't quite evenly distributed, we apply a fuzz factor of
4018 * 2.0 to that fraction. (If we used a larger fuzz factor, we'd have
4019 * to clamp inner_scan_frac to at most 1.0; but since match_count is
4020 * at least 1, no such clamp is needed now.)
4021 */
4029 /*
4030 * For unmatched outer-rel rows, the picture is quite a lot different.
4031 * In the first place, there is no reason to assume that these rows
4032 * preferentially hit heavily-populated buckets; instead assume they
4033 * are uncorrelated with the inner distribution and so they see an
4034 * average bucket size of inner_path_rows / virtualbuckets. In the
4035 * second place, it seems likely that they will have few if any exact
4036 * hash-code matches and so very few of the tuples in the bucket will
4037 * actually require eval of the hash quals. We don't have any good
4038 * way to estimate how many will, but for the moment assume that the
4039 * effective cost per bucket entry is one-tenth what it is for
4040 * matchable tuples.
4041 */
4046 /* Get # of tuples that will pass the basic join */
4049 else
4051 }
4052 else
4053 {
4054 /*
4055 * The number of tuple comparisons needed is the number of outer
4056 * tuples times the typical number of tuples in a hash bucket, which
4057 * is the inner relation size times its bucketsize fraction. At each
4058 * one, we need to evaluate the hashjoin quals. But actually,
4059 * charging the full qual eval cost at each tuple is pessimistic,
4060 * since we don't evaluate the quals unless the hash values match
4061 * exactly. For lack of a better idea, halve the cost estimate to
4062 * allow for that.
4063 */
4068 /*
4069 * Get approx # tuples passing the hashquals. We use
4070 * approx_tuple_count here because we need an estimate done with
4071 * JOIN_INNER semantics.
4072 */
4074 }
4076 /*
4077 * For each tuple that gets through the hashjoin proper, we charge
4078 * cpu_tuple_cost plus the cost of evaluating additional restriction
4079 * clauses that are to be applied at the join. (This is pessimistic since
4080 * not all of the quals may get evaluated at each tuple.)
4081 */
4086 /* tlist eval costs are paid per output row, not per tuple scanned */
4092 }
4095 /*
4096 * cost_subplan
4097 * Figure the costs for a SubPlan (or initplan).
4098 *
4099 * Note: we could dig the subplan's Plan out of the root list, but in practice
4100 * all callers have it handy already, so we make them pass it.
4101 */
4102 void
4104 {
4105 QualCost sp_cost;
4107 /* Figure any cost for evaluating the testexpr */
4110 root);
4113 {
4114 /*
4115 * If we are using a hash table for the subquery outputs, then the
4116 * cost of evaluating the query is a one-time cost. We charge one
4117 * cpu_operator_cost per tuple for the work of loading the hashtable,
4118 * too.
4119 */
4123 /*
4124 * The per-tuple costs include the cost of evaluating the lefthand
4125 * expressions, plus the cost of probing the hashtable. We already
4126 * accounted for the lefthand expressions as part of the testexpr, and
4127 * will also have counted one cpu_operator_cost for each comparison
4128 * operator. That is probably too low for the probing cost, but it's
4129 * hard to make a better estimate, so live with it for now.
4130 */
4131 }
4132 else
4133 {
4134 /*
4135 * Otherwise we will be rescanning the subplan output on each
4136 * evaluation. We need to estimate how much of the output we will
4137 * actually need to scan. NOTE: this logic should agree with the
4138 * tuple_fraction estimates used by make_subplan() in
4139 * plan/subselect.c.
4140 */
4144 {
4145 /* we only need to fetch 1 tuple; clamp to avoid zero divide */
4147 }
4150 {
4151 /* assume we need 50% of the tuples */
4153 /* also charge a cpu_operator_cost per row examined */
4155 }
4156 else
4157 {
4158 /* assume we need all tuples */
4160 }
4162 /*
4163 * Also account for subplan's startup cost. If the subplan is
4164 * uncorrelated or undirect correlated, AND its topmost node is one
4165 * that materializes its output, assume that we'll only need to pay
4166 * its startup cost once; otherwise assume we pay the startup cost
4167 * every time.
4168 */
4172 else
4174 }
4178 }
4181 /*
4182 * cost_rescan
4183 * Given a finished Path, estimate the costs of rescanning it after
4184 * having done so the first time. For some Path types a rescan is
4185 * cheaper than an original scan (if no parameters change), and this
4186 * function embodies knowledge about that. The default is to return
4187 * the same costs stored in the Path. (Note that the cost estimates
4188 * actually stored in Paths are always for first scans.)
4189 *
4190 * This function is not currently intended to model effects such as rescans
4191 * being cheaper due to disk block caching; what we are concerned with is
4192 * plan types wherein the executor caches results explicitly, or doesn't
4193 * redo startup calculations, etc.
4194 */
4195 static void
4199 {
4201 {
4204 /*
4205 * Currently, nodeFunctionscan.c always executes the function to
4206 * completion before returning any rows, and caches the results in
4207 * a tuplestore. So the function eval cost is all startup cost
4208 * and isn't paid over again on rescans. However, all run costs
4209 * will be paid over again.
4210 */
4216 /*
4217 * If it's a single-batch join, we don't need to rebuild the hash
4218 * table during a rescan.
4219 */
4221 {
4222 /* Startup cost is exactly the cost of hash table building */
4225 }
4226 else
4227 {
4228 /* Otherwise, no special treatment */
4231 }
4235 {
4236 /*
4237 * These plan types materialize their final result in a
4238 * tuplestore or tuplesort object. So the rescan cost is only
4239 * cpu_tuple_cost per tuple, unless the result is large enough
4240 * to spill to disk.
