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) 1996-2005, 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 1.140 2005/03/27 23:53:02 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 * 'index' is the index to be used
231 * 'indexQuals' is the list of applicable qual clauses (implicit AND semantics)
232 * 'is_injoin' is T if we are considering using the index scan as the inside
233 * of a nestloop join (hence, some of the indexQuals are join clauses)
235 * NOTE: 'indexQuals' must contain only clauses usable as index restrictions.
236 * Any additional quals evaluated as qpquals may reduce the number of returned
237 * tuples, but they won't reduce the number of tuples we have to fetch from
238 * the table, so they don't reduce the scan cost.
240 * NOTE: as of 8.0, indexQuals is a list of RestrictInfo nodes, where formerly
241 * it was a list of bare clause expressions.
244 cost_index(Path *path, Query *root,
249 RelOptInfo *baserel = index->rel;
250 Cost startup_cost = 0;
252 Cost indexStartupCost;
254 Selectivity indexSelectivity;
255 double indexCorrelation,
260 double tuples_fetched;
261 double pages_fetched;
265 /* Should only be applied to base relations */
266 Assert(IsA(baserel, RelOptInfo) &&
267 IsA(index, IndexOptInfo));
268 Assert(baserel->relid > 0);
269 Assert(baserel->rtekind == RTE_RELATION);
271 if (!enable_indexscan)
272 startup_cost += disable_cost;
275 * Call index-access-method-specific code to estimate the processing
276 * cost for scanning the index, as well as the selectivity of the
277 * index (ie, the fraction of main-table tuples we will have to
278 * retrieve) and its correlation to the main-table tuple order.
280 OidFunctionCall7(index->amcostestimate,
281 PointerGetDatum(root),
282 PointerGetDatum(index),
283 PointerGetDatum(indexQuals),
284 PointerGetDatum(&indexStartupCost),
285 PointerGetDatum(&indexTotalCost),
286 PointerGetDatum(&indexSelectivity),
287 PointerGetDatum(&indexCorrelation));
289 /* all costs for touching index itself included here */
290 startup_cost += indexStartupCost;
291 run_cost += indexTotalCost - indexStartupCost;
294 * Estimate number of main-table tuples and pages fetched.
296 * When the index ordering is uncorrelated with the table ordering,
297 * we use an approximation proposed by Mackert and Lohman, "Index Scans
298 * Using a Finite LRU Buffer: A Validated I/O Model", ACM Transactions
299 * on Database Systems, Vol. 14, No. 3, September 1989, Pages 401-424.
300 * The Mackert and Lohman approximation is that the number of pages
303 * min(2TNs/(2T+Ns), T) when T <= b
304 * 2TNs/(2T+Ns) when T > b and Ns <= 2Tb/(2T-b)
305 * b + (Ns - 2Tb/(2T-b))*(T-b)/T when T > b and Ns > 2Tb/(2T-b)
307 * T = # pages in table
308 * N = # tuples in table
309 * s = selectivity = fraction of table to be scanned
310 * b = # buffer pages available (we include kernel space here)
312 * When the index ordering is exactly correlated with the table ordering
313 * (just after a CLUSTER, for example), the number of pages fetched should
314 * be just sT. What's more, these will be sequential fetches, not the
315 * random fetches that occur in the uncorrelated case. So, depending on
316 * the extent of correlation, we should estimate the actual I/O cost
317 * somewhere between s * T * 1.0 and PF * random_cost. We currently
318 * interpolate linearly between these two endpoints based on the
319 * correlation squared (XXX is that appropriate?).
321 * In any case the number of tuples fetched is Ns.
325 tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
327 /* This part is the Mackert and Lohman formula */
329 T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
330 b = (effective_cache_size > 1) ? effective_cache_size : 1.0;
335 (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
336 if (pages_fetched > T)
343 lim = (2.0 * T * b) / (2.0 * T - b);
344 if (tuples_fetched <= lim)
347 (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
352 b + (tuples_fetched - lim) * (T - b) / T;
357 * min_IO_cost corresponds to the perfectly correlated case
358 * (csquared=1), max_IO_cost to the perfectly uncorrelated case
359 * (csquared=0). Note that we just charge random_page_cost per page
360 * in the uncorrelated case, rather than using
361 * cost_nonsequential_access, since we've already accounted for
362 * caching effects by using the Mackert model.
364 min_IO_cost = ceil(indexSelectivity * T);
365 max_IO_cost = pages_fetched * random_page_cost;
368 * Now interpolate based on estimated index order correlation to get
369 * total disk I/O cost for main table accesses.
371 csquared = indexCorrelation * indexCorrelation;
373 run_cost += max_IO_cost + csquared * (min_IO_cost - max_IO_cost);
376 * Estimate CPU costs per tuple.
378 * Normally the indexquals will be removed from the list of restriction
379 * clauses that we have to evaluate as qpquals, so we should subtract
380 * their costs from baserestrictcost. But if we are doing a join then
381 * some of the indexquals are join clauses and shouldn't be
382 * subtracted. Rather than work out exactly how much to subtract, we
383 * don't subtract anything.
385 startup_cost += baserel->baserestrictcost.startup;
386 cpu_per_tuple = cpu_tuple_cost + baserel->baserestrictcost.per_tuple;
390 QualCost index_qual_cost;
392 cost_qual_eval(&index_qual_cost, indexQuals);
393 /* any startup cost still has to be paid ... */
394 cpu_per_tuple -= index_qual_cost.per_tuple;
397 run_cost += cpu_per_tuple * tuples_fetched;
399 path->startup_cost = startup_cost;
400 path->total_cost = startup_cost + run_cost;
405 * Determines and returns the cost of scanning a relation using TIDs.
408 cost_tidscan(Path *path, Query *root,
409 RelOptInfo *baserel, List *tideval)
411 Cost startup_cost = 0;
414 int ntuples = list_length(tideval);
416 /* Should only be applied to base relations */
417 Assert(baserel->relid > 0);
418 Assert(baserel->rtekind == RTE_RELATION);
421 startup_cost += disable_cost;
423 /* disk costs --- assume each tuple on a different page */
424 run_cost += random_page_cost * ntuples;
427 startup_cost += baserel->baserestrictcost.startup;
428 cpu_per_tuple = cpu_tuple_cost + baserel->baserestrictcost.per_tuple;
429 run_cost += cpu_per_tuple * ntuples;
431 path->startup_cost = startup_cost;
432 path->total_cost = startup_cost + run_cost;
437 * Determines and returns the cost of scanning a subquery RTE.
