1 /*-------------------------------------------------------------------------
4 * Routines to compute (and set) relation sizes and path costs
6 * Path costs are measured in units of disk accesses: one sequential page
7 * fetch has cost 1. All else is scaled relative to a page fetch, using
8 * the scaling parameters
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 process a typical WHERE operator
15 * We also use a rough estimate "effective_cache_size" of the number of
16 * disk pages in Postgres + OS-level disk cache. (We can't simply use
17 * NBuffers for this purpose because that would ignore the effects of
18 * the kernel's disk cache.)
20 * Obviously, taking constants for these values is an oversimplification,
21 * but it's tough enough to get any useful estimates even at this level of
22 * detail. Note that all of these parameters are user-settable, in case
23 * the default values are drastically off for a particular platform.
25 * We compute two separate costs for each path:
26 * total_cost: total estimated cost to fetch all tuples
27 * startup_cost: cost that is expended before first tuple is fetched
28 * In some scenarios, such as when there is a LIMIT or we are implementing
29 * an EXISTS(...) sub-select, it is not necessary to fetch all tuples of the
30 * path's result. A caller can estimate the cost of fetching a partial
31 * result by interpolating between startup_cost and total_cost. In detail:
32 * actual_cost = startup_cost +
33 * (total_cost - startup_cost) * tuples_to_fetch / path->parent->rows;
34 * Note that a base relation's rows count (and, by extension, plan_rows for
35 * plan nodes below the LIMIT node) are set without regard to any LIMIT, so
36 * that this equation works properly. (Also, these routines guarantee not to
37 * set the rows count to zero, so there will be no zero divide.) The LIMIT is
38 * applied as a top-level plan node.
40 * For largely historical reasons, most of the routines in this module use
41 * the passed result Path only to store their startup_cost and total_cost
42 * results into. All the input data they need is passed as separate
43 * parameters, even though much of it could be extracted from the Path.
44 * An exception is made for the cost_XXXjoin() routines, which expect all
45 * the non-cost fields of the passed XXXPath to be filled in.
48 * Portions Copyright (c) 1996-2004, PostgreSQL Global Development Group
49 * Portions Copyright (c) 1994, Regents of the University of California
52 * $PostgreSQL: pgsql/src/backend/optimizer/path/costsize.c,v 1.135 2004/10/23 00:05:27 tgl Exp $
54 *-------------------------------------------------------------------------
61 #include "catalog/pg_statistic.h"
62 #include "executor/nodeHash.h"
63 #include "miscadmin.h"
64 #include "optimizer/clauses.h"
65 #include "optimizer/cost.h"
66 #include "optimizer/pathnode.h"
67 #include "optimizer/plancat.h"
68 #include "parser/parsetree.h"
69 #include "utils/selfuncs.h"
70 #include "utils/lsyscache.h"
71 #include "utils/syscache.h"
74 #define LOG2(x) (log(x) / 0.693147180559945)
75 #define LOG6(x) (log(x) / 1.79175946922805)
78 * Some Paths return less than the nominal number of rows of their parent
79 * relations; join nodes need to do this to get the correct input count:
81 #define PATH_ROWS(path) \
82 (IsA(path, UniquePath) ? \
83 ((UniquePath *) (path))->rows : \
87 double effective_cache_size = DEFAULT_EFFECTIVE_CACHE_SIZE;
88 double random_page_cost = DEFAULT_RANDOM_PAGE_COST;
89 double cpu_tuple_cost = DEFAULT_CPU_TUPLE_COST;
90 double cpu_index_tuple_cost = DEFAULT_CPU_INDEX_TUPLE_COST;
91 double cpu_operator_cost = DEFAULT_CPU_OPERATOR_COST;
93 Cost disable_cost = 100000000.0;
95 bool enable_seqscan = true;
96 bool enable_indexscan = true;
97 bool enable_tidscan = true;
98 bool enable_sort = true;
99 bool enable_hashagg = true;
100 bool enable_nestloop = true;
101 bool enable_mergejoin = true;
102 bool enable_hashjoin = true;
105 static bool cost_qual_eval_walker(Node *node, QualCost *total);
106 static Selectivity approx_selectivity(Query *root, List *quals,
108 static Selectivity join_in_selectivity(JoinPath *path, Query *root);
109 static void set_rel_width(Query *root, RelOptInfo *rel);
110 static double relation_byte_size(double tuples, int width);
111 static double page_size(double tuples, int width);
116 * Force a row-count estimate to a sane value.
119 clamp_row_est(double nrows)
122 * Force estimate to be at least one row, to make explain output look
123 * better and to avoid possible divide-by-zero when interpolating
124 * costs. Make it an integer, too.
137 * Determines and returns the cost of scanning a relation sequentially.
140 cost_seqscan(Path *path, Query *root,
143 Cost startup_cost = 0;
147 /* Should only be applied to base relations */
148 Assert(baserel->relid > 0);
149 Assert(baserel->rtekind == RTE_RELATION);
152 startup_cost += disable_cost;
157 * The cost of reading a page sequentially is 1.0, by definition. Note
158 * that the Unix kernel will typically do some amount of read-ahead
159 * optimization, so that this cost is less than the true cost of
160 * reading a page from disk. We ignore that issue here, but must take
161 * it into account when estimating the cost of non-sequential
164 run_cost += baserel->pages; /* sequential fetches with cost 1.0 */
167 startup_cost += baserel->baserestrictcost.startup;
168 cpu_per_tuple = cpu_tuple_cost + baserel->baserestrictcost.per_tuple;
169 run_cost += cpu_per_tuple * baserel->tuples;
171 path->startup_cost = startup_cost;
172 path->total_cost = startup_cost + run_cost;
176 * cost_nonsequential_access
177 * Estimate the cost of accessing one page at random from a relation
178 * (or sort temp file) of the given size in pages.
180 * The simplistic model that the cost is random_page_cost is what we want
181 * to use for large relations; but for small ones that is a serious
182 * overestimate because of the effects of caching. This routine tries to
185 * Unfortunately we don't have any good way of estimating the effective cache
186 * size we are working with --- we know that Postgres itself has NBuffers
187 * internal buffers, but the size of the kernel's disk cache is uncertain,
188 * and how much of it we get to use is even less certain. We punt the problem
189 * for now by assuming we are given an effective_cache_size parameter.
191 * Given a guesstimated cache size, we estimate the actual I/O cost per page
192 * with the entirely ad-hoc equations (writing relsize for
193 * relpages/effective_cache_size):
195 * random_page_cost - (random_page_cost-1)/2 * (1/relsize)
197 * 1 + ((random_page_cost-1)/2) * relsize ** 2
198 * These give the right asymptotic behavior (=> 1.0 as relpages becomes
199 * small, => random_page_cost as it becomes large) and meet in the middle
200 * with the estimate that the cache is about 50% effective for a relation
201 * of the same size as effective_cache_size. (XXX this is probably all
202 * wrong, but I haven't been able to find any theory about how effective
203 * a disk cache should be presumed to be.)