4241 */
4248 {
4249 /* It will spill, so account for re-read cost */
4253 }
4256 }
4260 {
4261 /*
4262 * These plan types not only materialize their results, but do
4263 * not implement qual filtering or projection. So they are
4264 * even cheaper to rescan than the ones above. We charge only
4265 * cpu_operator_cost per tuple. (Note: keep that in sync with
4266 * the run_cost charge in cost_sort, and also see comments in
4267 * cost_material before you change it.)
4268 */
4275 {
4276 /* It will spill, so account for re-read cost */
4280 }
4283 }
4286 /* All the hard work is done by cost_memoize_rescan */
4294 }
4295 }
4298 /*
4299 * cost_qual_eval
4300 * Estimate the CPU costs of evaluating a WHERE clause.
4301 * The input can be either an implicitly-ANDed list of boolean
4302 * expressions, or a list of RestrictInfo nodes. (The latter is
4303 * preferred since it allows caching of the results.)
4304 * The result includes both a one-time (startup) component,
4305 * and a per-evaluation component.
4306 */
4307 void
4309 {
4310 cost_qual_eval_context context;
4317 /* We don't charge any cost for the implicit ANDing at top level ... */
4320 {
4324 }
4327 }
4329 /*
4330 * cost_qual_eval_node
4331 * As above, for a single RestrictInfo or expression.
4332 */
4333 void
4335 {
4336 cost_qual_eval_context context;
4345 }
4347 static bool
4349 {
4353 /*
4354 * RestrictInfo nodes contain an eval_cost field reserved for this
4355 * routine's use, so that it's not necessary to evaluate the qual clause's
4356 * cost more than once. If the clause's cost hasn't been computed yet,
4357 * the field's startup value will contain -1.
4358 */
4360 {
4364 {
4365 cost_qual_eval_context locContext;
4371 /*
4372 * For an OR clause, recurse into the marked-up tree so that we
4373 * set the eval_cost for contained RestrictInfos too.
4374 */
4377 else
4380 /*
4381 * If the RestrictInfo is marked pseudoconstant, it will be tested
4382 * only once, so treat its cost as all startup cost.
4383 */
4385 {
4386 /* count one execution during startup */
4389 }
4391 }
4394 /* do NOT recurse into children */
4396 }
4398 /*
4399 * For each operator or function node in the given tree, we charge the
4400 * estimated execution cost given by pg_proc.procost (remember to multiply
4401 * this by cpu_operator_cost).
4402 *
4403 * Vars and Consts are charged zero, and so are boolean operators (AND,
4404 * OR, NOT). Simplistic, but a lot better than no model at all.
4405 *
4406 * Should we try to account for the possibility of short-circuit
4407 * evaluation of AND/OR? Probably *not*, because that would make the
4408 * results depend on the clause ordering, and we are not in any position
4409 * to expect that the current ordering of the clauses is the one that's
4410 * going to end up being used. The above per-RestrictInfo caching would
4411 * not mix well with trying to re-order clauses anyway.
4412 *
4413 * Another issue that is entirely ignored here is that if a set-returning
4414 * function is below top level in the tree, the functions/operators above
4415 * it will need to be evaluated multiple times. In practical use, such
4416 * cases arise so seldom as to not be worth the added complexity needed;
4417 * moreover, since our rowcount estimates for functions tend to be pretty
4418 * phony, the results would also be pretty phony.
4419 */
4421 {
4424 }
4428 {
4429 /* rely on struct equivalence to treat these all alike */
4433 }
4435 {
4438 QualCost sacosts;
4439 QualCost hcosts;
4448 {
4449 /* Handle costs for hashed ScalarArrayOpExpr */
4455 /* Estimate the cost of building the hashtable. */
4458 /*
4459 * XXX should we charge a little bit for sacosts.per_tuple when
4460 * building the table, or is it ok to assume there will be zero
4461 * hash collision?
4462 */
4464 /*
4465 * Charge for hashtable lookups. Charge a single hash and a
4466 * single comparison.
4467 */
4469 }
4470 else
4471 {
4472 /*
4473 * Estimate that the operator will be applied to about half of the
4474 * array elements before the answer is determined.
4475 */
4479 }
4480 }
4483 {
4484 /*
4485 * Aggref and WindowFunc nodes are (and should be) treated like Vars,
4486 * ie, zero execution cost in the current model, because they behave
4487 * essentially like Vars at execution. We disregard the costs of
4488 * their input expressions for the same reason. The actual execution
4489 * costs of the aggregate/window functions and their arguments have to
4490 * be factored into plan-node-specific costing of the Agg or WindowAgg
4491 * plan node.
4492 */
4494 }
4496 {
4498 Oid iofunc;
4499 Oid typioparam;
4502 /* check the result type's input function */
4507 /* check the input type's output function */
4512 }
4514 {
4516 QualCost perelemcost;
4524 }
4526 {
4527 /* Conservatively assume we will check all the columns */
4532 {
4537 }
4538 }
4544 {
4545 /* Treat all these as having cost 1 */
4547 }
4549 {
4550 /* Report high cost to prevent selection of anything but TID scan */
4552 }
4554 {
4555 /* This routine should not be applied to un-planned expressions */
4557 }
4559 {
4560 /*
4561 * A subplan node in an expression typically indicates that the
4562 * subplan will be executed on each evaluation, so charge accordingly.
4563 * (Sub-selects that can be executed as InitPlans have already been
4564 * removed from the expression.)
4565 */
4571 /*
4572 * We don't want to recurse into the testexpr, because it was already
4573 * counted in the SubPlan node's costs. So we're done.
4574 */
4576 }
4578 {
4579 /*
4580 * Arbitrarily use the first alternative plan for costing. (We should
4581 * certainly only include one alternative, and we don't yet have
4582 * enough information to know which one the executor is most likely to
4583 * use.)
