1 /*-------------------------------------------------------------------------
4 * Routines to compute (and set) relation sizes and path costs
6 * Path costs are measured in arbitrary units established by these basic
9 * seq_page_cost Cost of a sequential page fetch
10 * random_page_cost Cost of a non-sequential page fetch
11 * cpu_tuple_cost Cost of typical CPU time to process a tuple
12 * cpu_index_tuple_cost Cost of typical CPU time to process an index tuple
13 * cpu_operator_cost Cost of CPU time to execute an operator or function
15 * We expect that the kernel will typically do some amount of read-ahead
16 * optimization; this in conjunction with seek costs means that seq_page_cost
17 * is normally considerably less than random_page_cost. (However, if the
18 * database is fully cached in RAM, it is reasonable to set them equal.)
20 * We also use a rough estimate "effective_cache_size" of the number of
21 * disk pages in Postgres + OS-level disk cache. (We can't simply use
22 * NBuffers for this purpose because that would ignore the effects of
23 * the kernel's disk cache.)
25 * Obviously, taking constants for these values is an oversimplification,
26 * but it's tough enough to get any useful estimates even at this level of
27 * detail. Note that all of these parameters are user-settable, in case
28 * the default values are drastically off for a particular platform.
30 * We compute two separate costs for each path:
31 * total_cost: total estimated cost to fetch all tuples
32 * startup_cost: cost that is expended before first tuple is fetched
33 * In some scenarios, such as when there is a LIMIT or we are implementing
34 * an EXISTS(...) sub-select, it is not necessary to fetch all tuples of the
35 * path's result. A caller can estimate the cost of fetching a partial
36 * result by interpolating between startup_cost and total_cost. In detail:
37 * actual_cost = startup_cost +
38 * (total_cost - startup_cost) * tuples_to_fetch / path->parent->rows;
39 * Note that a base relation's rows count (and, by extension, plan_rows for
40 * plan nodes below the LIMIT node) are set without regard to any LIMIT, so
41 * that this equation works properly. (Also, these routines guarantee not to
42 * set the rows count to zero, so there will be no zero divide.) The LIMIT is
43 * applied as a top-level plan node.
45 * For largely historical reasons, most of the routines in this module use
46 * the passed result Path only to store their startup_cost and total_cost
47 * results into. All the input data they need is passed as separate
48 * parameters, even though much of it could be extracted from the Path.
49 * An exception is made for the cost_XXXjoin() routines, which expect all
50 * the non-cost fields of the passed XXXPath to be filled in.
53 * Portions Copyright (c) 1996-2007, PostgreSQL Global Development Group
54 * Portions Copyright (c) 1994, Regents of the University of California
57 * $PostgreSQL: pgsql/src/backend/optimizer/path/costsize.c,v 1.178 2007/02/22 22:00:24 tgl Exp $
59 *-------------------------------------------------------------------------
66 #include "executor/nodeHash.h"
67 #include "miscadmin.h"
68 #include "optimizer/clauses.h"
69 #include "optimizer/cost.h"
70 #include "optimizer/pathnode.h"
71 #include "optimizer/planmain.h"
72 #include "parser/parsetree.h"
73 #include "utils/lsyscache.h"
74 #include "utils/selfuncs.h"
75 #include "utils/tuplesort.h"
78 #define LOG2(x) (log(x) / 0.693147180559945)
81 * Some Paths return less than the nominal number of rows of their parent
82 * relations; join nodes need to do this to get the correct input count:
84 #define PATH_ROWS(path) \
85 (IsA(path, UniquePath) ? \
86 ((UniquePath *) (path))->rows : \
90 double seq_page_cost = DEFAULT_SEQ_PAGE_COST;
91 double random_page_cost = DEFAULT_RANDOM_PAGE_COST;
92 double cpu_tuple_cost = DEFAULT_CPU_TUPLE_COST;
93 double cpu_index_tuple_cost = DEFAULT_CPU_INDEX_TUPLE_COST;
94 double cpu_operator_cost = DEFAULT_CPU_OPERATOR_COST;
96 int effective_cache_size = DEFAULT_EFFECTIVE_CACHE_SIZE;
98 Cost disable_cost = 100000000.0;
100 bool enable_seqscan = true;
101 bool enable_indexscan = true;
102 bool enable_bitmapscan = true;
103 bool enable_tidscan = true;
104 bool enable_sort = true;
105 bool enable_hashagg = true;
106 bool enable_nestloop = true;
107 bool enable_mergejoin = true;
108 bool enable_hashjoin = true;
114 } cost_qual_eval_context;
116 static MergeScanSelCache *cached_scansel(PlannerInfo *root,
119 static bool cost_qual_eval_walker(Node *node, cost_qual_eval_context *context);
120 static Selectivity approx_selectivity(PlannerInfo *root, List *quals,
122 static Selectivity join_in_selectivity(JoinPath *path, PlannerInfo *root);
123 static void set_rel_width(PlannerInfo *root, RelOptInfo *rel);
124 static double relation_byte_size(double tuples, int width);
125 static double page_size(double tuples, int width);
130 * Force a row-count estimate to a sane value.