440 cost_subqueryscan(Path *path, RelOptInfo *baserel)
446 /* Should only be applied to base relations that are subqueries */
447 Assert(baserel->relid > 0);
448 Assert(baserel->rtekind == RTE_SUBQUERY);
451 * Cost of path is cost of evaluating the subplan, plus cost of
452 * evaluating any restriction clauses that will be attached to the
453 * SubqueryScan node, plus cpu_tuple_cost to account for selection and
454 * projection overhead.
456 path->startup_cost = baserel->subplan->startup_cost;
457 path->total_cost = baserel->subplan->total_cost;
459 startup_cost = baserel->baserestrictcost.startup;
460 cpu_per_tuple = cpu_tuple_cost + baserel->baserestrictcost.per_tuple;
461 run_cost = cpu_per_tuple * baserel->tuples;
463 path->startup_cost += startup_cost;
464 path->total_cost += startup_cost + run_cost;
469 * Determines and returns the cost of scanning a function RTE.
472 cost_functionscan(Path *path, Query *root, RelOptInfo *baserel)
474 Cost startup_cost = 0;
478 /* Should only be applied to base relations that are functions */
479 Assert(baserel->relid > 0);
480 Assert(baserel->rtekind == RTE_FUNCTION);
483 * For now, estimate function's cost at one operator eval per function
484 * call. Someday we should revive the function cost estimate columns
487 cpu_per_tuple = cpu_operator_cost;
489 /* Add scanning CPU costs */
490 startup_cost += baserel->baserestrictcost.startup;
491 cpu_per_tuple += cpu_tuple_cost + baserel->baserestrictcost.per_tuple;
492 run_cost += cpu_per_tuple * baserel->tuples;
494 path->startup_cost = startup_cost;
495 path->total_cost = startup_cost + run_cost;
500 * Determines and returns the cost of sorting a relation, including
501 * the cost of reading the input data.
503 * If the total volume of data to sort is less than work_mem, we will do
504 * an in-memory sort, which requires no I/O and about t*log2(t) tuple
505 * comparisons for t tuples.
507 * If the total volume exceeds work_mem, we switch to a tape-style merge
508 * algorithm. There will still be about t*log2(t) tuple comparisons in
509 * total, but we will also need to write and read each tuple once per
510 * merge pass. We expect about ceil(log6(r)) merge passes where r is the
511 * number of initial runs formed (log6 because tuplesort.c uses six-tape
512 * merging). Since the average initial run should be about twice work_mem,
514 * disk traffic = 2 * relsize * ceil(log6(p / (2*work_mem)))
515 * cpu = comparison_cost * t * log2(t)
517 * The disk traffic is assumed to be half sequential and half random
518 * accesses (XXX can't we refine that guess?)
520 * We charge two operator evals per tuple comparison, which should be in
521 * the right ballpark in most cases.
523 * 'pathkeys' is a list of sort keys
524 * 'input_cost' is the total cost for reading the input data
525 * 'tuples' is the number of tuples in the relation
526 * 'width' is the average tuple width in bytes
528 * NOTE: some callers currently pass NIL for pathkeys because they
529 * can't conveniently supply the sort keys. Since this routine doesn't
530 * currently do anything with pathkeys anyway, that doesn't matter...
531 * but if it ever does, it should react gracefully to lack of key data.
532 * (Actually, the thing we'd most likely be interested in is just the number
533 * of sort keys, which all callers *could* supply.)
536 cost_sort(Path *path, Query *root,
537 List *pathkeys, Cost input_cost, double tuples, int width)
539 Cost startup_cost = input_cost;
541 double nbytes = relation_byte_size(tuples, width);
542 long work_mem_bytes = work_mem * 1024L;
545 startup_cost += disable_cost;
548 * We want to be sure the cost of a sort is never estimated as zero,
549 * even if passed-in tuple count is zero. Besides, mustn't do
558 * Assume about two operator evals per tuple comparison and N log2 N
561 startup_cost += 2.0 * cpu_operator_cost * tuples * LOG2(tuples);
564 if (nbytes > work_mem_bytes)
566 double npages = ceil(nbytes / BLCKSZ);
567 double nruns = (nbytes / work_mem_bytes) * 0.5;
568 double log_runs = ceil(LOG6(nruns));
569 double npageaccesses;
573 npageaccesses = 2.0 * npages * log_runs;
574 /* Assume half are sequential (cost 1), half are not */
575 startup_cost += npageaccesses *
576 (1.0 + cost_nonsequential_access(npages)) * 0.5;
580 * Also charge a small amount (arbitrarily set equal to operator cost)
581 * per extracted tuple.
583 run_cost += cpu_operator_cost * tuples;
585 path->startup_cost = startup_cost;
586 path->total_cost = startup_cost + run_cost;
591 * Determines and returns the cost of materializing a relation, including
592 * the cost of reading the input data.
594 * If the total volume of data to materialize exceeds work_mem, we will need
595 * to write it to disk, so the cost is much higher in that case.
598 cost_material(Path *path,
599 Cost input_cost, double tuples, int width)
601 Cost startup_cost = input_cost;
603 double nbytes = relation_byte_size(tuples, width);
604 long work_mem_bytes = work_mem * 1024L;
607 if (nbytes > work_mem_bytes)
609 double npages = ceil(nbytes / BLCKSZ);
611 /* We'll write during startup and read during retrieval */
612 startup_cost += npages;
617 * Charge a very small amount per inserted tuple, to reflect bookkeeping
618 * costs. We use cpu_tuple_cost/10 for this. This is needed to break
619 * the tie that would otherwise exist between nestloop with A outer,
620 * materialized B inner and nestloop with B outer, materialized A inner.
621 * The extra cost ensures we'll prefer materializing the smaller rel.