206 cost_nonsequential_access(double relpages)
210 /* don't crash on bad input data */
211 if (relpages <= 0.0 || effective_cache_size <= 0.0)
212 return random_page_cost;
214 relsize = relpages / effective_cache_size;
217 return random_page_cost - (random_page_cost - 1.0) * 0.5 / relsize;
219 return 1.0 + (random_page_cost - 1.0) * 0.5 * relsize * relsize;
224 * Determines and returns the cost of scanning a relation using an index.
226 * NOTE: an indexscan plan node can actually represent several passes,
227 * but here we consider the cost of just one pass.
229 * 'root' is the query root
230 * 'baserel' is the base relation the index is for
231 * 'index' is the index to be used
232 * 'indexQuals' is the list of applicable qual clauses (implicit AND semantics)
233 * 'is_injoin' is T if we are considering using the index scan as the inside
234 * of a nestloop join (hence, some of the indexQuals are join clauses)
236 * NOTE: 'indexQuals' must contain only clauses usable as index restrictions.
237 * Any additional quals evaluated as qpquals may reduce the number of returned
238 * tuples, but they won't reduce the number of tuples we have to fetch from
239 * the table, so they don't reduce the scan cost.
241 * NOTE: as of 8.0, indexQuals is a list of RestrictInfo nodes, where formerly
242 * it was a list of bare clause expressions.
245 cost_index(Path *path, Query *root,
251 Cost startup_cost = 0;
253 Cost indexStartupCost;
255 Selectivity indexSelectivity;
256 double indexCorrelation,
261 double tuples_fetched;
262 double pages_fetched;
266 /* Should only be applied to base relations */
267 Assert(IsA(baserel, RelOptInfo) &&
268 IsA(index, IndexOptInfo));
269 Assert(baserel->relid > 0);
270 Assert(baserel->rtekind == RTE_RELATION);
272 if (!enable_indexscan)
273 startup_cost += disable_cost;
276 * Call index-access-method-specific code to estimate the processing
277 * cost for scanning the index, as well as the selectivity of the
278 * index (ie, the fraction of main-table tuples we will have to
279 * retrieve) and its correlation to the main-table tuple order.
281 OidFunctionCall8(index->amcostestimate,
282 PointerGetDatum(root),
283 PointerGetDatum(baserel),
284 PointerGetDatum(index),
285 PointerGetDatum(indexQuals),
286 PointerGetDatum(&indexStartupCost),
287 PointerGetDatum(&indexTotalCost),
288 PointerGetDatum(&indexSelectivity),
289 PointerGetDatum(&indexCorrelation));
291 /* all costs for touching index itself included here */
292 startup_cost += indexStartupCost;
293 run_cost += indexTotalCost - indexStartupCost;
296 * Estimate number of main-table tuples and pages fetched.
298 * When the index ordering is uncorrelated with the table ordering,
299 * we use an approximation proposed by Mackert and Lohman, "Index Scans
300 * Using a Finite LRU Buffer: A Validated I/O Model", ACM Transactions
301 * on Database Systems, Vol. 14, No. 3, September 1989, Pages 401-424.
302 * The Mackert and Lohman approximation is that the number of pages
305 * min(2TNs/(2T+Ns), T) when T <= b
306 * 2TNs/(2T+Ns) when T > b and Ns <= 2Tb/(2T-b)
307 * b + (Ns - 2Tb/(2T-b))*(T-b)/T when T > b and Ns > 2Tb/(2T-b)
309 * T = # pages in table
310 * N = # tuples in table
311 * s = selectivity = fraction of table to be scanned
312 * b = # buffer pages available (we include kernel space here)
314 * When the index ordering is exactly correlated with the table ordering
315 * (just after a CLUSTER, for example), the number of pages fetched should
316 * be just sT. What's more, these will be sequential fetches, not the
317 * random fetches that occur in the uncorrelated case. So, depending on
318 * the extent of correlation, we should estimate the actual I/O cost
319 * somewhere between s * T * 1.0 and PF * random_cost. We currently
320 * interpolate linearly between these two endpoints based on the
321 * correlation squared (XXX is that appropriate?).
323 * In any case the number of tuples fetched is Ns.
327 tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
329 /* This part is the Mackert and Lohman formula */
331 T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
332 b = (effective_cache_size > 1) ? effective_cache_size : 1.0;
337 (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
338 if (pages_fetched > T)
345 lim = (2.0 * T * b) / (2.0 * T - b);
346 if (tuples_fetched <= lim)
349 (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
354 b + (tuples_fetched - lim) * (T - b) / T;
359 * min_IO_cost corresponds to the perfectly correlated case
360 * (csquared=1), max_IO_cost to the perfectly uncorrelated case
361 * (csquared=0). Note that we just charge random_page_cost per page
362 * in the uncorrelated case, rather than using
363 * cost_nonsequential_access, since we've already accounted for
364 * caching effects by using the Mackert model.
366 min_IO_cost = ceil(indexSelectivity * T);
367 max_IO_cost = pages_fetched * random_page_cost;
370 * Now interpolate based on estimated index order correlation to get
371 * total disk I/O cost for main table accesses.
373 csquared = indexCorrelation * indexCorrelation;
375 run_cost += max_IO_cost + csquared * (min_IO_cost - max_IO_cost);
378 * Estimate CPU costs per tuple.
380 * Normally the indexquals will be removed from the list of restriction
381 * clauses that we have to evaluate as qpquals, so we should subtract
382 * their costs from baserestrictcost. But if we are doing a join then
383 * some of the indexquals are join clauses and shouldn't be
384 * subtracted. Rather than work out exactly how much to subtract, we
385 * don't subtract anything.
387 startup_cost += baserel->baserestrictcost.startup;
388 cpu_per_tuple = cpu_tuple_cost + baserel->baserestrictcost.per_tuple;
392 QualCost index_qual_cost;
394 cost_qual_eval(&index_qual_cost, indexQuals);
395 /* any startup cost still has to be paid ... */
396 cpu_per_tuple -= index_qual_cost.per_tuple;
399 run_cost += cpu_per_tuple * tuples_fetched;
401 path->startup_cost = startup_cost;
402 path->total_cost = startup_cost + run_cost;
407 * Determines and returns the cost of scanning a relation using TIDs.
410 cost_tidscan(Path *path, Query *root,
411 RelOptInfo *baserel, List *tideval)
413 Cost startup_cost = 0;
416 int ntuples = list_length(tideval);
418 /* Should only be applied to base relations */
419 Assert(baserel->relid > 0);
420 Assert(baserel->rtekind == RTE_RELATION);
423 startup_cost += disable_cost;
425 /* disk costs --- assume each tuple on a different page */
426 run_cost += random_page_cost * ntuples;
429 startup_cost += baserel->baserestrictcost.startup;
430 cpu_per_tuple = cpu_tuple_cost + baserel->baserestrictcost.per_tuple;
431 run_cost += cpu_per_tuple * ntuples;
433 path->startup_cost = startup_cost;
434 path->total_cost = startup_cost + run_cost;
439 * Determines and returns the cost of scanning a subquery RTE.