133 clamp_row_est(double nrows)
136 * Force estimate to be at least one row, to make explain output look
137 * better and to avoid possible divide-by-zero when interpolating costs.
138 * Make it an integer, too.
151 * Determines and returns the cost of scanning a relation sequentially.
154 cost_seqscan(Path *path, PlannerInfo *root,
157 Cost startup_cost = 0;
161 /* Should only be applied to base relations */
162 Assert(baserel->relid > 0);
163 Assert(baserel->rtekind == RTE_RELATION);
166 startup_cost += disable_cost;
171 run_cost += seq_page_cost * baserel->pages;
174 startup_cost += baserel->baserestrictcost.startup;
175 cpu_per_tuple = cpu_tuple_cost + baserel->baserestrictcost.per_tuple;
176 run_cost += cpu_per_tuple * baserel->tuples;
178 path->startup_cost = startup_cost;
179 path->total_cost = startup_cost + run_cost;
184 * Determines and returns the cost of scanning a relation using an index.
186 * 'index' is the index to be used
187 * 'indexQuals' is the list of applicable qual clauses (implicit AND semantics)
188 * 'outer_rel' is the outer relation when we are considering using the index
189 * scan as the inside of a nestloop join (hence, some of the indexQuals
190 * are join clauses, and we should expect repeated scans of the index);
191 * NULL for a plain index scan
193 * cost_index() takes an IndexPath not just a Path, because it sets a few
194 * additional fields of the IndexPath besides startup_cost and total_cost.
195 * These fields are needed if the IndexPath is used in a BitmapIndexScan.
197 * NOTE: 'indexQuals' must contain only clauses usable as index restrictions.
198 * Any additional quals evaluated as qpquals may reduce the number of returned
199 * tuples, but they won't reduce the number of tuples we have to fetch from
200 * the table, so they don't reduce the scan cost.
202 * NOTE: as of 8.0, indexQuals is a list of RestrictInfo nodes, where formerly
203 * it was a list of bare clause expressions.
206 cost_index(IndexPath *path, PlannerInfo *root,
209 RelOptInfo *outer_rel)
211 RelOptInfo *baserel = index->rel;
212 Cost startup_cost = 0;
214 Cost indexStartupCost;
216 Selectivity indexSelectivity;
217 double indexCorrelation,
222 double tuples_fetched;
223 double pages_fetched;
225 /* Should only be applied to base relations */
226 Assert(IsA(baserel, RelOptInfo) &&
227 IsA(index, IndexOptInfo));
228 Assert(baserel->relid > 0);
229 Assert(baserel->rtekind == RTE_RELATION);
231 if (!enable_indexscan)
232 startup_cost += disable_cost;
235 * Call index-access-method-specific code to estimate the processing cost
236 * for scanning the index, as well as the selectivity of the index (ie,
237 * the fraction of main-table tuples we will have to retrieve) and its
238 * correlation to the main-table tuple order.
240 OidFunctionCall8(index->amcostestimate,
241 PointerGetDatum(root),
242 PointerGetDatum(index),
243 PointerGetDatum(indexQuals),
244 PointerGetDatum(outer_rel),
245 PointerGetDatum(&indexStartupCost),
246 PointerGetDatum(&indexTotalCost),
247 PointerGetDatum(&indexSelectivity),
248 PointerGetDatum(&indexCorrelation));
251 * Save amcostestimate's results for possible use in bitmap scan planning.
252 * We don't bother to save indexStartupCost or indexCorrelation, because a
253 * bitmap scan doesn't care about either.
255 path->indextotalcost = indexTotalCost;
256 path->indexselectivity = indexSelectivity;
258 /* all costs for touching index itself included here */
259 startup_cost += indexStartupCost;
260 run_cost += indexTotalCost - indexStartupCost;
262 /* estimate number of main-table tuples fetched */
263 tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
266 * Estimate number of main-table pages fetched, and compute I/O cost.
268 * When the index ordering is uncorrelated with the table ordering,
269 * we use an approximation proposed by Mackert and Lohman (see
270 * index_pages_fetched() for details) to compute the number of pages
271 * fetched, and then charge random_page_cost per page fetched.
273 * When the index ordering is exactly correlated with the table ordering
274 * (just after a CLUSTER, for example), the number of pages fetched should
275 * be exactly selectivity * table_size. What's more, all but the first
276 * will be sequential fetches, not the random fetches that occur in the
277 * uncorrelated case. So if the number of pages is more than 1, we
279 * random_page_cost + (pages_fetched - 1) * seq_page_cost
280 * For partially-correlated indexes, we ought to charge somewhere between
281 * these two estimates. We currently interpolate linearly between the
282 * estimates based on the correlation squared (XXX is that appropriate?).
285 if (outer_rel != NULL && outer_rel->rows > 1)
288 * For repeated indexscans, the appropriate estimate for the
289 * uncorrelated case is to scale up the number of tuples fetched in
290 * the Mackert and Lohman formula by the number of scans, so that we
291 * estimate the number of pages fetched by all the scans; then
292 * pro-rate the costs for one scan. In this case we assume all the
293 * fetches are random accesses.