623 startup_cost += cpu_tuple_cost * 0.1 * tuples;
626 * Also charge a small amount per extracted tuple. We use
627 * cpu_tuple_cost so that it doesn't appear worthwhile to materialize
630 run_cost += cpu_tuple_cost * tuples;
632 path->startup_cost = startup_cost;
633 path->total_cost = startup_cost + run_cost;
638 * Determines and returns the cost of performing an Agg plan node,
639 * including the cost of its input.
641 * Note: when aggstrategy == AGG_SORTED, caller must ensure that input costs
642 * are for appropriately-sorted input.
645 cost_agg(Path *path, Query *root,
646 AggStrategy aggstrategy, int numAggs,
647 int numGroupCols, double numGroups,
648 Cost input_startup_cost, Cost input_total_cost,
655 * We charge one cpu_operator_cost per aggregate function per input
656 * tuple, and another one per output tuple (corresponding to transfn
657 * and finalfn calls respectively). If we are grouping, we charge an
658 * additional cpu_operator_cost per grouping column per input tuple
659 * for grouping comparisons.
661 * We will produce a single output tuple if not grouping, and a tuple per
664 * Note: in this cost model, AGG_SORTED and AGG_HASHED have exactly the
665 * same total CPU cost, but AGG_SORTED has lower startup cost. If the
666 * input path is already sorted appropriately, AGG_SORTED should be
667 * preferred (since it has no risk of memory overflow). This will
668 * happen as long as the computed total costs are indeed exactly equal
669 * --- but if there's roundoff error we might do the wrong thing. So
670 * be sure that the computations below form the same intermediate
671 * values in the same order.
673 if (aggstrategy == AGG_PLAIN)
675 startup_cost = input_total_cost;
676 startup_cost += cpu_operator_cost * (input_tuples + 1) * numAggs;
677 /* we aren't grouping */
678 total_cost = startup_cost;
680 else if (aggstrategy == AGG_SORTED)
682 /* Here we are able to deliver output on-the-fly */
683 startup_cost = input_startup_cost;
684 total_cost = input_total_cost;
685 /* calcs phrased this way to match HASHED case, see note above */
686 total_cost += cpu_operator_cost * input_tuples * numGroupCols;
687 total_cost += cpu_operator_cost * input_tuples * numAggs;
688 total_cost += cpu_operator_cost * numGroups * numAggs;
692 /* must be AGG_HASHED */
693 startup_cost = input_total_cost;
694 startup_cost += cpu_operator_cost * input_tuples * numGroupCols;
695 startup_cost += cpu_operator_cost * input_tuples * numAggs;
696 total_cost = startup_cost;
697 total_cost += cpu_operator_cost * numGroups * numAggs;
700 path->startup_cost = startup_cost;
701 path->total_cost = total_cost;
706 * Determines and returns the cost of performing a Group plan node,
707 * including the cost of its input.
709 * Note: caller must ensure that input costs are for appropriately-sorted
713 cost_group(Path *path, Query *root,
714 int numGroupCols, double numGroups,
715 Cost input_startup_cost, Cost input_total_cost,
721 startup_cost = input_startup_cost;
722 total_cost = input_total_cost;
725 * Charge one cpu_operator_cost per comparison per input tuple. We
726 * assume all columns get compared at most of the tuples.
728 total_cost += cpu_operator_cost * input_tuples * numGroupCols;
730 path->startup_cost = startup_cost;
731 path->total_cost = total_cost;
736 * Determines and returns the cost of joining two relations using the
737 * nested loop algorithm.
739 * 'path' is already filled in except for the cost fields
742 cost_nestloop(NestPath *path, Query *root)
744 Path *outer_path = path->outerjoinpath;
745 Path *inner_path = path->innerjoinpath;
746 Cost startup_cost = 0;
749 QualCost restrict_qual_cost;
750 double outer_path_rows = PATH_ROWS(outer_path);
751 double inner_path_rows = PATH_ROWS(inner_path);
753 Selectivity joininfactor;
756 * If inner path is an indexscan, be sure to use its estimated output
757 * row count, which may be lower than the restriction-clause-only row
758 * count of its parent. (We don't include this case in the PATH_ROWS
759 * macro because it applies *only* to a nestloop's inner relation.)
761 if (IsA(inner_path, IndexPath))
762 inner_path_rows = ((IndexPath *) inner_path)->rows;
764 if (!enable_nestloop)
765 startup_cost += disable_cost;
768 * If we're doing JOIN_IN then we will stop scanning inner tuples for
769 * an outer tuple as soon as we have one match. Account for the
770 * effects of this by scaling down the cost estimates in proportion to
771 * the JOIN_IN selectivity. (This assumes that all the quals attached
772 * to the join are IN quals, which should be true.)
774 joininfactor = join_in_selectivity(path, root);
776 /* cost of source data */
779 * NOTE: clearly, we must pay both outer and inner paths' startup_cost
780 * before we can start returning tuples, so the join's startup cost is
781 * their sum. What's not so clear is whether the inner path's
782 * startup_cost must be paid again on each rescan of the inner path.
783 * This is not true if the inner path is materialized or is a
784 * hashjoin, but probably is true otherwise.
786 startup_cost += outer_path->startup_cost + inner_path->startup_cost;
787 run_cost += outer_path->total_cost - outer_path->startup_cost;
788 if (IsA(inner_path, MaterialPath) ||
789 IsA(inner_path, HashPath))
791 /* charge only run cost for each iteration of inner path */
796 * charge startup cost for each iteration of inner path, except we
797 * already charged the first startup_cost in our own startup
799 run_cost += (outer_path_rows - 1) * inner_path->startup_cost;
801 run_cost += outer_path_rows *
802 (inner_path->total_cost - inner_path->startup_cost) * joininfactor;
805 * Compute number of tuples processed (not number emitted!)
807 ntuples = outer_path_rows * inner_path_rows * joininfactor;
810 cost_qual_eval(&restrict_qual_cost, path->joinrestrictinfo);
811 startup_cost += restrict_qual_cost.startup;
812 cpu_per_tuple = cpu_tuple_cost + restrict_qual_cost.per_tuple;
813 run_cost += cpu_per_tuple * ntuples;
815 path->path.startup_cost = startup_cost;
816 path->path.total_cost = startup_cost + run_cost;
821 * Determines and returns the cost of joining two relations using the
822 * merge join algorithm.