442 cost_subqueryscan(Path *path, RelOptInfo *baserel)
448 /* Should only be applied to base relations that are subqueries */
449 Assert(baserel->relid > 0);
450 Assert(baserel->rtekind == RTE_SUBQUERY);
453 * Cost of path is cost of evaluating the subplan, plus cost of
454 * evaluating any restriction clauses that will be attached to the
455 * SubqueryScan node, plus cpu_tuple_cost to account for selection and
456 * projection overhead.
458 path->startup_cost = baserel->subplan->startup_cost;
459 path->total_cost = baserel->subplan->total_cost;
461 startup_cost = baserel->baserestrictcost.startup;
462 cpu_per_tuple = cpu_tuple_cost + baserel->baserestrictcost.per_tuple;
463 run_cost = cpu_per_tuple * baserel->tuples;
465 path->startup_cost += startup_cost;
466 path->total_cost += startup_cost + run_cost;
471 * Determines and returns the cost of scanning a function RTE.
474 cost_functionscan(Path *path, Query *root, RelOptInfo *baserel)
476 Cost startup_cost = 0;
480 /* Should only be applied to base relations that are functions */
481 Assert(baserel->relid > 0);
482 Assert(baserel->rtekind == RTE_FUNCTION);
485 * For now, estimate function's cost at one operator eval per function
486 * call. Someday we should revive the function cost estimate columns
489 cpu_per_tuple = cpu_operator_cost;
491 /* Add scanning CPU costs */
492 startup_cost += baserel->baserestrictcost.startup;
493 cpu_per_tuple += cpu_tuple_cost + baserel->baserestrictcost.per_tuple;
494 run_cost += cpu_per_tuple * baserel->tuples;
496 path->startup_cost = startup_cost;
497 path->total_cost = startup_cost + run_cost;
502 * Determines and returns the cost of sorting a relation, including
503 * the cost of reading the input data.
505 * If the total volume of data to sort is less than work_mem, we will do
506 * an in-memory sort, which requires no I/O and about t*log2(t) tuple
507 * comparisons for t tuples.
509 * If the total volume exceeds work_mem, we switch to a tape-style merge
510 * algorithm. There will still be about t*log2(t) tuple comparisons in
511 * total, but we will also need to write and read each tuple once per
512 * merge pass. We expect about ceil(log6(r)) merge passes where r is the
513 * number of initial runs formed (log6 because tuplesort.c uses six-tape
514 * merging). Since the average initial run should be about twice work_mem,
516 * disk traffic = 2 * relsize * ceil(log6(p / (2*work_mem)))
517 * cpu = comparison_cost * t * log2(t)
519 * The disk traffic is assumed to be half sequential and half random
520 * accesses (XXX can't we refine that guess?)
522 * We charge two operator evals per tuple comparison, which should be in
523 * the right ballpark in most cases.
525 * 'pathkeys' is a list of sort keys
526 * 'input_cost' is the total cost for reading the input data
527 * 'tuples' is the number of tuples in the relation
528 * 'width' is the average tuple width in bytes
530 * NOTE: some callers currently pass NIL for pathkeys because they
531 * can't conveniently supply the sort keys. Since this routine doesn't
532 * currently do anything with pathkeys anyway, that doesn't matter...
533 * but if it ever does, it should react gracefully to lack of key data.
534 * (Actually, the thing we'd most likely be interested in is just the number
535 * of sort keys, which all callers *could* supply.)
538 cost_sort(Path *path, Query *root,
539 List *pathkeys, Cost input_cost, double tuples, int width)
541 Cost startup_cost = input_cost;
543 double nbytes = relation_byte_size(tuples, width);
544 long work_mem_bytes = work_mem * 1024L;
547 startup_cost += disable_cost;
550 * We want to be sure the cost of a sort is never estimated as zero,
551 * even if passed-in tuple count is zero. Besides, mustn't do
560 * Assume about two operator evals per tuple comparison and N log2 N
563 startup_cost += 2.0 * cpu_operator_cost * tuples * LOG2(tuples);
566 if (nbytes > work_mem_bytes)
568 double npages = ceil(nbytes / BLCKSZ);
569 double nruns = (nbytes / work_mem_bytes) * 0.5;
570 double log_runs = ceil(LOG6(nruns));
571 double npageaccesses;
575 npageaccesses = 2.0 * npages * log_runs;
576 /* Assume half are sequential (cost 1), half are not */
577 startup_cost += npageaccesses *
578 (1.0 + cost_nonsequential_access(npages)) * 0.5;
582 * Also charge a small amount (arbitrarily set equal to operator cost)
583 * per extracted tuple.
585 run_cost += cpu_operator_cost * tuples;
587 path->startup_cost = startup_cost;
588 path->total_cost = startup_cost + run_cost;
593 * Determines and returns the cost of materializing a relation, including
594 * the cost of reading the input data.
596 * If the total volume of data to materialize exceeds work_mem, we will need
597 * to write it to disk, so the cost is much higher in that case.
600 cost_material(Path *path,
601 Cost input_cost, double tuples, int width)
603 Cost startup_cost = input_cost;
605 double nbytes = relation_byte_size(tuples, width);
606 long work_mem_bytes = work_mem * 1024L;
609 if (nbytes > work_mem_bytes)
611 double npages = ceil(nbytes / BLCKSZ);
613 /* We'll write during startup and read during retrieval */
614 startup_cost += npages;
619 * Also charge a small amount per extracted tuple. We use
620 * cpu_tuple_cost so that it doesn't appear worthwhile to materialize
623 run_cost += cpu_tuple_cost * tuples;
625 path->startup_cost = startup_cost;
626 path->total_cost = startup_cost + run_cost;
631 * Determines and returns the cost of performing an Agg plan node,
632 * including the cost of its input.
634 * Note: when aggstrategy == AGG_SORTED, caller must ensure that input costs
635 * are for appropriately-sorted input.
638 cost_agg(Path *path, Query *root,
639 AggStrategy aggstrategy, int numAggs,
640 int numGroupCols, double numGroups,
641 Cost input_startup_cost, Cost input_total_cost,
648 * We charge one cpu_operator_cost per aggregate function per input
649 * tuple, and another one per output tuple (corresponding to transfn
650 * and finalfn calls respectively). If we are grouping, we charge an
651 * additional cpu_operator_cost per grouping column per input tuple
652 * for grouping comparisons.
654 * We will produce a single output tuple if not grouping, and a tuple per
657 * Note: in this cost model, AGG_SORTED and AGG_HASHED have exactly the
658 * same total CPU cost, but AGG_SORTED has lower startup cost. If the
659 * input path is already sorted appropriately, AGG_SORTED should be
660 * preferred (since it has no risk of memory overflow). This will
661 * happen as long as the computed total costs are indeed exactly equal
662 * --- but if there's roundoff error we might do the wrong thing. So
663 * be sure that the computations below form the same intermediate
664 * values in the same order.