4584 */
4588 context);
4589 }
4591 {
4592 /*
4593 * A PlaceHolderVar should be given cost zero when considering general
4594 * expression evaluation costs. The expense of doing the contained
4595 * expression is charged as part of the tlist eval costs of the scan
4596 * or join where the PHV is first computed (see set_rel_width and
4597 * add_placeholders_to_joinrel). If we charged it again here, we'd be
4598 * double-counting the cost for each level of plan that the PHV
4599 * bubbles up through. Hence, return without recursing into the
4600 * phexpr.
4601 */
4603 }
4605 /* recurse into children */
4608 }
4610 /*
4611 * get_restriction_qual_cost
4612 * Compute evaluation costs of a baserel's restriction quals, plus any
4613 * movable join quals that have been pushed down to the scan.
4614 * Results are returned into *qpqual_cost.
4615 *
4616 * This is a convenience subroutine that works for seqscans and other cases
4617 * where all the given quals will be evaluated the hard way. It's not useful
4618 * for cost_index(), for example, where the index machinery takes care of
4619 * some of the quals. We assume baserestrictcost was previously set by
4620 * set_baserel_size_estimates().
4621 */
4622 static void
4626 {
4628 {
4629 /* Include costs of pushed-down clauses */
4634 }
4635 else
4637 }
4640 /*
4641 * compute_semi_anti_join_factors
4642 * Estimate how much of the inner input a SEMI, ANTI, or inner_unique join
4643 * can be expected to scan.
4644 *
4645 * In a hash or nestloop SEMI/ANTI join, the executor will stop scanning
4646 * inner rows as soon as it finds a match to the current outer row.
4647 * The same happens if we have detected the inner rel is unique.
4648 * We should therefore adjust some of the cost components for this effect.
4649 * This function computes some estimates needed for these adjustments.
4650 * These estimates will be the same regardless of the particular paths used
4651 * for the outer and inner relation, so we compute these once and then pass
4652 * them to all the join cost estimation functions.
4653 *
4654 * Input parameters:
4655 * joinrel: join relation under consideration
4656 * outerrel: outer relation under consideration
4657 * innerrel: inner relation under consideration
4658 * jointype: if not JOIN_SEMI or JOIN_ANTI, we assume it's inner_unique
4659 * sjinfo: SpecialJoinInfo relevant to this join
4660 * restrictlist: join quals
4661 * Output parameters:
4662 * *semifactors is filled in (see pathnodes.h for field definitions)
4663 */
4664 void
4669 JoinType jointype,
4673 {
4674 Selectivity jselec;
4675 Selectivity nselec;
4676 Selectivity avgmatch;
4677 SpecialJoinInfo norm_sjinfo;
4681 /*
4682 * In an ANTI join, we must ignore clauses that are "pushed down", since
4683 * those won't affect the match logic. In a SEMI join, we do not
4684 * distinguish joinquals from "pushed down" quals, so just use the whole
4685 * restrictinfo list. For other outer join types, we should consider only
4686 * non-pushed-down quals, so that this devolves to an IS_OUTER_JOIN check.
4687 */
4689 {
4692 {
4697 }
4698 }
4699 else
4702 /*
4703 * Get the JOIN_SEMI or JOIN_ANTI selectivity of the join clauses.
4704 */
4706 joinquals,
4709 sjinfo);
4711 /*
4712 * Also get the normal inner-join selectivity of the join clauses.
4713 */
4720 /* we don't bother trying to make the remaining fields valid */
4729 joinquals,
4731 JOIN_INNER,
4734 /* Avoid leaking a lot of ListCells */
4738 /*
4739 * jselec can be interpreted as the fraction of outer-rel rows that have
4740 * any matches (this is true for both SEMI and ANTI cases). And nselec is
4741 * the fraction of the Cartesian product that matches. So, the average
4742 * number of matches for each outer-rel row that has at least one match is
4743 * nselec * inner_rows / jselec.
4744 *
4745 * Note: it is correct to use the inner rel's "rows" count here, even
4746 * though we might later be considering a parameterized inner path with
4747 * fewer rows. This is because we have included all the join clauses in
4748 * the selectivity estimate.
4749 */
4751 {
4753 /* Clamp to sane range */
4755 }
4756 else
4761 }
4763 /*
4764 * has_indexed_join_quals
4765 * Check whether all the joinquals of a nestloop join are used as
4766 * inner index quals.
4767 *
4768 * If the inner path of a SEMI/ANTI join is an indexscan (including bitmap
4769 * indexscan) that uses all the joinquals as indexquals, we can assume that an
4770 * unmatched outer tuple is cheap to process, whereas otherwise it's probably
4771 * expensive.
4772 */
4773 static bool
4775 {
4783 /* If join still has quals to evaluate, it's not fast */
4786 /* Nor if the inner path isn't parameterized at all */
4790 /* Find the indexclauses list for the inner scan */
4792 {
4798 {
4799 /* Accept only a simple bitmap scan, not AND/OR cases */
4804 else
4807 }
4810 /*
4811 * If it's not a simple indexscan, it probably doesn't run quickly
4812 * for zero rows out, even if it's a parameterized path using all
4813 * the joinquals.
4814 */
4816 }
4818 /*
4819 * Examine the inner path's param clauses. Any that are from the outer
4820 * path must be found in the indexclauses list, either exactly or in an
4821 * equivalent form generated by equivclass.c. Also, we must find at least
4822 * one such clause, else it's a clauseless join which isn't fast.
4823 */
4826 {
4831 joinrelids))
4832 {
4836 }
4837 }
4839 }
4842 /*
4843 * approx_tuple_count
4844 * Quick-and-dirty estimation of the number of join rows passing
4845 * a set of qual conditions.