824 * 'path' is already filled in except for the cost fields
826 * Notes: path's mergeclauses should be a subset of the joinrestrictinfo list;
827 * outersortkeys and innersortkeys are lists of the keys to be used
828 * to sort the outer and inner relations, or NIL if no explicit
829 * sort is needed because the source path is already ordered.
832 cost_mergejoin(MergePath *path, Query *root)
834 Path *outer_path = path->jpath.outerjoinpath;
835 Path *inner_path = path->jpath.innerjoinpath;
836 List *mergeclauses = path->path_mergeclauses;
837 List *outersortkeys = path->outersortkeys;
838 List *innersortkeys = path->innersortkeys;
839 Cost startup_cost = 0;
842 Selectivity merge_selec;
843 QualCost merge_qual_cost;
844 QualCost qp_qual_cost;
845 RestrictInfo *firstclause;
846 double outer_path_rows = PATH_ROWS(outer_path);
847 double inner_path_rows = PATH_ROWS(inner_path);
850 double mergejointuples,
853 Selectivity outerscansel,
855 Selectivity joininfactor;
856 Path sort_path; /* dummy for result of cost_sort */
858 if (!enable_mergejoin)
859 startup_cost += disable_cost;
862 * Compute cost and selectivity of the mergequals and qpquals (other
863 * restriction clauses) separately. We use approx_selectivity here
864 * for speed --- in most cases, any errors won't affect the result
867 * Note: it's probably bogus to use the normal selectivity calculation
868 * here when either the outer or inner path is a UniquePath.
870 merge_selec = approx_selectivity(root, mergeclauses,
871 path->jpath.jointype);
872 cost_qual_eval(&merge_qual_cost, mergeclauses);
873 cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo);
874 qp_qual_cost.startup -= merge_qual_cost.startup;
875 qp_qual_cost.per_tuple -= merge_qual_cost.per_tuple;
877 /* approx # tuples passing the merge quals */
878 mergejointuples = clamp_row_est(merge_selec * outer_path_rows * inner_path_rows);
881 * When there are equal merge keys in the outer relation, the
882 * mergejoin must rescan any matching tuples in the inner relation.
883 * This means re-fetching inner tuples. Our cost model for this is
884 * that a re-fetch costs the same as an original fetch, which is
885 * probably an overestimate; but on the other hand we ignore the
886 * bookkeeping costs of mark/restore. Not clear if it's worth
887 * developing a more refined model.
889 * The number of re-fetches can be estimated approximately as size of
890 * merge join output minus size of inner relation. Assume that the
891 * distinct key values are 1, 2, ..., and denote the number of values
892 * of each key in the outer relation as m1, m2, ...; in the inner
893 * relation, n1, n2, ... Then we have
895 * size of join = m1 * n1 + m2 * n2 + ...
897 * number of rescanned tuples = (m1 - 1) * n1 + (m2 - 1) * n2 + ... = m1 *
898 * n1 + m2 * n2 + ... - (n1 + n2 + ...) = size of join - size of inner
901 * This equation works correctly for outer tuples having no inner match
902 * (nk = 0), but not for inner tuples having no outer match (mk = 0);
903 * we are effectively subtracting those from the number of rescanned
904 * tuples, when we should not. Can we do better without expensive
905 * selectivity computations?
907 if (IsA(outer_path, UniquePath))
911 rescannedtuples = mergejointuples - inner_path_rows;
912 /* Must clamp because of possible underestimate */
913 if (rescannedtuples < 0)
916 /* We'll inflate inner run cost this much to account for rescanning */
917 rescanratio = 1.0 + (rescannedtuples / inner_path_rows);
920 * A merge join will stop as soon as it exhausts either input stream.
921 * Estimate fraction of the left and right inputs that will actually
922 * need to be scanned. We use only the first (most significant) merge
923 * clause for this purpose.
925 * Since this calculation is somewhat expensive, and will be the same for
926 * all mergejoin paths associated with the merge clause, we cache the
927 * results in the RestrictInfo node.
931 firstclause = (RestrictInfo *) linitial(mergeclauses);
932 if (firstclause->left_mergescansel < 0) /* not computed yet? */
933 mergejoinscansel(root, (Node *) firstclause->clause,
934 &firstclause->left_mergescansel,
935 &firstclause->right_mergescansel);
937 if (bms_is_subset(firstclause->left_relids, outer_path->parent->relids))
939 /* left side of clause is outer */
940 outerscansel = firstclause->left_mergescansel;
941 innerscansel = firstclause->right_mergescansel;
945 /* left side of clause is inner */
946 outerscansel = firstclause->right_mergescansel;
947 innerscansel = firstclause->left_mergescansel;
952 /* cope with clauseless mergejoin */
953 outerscansel = innerscansel = 1.0;
956 /* convert selectivity to row count; must scan at least one row */
957 outer_rows = clamp_row_est(outer_path_rows * outerscansel);
958 inner_rows = clamp_row_est(inner_path_rows * innerscansel);
961 * Readjust scan selectivities to account for above rounding. This is
962 * normally an insignificant effect, but when there are only a few
963 * rows in the inputs, failing to do this makes for a large percentage
966 outerscansel = outer_rows / outer_path_rows;
967 innerscansel = inner_rows / inner_path_rows;
969 /* cost of source data */
971 if (outersortkeys) /* do we need to sort outer? */
973 cost_sort(&sort_path,
976 outer_path->total_cost,
978 outer_path->parent->width);
979 startup_cost += sort_path.startup_cost;
980 run_cost += (sort_path.total_cost - sort_path.startup_cost)
985 startup_cost += outer_path->startup_cost;
986 run_cost += (outer_path->total_cost - outer_path->startup_cost)
990 if (innersortkeys) /* do we need to sort inner? */
992 cost_sort(&sort_path,
995 inner_path->total_cost,
997 inner_path->parent->width);
998 startup_cost += sort_path.startup_cost;
999 run_cost += (sort_path.total_cost - sort_path.startup_cost)
1000 * innerscansel * rescanratio;
1004 startup_cost += inner_path->startup_cost;
1005 run_cost += (inner_path->total_cost - inner_path->startup_cost)
1006 * innerscansel * rescanratio;
1012 * If we're doing JOIN_IN then we will stop outputting inner tuples
1013 * for an outer tuple as soon as we have one match. Account for the
1014 * effects of this by scaling down the cost estimates in proportion to
1015 * the expected output size. (This assumes that all the quals
1016 * attached to the join are IN quals, which should be true.)