666 if (aggstrategy == AGG_PLAIN)
668 startup_cost = input_total_cost;
669 startup_cost += cpu_operator_cost * (input_tuples + 1) * numAggs;
670 /* we aren't grouping */
671 total_cost = startup_cost;
673 else if (aggstrategy == AGG_SORTED)
675 /* Here we are able to deliver output on-the-fly */
676 startup_cost = input_startup_cost;
677 total_cost = input_total_cost;
678 /* calcs phrased this way to match HASHED case, see note above */
679 total_cost += cpu_operator_cost * input_tuples * numGroupCols;
680 total_cost += cpu_operator_cost * input_tuples * numAggs;
681 total_cost += cpu_operator_cost * numGroups * numAggs;
685 /* must be AGG_HASHED */
686 startup_cost = input_total_cost;
687 startup_cost += cpu_operator_cost * input_tuples * numGroupCols;
688 startup_cost += cpu_operator_cost * input_tuples * numAggs;
689 total_cost = startup_cost;
690 total_cost += cpu_operator_cost * numGroups * numAggs;
693 path->startup_cost = startup_cost;
694 path->total_cost = total_cost;
699 * Determines and returns the cost of performing a Group plan node,
700 * including the cost of its input.
702 * Note: caller must ensure that input costs are for appropriately-sorted
706 cost_group(Path *path, Query *root,
707 int numGroupCols, double numGroups,
708 Cost input_startup_cost, Cost input_total_cost,
714 startup_cost = input_startup_cost;
715 total_cost = input_total_cost;
718 * Charge one cpu_operator_cost per comparison per input tuple. We
719 * assume all columns get compared at most of the tuples.
721 total_cost += cpu_operator_cost * input_tuples * numGroupCols;
723 path->startup_cost = startup_cost;
724 path->total_cost = total_cost;
729 * Determines and returns the cost of joining two relations using the
730 * nested loop algorithm.
732 * 'path' is already filled in except for the cost fields
735 cost_nestloop(NestPath *path, Query *root)
737 Path *outer_path = path->outerjoinpath;
738 Path *inner_path = path->innerjoinpath;
739 Cost startup_cost = 0;
742 QualCost restrict_qual_cost;
743 double outer_path_rows = PATH_ROWS(outer_path);
744 double inner_path_rows = PATH_ROWS(inner_path);
746 Selectivity joininfactor;
749 * If inner path is an indexscan, be sure to use its estimated output
750 * row count, which may be lower than the restriction-clause-only row
751 * count of its parent. (We don't include this case in the PATH_ROWS
752 * macro because it applies *only* to a nestloop's inner relation.)
754 if (IsA(inner_path, IndexPath))
755 inner_path_rows = ((IndexPath *) inner_path)->rows;
757 if (!enable_nestloop)
758 startup_cost += disable_cost;
761 * If we're doing JOIN_IN then we will stop scanning inner tuples for
762 * an outer tuple as soon as we have one match. Account for the
763 * effects of this by scaling down the cost estimates in proportion to
764 * the JOIN_IN selectivity. (This assumes that all the quals attached
765 * to the join are IN quals, which should be true.)
767 joininfactor = join_in_selectivity(path, root);
769 /* cost of source data */
772 * NOTE: clearly, we must pay both outer and inner paths' startup_cost
773 * before we can start returning tuples, so the join's startup cost is
774 * their sum. What's not so clear is whether the inner path's
775 * startup_cost must be paid again on each rescan of the inner path.
776 * This is not true if the inner path is materialized or is a
777 * hashjoin, but probably is true otherwise.
779 startup_cost += outer_path->startup_cost + inner_path->startup_cost;
780 run_cost += outer_path->total_cost - outer_path->startup_cost;
781 if (IsA(inner_path, MaterialPath) ||
782 IsA(inner_path, HashPath))
784 /* charge only run cost for each iteration of inner path */
789 * charge startup cost for each iteration of inner path, except we
790 * already charged the first startup_cost in our own startup
792 run_cost += (outer_path_rows - 1) * inner_path->startup_cost;
794 run_cost += outer_path_rows *
795 (inner_path->total_cost - inner_path->startup_cost) * joininfactor;
798 * Compute number of tuples processed (not number emitted!)
800 ntuples = outer_path_rows * inner_path_rows * joininfactor;
803 cost_qual_eval(&restrict_qual_cost, path->joinrestrictinfo);
804 startup_cost += restrict_qual_cost.startup;
805 cpu_per_tuple = cpu_tuple_cost + restrict_qual_cost.per_tuple;
806 run_cost += cpu_per_tuple * ntuples;
808 path->path.startup_cost = startup_cost;
809 path->path.total_cost = startup_cost + run_cost;
814 * Determines and returns the cost of joining two relations using the
815 * merge join algorithm.
817 * 'path' is already filled in except for the cost fields
819 * Notes: path's mergeclauses should be a subset of the joinrestrictinfo list;
820 * outersortkeys and innersortkeys are lists of the keys to be used
821 * to sort the outer and inner relations, or NIL if no explicit
822 * sort is needed because the source path is already ordered.
825 cost_mergejoin(MergePath *path, Query *root)
827 Path *outer_path = path->jpath.outerjoinpath;
828 Path *inner_path = path->jpath.innerjoinpath;
829 List *mergeclauses = path->path_mergeclauses;
830 List *outersortkeys = path->outersortkeys;
831 List *innersortkeys = path->innersortkeys;
832 Cost startup_cost = 0;
835 Selectivity merge_selec;
836 QualCost merge_qual_cost;
837 QualCost qp_qual_cost;
838 RestrictInfo *firstclause;
839 double outer_path_rows = PATH_ROWS(outer_path);
840 double inner_path_rows = PATH_ROWS(inner_path);
843 double mergejointuples,
846 Selectivity outerscansel,
848 Selectivity joininfactor;
849 Path sort_path; /* dummy for result of cost_sort */
851 if (!enable_mergejoin)
852 startup_cost += disable_cost;
855 * Compute cost and selectivity of the mergequals and qpquals (other
856 * restriction clauses) separately. We use approx_selectivity here
857 * for speed --- in most cases, any errors won't affect the result
860 * Note: it's probably bogus to use the normal selectivity calculation
861 * here when either the outer or inner path is a UniquePath.
863 merge_selec = approx_selectivity(root, mergeclauses,
864 path->jpath.jointype);
865 cost_qual_eval(&merge_qual_cost, mergeclauses);
866 cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo);
867 qp_qual_cost.startup -= merge_qual_cost.startup;
868 qp_qual_cost.per_tuple -= merge_qual_cost.per_tuple;
870 /* approx # tuples passing the merge quals */
871 mergejointuples = clamp_row_est(merge_selec * outer_path_rows * inner_path_rows);
874 * When there are equal merge keys in the outer relation, the
875 * mergejoin must rescan any matching tuples in the inner relation.
876 * This means re-fetching inner tuples. Our cost model for this is
877 * that a re-fetch costs the same as an original fetch, which is
878 * probably an overestimate; but on the other hand we ignore the
879 * bookkeeping costs of mark/restore. Not clear if it's worth
880 * developing a more refined model.