295 double num_scans = outer_rel->rows;
297 pages_fetched = index_pages_fetched(tuples_fetched * num_scans,
299 (double) index->pages,
302 max_IO_cost = (pages_fetched * random_page_cost) / num_scans;
305 * In the perfectly correlated case, the number of pages touched
306 * by each scan is selectivity * table_size, and we can use the
307 * Mackert and Lohman formula at the page level to estimate how
308 * much work is saved by caching across scans. We still assume
309 * all the fetches are random, though, which is an overestimate
310 * that's hard to correct for without double-counting the cache
311 * effects. (But in most cases where such a plan is actually
312 * interesting, only one page would get fetched per scan anyway,
313 * so it shouldn't matter much.)
315 pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
317 pages_fetched = index_pages_fetched(pages_fetched * num_scans,
319 (double) index->pages,
322 min_IO_cost = (pages_fetched * random_page_cost) / num_scans;
327 * Normal case: apply the Mackert and Lohman formula, and then
328 * interpolate between that and the correlation-derived result.
330 pages_fetched = index_pages_fetched(tuples_fetched,
332 (double) index->pages,
335 /* max_IO_cost is for the perfectly uncorrelated case (csquared=0) */
336 max_IO_cost = pages_fetched * random_page_cost;
338 /* min_IO_cost is for the perfectly correlated case (csquared=1) */
339 pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
340 min_IO_cost = random_page_cost;
341 if (pages_fetched > 1)
342 min_IO_cost += (pages_fetched - 1) * seq_page_cost;
346 * Now interpolate based on estimated index order correlation to get
347 * total disk I/O cost for main table accesses.
349 csquared = indexCorrelation * indexCorrelation;
351 run_cost += max_IO_cost + csquared * (min_IO_cost - max_IO_cost);
354 * Estimate CPU costs per tuple.
356 * Normally the indexquals will be removed from the list of restriction
357 * clauses that we have to evaluate as qpquals, so we should subtract
358 * their costs from baserestrictcost. But if we are doing a join then
359 * some of the indexquals are join clauses and shouldn't be subtracted.
360 * Rather than work out exactly how much to subtract, we don't subtract
363 startup_cost += baserel->baserestrictcost.startup;
364 cpu_per_tuple = cpu_tuple_cost + baserel->baserestrictcost.per_tuple;
366 if (outer_rel == NULL)
368 QualCost index_qual_cost;
370 cost_qual_eval(&index_qual_cost, indexQuals, root);
371 /* any startup cost still has to be paid ... */
372 cpu_per_tuple -= index_qual_cost.per_tuple;
375 run_cost += cpu_per_tuple * tuples_fetched;
377 path->path.startup_cost = startup_cost;
378 path->path.total_cost = startup_cost + run_cost;
382 * index_pages_fetched
383 * Estimate the number of pages actually fetched after accounting for
386 * We use an approximation proposed by Mackert and Lohman, "Index Scans
387 * Using a Finite LRU Buffer: A Validated I/O Model", ACM Transactions
388 * on Database Systems, Vol. 14, No. 3, September 1989, Pages 401-424.
389 * The Mackert and Lohman approximation is that the number of pages
392 * min(2TNs/(2T+Ns), T) when T <= b
393 * 2TNs/(2T+Ns) when T > b and Ns <= 2Tb/(2T-b)
394 * b + (Ns - 2Tb/(2T-b))*(T-b)/T when T > b and Ns > 2Tb/(2T-b)
396 * T = # pages in table
397 * N = # tuples in table
398 * s = selectivity = fraction of table to be scanned
399 * b = # buffer pages available (we include kernel space here)
401 * We assume that effective_cache_size is the total number of buffer pages
402 * available for the whole query, and pro-rate that space across all the
403 * tables in the query and the index currently under consideration. (This
404 * ignores space needed for other indexes used by the query, but since we
405 * don't know which indexes will get used, we can't estimate that very well;
406 * and in any case counting all the tables may well be an overestimate, since
407 * depending on the join plan not all the tables may be scanned concurrently.)
409 * The product Ns is the number of tuples fetched; we pass in that
410 * product rather than calculating it here. "pages" is the number of pages
411 * in the object under consideration (either an index or a table).
412 * "index_pages" is the amount to add to the total table space, which was
413 * computed for us by query_planner.
415 * Caller is expected to have ensured that tuples_fetched is greater than zero
416 * and rounded to integer (see clamp_row_est). The result will likewise be
417 * greater than zero and integral.
420 index_pages_fetched(double tuples_fetched, BlockNumber pages,
421 double index_pages, PlannerInfo *root)
423 double pages_fetched;
428 /* T is # pages in table, but don't allow it to be zero */
429 T = (pages > 1) ? (double) pages : 1.0;
431 /* Compute number of pages assumed to be competing for cache space */
432 total_pages = root->total_table_pages + index_pages;
433 total_pages = Max(total_pages, 1.0);
434 Assert(T <= total_pages);
436 /* b is pro-rated share of effective_cache_size */
437 b = (double) effective_cache_size *T / total_pages;
439 /* force it positive and integral */
445 /* This part is the Mackert and Lohman formula */
449 (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
450 if (pages_fetched >= T)
453 pages_fetched = ceil(pages_fetched);
459 lim = (2.0 * T * b) / (2.0 * T - b);
460 if (tuples_fetched <= lim)
463 (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
468 b + (tuples_fetched - lim) * (T - b) / T;
470 pages_fetched = ceil(pages_fetched);
472 return pages_fetched;
476 * get_indexpath_pages
477 * Determine the total size of the indexes used in a bitmap index path.