1018 joininfactor = join_in_selectivity(&path->jpath, root);
1021 * The number of tuple comparisons needed is approximately number of
1022 * outer rows plus number of inner rows plus number of rescanned
1023 * tuples (can we refine this?). At each one, we need to evaluate the
1024 * mergejoin quals. NOTE: JOIN_IN mode does not save any work here,
1025 * so do NOT include joininfactor.
1027 startup_cost += merge_qual_cost.startup;
1028 run_cost += merge_qual_cost.per_tuple *
1029 (outer_rows + inner_rows * rescanratio);
1032 * For each tuple that gets through the mergejoin proper, we charge
1033 * cpu_tuple_cost plus the cost of evaluating additional restriction
1034 * clauses that are to be applied at the join. (This is pessimistic
1035 * since not all of the quals may get evaluated at each tuple.) This
1036 * work is skipped in JOIN_IN mode, so apply the factor.
1038 startup_cost += qp_qual_cost.startup;
1039 cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
1040 run_cost += cpu_per_tuple * mergejointuples * joininfactor;
1042 path->jpath.path.startup_cost = startup_cost;
1043 path->jpath.path.total_cost = startup_cost + run_cost;
1048 * Determines and returns the cost of joining two relations using the
1049 * hash join algorithm.
1051 * 'path' is already filled in except for the cost fields
1053 * Note: path's hashclauses should be a subset of the joinrestrictinfo list
1056 cost_hashjoin(HashPath *path, Query *root)
1058 Path *outer_path = path->jpath.outerjoinpath;
1059 Path *inner_path = path->jpath.innerjoinpath;
1060 List *hashclauses = path->path_hashclauses;
1061 Cost startup_cost = 0;
1064 Selectivity hash_selec;
1065 QualCost hash_qual_cost;
1066 QualCost qp_qual_cost;
1067 double hashjointuples;
1068 double outer_path_rows = PATH_ROWS(outer_path);
1069 double inner_path_rows = PATH_ROWS(inner_path);
1070 double outerbytes = relation_byte_size(outer_path_rows,
1071 outer_path->parent->width);
1072 double innerbytes = relation_byte_size(inner_path_rows,
1073 inner_path->parent->width);
1074 int num_hashclauses = list_length(hashclauses);
1077 double virtualbuckets;
1078 Selectivity innerbucketsize;
1079 Selectivity joininfactor;
1082 if (!enable_hashjoin)
1083 startup_cost += disable_cost;
1086 * Compute cost and selectivity of the hashquals and qpquals (other
1087 * restriction clauses) separately. We use approx_selectivity here
1088 * for speed --- in most cases, any errors won't affect the result
1091 * Note: it's probably bogus to use the normal selectivity calculation
1092 * here when either the outer or inner path is a UniquePath.
1094 hash_selec = approx_selectivity(root, hashclauses,
1095 path->jpath.jointype);
1096 cost_qual_eval(&hash_qual_cost, hashclauses);
1097 cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo);
1098 qp_qual_cost.startup -= hash_qual_cost.startup;
1099 qp_qual_cost.per_tuple -= hash_qual_cost.per_tuple;
1101 /* approx # tuples passing the hash quals */
1102 hashjointuples = clamp_row_est(hash_selec * outer_path_rows * inner_path_rows);
1104 /* cost of source data */
1105 startup_cost += outer_path->startup_cost;
1106 run_cost += outer_path->total_cost - outer_path->startup_cost;
1107 startup_cost += inner_path->total_cost;
1110 * Cost of computing hash function: must do it once per input tuple.
1111 * We charge one cpu_operator_cost for each column's hash function.
1113 * XXX when a hashclause is more complex than a single operator, we
1114 * really should charge the extra eval costs of the left or right
1115 * side, as appropriate, here. This seems more work than it's worth
1118 startup_cost += cpu_operator_cost * num_hashclauses * inner_path_rows;
1119 run_cost += cpu_operator_cost * num_hashclauses * outer_path_rows;
1121 /* Get hash table size that executor would use for inner relation */
1122 ExecChooseHashTableSize(inner_path_rows,
1123 inner_path->parent->width,
1126 virtualbuckets = (double) numbuckets * (double) numbatches;
1129 * Determine bucketsize fraction for inner relation. We use the
1130 * smallest bucketsize estimated for any individual hashclause; this
1131 * is undoubtedly conservative.
1133 * BUT: if inner relation has been unique-ified, we can assume it's good
1134 * for hashing. This is important both because it's the right answer,
1135 * and because we avoid contaminating the cache with a value that's
1136 * wrong for non-unique-ified paths.
1138 if (IsA(inner_path, UniquePath))
1139 innerbucketsize = 1.0 / virtualbuckets;
1142 innerbucketsize = 1.0;
1143 foreach(hcl, hashclauses)
1145 RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(hcl);
1146 Selectivity thisbucketsize;
1148 Assert(IsA(restrictinfo, RestrictInfo));
1151 * First we have to figure out which side of the hashjoin
1152 * clause is the inner side.