4846 *
4847 * The quals can be either an implicitly-ANDed list of boolean expressions,
4848 * or a list of RestrictInfo nodes (typically the latter).
4849 *
4850 * We intentionally compute the selectivity under JOIN_INNER rules, even
4851 * if it's some type of outer join. This is appropriate because we are
4852 * trying to figure out how many tuples pass the initial merge or hash
4853 * join step.
4854 *
4855 * This is quick-and-dirty because we bypass clauselist_selectivity, and
4856 * simply multiply the independent clause selectivities together. Now
4857 * clauselist_selectivity often can't do any better than that anyhow, but
4858 * for some situations (such as range constraints) it is smarter. However,
4859 * we can't effectively cache the results of clauselist_selectivity, whereas
4860 * the individual clause selectivities can be and are cached.
4861 *
4862 * Since we are only using the results to estimate how many potential
4863 * output tuples are generated and passed through qpqual checking, it
4864 * seems OK to live with the approximation.
4865 */
4866 static double
4868 {
4872 SpecialJoinInfo sjinfo;
4876 /*
4877 * Make up a SpecialJoinInfo for JOIN_INNER semantics.
4878 */
4885 /* we don't bother trying to make the remaining fields valid */
4893 /* Get the approximate selectivity */
4895 {
4898 /* Note that clause_selectivity will be able to cache its result */
4900 }
4902 /* Apply it to the input relation sizes */
4906 }
4909 /*
4910 * set_baserel_size_estimates
4911 * Set the size estimates for the given base relation.
4912 *
4913 * The rel's targetlist and restrictinfo list must have been constructed
4914 * already, and rel->tuples must be set.
4915 *
4916 * We set the following fields of the rel node:
4917 * rows: the estimated number of output tuples (after applying
4918 * restriction clauses).
4919 * width: the estimated average output tuple width in bytes.
4920 * baserestrictcost: estimated cost of evaluating baserestrictinfo clauses.
4921 */
4922 void
4924 {
4927 /* Should only be applied to base relations */
4934 JOIN_INNER,
4935 NULL);
4942 }
4944 /*
4945 * get_parameterized_baserel_size
4946 * Make a size estimate for a parameterized scan of a base relation.
4947 *
4948 * 'param_clauses' lists the additional join clauses to be used.
4949 *
4950 * set_baserel_size_estimates must have been applied already.
4951 */
4952 double
4955 {
4959 /*
4960 * Estimate the number of rows returned by the parameterized scan, knowing
4961 * that it will apply all the extra join clauses as well as the rel's own
4962 * restriction clauses. Note that we force the clauses to be treated as
4963 * non-join clauses during selectivity estimation.
4964 */
4968 allclauses,
4970 JOIN_INNER,
4971 NULL);
4973 /* For safety, make sure result is not more than the base estimate */
4977 }
4979 /*
4980 * set_joinrel_size_estimates
4981 * Set the size estimates for the given join relation.
4982 *
4983 * The rel's targetlist must have been constructed already, and a
4984 * restriction clause list that matches the given component rels must
4985 * be provided.
4986 *
4987 * Since there is more than one way to make a joinrel for more than two
4988 * base relations, the results we get here could depend on which component
4989 * rel pair is provided. In theory we should get the same answers no matter
4990 * which pair is provided; in practice, since the selectivity estimation
4991 * routines don't handle all cases equally well, we might not. But there's
4992 * not much to be done about it. (Would it make sense to repeat the
4993 * calculations for each pair of input rels that's encountered, and somehow
4994 * average the results? Probably way more trouble than it's worth, and
4995 * anyway we must keep the rowcount estimate the same for all paths for the
4996 * joinrel.)
4997 *
4998 * We set only the rows field here. The reltarget field was already set by
4999 * build_joinrel_tlist, and baserestrictcost is not used for join rels.
5000 */
5001 void
5007 {
5009 rel,
5010 outer_rel,
5011 inner_rel,
5014 sjinfo,
5015 restrictlist);
5016 }
5018 /*
5019 * get_parameterized_joinrel_size
5020 * Make a size estimate for a parameterized scan of a join relation.
5021 *
5022 * 'rel' is the joinrel under consideration.
5023 * 'outer_path', 'inner_path' are (probably also parameterized) Paths that
5024 * produce the relations being joined.
5025 * 'sjinfo' is any SpecialJoinInfo relevant to this join.
5026 * 'restrict_clauses' lists the join clauses that need to be applied at the
5027 * join node (including any movable clauses that were moved down to this join,
5028 * and not including any movable clauses that were pushed down into the
5029 * child paths).
5030 *
5031 * set_joinrel_size_estimates must have been applied already.
5032 */
5033 double
5039 {
5042 /*
5043 * Estimate the number of rows returned by the parameterized join as the
5044 * sizes of the input paths times the selectivity of the clauses that have
5045 * ended up at this join node.
5046 *
5047 * As with set_joinrel_size_estimates, the rowcount estimate could depend
5048 * on the pair of input paths provided, though ideally we'd get the same
5049 * estimate for any pair with the same parameterization.
5050 */
5052 rel,
5057 sjinfo,
5058 restrict_clauses);
5059 /* For safety, make sure result is not more than the base estimate */
5063 }
5065 /*
5066 * calc_joinrel_size_estimate
5067 * Workhorse for set_joinrel_size_estimates and
5068 * get_parameterized_joinrel_size.
5069 *
5070 * outer_rel/inner_rel are the relations being joined, but they should be
5071 * assumed to have sizes outer_rows/inner_rows; those numbers might be less
5072 * than what rel->rows says, when we are considering parameterized paths.