479 * Note: if the same index is used more than once in a bitmap tree, we will
480 * count it multiple times, which perhaps is the wrong thing ... but it's
481 * not completely clear, and detecting duplicates is difficult, so ignore it
485 get_indexpath_pages(Path *bitmapqual)
490 if (IsA(bitmapqual, BitmapAndPath))
492 BitmapAndPath *apath = (BitmapAndPath *) bitmapqual;
494 foreach(l, apath->bitmapquals)
496 result += get_indexpath_pages((Path *) lfirst(l));
499 else if (IsA(bitmapqual, BitmapOrPath))
501 BitmapOrPath *opath = (BitmapOrPath *) bitmapqual;
503 foreach(l, opath->bitmapquals)
505 result += get_indexpath_pages((Path *) lfirst(l));
508 else if (IsA(bitmapqual, IndexPath))
510 IndexPath *ipath = (IndexPath *) bitmapqual;
512 result = (double) ipath->indexinfo->pages;
515 elog(ERROR, "unrecognized node type: %d", nodeTag(bitmapqual));
521 * cost_bitmap_heap_scan
522 * Determines and returns the cost of scanning a relation using a bitmap
523 * index-then-heap plan.
525 * 'baserel' is the relation to be scanned
526 * 'bitmapqual' is a tree of IndexPaths, BitmapAndPaths, and BitmapOrPaths
527 * 'outer_rel' is the outer relation when we are considering using the bitmap
528 * scan as the inside of a nestloop join (hence, some of the indexQuals
529 * are join clauses, and we should expect repeated scans of the table);
530 * NULL for a plain bitmap scan
532 * Note: if this is a join inner path, the component IndexPaths in bitmapqual
533 * should have been costed accordingly.
536 cost_bitmap_heap_scan(Path *path, PlannerInfo *root, RelOptInfo *baserel,
537 Path *bitmapqual, RelOptInfo *outer_rel)
539 Cost startup_cost = 0;
542 Selectivity indexSelectivity;
545 double tuples_fetched;
546 double pages_fetched;
549 /* Should only be applied to base relations */
550 Assert(IsA(baserel, RelOptInfo));
551 Assert(baserel->relid > 0);
552 Assert(baserel->rtekind == RTE_RELATION);
554 if (!enable_bitmapscan)
555 startup_cost += disable_cost;
558 * Fetch total cost of obtaining the bitmap, as well as its total
561 cost_bitmap_tree_node(bitmapqual, &indexTotalCost, &indexSelectivity);
563 startup_cost += indexTotalCost;
566 * Estimate number of main-table pages fetched.
568 tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
570 T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
572 if (outer_rel != NULL && outer_rel->rows > 1)
575 * For repeated bitmap scans, scale up the number of tuples fetched in
576 * the Mackert and Lohman formula by the number of scans, so that we
577 * estimate the number of pages fetched by all the scans. Then
578 * pro-rate for one scan.
580 double num_scans = outer_rel->rows;
582 pages_fetched = index_pages_fetched(tuples_fetched * num_scans,
584 get_indexpath_pages(bitmapqual),
586 pages_fetched /= num_scans;
591 * For a single scan, the number of heap pages that need to be fetched
592 * is the same as the Mackert and Lohman formula for the case T <= b
593 * (ie, no re-reads needed).
595 pages_fetched = (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
597 if (pages_fetched >= T)
600 pages_fetched = ceil(pages_fetched);
603 * For small numbers of pages we should charge random_page_cost apiece,
604 * while if nearly all the table's pages are being read, it's more
605 * appropriate to charge seq_page_cost apiece. The effect is nonlinear,
606 * too. For lack of a better idea, interpolate like this to determine the
609 if (pages_fetched >= 2.0)
610 cost_per_page = random_page_cost -
611 (random_page_cost - seq_page_cost) * sqrt(pages_fetched / T);
613 cost_per_page = random_page_cost;
615 run_cost += pages_fetched * cost_per_page;
618 * Estimate CPU costs per tuple.