1154 * Since we tend to visit the same clauses over and over when
1155 * planning a large query, we cache the bucketsize estimate in
1156 * the RestrictInfo node to avoid repeated lookups of
1159 if (bms_is_subset(restrictinfo->right_relids,
1160 inner_path->parent->relids))
1162 /* righthand side is inner */
1163 thisbucketsize = restrictinfo->right_bucketsize;
1164 if (thisbucketsize < 0)
1166 /* not cached yet */
1168 estimate_hash_bucketsize(root,
1169 get_rightop(restrictinfo->clause),
1171 restrictinfo->right_bucketsize = thisbucketsize;
1176 Assert(bms_is_subset(restrictinfo->left_relids,
1177 inner_path->parent->relids));
1178 /* lefthand side is inner */
1179 thisbucketsize = restrictinfo->left_bucketsize;
1180 if (thisbucketsize < 0)
1182 /* not cached yet */
1184 estimate_hash_bucketsize(root,
1185 get_leftop(restrictinfo->clause),
1187 restrictinfo->left_bucketsize = thisbucketsize;
1191 if (innerbucketsize > thisbucketsize)
1192 innerbucketsize = thisbucketsize;
1197 * If inner relation is too big then we will need to "batch" the join,
1198 * which implies writing and reading most of the tuples to disk an
1199 * extra time. Charge one cost unit per page of I/O (correct since it
1200 * should be nice and sequential...). Writing the inner rel counts as
1201 * startup cost, all the rest as run cost.
1205 double outerpages = page_size(outer_path_rows,
1206 outer_path->parent->width);
1207 double innerpages = page_size(inner_path_rows,
1208 inner_path->parent->width);
1210 startup_cost += innerpages;
1211 run_cost += innerpages + 2 * outerpages;
1217 * If we're doing JOIN_IN then we will stop comparing inner tuples to
1218 * an outer tuple as soon as we have one match. Account for the
1219 * effects of this by scaling down the cost estimates in proportion to
1220 * the expected output size. (This assumes that all the quals
1221 * attached to the join are IN quals, which should be true.)
1223 joininfactor = join_in_selectivity(&path->jpath, root);
1226 * The number of tuple comparisons needed is the number of outer
1227 * tuples times the typical number of tuples in a hash bucket, which
1228 * is the inner relation size times its bucketsize fraction. At each
1229 * one, we need to evaluate the hashjoin quals. (Note: charging the
1230 * full qual eval cost at each tuple is pessimistic, since we don't
1231 * evaluate the quals unless the hash values match exactly.)
1233 startup_cost += hash_qual_cost.startup;
1234 run_cost += hash_qual_cost.per_tuple *
1235 outer_path_rows * clamp_row_est(inner_path_rows * innerbucketsize) *
1239 * For each tuple that gets through the hashjoin proper, we charge
1240 * cpu_tuple_cost plus the cost of evaluating additional restriction
1241 * clauses that are to be applied at the join. (This is pessimistic
1242 * since not all of the quals may get evaluated at each tuple.)
1244 startup_cost += qp_qual_cost.startup;
1245 cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
1246 run_cost += cpu_per_tuple * hashjointuples * joininfactor;
1249 * Bias against putting larger relation on inside. We don't want an
1250 * absolute prohibition, though, since larger relation might have
1251 * better bucketsize --- and we can't trust the size estimates
1252 * unreservedly, anyway. Instead, inflate the run cost by the square
1253 * root of the size ratio. (Why square root? No real good reason,
1254 * but it seems reasonable...)
1256 * Note: before 7.4 we implemented this by inflating startup cost; but if
1257 * there's a disable_cost component in the input paths' startup cost,
1258 * that unfairly penalizes the hash. Probably it'd be better to keep
1259 * track of disable penalty separately from cost.
1261 if (innerbytes > outerbytes && outerbytes > 0)
1262 run_cost *= sqrt(innerbytes / outerbytes);
1264 path->jpath.path.startup_cost = startup_cost;
1265 path->jpath.path.total_cost = startup_cost + run_cost;
1271 * Estimate the CPU costs of evaluating a WHERE clause.
1272 * The input can be either an implicitly-ANDed list of boolean
1273 * expressions, or a list of RestrictInfo nodes.
1274 * The result includes both a one-time (startup) component,
1275 * and a per-evaluation component.
1278 cost_qual_eval(QualCost *cost, List *quals)
1283 cost->per_tuple = 0;
1285 /* We don't charge any cost for the implicit ANDing at top level ... */
1289 Node *qual = (Node *) lfirst(l);
1292 * RestrictInfo nodes contain an eval_cost field reserved for this
1293 * routine's use, so that it's not necessary to evaluate the qual
1294 * clause's cost more than once. If the clause's cost hasn't been
1295 * computed yet, the field's startup value will contain -1.
1297 if (qual && IsA(qual, RestrictInfo))
1299 RestrictInfo *restrictinfo = (RestrictInfo *) qual;
1301 if (restrictinfo->eval_cost.startup < 0)
1303 restrictinfo->eval_cost.startup = 0;
1304 restrictinfo->eval_cost.per_tuple = 0;
1305 cost_qual_eval_walker((Node *) restrictinfo->clause,
1306 &restrictinfo->eval_cost);
1308 cost->startup += restrictinfo->eval_cost.startup;
1309 cost->per_tuple += restrictinfo->eval_cost.per_tuple;
1313 /* If it's a bare expression, must always do it the hard way */
1314 cost_qual_eval_walker(qual, cost);
1320 cost_qual_eval_walker(Node *node, QualCost *total)
1326 * Our basic strategy is to charge one cpu_operator_cost for each
1327 * operator or function node in the given tree. Vars and Consts are
1328 * charged zero, and so are boolean operators (AND, OR, NOT).
1329 * Simplistic, but a lot better than no model at all.
1331 * Should we try to account for the possibility of short-circuit
1332 * evaluation of AND/OR?
1334 if (IsA(node, FuncExpr) ||
1335 IsA(node, OpExpr) ||
1336 IsA(node, DistinctExpr) ||
1337 IsA(node, NullIfExpr))
1338 total->per_tuple += cpu_operator_cost;
1339 else if (IsA(node, ScalarArrayOpExpr))
1341 /* should charge more than 1 op cost, but how many? */
1342 total->per_tuple += cpu_operator_cost * 10;
1344 else if (IsA(node, SubLink))
1346 /* This routine should not be applied to un-planned expressions */
1347 elog(ERROR, "cannot handle unplanned sub-select");
1349 else if (IsA(node, SubPlan))
1352 * A subplan node in an expression typically indicates that the
1353 * subplan will be executed on each evaluation, so charge
1354 * accordingly. (Sub-selects that can be executed as InitPlans
1355 * have already been removed from the expression.)
1357 * An exception occurs when we have decided we can implement the
1358 * subplan by hashing.