882 * The number of re-fetches can be estimated approximately as size of
883 * merge join output minus size of inner relation. Assume that the
884 * distinct key values are 1, 2, ..., and denote the number of values
885 * of each key in the outer relation as m1, m2, ...; in the inner
886 * relation, n1, n2, ... Then we have
888 * size of join = m1 * n1 + m2 * n2 + ...
890 * number of rescanned tuples = (m1 - 1) * n1 + (m2 - 1) * n2 + ... = m1 *
891 * n1 + m2 * n2 + ... - (n1 + n2 + ...) = size of join - size of inner
894 * This equation works correctly for outer tuples having no inner match
895 * (nk = 0), but not for inner tuples having no outer match (mk = 0);
896 * we are effectively subtracting those from the number of rescanned
897 * tuples, when we should not. Can we do better without expensive
898 * selectivity computations?
900 if (IsA(outer_path, UniquePath))
904 rescannedtuples = mergejointuples - inner_path_rows;
905 /* Must clamp because of possible underestimate */
906 if (rescannedtuples < 0)
909 /* We'll inflate inner run cost this much to account for rescanning */
910 rescanratio = 1.0 + (rescannedtuples / inner_path_rows);
913 * A merge join will stop as soon as it exhausts either input stream.
914 * Estimate fraction of the left and right inputs that will actually
915 * need to be scanned. We use only the first (most significant) merge
916 * clause for this purpose.
918 * Since this calculation is somewhat expensive, and will be the same for
919 * all mergejoin paths associated with the merge clause, we cache the
920 * results in the RestrictInfo node.
924 firstclause = (RestrictInfo *) linitial(mergeclauses);
925 if (firstclause->left_mergescansel < 0) /* not computed yet? */
926 mergejoinscansel(root, (Node *) firstclause->clause,
927 &firstclause->left_mergescansel,
928 &firstclause->right_mergescansel);
930 if (bms_is_subset(firstclause->left_relids, outer_path->parent->relids))
932 /* left side of clause is outer */
933 outerscansel = firstclause->left_mergescansel;
934 innerscansel = firstclause->right_mergescansel;
938 /* left side of clause is inner */
939 outerscansel = firstclause->right_mergescansel;
940 innerscansel = firstclause->left_mergescansel;
945 /* cope with clauseless mergejoin */
946 outerscansel = innerscansel = 1.0;
949 /* convert selectivity to row count; must scan at least one row */
950 outer_rows = clamp_row_est(outer_path_rows * outerscansel);
951 inner_rows = clamp_row_est(inner_path_rows * innerscansel);
954 * Readjust scan selectivities to account for above rounding. This is
955 * normally an insignificant effect, but when there are only a few
956 * rows in the inputs, failing to do this makes for a large percentage
959 outerscansel = outer_rows / outer_path_rows;
960 innerscansel = inner_rows / inner_path_rows;
962 /* cost of source data */
964 if (outersortkeys) /* do we need to sort outer? */
966 cost_sort(&sort_path,
969 outer_path->total_cost,
971 outer_path->parent->width);
972 startup_cost += sort_path.startup_cost;
973 run_cost += (sort_path.total_cost - sort_path.startup_cost)
978 startup_cost += outer_path->startup_cost;
979 run_cost += (outer_path->total_cost - outer_path->startup_cost)
983 if (innersortkeys) /* do we need to sort inner? */
985 cost_sort(&sort_path,
988 inner_path->total_cost,
990 inner_path->parent->width);
991 startup_cost += sort_path.startup_cost;
992 run_cost += (sort_path.total_cost - sort_path.startup_cost)
993 * innerscansel * rescanratio;
997 startup_cost += inner_path->startup_cost;
998 run_cost += (inner_path->total_cost - inner_path->startup_cost)
999 * innerscansel * rescanratio;
1005 * If we're doing JOIN_IN then we will stop outputting inner tuples
1006 * for an outer tuple as soon as we have one match. Account for the
1007 * effects of this by scaling down the cost estimates in proportion to
1008 * the expected output size. (This assumes that all the quals
1009 * attached to the join are IN quals, which should be true.)
1011 joininfactor = join_in_selectivity(&path->jpath, root);
1014 * The number of tuple comparisons needed is approximately number of
1015 * outer rows plus number of inner rows plus number of rescanned
1016 * tuples (can we refine this?). At each one, we need to evaluate the
1017 * mergejoin quals. NOTE: JOIN_IN mode does not save any work here,
1018 * so do NOT include joininfactor.
1020 startup_cost += merge_qual_cost.startup;
1021 run_cost += merge_qual_cost.per_tuple *
1022 (outer_rows + inner_rows * rescanratio);
1025 * For each tuple that gets through the mergejoin proper, we charge
1026 * cpu_tuple_cost plus the cost of evaluating additional restriction
1027 * clauses that are to be applied at the join. (This is pessimistic
1028 * since not all of the quals may get evaluated at each tuple.) This
1029 * work is skipped in JOIN_IN mode, so apply the factor.
1031 startup_cost += qp_qual_cost.startup;
1032 cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
1033 run_cost += cpu_per_tuple * mergejointuples * joininfactor;
1035 path->jpath.path.startup_cost = startup_cost;
1036 path->jpath.path.total_cost = startup_cost + run_cost;
1041 * Determines and returns the cost of joining two relations using the
1042 * hash join algorithm.
1044 * 'path' is already filled in except for the cost fields
1046 * Note: path's hashclauses should be a subset of the joinrestrictinfo list
1049 cost_hashjoin(HashPath *path, Query *root)
1051 Path *outer_path = path->jpath.outerjoinpath;
1052 Path *inner_path = path->jpath.innerjoinpath;
1053 List *hashclauses = path->path_hashclauses;
1054 Cost startup_cost = 0;
1057 Selectivity hash_selec;
1058 QualCost hash_qual_cost;
1059 QualCost qp_qual_cost;
1060 double hashjointuples;
1061 double outer_path_rows = PATH_ROWS(outer_path);
1062 double inner_path_rows = PATH_ROWS(inner_path);
1063 double outerbytes = relation_byte_size(outer_path_rows,
1064 outer_path->parent->width);
1065 double innerbytes = relation_byte_size(inner_path_rows,
1066 inner_path->parent->width);
1067 int num_hashclauses = list_length(hashclauses);
1069 int physicalbuckets;
1071 Selectivity innerbucketsize;
1072 Selectivity joininfactor;
1075 if (!enable_hashjoin)
1076 startup_cost += disable_cost;
1079 * Compute cost and selectivity of the hashquals and qpquals (other
1080 * restriction clauses) separately. We use approx_selectivity here
1081 * for speed --- in most cases, any errors won't affect the result
1084 * Note: it's probably bogus to use the normal selectivity calculation
1085 * here when either the outer or inner path is a UniquePath.