620 * Often the indexquals don't need to be rechecked at each tuple ... but
621 * not always, especially not if there are enough tuples involved that the
622 * bitmaps become lossy. For the moment, just assume they will be
625 startup_cost += baserel->baserestrictcost.startup;
626 cpu_per_tuple = cpu_tuple_cost + baserel->baserestrictcost.per_tuple;
628 run_cost += cpu_per_tuple * tuples_fetched;
630 path->startup_cost = startup_cost;
631 path->total_cost = startup_cost + run_cost;
635 * cost_bitmap_tree_node
636 * Extract cost and selectivity from a bitmap tree node (index/and/or)
639 cost_bitmap_tree_node(Path *path, Cost *cost, Selectivity *selec)
641 if (IsA(path, IndexPath))
643 *cost = ((IndexPath *) path)->indextotalcost;
644 *selec = ((IndexPath *) path)->indexselectivity;
646 * Charge a small amount per retrieved tuple to reflect the costs of
647 * manipulating the bitmap. This is mostly to make sure that a bitmap
648 * scan doesn't look to be the same cost as an indexscan to retrieve
651 *cost += 0.1 * cpu_operator_cost * ((IndexPath *) path)->rows;
653 else if (IsA(path, BitmapAndPath))
655 *cost = path->total_cost;
656 *selec = ((BitmapAndPath *) path)->bitmapselectivity;
658 else if (IsA(path, BitmapOrPath))
660 *cost = path->total_cost;
661 *selec = ((BitmapOrPath *) path)->bitmapselectivity;
665 elog(ERROR, "unrecognized node type: %d", nodeTag(path));
666 *cost = *selec = 0; /* keep compiler quiet */
671 * cost_bitmap_and_node
672 * Estimate the cost of a BitmapAnd node
674 * Note that this considers only the costs of index scanning and bitmap
675 * creation, not the eventual heap access. In that sense the object isn't
676 * truly a Path, but it has enough path-like properties (costs in particular)
677 * to warrant treating it as one.
680 cost_bitmap_and_node(BitmapAndPath *path, PlannerInfo *root)
687 * We estimate AND selectivity on the assumption that the inputs are
688 * independent. This is probably often wrong, but we don't have the info
691 * The runtime cost of the BitmapAnd itself is estimated at 100x
692 * cpu_operator_cost for each tbm_intersect needed. Probably too small,
693 * definitely too simplistic?
697 foreach(l, path->bitmapquals)
699 Path *subpath = (Path *) lfirst(l);
701 Selectivity subselec;
703 cost_bitmap_tree_node(subpath, &subCost, &subselec);
707 totalCost += subCost;
708 if (l != list_head(path->bitmapquals))
709 totalCost += 100.0 * cpu_operator_cost;
711 path->bitmapselectivity = selec;
712 path->path.startup_cost = totalCost;
713 path->path.total_cost = totalCost;
717 * cost_bitmap_or_node
718 * Estimate the cost of a BitmapOr node
720 * See comments for cost_bitmap_and_node.
723 cost_bitmap_or_node(BitmapOrPath *path, PlannerInfo *root)
730 * We estimate OR selectivity on the assumption that the inputs are
731 * non-overlapping, since that's often the case in "x IN (list)" type
732 * situations. Of course, we clamp to 1.0 at the end.
734 * The runtime cost of the BitmapOr itself is estimated at 100x
735 * cpu_operator_cost for each tbm_union needed. Probably too small,
736 * definitely too simplistic? We are aware that the tbm_unions are
737 * optimized out when the inputs are BitmapIndexScans.
741 foreach(l, path->bitmapquals)
743 Path *subpath = (Path *) lfirst(l);
745 Selectivity subselec;
747 cost_bitmap_tree_node(subpath, &subCost, &subselec);
751 totalCost += subCost;
752 if (l != list_head(path->bitmapquals) &&
753 !IsA(subpath, IndexPath))
754 totalCost += 100.0 * cpu_operator_cost;
756 path->bitmapselectivity = Min(selec, 1.0);
757 path->path.startup_cost = totalCost;
758 path->path.total_cost = totalCost;
763 * Determines and returns the cost of scanning a relation using TIDs.
766 cost_tidscan(Path *path, PlannerInfo *root,
767 RelOptInfo *baserel, List *tidquals)
769 Cost startup_cost = 0;
775 /* Should only be applied to base relations */
776 Assert(baserel->relid > 0);
777 Assert(baserel->rtekind == RTE_RELATION);
780 startup_cost += disable_cost;
782 /* Count how many tuples we expect to retrieve */
786 if (IsA(lfirst(l), ScalarArrayOpExpr))
788 /* Each element of the array yields 1 tuple */
789 ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) lfirst(l);
790 Node *arraynode = (Node *) lsecond(saop->args);
792 ntuples += estimate_array_length(arraynode);
796 /* It's just CTID = something, count 1 tuple */
801 /* disk costs --- assume each tuple on a different page */
802 run_cost += random_page_cost * ntuples;
805 startup_cost += baserel->baserestrictcost.startup;
806 cpu_per_tuple = cpu_tuple_cost + baserel->baserestrictcost.per_tuple;
807 run_cost += cpu_per_tuple * ntuples;
809 path->startup_cost = startup_cost;
810 path->total_cost = startup_cost + run_cost;
815 * Determines and returns the cost of scanning a subquery RTE.
818 cost_subqueryscan(Path *path, RelOptInfo *baserel)
824 /* Should only be applied to base relations that are subqueries */
825 Assert(baserel->relid > 0);
826 Assert(baserel->rtekind == RTE_SUBQUERY);
829 * Cost of path is cost of evaluating the subplan, plus cost of evaluating
830 * any restriction clauses that will be attached to the SubqueryScan node,
831 * plus cpu_tuple_cost to account for selection and projection overhead.
833 path->startup_cost = baserel->subplan->startup_cost;
834 path->total_cost = baserel->subplan->total_cost;
836 startup_cost = baserel->baserestrictcost.startup;
837 cpu_per_tuple = cpu_tuple_cost + baserel->baserestrictcost.per_tuple;
838 run_cost = cpu_per_tuple * baserel->tuples;
840 path->startup_cost += startup_cost;
841 path->total_cost += startup_cost + run_cost;
846 * Determines and returns the cost of scanning a function RTE.