1361 SubPlan *subplan = (SubPlan *) node;
1362 Plan *plan = subplan->plan;
1364 if (subplan->useHashTable)
1367 * If we are using a hash table for the subquery outputs, then
1368 * the cost of evaluating the query is a one-time cost. We
1369 * charge one cpu_operator_cost per tuple for the work of
1370 * loading the hashtable, too.
1372 total->startup += plan->total_cost +
1373 cpu_operator_cost * plan->plan_rows;
1376 * The per-tuple costs include the cost of evaluating the
1377 * lefthand expressions, plus the cost of probing the
1378 * hashtable. Recursion into the exprs list will handle the
1379 * lefthand expressions properly, and will count one
1380 * cpu_operator_cost for each comparison operator. That is
1381 * probably too low for the probing cost, but it's hard to
1382 * make a better estimate, so live with it for now.
1388 * Otherwise we will be rescanning the subplan output on each
1389 * evaluation. We need to estimate how much of the output we
1390 * will actually need to scan. NOTE: this logic should agree
1391 * with the estimates used by make_subplan() in
1394 Cost plan_run_cost = plan->total_cost - plan->startup_cost;
1396 if (subplan->subLinkType == EXISTS_SUBLINK)
1398 /* we only need to fetch 1 tuple */
1399 total->per_tuple += plan_run_cost / plan->plan_rows;
1401 else if (subplan->subLinkType == ALL_SUBLINK ||
1402 subplan->subLinkType == ANY_SUBLINK)
1404 /* assume we need 50% of the tuples */
1405 total->per_tuple += 0.50 * plan_run_cost;
1406 /* also charge a cpu_operator_cost per row examined */
1407 total->per_tuple += 0.50 * plan->plan_rows * cpu_operator_cost;
1411 /* assume we need all tuples */
1412 total->per_tuple += plan_run_cost;
1416 * Also account for subplan's startup cost. If the subplan is
1417 * uncorrelated or undirect correlated, AND its topmost node
1418 * is a Sort or Material node, assume that we'll only need to
1419 * pay its startup cost once; otherwise assume we pay the
1420 * startup cost every time.
1422 if (subplan->parParam == NIL &&
1424 IsA(plan, Material)))
1425 total->startup += plan->startup_cost;
1427 total->per_tuple += plan->startup_cost;
1431 return expression_tree_walker(node, cost_qual_eval_walker,
1437 * approx_selectivity
1438 * Quick-and-dirty estimation of clause selectivities.
1439 * The input can be either an implicitly-ANDed list of boolean
1440 * expressions, or a list of RestrictInfo nodes (typically the latter).
1442 * This is quick-and-dirty because we bypass clauselist_selectivity, and
1443 * simply multiply the independent clause selectivities together. Now
1444 * clauselist_selectivity often can't do any better than that anyhow, but
1445 * for some situations (such as range constraints) it is smarter. However,
1446 * we can't effectively cache the results of clauselist_selectivity, whereas
1447 * the individual clause selectivities can be and are cached.
1449 * Since we are only using the results to estimate how many potential
1450 * output tuples are generated and passed through qpqual checking, it
1451 * seems OK to live with the approximation.
1454 approx_selectivity(Query *root, List *quals, JoinType jointype)
1456 Selectivity total = 1.0;
1461 Node *qual = (Node *) lfirst(l);
1463 /* Note that clause_selectivity will be able to cache its result */
1464 total *= clause_selectivity(root, qual, 0, jointype);
1471 * set_baserel_size_estimates
1472 * Set the size estimates for the given base relation.
1474 * The rel's targetlist and restrictinfo list must have been constructed
1477 * We set the following fields of the rel node:
1478 * rows: the estimated number of output tuples (after applying
1479 * restriction clauses).
1480 * width: the estimated average output tuple width in bytes.
1481 * baserestrictcost: estimated cost of evaluating baserestrictinfo clauses.
1484 set_baserel_size_estimates(Query *root, RelOptInfo *rel)
1488 /* Should only be applied to base relations */
1489 Assert(rel->relid > 0);
1491 nrows = rel->tuples *
1492 clauselist_selectivity(root,
1493 rel->baserestrictinfo,
1497 rel->rows = clamp_row_est(nrows);
1499 cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo);
1501 set_rel_width(root, rel);
1505 * set_joinrel_size_estimates
1506 * Set the size estimates for the given join relation.
1508 * The rel's targetlist must have been constructed already, and a
1509 * restriction clause list that matches the given component rels must
1512 * Since there is more than one way to make a joinrel for more than two
1513 * base relations, the results we get here could depend on which component
1514 * rel pair is provided. In theory we should get the same answers no matter
1515 * which pair is provided; in practice, since the selectivity estimation
1516 * routines don't handle all cases equally well, we might not. But there's
1517 * not much to be done about it. (Would it make sense to repeat the
1518 * calculations for each pair of input rels that's encountered, and somehow
1519 * average the results? Probably way more trouble than it's worth.)
1521 * It's important that the results for symmetric JoinTypes be symmetric,
1522 * eg, (rel1, rel2, JOIN_LEFT) should produce the same result as (rel2,
1523 * rel1, JOIN_RIGHT). Also, JOIN_IN should produce the same result as
1524 * JOIN_UNIQUE_INNER, likewise JOIN_REVERSE_IN == JOIN_UNIQUE_OUTER.
1526 * We set only the rows field here. The width field was already set by
1527 * build_joinrel_tlist, and baserestrictcost is not used for join rels.
1530 set_joinrel_size_estimates(Query *root, RelOptInfo *rel,
1531 RelOptInfo *outer_rel,
1532 RelOptInfo *inner_rel,
1541 * Compute joinclause selectivity. Note that we are only considering
1542 * clauses that become restriction clauses at this join level; we are
1543 * not double-counting them because they were not considered in
1544 * estimating the sizes of the component rels.
1546 selec = clauselist_selectivity(root,
1552 * Basically, we multiply size of Cartesian product by selectivity.
1554 * If we are doing an outer join, take that into account: the output must
1555 * be at least as large as the non-nullable input. (Is there any
1556 * chance of being even smarter?)