1087 hash_selec = approx_selectivity(root, hashclauses,
1088 path->jpath.jointype);
1089 cost_qual_eval(&hash_qual_cost, hashclauses);
1090 cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo);
1091 qp_qual_cost.startup -= hash_qual_cost.startup;
1092 qp_qual_cost.per_tuple -= hash_qual_cost.per_tuple;
1094 /* approx # tuples passing the hash quals */
1095 hashjointuples = clamp_row_est(hash_selec * outer_path_rows * inner_path_rows);
1097 /* cost of source data */
1098 startup_cost += outer_path->startup_cost;
1099 run_cost += outer_path->total_cost - outer_path->startup_cost;
1100 startup_cost += inner_path->total_cost;
1103 * Cost of computing hash function: must do it once per input tuple.
1104 * We charge one cpu_operator_cost for each column's hash function.
1106 * XXX when a hashclause is more complex than a single operator, we
1107 * really should charge the extra eval costs of the left or right
1108 * side, as appropriate, here. This seems more work than it's worth
1111 startup_cost += cpu_operator_cost * num_hashclauses * inner_path_rows;
1112 run_cost += cpu_operator_cost * num_hashclauses * outer_path_rows;
1114 /* Get hash table size that executor would use for inner relation */
1115 ExecChooseHashTableSize(inner_path_rows,
1116 inner_path->parent->width,
1122 * Determine bucketsize fraction for inner relation. We use the
1123 * smallest bucketsize estimated for any individual hashclause; this
1124 * is undoubtedly conservative.
1126 * BUT: if inner relation has been unique-ified, we can assume it's good
1127 * for hashing. This is important both because it's the right answer,
1128 * and because we avoid contaminating the cache with a value that's
1129 * wrong for non-unique-ified paths.
1131 if (IsA(inner_path, UniquePath))
1132 innerbucketsize = 1.0 / virtualbuckets;
1135 innerbucketsize = 1.0;
1136 foreach(hcl, hashclauses)
1138 RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(hcl);
1139 Selectivity thisbucketsize;
1141 Assert(IsA(restrictinfo, RestrictInfo));
1144 * First we have to figure out which side of the hashjoin
1145 * clause is the inner side.
1147 * Since we tend to visit the same clauses over and over when
1148 * planning a large query, we cache the bucketsize estimate in
1149 * the RestrictInfo node to avoid repeated lookups of
1152 if (bms_is_subset(restrictinfo->right_relids,
1153 inner_path->parent->relids))
1155 /* righthand side is inner */
1156 thisbucketsize = restrictinfo->right_bucketsize;
1157 if (thisbucketsize < 0)
1159 /* not cached yet */
1161 estimate_hash_bucketsize(root,
1162 get_rightop(restrictinfo->clause),
1164 restrictinfo->right_bucketsize = thisbucketsize;
1169 Assert(bms_is_subset(restrictinfo->left_relids,
1170 inner_path->parent->relids));
1171 /* lefthand side is inner */
1172 thisbucketsize = restrictinfo->left_bucketsize;
1173 if (thisbucketsize < 0)
1175 /* not cached yet */
1177 estimate_hash_bucketsize(root,
1178 get_leftop(restrictinfo->clause),
1180 restrictinfo->left_bucketsize = thisbucketsize;
1184 if (innerbucketsize > thisbucketsize)
1185 innerbucketsize = thisbucketsize;
1190 * if inner relation is too big then we will need to "batch" the join,
1191 * which implies writing and reading most of the tuples to disk an
1192 * extra time. Charge one cost unit per page of I/O (correct since it
1193 * should be nice and sequential...). Writing the inner rel counts as
1194 * startup cost, all the rest as run cost.
1198 double outerpages = page_size(outer_path_rows,
1199 outer_path->parent->width);
1200 double innerpages = page_size(inner_path_rows,
1201 inner_path->parent->width);
1203 startup_cost += innerpages;
1204 run_cost += innerpages + 2 * outerpages;
1210 * If we're doing JOIN_IN then we will stop comparing inner tuples to
1211 * an outer tuple as soon as we have one match. Account for the
1212 * effects of this by scaling down the cost estimates in proportion to
1213 * the expected output size. (This assumes that all the quals
1214 * attached to the join are IN quals, which should be true.)
1216 joininfactor = join_in_selectivity(&path->jpath, root);
1219 * The number of tuple comparisons needed is the number of outer
1220 * tuples times the typical number of tuples in a hash bucket, which
1221 * is the inner relation size times its bucketsize fraction. At each
1222 * one, we need to evaluate the hashjoin quals.
1224 startup_cost += hash_qual_cost.startup;
1225 run_cost += hash_qual_cost.per_tuple *
1226 outer_path_rows * clamp_row_est(inner_path_rows * innerbucketsize) *
1230 * For each tuple that gets through the hashjoin proper, we charge
1231 * cpu_tuple_cost plus the cost of evaluating additional restriction
1232 * clauses that are to be applied at the join. (This is pessimistic
1233 * since not all of the quals may get evaluated at each tuple.)
1235 startup_cost += qp_qual_cost.startup;
1236 cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
1237 run_cost += cpu_per_tuple * hashjointuples * joininfactor;
1240 * Bias against putting larger relation on inside. We don't want an
1241 * absolute prohibition, though, since larger relation might have
1242 * better bucketsize --- and we can't trust the size estimates
1243 * unreservedly, anyway. Instead, inflate the run cost by the square
1244 * root of the size ratio. (Why square root? No real good reason,
1245 * but it seems reasonable...)
1247 * Note: before 7.4 we implemented this by inflating startup cost; but if
1248 * there's a disable_cost component in the input paths' startup cost,
1249 * that unfairly penalizes the hash. Probably it'd be better to keep
1250 * track of disable penalty separately from cost.
1252 if (innerbytes > outerbytes && outerbytes > 0)
1253 run_cost *= sqrt(innerbytes / outerbytes);
1255 path->jpath.path.startup_cost = startup_cost;
1256 path->jpath.path.total_cost = startup_cost + run_cost;
1262 * Estimate the CPU costs of evaluating a WHERE clause.
1263 * The input can be either an implicitly-ANDed list of boolean
1264 * expressions, or a list of RestrictInfo nodes.
1265 * The result includes both a one-time (startup) component,
1266 * and a per-evaluation component.
1269 cost_qual_eval(QualCost *cost, List *quals)
1274 cost->per_tuple = 0;
1276 /* We don't charge any cost for the implicit ANDing at top level ... */
1280 Node *qual = (Node *) lfirst(l);
1283 * RestrictInfo nodes contain an eval_cost field reserved for this
1284 * routine's use, so that it's not necessary to evaluate the qual
1285 * clause's cost more than once. If the clause's cost hasn't been
1286 * computed yet, the field's startup value will contain -1.