849 cost_functionscan(Path *path, PlannerInfo *root, RelOptInfo *baserel)
851 Cost startup_cost = 0;
857 /* Should only be applied to base relations that are functions */
858 Assert(baserel->relid > 0);
859 rte = rt_fetch(baserel->relid, root->parse->rtable);
860 Assert(rte->rtekind == RTE_FUNCTION);
862 /* Estimate costs of executing the function expression */
863 cost_qual_eval_node(&exprcost, rte->funcexpr, root);
865 startup_cost += exprcost.startup;
866 cpu_per_tuple = exprcost.per_tuple;
868 /* Add scanning CPU costs */
869 startup_cost += baserel->baserestrictcost.startup;
870 cpu_per_tuple += cpu_tuple_cost + baserel->baserestrictcost.per_tuple;
871 run_cost += cpu_per_tuple * baserel->tuples;
873 path->startup_cost = startup_cost;
874 path->total_cost = startup_cost + run_cost;
879 * Determines and returns the cost of scanning a VALUES RTE.
882 cost_valuesscan(Path *path, PlannerInfo *root, RelOptInfo *baserel)
884 Cost startup_cost = 0;
888 /* Should only be applied to base relations that are values lists */
889 Assert(baserel->relid > 0);
890 Assert(baserel->rtekind == RTE_VALUES);
893 * For now, estimate list evaluation cost at one operator eval per list
894 * (probably pretty bogus, but is it worth being smarter?)
896 cpu_per_tuple = cpu_operator_cost;
898 /* Add scanning CPU costs */
899 startup_cost += baserel->baserestrictcost.startup;
900 cpu_per_tuple += cpu_tuple_cost + baserel->baserestrictcost.per_tuple;
901 run_cost += cpu_per_tuple * baserel->tuples;
903 path->startup_cost = startup_cost;
904 path->total_cost = startup_cost + run_cost;
909 * Determines and returns the cost of sorting a relation, including
910 * the cost of reading the input data.
912 * If the total volume of data to sort is less than work_mem, we will do
913 * an in-memory sort, which requires no I/O and about t*log2(t) tuple
914 * comparisons for t tuples.
916 * If the total volume exceeds work_mem, we switch to a tape-style merge
917 * algorithm. There will still be about t*log2(t) tuple comparisons in
918 * total, but we will also need to write and read each tuple once per
919 * merge pass. We expect about ceil(logM(r)) merge passes where r is the
920 * number of initial runs formed and M is the merge order used by tuplesort.c.
921 * Since the average initial run should be about twice work_mem, we have
922 * disk traffic = 2 * relsize * ceil(logM(p / (2*work_mem)))
923 * cpu = comparison_cost * t * log2(t)
925 * The disk traffic is assumed to be 3/4ths sequential and 1/4th random
926 * accesses (XXX can't we refine that guess?)
928 * We charge two operator evals per tuple comparison, which should be in
929 * the right ballpark in most cases.
931 * 'pathkeys' is a list of sort keys
932 * 'input_cost' is the total cost for reading the input data
933 * 'tuples' is the number of tuples in the relation
934 * 'width' is the average tuple width in bytes
936 * NOTE: some callers currently pass NIL for pathkeys because they
937 * can't conveniently supply the sort keys. Since this routine doesn't
938 * currently do anything with pathkeys anyway, that doesn't matter...
939 * but if it ever does, it should react gracefully to lack of key data.
940 * (Actually, the thing we'd most likely be interested in is just the number
941 * of sort keys, which all callers *could* supply.)
944 cost_sort(Path *path, PlannerInfo *root,
945 List *pathkeys, Cost input_cost, double tuples, int width)
947 Cost startup_cost = input_cost;
949 double nbytes = relation_byte_size(tuples, width);
950 long work_mem_bytes = work_mem * 1024L;
953 startup_cost += disable_cost;
956 * We want to be sure the cost of a sort is never estimated as zero, even
957 * if passed-in tuple count is zero. Besides, mustn't do log(0)...
965 * Assume about two operator evals per tuple comparison and N log2 N
968 startup_cost += 2.0 * cpu_operator_cost * tuples * LOG2(tuples);
971 if (nbytes > work_mem_bytes)
973 double npages = ceil(nbytes / BLCKSZ);
974 double nruns = (nbytes / work_mem_bytes) * 0.5;
975 double mergeorder = tuplesort_merge_order(work_mem_bytes);
977 double npageaccesses;
979 /* Compute logM(r) as log(r) / log(M) */
980 if (nruns > mergeorder)
981 log_runs = ceil(log(nruns) / log(mergeorder));
984 npageaccesses = 2.0 * npages * log_runs;
985 /* Assume 3/4ths of accesses are sequential, 1/4th are not */
986 startup_cost += npageaccesses *
987 (seq_page_cost * 0.75 + random_page_cost * 0.25);
991 * Also charge a small amount (arbitrarily set equal to operator cost) per
994 run_cost += cpu_operator_cost * tuples;
996 path->startup_cost = startup_cost;
997 path->total_cost = startup_cost + run_cost;
1002 * Determines and returns the cost of materializing a relation, including
1003 * the cost of reading the input data.