5073 */
5074 static double
5083 {
5084 /* This apparently-useless variable dodges a compiler bug in VS2013: */
5087 Selectivity fkselec;
5088 Selectivity jselec;
5089 Selectivity pselec;
5092 /*
5093 * Compute joinclause selectivity. Note that we are only considering
5094 * clauses that become restriction clauses at this join level; we are not
5095 * double-counting them because they were not considered in estimating the
5096 * sizes of the component rels.
5097 *
5098 * First, see whether any of the joinclauses can be matched to known FK
5099 * constraints. If so, drop those clauses from the restrictlist, and
5100 * instead estimate their selectivity using FK semantics. (We do this
5101 * without regard to whether said clauses are local or "pushed down".
5102 * Probably, an FK-matching clause could never be seen as pushed down at
5103 * an outer join, since it would be strict and hence would be grounds for
5104 * join strength reduction.) fkselec gets the net selectivity for
5105 * FK-matching clauses, or 1.0 if there are none.
5106 */
5110 sjinfo,
5113 /*
5114 * For an outer join, we have to distinguish the selectivity of the join's
5115 * own clauses (JOIN/ON conditions) from any clauses that were "pushed
5116 * down". For inner joins we just count them all as joinclauses.
5117 */
5119 {
5124 /* Grovel through the clauses to separate into two lists */
5126 {
5131 else
5133 }
5135 /* Get the separate selectivities */
5137 joinquals,
5139 jointype,
5140 sjinfo);
5142 pushedquals,
5144 jointype,
5145 sjinfo);
5147 /* Avoid leaking a lot of ListCells */
5150 }
5151 else
5152 {
5154 restrictlist,
5156 jointype,
5157 sjinfo);
5159 }
5161 /*
5162 * Basically, we multiply size of Cartesian product by selectivity.
5163 *
5164 * If we are doing an outer join, take that into account: the joinqual
5165 * selectivity has to be clamped using the knowledge that the output must
5166 * be at least as large as the non-nullable input. However, any
5167 * pushed-down quals are applied after the outer join, so their
5168 * selectivity applies fully.
5169 *
5170 * For JOIN_SEMI and JOIN_ANTI, the selectivity is defined as the fraction
5171 * of LHS rows that have matches, and we apply that straightforwardly.
5172 */
5174 {
5177 /* pselec not used */
5195 /* pselec not used */
5202 /* other values not expected here */
5206 }
5209 }
5211 /*
5212 * get_foreign_key_join_selectivity
5213 * Estimate join selectivity for foreign-key-related clauses.
5214 *
5215 * Remove any clauses that can be matched to FK constraints from *restrictlist,
5216 * and return a substitute estimate of their selectivity. 1.0 is returned
5217 * when there are no such clauses.
5218 *
5219 * The reason for treating such clauses specially is that we can get better
5220 * estimates this way than by relying on clauselist_selectivity(), especially
5221 * for multi-column FKs where that function's assumption that the clauses are
5222 * independent falls down badly. But even with single-column FKs, we may be
5223 * able to get a better answer when the pg_statistic stats are missing or out
5224 * of date.
5225 */
5226 static Selectivity
5228 Relids outer_relids,
5229 Relids inner_relids,
5232 {
5238 /* Consider each FK constraint that is known to match the query */
5240 {
5246 /*
5247 * This FK is not relevant unless it connects a baserel on one side of
5248 * this join to a baserel on the other side.
5249 */
5256 else
5259 /*
5260 * If we're dealing with a semi/anti join, and the FK's referenced
5261 * relation is on the outside, then knowledge of the FK doesn't help
5262 * us figure out what we need to know (which is the fraction of outer
5263 * rows that have matches). On the other hand, if the referenced rel
5264 * is on the inside, then all outer rows must have matches in the
5265 * referenced table (ignoring nulls). But any restriction or join
5266 * clauses that filter that table will reduce the fraction of matches.
5267 * We can account for restriction clauses, but it's too hard to guess
5268 * how many table rows would get through a join that's inside the RHS.
5269 * Hence, if either case applies, punt and ignore the FK.
5270 */
5275 /*
5276 * Modify the restrictlist by removing clauses that match the FK (and
5277 * putting them into removedlist instead). It seems unsafe to modify
5278 * the originally-passed List structure, so we make a shallow copy the
5279 * first time through.
5280 */
5286 {
5291 /* Drop this clause if it matches any column of the FK */
5293 {
5295 {
5296 /*
5297 * EC-derived clauses can only match by EC. It is okay to
5298 * consider any clause derived from the same EC as
5299 * matching the FK: even if equivclass.c chose to generate
5300 * a clause equating some other pair of Vars, it could
5301 * have generated one equating the FK's Vars. So for
5302 * purposes of estimation, we can act as though it did so.
5303 *
5304 * Note: checking parent_ec is a bit of a cheat because
5305 * there are EC-derived clauses that don't have parent_ec
5306 * set; but such clauses must compare expressions that
5307 * aren't just Vars, so they cannot match the FK anyway.
5308 */
5310 {
5313 }
5314 }
5315 else
5316 {
5317 /*
5318 * Otherwise, see if rinfo was previously matched to FK as
5319 * a "loose" clause.
5320 */
5322 {
5325 }
5326 }
5327 }
5329 {
5332 }
5333 }
5335 /*
5336 * If we failed to remove all the matching clauses we expected to
5337 * find, chicken out and ignore this FK; applying its selectivity
5338 * might result in double-counting. Put any clauses we did manage to
5339 * remove back into the worklist.
5340 *
5341 * Since the matching clauses are known not outerjoin-delayed, they
5342 * would normally have appeared in the initial joinclause list. If we
5343 * didn't find them, there are two possibilities:
5344 *
5345 * 1. If the FK match is based on an EC that is ec_has_const, it won't
5346 * have generated any join clauses at all. We discount such ECs while
5347 * checking to see if we have "all" the clauses. (Below, we'll adjust
5348 * the selectivity estimate for this case.)