1558 * For JOIN_IN and variants, the Cartesian product is figured with
1559 * respect to a unique-ified input, and then we can clamp to the size
1560 * of the other input.
1565 nrows = outer_rel->rows * inner_rel->rows * selec;
1568 nrows = outer_rel->rows * inner_rel->rows * selec;
1569 if (nrows < outer_rel->rows)
1570 nrows = outer_rel->rows;
1573 nrows = outer_rel->rows * inner_rel->rows * selec;
1574 if (nrows < inner_rel->rows)
1575 nrows = inner_rel->rows;
1578 nrows = outer_rel->rows * inner_rel->rows * selec;
1579 if (nrows < outer_rel->rows)
1580 nrows = outer_rel->rows;
1581 if (nrows < inner_rel->rows)
1582 nrows = inner_rel->rows;
1585 case JOIN_UNIQUE_INNER:
1586 upath = create_unique_path(root, inner_rel,
1587 inner_rel->cheapest_total_path);
1588 nrows = outer_rel->rows * upath->rows * selec;
1589 if (nrows > outer_rel->rows)
1590 nrows = outer_rel->rows;
1592 case JOIN_REVERSE_IN:
1593 case JOIN_UNIQUE_OUTER:
1594 upath = create_unique_path(root, outer_rel,
1595 outer_rel->cheapest_total_path);
1596 nrows = upath->rows * inner_rel->rows * selec;
1597 if (nrows > inner_rel->rows)
1598 nrows = inner_rel->rows;
1601 elog(ERROR, "unrecognized join type: %d", (int) jointype);
1602 nrows = 0; /* keep compiler quiet */
1606 rel->rows = clamp_row_est(nrows);
1610 * join_in_selectivity
1611 * Determines the factor by which a JOIN_IN join's result is expected
1612 * to be smaller than an ordinary inner join.
1614 * 'path' is already filled in except for the cost fields
1617 join_in_selectivity(JoinPath *path, Query *root)
1619 RelOptInfo *innerrel;
1620 UniquePath *innerunique;
1624 /* Return 1.0 whenever it's not JOIN_IN */
1625 if (path->jointype != JOIN_IN)
1629 * Return 1.0 if the inner side is already known unique. The case
1630 * where the inner path is already a UniquePath probably cannot happen
1631 * in current usage, but check it anyway for completeness. The
1632 * interesting case is where we've determined the inner relation
1633 * itself is unique, which we can check by looking at the rows
1634 * estimate for its UniquePath.
1636 if (IsA(path->innerjoinpath, UniquePath))
1638 innerrel = path->innerjoinpath->parent;
1639 innerunique = create_unique_path(root,
1641 innerrel->cheapest_total_path);
1642 if (innerunique->rows >= innerrel->rows)
1646 * Compute same result set_joinrel_size_estimates would compute for
1647 * JOIN_INNER. Note that we use the input rels' absolute size
1648 * estimates, not PATH_ROWS() which might be less; if we used
1649 * PATH_ROWS() we'd be double-counting the effects of any join clauses
1650 * used in input scans.
1652 selec = clauselist_selectivity(root,
1653 path->joinrestrictinfo,
1656 nrows = path->outerjoinpath->parent->rows * innerrel->rows * selec;
1658 nrows = clamp_row_est(nrows);
1660 /* See if it's larger than the actual JOIN_IN size estimate */
1661 if (nrows > path->path.parent->rows)
1662 return path->path.parent->rows / nrows;
1668 * set_function_size_estimates
1669 * Set the size estimates for a base relation that is a function call.
1671 * The rel's targetlist and restrictinfo list must have been constructed
1674 * We set the same fields as set_baserel_size_estimates.
1677 set_function_size_estimates(Query *root, RelOptInfo *rel)
1679 /* Should only be applied to base relations that are functions */
1680 Assert(rel->relid > 0);
1681 Assert(rel->rtekind == RTE_FUNCTION);
1684 * Estimate number of rows the function itself will return.
1686 * XXX no idea how to do this yet; but should at least check whether
1687 * function returns set or not...
1691 /* Now estimate number of output rows, etc */
1692 set_baserel_size_estimates(root, rel);
1698 * Set the estimated output width of a base relation.
1700 * NB: this works best on plain relations because it prefers to look at
1701 * real Vars. It will fail to make use of pg_statistic info when applied
1702 * to a subquery relation, even if the subquery outputs are simple vars
1703 * that we could have gotten info for. Is it worth trying to be smarter
1706 * The per-attribute width estimates are cached for possible re-use while
1707 * building join relations.
1710 set_rel_width(Query *root, RelOptInfo *rel)
1712 int32 tuple_width = 0;
1715 foreach(tllist, rel->reltargetlist)
1717 Var *var = (Var *) lfirst(tllist);
1722 /* For now, punt on whole-row child Vars */
1725 tuple_width += 32; /* arbitrary */
1729 ndx = var->varattno - rel->min_attr;
1732 * The width probably hasn't been cached yet, but may as well
1735 if (rel->attr_widths[ndx] > 0)
1737 tuple_width += rel->attr_widths[ndx];
1741 relid = getrelid(var->varno, root->rtable);
1742 if (relid != InvalidOid)
1744 item_width = get_attavgwidth(relid, var->varattno);
1747 rel->attr_widths[ndx] = item_width;
1748 tuple_width += item_width;
1754 * Not a plain relation, or can't find statistics for it. Estimate
1755 * using just the type info.
1757 item_width = get_typavgwidth(var->vartype, var->vartypmod);
1758 Assert(item_width > 0);
1759 rel->attr_widths[ndx] = item_width;
1760 tuple_width += item_width;
1762 Assert(tuple_width >= 0);
1763 rel->width = tuple_width;
1767 * relation_byte_size
1768 * Estimate the storage space in bytes for a given number of tuples
1769 * of a given width (size in bytes).
1772 relation_byte_size(double tuples, int width)
1774 return tuples * (MAXALIGN(width) + MAXALIGN(sizeof(HeapTupleHeaderData)));
1779 * Returns an estimate of the number of pages covered by a given
1780 * number of tuples of a given width (size in bytes).
1783 page_size(double tuples, int width)
1785 return ceil(relation_byte_size(tuples, width) / BLCKSZ);