1005 * If the total volume of data to materialize exceeds work_mem, we will need
1006 * to write it to disk, so the cost is much higher in that case.
1009 cost_material(Path *path,
1010 Cost input_cost, double tuples, int width)
1012 Cost startup_cost = input_cost;
1014 double nbytes = relation_byte_size(tuples, width);
1015 long work_mem_bytes = work_mem * 1024L;
1018 if (nbytes > work_mem_bytes)
1020 double npages = ceil(nbytes / BLCKSZ);
1022 /* We'll write during startup and read during retrieval */
1023 startup_cost += seq_page_cost * npages;
1024 run_cost += seq_page_cost * npages;
1028 * Charge a very small amount per inserted tuple, to reflect bookkeeping
1029 * costs. We use cpu_tuple_cost/10 for this. This is needed to break the
1030 * tie that would otherwise exist between nestloop with A outer,
1031 * materialized B inner and nestloop with B outer, materialized A inner.
1032 * The extra cost ensures we'll prefer materializing the smaller rel.
1034 startup_cost += cpu_tuple_cost * 0.1 * tuples;
1037 * Also charge a small amount per extracted tuple. We use cpu_tuple_cost
1038 * so that it doesn't appear worthwhile to materialize a bare seqscan.
1040 run_cost += cpu_tuple_cost * tuples;
1042 path->startup_cost = startup_cost;
1043 path->total_cost = startup_cost + run_cost;
1048 * Determines and returns the cost of performing an Agg plan node,
1049 * including the cost of its input.
1051 * Note: when aggstrategy == AGG_SORTED, caller must ensure that input costs
1052 * are for appropriately-sorted input.
1055 cost_agg(Path *path, PlannerInfo *root,
1056 AggStrategy aggstrategy, int numAggs,
1057 int numGroupCols, double numGroups,
1058 Cost input_startup_cost, Cost input_total_cost,
1059 double input_tuples)
1065 * We charge one cpu_operator_cost per aggregate function per input tuple,
1066 * and another one per output tuple (corresponding to transfn and finalfn
1067 * calls respectively). If we are grouping, we charge an additional
1068 * cpu_operator_cost per grouping column per input tuple for grouping
1071 * We will produce a single output tuple if not grouping, and a tuple per
1072 * group otherwise. We charge cpu_tuple_cost for each output tuple.
1074 * Note: in this cost model, AGG_SORTED and AGG_HASHED have exactly the
1075 * same total CPU cost, but AGG_SORTED has lower startup cost. If the
1076 * input path is already sorted appropriately, AGG_SORTED should be
1077 * preferred (since it has no risk of memory overflow). This will happen
1078 * as long as the computed total costs are indeed exactly equal --- but if
1079 * there's roundoff error we might do the wrong thing. So be sure that
1080 * the computations below form the same intermediate values in the same
1083 * Note: ideally we should use the pg_proc.procost costs of each
1084 * aggregate's component functions, but for now that seems like an
1085 * excessive amount of work.
1087 if (aggstrategy == AGG_PLAIN)
1089 startup_cost = input_total_cost;
1090 startup_cost += cpu_operator_cost * (input_tuples + 1) * numAggs;
1091 /* we aren't grouping */
1092 total_cost = startup_cost + cpu_tuple_cost;
1094 else if (aggstrategy == AGG_SORTED)
1096 /* Here we are able to deliver output on-the-fly */
1097 startup_cost = input_startup_cost;
1098 total_cost = input_total_cost;
1099 /* calcs phrased this way to match HASHED case, see note above */
1100 total_cost += cpu_operator_cost * input_tuples * numGroupCols;
1101 total_cost += cpu_operator_cost * input_tuples * numAggs;
1102 total_cost += cpu_operator_cost * numGroups * numAggs;
1103 total_cost += cpu_tuple_cost * numGroups;
1107 /* must be AGG_HASHED */
1108 startup_cost = input_total_cost;
1109 startup_cost += cpu_operator_cost * input_tuples * numGroupCols;
1110 startup_cost += cpu_operator_cost * input_tuples * numAggs;
1111 total_cost = startup_cost;
1112 total_cost += cpu_operator_cost * numGroups * numAggs;
1113 total_cost += cpu_tuple_cost * numGroups;
1116 path->startup_cost = startup_cost;
1117 path->total_cost = total_cost;
1122 * Determines and returns the cost of performing a Group plan node,
1123 * including the cost of its input.
1125 * Note: caller must ensure that input costs are for appropriately-sorted
1129 cost_group(Path *path, PlannerInfo *root,
1130 int numGroupCols, double numGroups,
1131 Cost input_startup_cost, Cost input_total_cost,
1132 double input_tuples)
1137 startup_cost = input_startup_cost;
1138 total_cost = input_total_cost;
1141 * Charge one cpu_operator_cost per comparison per input tuple. We assume
1142 * all columns get compared at most of the tuples.