5349 *
5350 * 2. The clauses were matched to some other FK in a previous
5351 * iteration of this loop, and thus removed from worklist. (A likely
5352 * case is that two FKs are matched to the same EC; there will be only
5353 * one EC-derived clause in the initial list, so the first FK will
5354 * consume it.) Applying both FKs' selectivity independently risks
5355 * underestimating the join size; in particular, this would undo one
5356 * of the main things that ECs were invented for, namely to avoid
5357 * double-counting the selectivity of redundant equality conditions.
5358 * Later we might think of a reasonable way to combine the estimates,
5359 * but for now, just punt, since this is a fairly uncommon situation.
5360 */
5364 {
5367 }
5369 /*
5370 * Finally we get to the payoff: estimate selectivity using the
5371 * knowledge that each referencing row will match exactly one row in
5372 * the referenced table.
5373 *
5374 * XXX that's not true in the presence of nulls in the referencing
5375 * column(s), so in principle we should derate the estimate for those.
5376 * However (1) if there are any strict restriction clauses for the
5377 * referencing column(s) elsewhere in the query, derating here would
5378 * be double-counting the null fraction, and (2) it's not very clear
5379 * how to combine null fractions for multiple referencing columns. So
5380 * we do nothing for now about correcting for nulls.
5381 *
5382 * XXX another point here is that if either side of an FK constraint
5383 * is an inheritance parent, we estimate as though the constraint
5384 * covers all its children as well. This is not an unreasonable
5385 * assumption for a referencing table, ie the user probably applied
5386 * identical constraints to all child tables (though perhaps we ought
5387 * to check that). But it's not possible to have done that for a
5388 * referenced table. Fortunately, precisely because that doesn't
5389 * work, it is uncommon in practice to have an FK referencing a parent
5390 * table. So, at least for now, disregard inheritance here.
5391 */
5393 {
5394 /*
5395 * For JOIN_SEMI and JOIN_ANTI, we only get here when the FK's
5396 * referenced table is exactly the inside of the join. The join
5397 * selectivity is defined as the fraction of LHS rows that have
5398 * matches. The FK implies that every LHS row has a match *in the
5399 * referenced table*; but any restriction clauses on it will
5400 * reduce the number of matches. Hence we take the join
5401 * selectivity as equal to the selectivity of the table's
5402 * restriction clauses, which is rows / tuples; but we must guard
5403 * against tuples == 0.
5404 */
5409 }
5410 else
5411 {
5412 /*
5413 * Otherwise, selectivity is exactly 1/referenced-table-size; but
5414 * guard against tuples == 0. Note we should use the raw table
5415 * tuple count, not any estimate of its filtered or joined size.
5416 */
5421 }
5423 /*
5424 * If any of the FK columns participated in ec_has_const ECs, then
5425 * equivclass.c will have generated "var = const" restrictions for
5426 * each side of the join, thus reducing the sizes of both input
5427 * relations. Taking the fkselec at face value would amount to
5428 * double-counting the selectivity of the constant restriction for the
5429 * referencing Var. Hence, look for the restriction clause(s) that
5430 * were applied to the referencing Var(s), and divide out their
5431 * selectivity to correct for this.
5432 */
5434 {
5436 {
5440 {
5443 em);
5446 {
5447 Selectivity s0;
5452 jointype,
5453 sjinfo);
5456 }
5457 }
5458 }
5459 }
5460 }
5465 }
5467 /*
5468 * set_subquery_size_estimates
5469 * Set the size estimates for a base relation that is a subquery.
5470 *
5471 * The rel's targetlist and restrictinfo list must have been constructed
5472 * already, and the Paths for the subquery must have been completed.
5473 * We look at the subquery's PlannerInfo to extract data.
5474 *
5475 * We set the same fields as set_baserel_size_estimates.
5476 */
5477 void
5479 {
5484 /* Should only be applied to base relations that are subqueries */
5488 /*
5489 * Copy raw number of output rows from subquery. All of its paths should
5490 * have the same output rowcount, so just look at cheapest-total.
5491 */
5495 /*
5496 * Compute per-output-column width estimates by examining the subquery's
5497 * targetlist. For any output that is a plain Var, get the width estimate
5498 * that was made while planning the subquery. Otherwise, we leave it to
5499 * set_rel_width to fill in a datatype-based default estimate.
5500 */
5502 {
5507 /* junk columns aren't visible to upper query */
5511 /*
5512 * The subquery could be an expansion of a view that's had columns
5513 * added to it since the current query was parsed, so that there are
5514 * non-junk tlist columns in it that don't correspond to any column
5515 * visible at our query level. Ignore such columns.
5516 */
5520 /*
5521 * XXX This currently doesn't work for subqueries containing set
5522 * operations, because the Vars in their tlists are bogus references
5523 * to the first leaf subquery, which wouldn't give the right answer
5524 * even if we could still get to its PlannerInfo.
5525 *
5526 * Also, the subquery could be an appendrel for which all branches are
5527 * known empty due to constraint exclusion, in which case
5528 * set_append_rel_pathlist will have left the attr_widths set to zero.
5529 *
5530 * In either case, we just leave the width estimate zero until
5531 * set_rel_width fixes it.
5532 */
5535 {
5540 }
5542 }
5544 /* Now estimate number of output rows, etc */
5546 }
5548 /*
5549 * set_function_size_estimates
5550 * Set the size estimates for a base relation that is a function call.
5551 *
5552 * The rel's targetlist and restrictinfo list must have been constructed
5553 * already.
5554 *
5555 * We set the same fields as set_baserel_size_estimates.
5556 */
5557 void
5559 {
5563 /* Should only be applied to base relations that are functions */
5568 /*
5569 * Estimate number of rows the functions will return. The rowcount of the
5570 * node is that of the largest function result.