1288 if (qual && IsA(qual, RestrictInfo))
1290 RestrictInfo *restrictinfo = (RestrictInfo *) qual;
1292 if (restrictinfo->eval_cost.startup < 0)
1294 restrictinfo->eval_cost.startup = 0;
1295 restrictinfo->eval_cost.per_tuple = 0;
1296 cost_qual_eval_walker((Node *) restrictinfo->clause,
1297 &restrictinfo->eval_cost);
1299 cost->startup += restrictinfo->eval_cost.startup;
1300 cost->per_tuple += restrictinfo->eval_cost.per_tuple;
1304 /* If it's a bare expression, must always do it the hard way */
1305 cost_qual_eval_walker(qual, cost);
1311 cost_qual_eval_walker(Node *node, QualCost *total)
1317 * Our basic strategy is to charge one cpu_operator_cost for each
1318 * operator or function node in the given tree. Vars and Consts are
1319 * charged zero, and so are boolean operators (AND, OR, NOT).
1320 * Simplistic, but a lot better than no model at all.
1322 * Should we try to account for the possibility of short-circuit
1323 * evaluation of AND/OR?
1325 if (IsA(node, FuncExpr) ||
1326 IsA(node, OpExpr) ||
1327 IsA(node, DistinctExpr) ||
1328 IsA(node, NullIfExpr))
1329 total->per_tuple += cpu_operator_cost;
1330 else if (IsA(node, ScalarArrayOpExpr))
1332 /* should charge more than 1 op cost, but how many? */
1333 total->per_tuple += cpu_operator_cost * 10;
1335 else if (IsA(node, SubLink))
1337 /* This routine should not be applied to un-planned expressions */
1338 elog(ERROR, "cannot handle unplanned sub-select");
1340 else if (IsA(node, SubPlan))
1343 * A subplan node in an expression typically indicates that the
1344 * subplan will be executed on each evaluation, so charge
1345 * accordingly. (Sub-selects that can be executed as InitPlans
1346 * have already been removed from the expression.)
1348 * An exception occurs when we have decided we can implement the
1349 * subplan by hashing.
1352 SubPlan *subplan = (SubPlan *) node;
1353 Plan *plan = subplan->plan;
1355 if (subplan->useHashTable)
1358 * If we are using a hash table for the subquery outputs, then
1359 * the cost of evaluating the query is a one-time cost. We
1360 * charge one cpu_operator_cost per tuple for the work of
1361 * loading the hashtable, too.
1363 total->startup += plan->total_cost +
1364 cpu_operator_cost * plan->plan_rows;
1367 * The per-tuple costs include the cost of evaluating the
1368 * lefthand expressions, plus the cost of probing the
1369 * hashtable. Recursion into the exprs list will handle the
1370 * lefthand expressions properly, and will count one
1371 * cpu_operator_cost for each comparison operator. That is
1372 * probably too low for the probing cost, but it's hard to
1373 * make a better estimate, so live with it for now.
1379 * Otherwise we will be rescanning the subplan output on each
1380 * evaluation. We need to estimate how much of the output we
1381 * will actually need to scan. NOTE: this logic should agree
1382 * with the estimates used by make_subplan() in
1385 Cost plan_run_cost = plan->total_cost - plan->startup_cost;
1387 if (subplan->subLinkType == EXISTS_SUBLINK)
1389 /* we only need to fetch 1 tuple */
1390 total->per_tuple += plan_run_cost / plan->plan_rows;
1392 else if (subplan->subLinkType == ALL_SUBLINK ||
1393 subplan->subLinkType == ANY_SUBLINK)
1395 /* assume we need 50% of the tuples */
1396 total->per_tuple += 0.50 * plan_run_cost;
1397 /* also charge a cpu_operator_cost per row examined */
1398 total->per_tuple += 0.50 * plan->plan_rows * cpu_operator_cost;
1402 /* assume we need all tuples */
1403 total->per_tuple += plan_run_cost;
1407 * Also account for subplan's startup cost. If the subplan is
1408 * uncorrelated or undirect correlated, AND its topmost node
1409 * is a Sort or Material node, assume that we'll only need to
1410 * pay its startup cost once; otherwise assume we pay the
1411 * startup cost every time.
1413 if (subplan->parParam == NIL &&
1415 IsA(plan, Material)))
1416 total->startup += plan->startup_cost;
1418 total->per_tuple += plan->startup_cost;
1422 return expression_tree_walker(node, cost_qual_eval_walker,
1428 * approx_selectivity
1429 * Quick-and-dirty estimation of clause selectivities.
1430 * The input can be either an implicitly-ANDed list of boolean
1431 * expressions, or a list of RestrictInfo nodes (typically the latter).
1433 * This is quick-and-dirty because we bypass clauselist_selectivity, and
1434 * simply multiply the independent clause selectivities together. Now
1435 * clauselist_selectivity often can't do any better than that anyhow, but
1436 * for some situations (such as range constraints) it is smarter. However,
1437 * we can't effectively cache the results of clauselist_selectivity, whereas
1438 * the individual clause selectivities can be and are cached.
1440 * Since we are only using the results to estimate how many potential
1441 * output tuples are generated and passed through qpqual checking, it
1442 * seems OK to live with the approximation.
1445 approx_selectivity(Query *root, List *quals, JoinType jointype)
1447 Selectivity total = 1.0;
1452 Node *qual = (Node *) lfirst(l);
1454 /* Note that clause_selectivity will be able to cache its result */
1455 total *= clause_selectivity(root, qual, 0, jointype);
1462 * set_baserel_size_estimates
1463 * Set the size estimates for the given base relation.
1465 * The rel's targetlist and restrictinfo list must have been constructed
1468 * We set the following fields of the rel node:
1469 * rows: the estimated number of output tuples (after applying
1470 * restriction clauses).
1471 * width: the estimated average output tuple width in bytes.
1472 * baserestrictcost: estimated cost of evaluating baserestrictinfo clauses.
1475 set_baserel_size_estimates(Query *root, RelOptInfo *rel)
1479 /* Should only be applied to base relations */
1480 Assert(rel->relid > 0);
1482 nrows = rel->tuples *
1483 clauselist_selectivity(root,
1484 rel->baserestrictinfo,
1488 rel->rows = clamp_row_est(nrows);
1490 cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo);
1492 set_rel_width(root, rel);
1496 * set_joinrel_size_estimates
1497 * Set the size estimates for the given join relation.
1499 * The rel's targetlist must have been constructed already, and a
1500 * restriction clause list that matches the given component rels must
1503 * Since there is more than one way to make a joinrel for more than two
1504 * base relations, the results we get here could depend on which component
1505 * rel pair is provided. In theory we should get the same answers no matter
1506 * which pair is provided; in practice, since the selectivity estimation
1507 * routines don't handle all cases equally well, we might not. But there's
1508 * not much to be done about it. (Would it make sense to repeat the
1509 * calculations for each pair of input rels that's encountered, and somehow
1510 * average the results? Probably way more trouble than it's worth.)
1512 * It's important that the results for symmetric JoinTypes be symmetric,
1513 * eg, (rel1, rel2, JOIN_LEFT) should produce the same result as (rel2,
1514 * rel1, JOIN_RIGHT). Also, JOIN_IN should produce the same result as
1515 * JOIN_UNIQUE_INNER, likewise JOIN_REVERSE_IN == JOIN_UNIQUE_OUTER.