5571 */
5574 {
5580 }
5582 /* Now estimate number of output rows, etc */
5584 }
5586 /*
5587 * set_function_size_estimates
5588 * Set the size estimates for a base relation that is a function call.
5589 *
5590 * The rel's targetlist and restrictinfo list must have been constructed
5591 * already.
5592 *
5593 * We set the same fields as set_tablefunc_size_estimates.
5594 */
5595 void
5597 {
5598 /* Should only be applied to base relations that are functions */
5604 /* Now estimate number of output rows, etc */
5606 }
5608 /*
5609 * set_values_size_estimates
5610 * Set the size estimates for a base relation that is a values list.
5611 *
5612 * The rel's targetlist and restrictinfo list must have been constructed
5613 * already.
5614 *
5615 * We set the same fields as set_baserel_size_estimates.
5616 */
5617 void
5619 {
5622 /* Should only be applied to base relations that are values lists */
5627 /*
5628 * Estimate number of rows the values list will return. We know this
5629 * precisely based on the list length (well, barring set-returning
5630 * functions in list items, but that's a refinement not catered for
5631 * anywhere else either).
5632 */
5635 /* Now estimate number of output rows, etc */
5637 }
5639 /*
5640 * set_cte_size_estimates
5641 * Set the size estimates for a base relation that is a CTE reference.
5642 *
5643 * The rel's targetlist and restrictinfo list must have been constructed
5644 * already, and we need an estimate of the number of rows returned by the CTE
5645 * (if a regular CTE) or the non-recursive term (if a self-reference).
5646 *
5647 * We set the same fields as set_baserel_size_estimates.
5648 */
5649 void
5651 {
5654 /* Should only be applied to base relations that are CTE references */
5660 {
5661 /*
5662 * In a self-reference, arbitrarily assume the average worktable size
5663 * is about 10 times the nonrecursive term's size.
5664 */
5666 }
5667 else
5668 {
5669 /* Otherwise just believe the CTE's rowcount estimate */
5671 }
5673 /* Now estimate number of output rows, etc */
5675 }
5677 /*
5678 * set_namedtuplestore_size_estimates
5679 * Set the size estimates for a base relation that is a tuplestore reference.
5680 *
5681 * The rel's targetlist and restrictinfo list must have been constructed
5682 * already.
5683 *
5684 * We set the same fields as set_baserel_size_estimates.
5685 */
5686 void
5688 {
5691 /* Should only be applied to base relations that are tuplestore references */
5696 /*
5697 * Use the estimate provided by the code which is generating the named
5698 * tuplestore. In some cases, the actual number might be available; in
5699 * others the same plan will be re-used, so a "typical" value might be
5700 * estimated and used.
5701 */
5706 /* Now estimate number of output rows, etc */
5708 }
5710 /*
5711 * set_result_size_estimates
5712 * Set the size estimates for an RTE_RESULT base relation
5713 *
5714 * The rel's targetlist and restrictinfo list must have been constructed
5715 * already.
5716 *
5717 * We set the same fields as set_baserel_size_estimates.
5718 */
5719 void
5721 {
5722 /* Should only be applied to RTE_RESULT base relations */
5726 /* RTE_RESULT always generates a single row, natively */
5729 /* Now estimate number of output rows, etc */
5731 }
5733 /*
5734 * set_foreign_size_estimates
5735 * Set the size estimates for a base relation that is a foreign table.
5736 *
5737 * There is not a whole lot that we can do here; the foreign-data wrapper
5738 * is responsible for producing useful estimates. We can do a decent job
5739 * of estimating baserestrictcost, so we set that, and we also set up width
5740 * using what will be purely datatype-driven estimates from the targetlist.
5741 * There is no way to do anything sane with the rows value, so we just put
5742 * a default estimate and hope that the wrapper can improve on it. The
5743 * wrapper's GetForeignRelSize function will be called momentarily.
5744 *
5745 * The rel's targetlist and restrictinfo list must have been constructed
5746 * already.
5747 */
5748 void
5750 {
5751 /* Should only be applied to base relations */
5759 }
5762 /*
5763 * set_rel_width
5764 * Set the estimated output width of a base relation.
5765 *
5766 * The estimated output width is the sum of the per-attribute width estimates
5767 * for the actually-referenced columns, plus any PHVs or other expressions
5768 * that have to be calculated at this relation. This is the amount of data
5769 * we'd need to pass upwards in case of a sort, hash, etc.
5770 *
5771 * This function also sets reltarget->cost, so it's a bit misnamed now.
5772 *
5773 * NB: this works best on plain relations because it prefers to look at
5774 * real Vars. For subqueries, set_subquery_size_estimates will already have
5775 * copied up whatever per-column estimates were made within the subquery,
5776 * and for other types of rels there isn't much we can do anyway. We fall
5777 * back on (fairly stupid) datatype-based width estimates if we can't get
5778 * any better number.
5779 *
5780 * The per-attribute width estimates are cached for possible re-use while
5781 * building join relations or post-scan/join pathtargets.
5782 */
5783 static void
5785 {
5791 /* Vars are assumed to have cost zero, but other exprs do not */
5796 {
5799 /*
5800 * Ordinarily, a Var in a rel's targetlist must belong to that rel;
5801 * but there are corner cases involving LATERAL references where that
5802 * isn't so. If the Var has the wrong varno, fall through to the
5803 * generic case (it doesn't seem worth the trouble to be any smarter).
5804 */
5807 {
5810 int32 item_width;
5817 /*
5818 * If it's a whole-row Var, we'll deal with it below after we have
5819 * already cached as many attr widths as possible.
5820 */
5822 {
5825 }
5827 /*
5828 * The width may have been cached already (especially if it's a
5829 * subquery), so don't duplicate effort.
5830 */
5832 {
5835 }
5837 /* Try to get column width from statistics */
5839 {
5842 {
5846 }
5847 }
5849 /*
5850 * Not a plain relation, or can't find statistics for it. Estimate
5851 * using just the type info.
5852 */
5857 }
5859 {
5860 /*
5861 * We will need to evaluate the PHV's contained expression while
5862 * scanning this rel, so be sure to include it in reltarget->cost.
5863 */
5866 QualCost cost;
5872 }
5873 else
5874 {
5875 /*
5876 * We could be looking at an expression pulled up from a subquery,
5877 * or a ROW() representing a whole-row child Var, etc. Do what we
5878 * can using the expression type information.
5879 */
5880 int32 item_width;
5881 QualCost cost;
5886 /* Not entirely clear if we need to account for cost, but do so */
5890 }
5891 }
5893 /*
5894 * If we have a whole-row reference, estimate its width as the sum of
5895 * per-column widths plus heap tuple header overhead.
5896 */
5898 {
5902 {
5903 /* Real relation, so estimate true tuple width */
5906 }
5907 else
5908 {
5909 /* Do what we can with info for a phony rel */
5910 AttrNumber i;
5914 }
5918 /*
5919 * Include the whole-row Var as part of the output tuple. Yes, that
5920 * really is what happens at runtime.
5921 */
5923 }
5927 }
5929 /*
5930 * set_pathtarget_cost_width
5931 * Set the estimated eval cost and output width of a PathTarget tlist.
5932 *
5933 * As a notational convenience, returns the same PathTarget pointer passed in.
5934 *
5935 * Most, though not quite all, uses of this function occur after we've run
5936 * set_rel_width() for base relations; so we can usually obtain cached width
5937 * estimates for Vars. If we can't, fall back on datatype-based width
5938 * estimates. Present early-planning uses of PathTargets don't need accurate
5939 * widths badly enough to justify going to the catalogs for better data.
5940 */
5941 PathTarget *
5943 {
5947 /* Vars are assumed to have cost zero, but other exprs do not */
5952 {
5956 {
5958 int32 item_width;
5960 /* We should not see any upper-level Vars here */
5963 /* Try to get data from RelOptInfo cache */
5966 {
5972 {
5976 {
5979 }
5980 }
5981 }
5983 /*
5984 * No cached data available, so estimate using just the type info.
5985 */
5989 }
5990 else
5991 {
5992 /*
5993 * Handle general expressions using type info.
5994 */
5995 int32 item_width;
5996 QualCost cost;
6002 /* Account for cost, too */
6006 }
6007 }
6013 }
6015 /*
6016 * relation_byte_size
6017 * Estimate the storage space in bytes for a given number of tuples
6018 * of a given width (size in bytes).
6019 */
6020 static double
6022 {
6024 }
6026 /*
6027 * page_size
6028 * Returns an estimate of the number of pages covered by a given
6029 * number of tuples of a given width (size in bytes).
6030 */
6031 static double
6033 {
6035 }
6037 /*
6038 * Estimate the fraction of the work that each worker will do given the
6039 * number of workers budgeted for the path.
6040 */
6041 static double
6043 {
6046 /*
6047 * Early experience with parallel query suggests that when there is only
6048 * one worker, the leader often makes a very substantial contribution to
6049 * executing the parallel portion of the plan, but as more workers are
6050 * added, it does less and less, because it's busy reading tuples from the
6051 * workers and doing whatever non-parallel post-processing is needed. By
6052 * the time we reach 4 workers, the leader no longer makes a meaningful
6053 * contribution. Thus, for now, estimate that the leader spends 30% of
6054 * its time servicing each worker, and the remainder executing the
6055 * parallel plan.
6056 */
6058 {
6064 }
6067 }
6069 /*
6070 * compute_bitmap_pages
6071 *
6072 * compute number of pages fetched from heap in bitmap heap scan.
6073 */
6074 double
6077 {
6078 Cost indexTotalCost;
6079 Selectivity indexSelectivity;
6086 /*
6087 * Fetch total cost of obtaining the bitmap, as well as its total
6088 * selectivity.
6089 */
6092 /*
6093 * Estimate number of main-table pages fetched.
6094 */
6099 /*
6100 * For a single scan, the number of heap pages that need to be fetched is
6101 * the same as the Mackert and Lohman formula for the case T <= b (ie, no
6102 * re-reads needed).
6103 */
6106 /*
6107 * Calculate the number of pages fetched from the heap. Then based on
6108 * current work_mem estimate get the estimated maxentries in the bitmap.
6109 * (Note that we always do this calculation based on the number of pages
6110 * that would be fetched in a single iteration, even if loop_count > 1.
6111 * That's correct, because only that number of entries will be stored in
6112 * the bitmap at one time.)
6113 */
6118 {
6119 /*
6120 * For repeated bitmap scans, scale up the number of tuples fetched in
6121 * the Mackert and Lohman formula by the number of scans, so that we
6122 * estimate the number of pages fetched by all the scans. Then
6123 * pro-rate for one scan.
6124 */
6128 root);
6130 }
6134 else
6138 {
6142 /*
6143 * Crude approximation of the number of lossy pages. Because of the
6144 * way tbm_lossify() is coded, the number of lossy pages increases
6145 * very sharply as soon as we run short of memory; this formula has
6146 * that property and seems to perform adequately in testing, but it's
6147 * possible we could do better somehow.
6148 */
6152 /*
6153 * If there are lossy pages then recompute the number of tuples
6154 * processed by the bitmap heap node. We assume here that the chance
6155 * of a given tuple coming from an exact page is the same as the
6156 * chance that a given page is exact. This might not be true, but
6157 * it's not clear how we can do any better.
6158 */
6160 tuples_fetched =
6164 }
6172 }