1517 * We set only the rows field here. The width field was already set by
1518 * build_joinrel_tlist, and baserestrictcost is not used for join rels.
1521 set_joinrel_size_estimates(Query *root, RelOptInfo *rel,
1522 RelOptInfo *outer_rel,
1523 RelOptInfo *inner_rel,
1532 * Compute joinclause selectivity. Note that we are only considering
1533 * clauses that become restriction clauses at this join level; we are
1534 * not double-counting them because they were not considered in
1535 * estimating the sizes of the component rels.
1537 selec = clauselist_selectivity(root,
1543 * Basically, we multiply size of Cartesian product by selectivity.
1545 * If we are doing an outer join, take that into account: the output must
1546 * be at least as large as the non-nullable input. (Is there any
1547 * chance of being even smarter?)
1549 * For JOIN_IN and variants, the Cartesian product is figured with
1550 * respect to a unique-ified input, and then we can clamp to the size
1551 * of the other input.
1556 nrows = outer_rel->rows * inner_rel->rows * selec;
1559 nrows = outer_rel->rows * inner_rel->rows * selec;
1560 if (nrows < outer_rel->rows)
1561 nrows = outer_rel->rows;
1564 nrows = outer_rel->rows * inner_rel->rows * selec;
1565 if (nrows < inner_rel->rows)
1566 nrows = inner_rel->rows;
1569 nrows = outer_rel->rows * inner_rel->rows * selec;
1570 if (nrows < outer_rel->rows)
1571 nrows = outer_rel->rows;
1572 if (nrows < inner_rel->rows)
1573 nrows = inner_rel->rows;
1576 case JOIN_UNIQUE_INNER:
1577 upath = create_unique_path(root, inner_rel,
1578 inner_rel->cheapest_total_path);
1579 nrows = outer_rel->rows * upath->rows * selec;
1580 if (nrows > outer_rel->rows)
1581 nrows = outer_rel->rows;
1583 case JOIN_REVERSE_IN:
1584 case JOIN_UNIQUE_OUTER:
1585 upath = create_unique_path(root, outer_rel,
1586 outer_rel->cheapest_total_path);
1587 nrows = upath->rows * inner_rel->rows * selec;
1588 if (nrows > inner_rel->rows)
1589 nrows = inner_rel->rows;
1592 elog(ERROR, "unrecognized join type: %d", (int) jointype);
1593 nrows = 0; /* keep compiler quiet */
1597 rel->rows = clamp_row_est(nrows);
1601 * join_in_selectivity
1602 * Determines the factor by which a JOIN_IN join's result is expected
1603 * to be smaller than an ordinary inner join.
1605 * 'path' is already filled in except for the cost fields
1608 join_in_selectivity(JoinPath *path, Query *root)
1610 RelOptInfo *innerrel;
1611 UniquePath *innerunique;
1615 /* Return 1.0 whenever it's not JOIN_IN */
1616 if (path->jointype != JOIN_IN)
1620 * Return 1.0 if the inner side is already known unique. The case
1621 * where the inner path is already a UniquePath probably cannot happen
1622 * in current usage, but check it anyway for completeness. The
1623 * interesting case is where we've determined the inner relation
1624 * itself is unique, which we can check by looking at the rows
1625 * estimate for its UniquePath.
1627 if (IsA(path->innerjoinpath, UniquePath))
1629 innerrel = path->innerjoinpath->parent;
1630 innerunique = create_unique_path(root,
1632 innerrel->cheapest_total_path);
1633 if (innerunique->rows >= innerrel->rows)
1637 * Compute same result set_joinrel_size_estimates would compute for
1638 * JOIN_INNER. Note that we use the input rels' absolute size
1639 * estimates, not PATH_ROWS() which might be less; if we used
1640 * PATH_ROWS() we'd be double-counting the effects of any join clauses
1641 * used in input scans.
1643 selec = clauselist_selectivity(root,
1644 path->joinrestrictinfo,
1647 nrows = path->outerjoinpath->parent->rows * innerrel->rows * selec;
1649 nrows = clamp_row_est(nrows);
1651 /* See if it's larger than the actual JOIN_IN size estimate */
1652 if (nrows > path->path.parent->rows)
1653 return path->path.parent->rows / nrows;
1659 * set_function_size_estimates
1660 * Set the size estimates for a base relation that is a function call.
1662 * The rel's targetlist and restrictinfo list must have been constructed
1665 * We set the same fields as set_baserel_size_estimates.
1668 set_function_size_estimates(Query *root, RelOptInfo *rel)
1670 /* Should only be applied to base relations that are functions */
1671 Assert(rel->relid > 0);
1672 Assert(rel->rtekind == RTE_FUNCTION);
1675 * Estimate number of rows the function itself will return.
1677 * XXX no idea how to do this yet; but should at least check whether
1678 * function returns set or not...
1682 /* Now estimate number of output rows, etc */
1683 set_baserel_size_estimates(root, rel);
1689 * Set the estimated output width of a base relation.
1691 * NB: this works best on plain relations because it prefers to look at
1692 * real Vars. It will fail to make use of pg_statistic info when applied
1693 * to a subquery relation, even if the subquery outputs are simple vars
1694 * that we could have gotten info for. Is it worth trying to be smarter
1697 * The per-attribute width estimates are cached for possible re-use while
1698 * building join relations.
1701 set_rel_width(Query *root, RelOptInfo *rel)
1703 int32 tuple_width = 0;
1706 foreach(tllist, rel->reltargetlist)
1708 Var *var = (Var *) lfirst(tllist);
1713 /* For now, punt on whole-row child Vars */
1716 tuple_width += 32; /* arbitrary */
1720 ndx = var->varattno - rel->min_attr;
1723 * The width probably hasn't been cached yet, but may as well
1726 if (rel->attr_widths[ndx] > 0)
1728 tuple_width += rel->attr_widths[ndx];
1732 relid = getrelid(var->varno, root->rtable);
1733 if (relid != InvalidOid)
1735 item_width = get_attavgwidth(relid, var->varattno);
1738 rel->attr_widths[ndx] = item_width;
1739 tuple_width += item_width;
1745 * Not a plain relation, or can't find statistics for it. Estimate
1746 * using just the type info.
1748 item_width = get_typavgwidth(var->vartype, var->vartypmod);
1749 Assert(item_width > 0);
1750 rel->attr_widths[ndx] = item_width;
1751 tuple_width += item_width;
1753 Assert(tuple_width >= 0);
1754 rel->width = tuple_width;
1758 * relation_byte_size
1759 * Estimate the storage space in bytes for a given number of tuples
1760 * of a given width (size in bytes).
1763 relation_byte_size(double tuples, int width)
1765 return tuples * (MAXALIGN(width) + MAXALIGN(sizeof(HeapTupleHeaderData)));
1770 * Returns an estimate of the number of pages covered by a given
1771 * number of tuples of a given width (size in bytes).
1774 page_size(double tuples, int width)
1776 return ceil(relation_byte_size(tuples, width) / BLCKSZ);