1144 total_cost += cpu_operator_cost * input_tuples * numGroupCols;
1146 path->startup_cost = startup_cost;
1147 path->total_cost = total_cost;
1151 * If a nestloop's inner path is an indexscan, be sure to use its estimated
1152 * output row count, which may be lower than the restriction-clause-only row
1153 * count of its parent. (We don't include this case in the PATH_ROWS macro
1154 * because it applies *only* to a nestloop's inner relation.) We have to
1155 * be prepared to recurse through Append nodes in case of an appendrel.
1158 nestloop_inner_path_rows(Path *path)
1162 if (IsA(path, IndexPath))
1163 result = ((IndexPath *) path)->rows;
1164 else if (IsA(path, BitmapHeapPath))
1165 result = ((BitmapHeapPath *) path)->rows;
1166 else if (IsA(path, AppendPath))
1171 foreach(l, ((AppendPath *) path)->subpaths)
1173 result += nestloop_inner_path_rows((Path *) lfirst(l));
1177 result = PATH_ROWS(path);
1184 * Determines and returns the cost of joining two relations using the
1185 * nested loop algorithm.
1187 * 'path' is already filled in except for the cost fields
1190 cost_nestloop(NestPath *path, PlannerInfo *root)
1192 Path *outer_path = path->outerjoinpath;
1193 Path *inner_path = path->innerjoinpath;
1194 Cost startup_cost = 0;
1197 QualCost restrict_qual_cost;
1198 double outer_path_rows = PATH_ROWS(outer_path);
1199 double inner_path_rows = nestloop_inner_path_rows(inner_path);
1201 Selectivity joininfactor;
1203 if (!enable_nestloop)
1204 startup_cost += disable_cost;
1207 * If we're doing JOIN_IN then we will stop scanning inner tuples for an
1208 * outer tuple as soon as we have one match. Account for the effects of
1209 * this by scaling down the cost estimates in proportion to the JOIN_IN
1210 * selectivity. (This assumes that all the quals attached to the join are
1211 * IN quals, which should be true.)
1213 joininfactor = join_in_selectivity(path, root);
1215 /* cost of source data */
1218 * NOTE: clearly, we must pay both outer and inner paths' startup_cost
1219 * before we can start returning tuples, so the join's startup cost is
1220 * their sum. What's not so clear is whether the inner path's
1221 * startup_cost must be paid again on each rescan of the inner path. This
1222 * is not true if the inner path is materialized or is a hashjoin, but
1223 * probably is true otherwise.
1225 startup_cost += outer_path->startup_cost + inner_path->startup_cost;
1226 run_cost += outer_path->total_cost - outer_path->startup_cost;
1227 if (IsA(inner_path, MaterialPath) ||
1228 IsA(inner_path, HashPath))
1230 /* charge only run cost for each iteration of inner path */
1235 * charge startup cost for each iteration of inner path, except we
1236 * already charged the first startup_cost in our own startup
1238 run_cost += (outer_path_rows - 1) * inner_path->startup_cost;
1240 run_cost += outer_path_rows *
1241 (inner_path->total_cost - inner_path->startup_cost) * joininfactor;
1244 * Compute number of tuples processed (not number emitted!)
1246 ntuples = outer_path_rows * inner_path_rows * joininfactor;
1249 cost_qual_eval(&restrict_qual_cost, path->joinrestrictinfo, root);
1250 startup_cost += restrict_qual_cost.startup;
1251 cpu_per_tuple = cpu_tuple_cost + restrict_qual_cost.per_tuple;
1252 run_cost += cpu_per_tuple * ntuples;
1254 path->path.startup_cost = startup_cost;
1255 path->path.total_cost = startup_cost + run_cost;
1260 * Determines and returns the cost of joining two relations using the
1261 * merge join algorithm.
1263 * 'path' is already filled in except for the cost fields
1265 * Notes: path's mergeclauses should be a subset of the joinrestrictinfo list;
1266 * outersortkeys and innersortkeys are lists of the keys to be used
1267 * to sort the outer and inner relations, or NIL if no explicit
1268 * sort is needed because the source path is already ordered.
1271 cost_mergejoin(MergePath *path, PlannerInfo *root)
1273 Path *outer_path = path->jpath.outerjoinpath;
1274 Path *inner_path = path->jpath.innerjoinpath;
1275 List *mergeclauses = path->path_mergeclauses;
1276 List *outersortkeys = path->outersortkeys;
1277 List *innersortkeys = path->innersortkeys;
1278 Cost startup_cost = 0;
1281 Selectivity merge_selec;
1282 QualCost merge_qual_cost;
1283 QualCost qp_qual_cost;
1284 double outer_path_rows = PATH_ROWS(outer_path);
1285 double inner_path_rows = PATH_ROWS(inner_path);
1288 double mergejointuples,
1291 Selectivity outerscansel,
1293 Selectivity joininfactor;
1294 Path sort_path; /* dummy for result of cost_sort */
1296 if (!enable_mergejoin)
1297 startup_cost += disable_cost;
1300 * Compute cost and selectivity of the mergequals and qpquals (other
1301 * restriction clauses) separately. We use approx_selectivity here for
1302 * speed --- in most cases, any errors won't affect the result much.
1304 * Note: it's probably bogus to use the normal selectivity calculation
1305 * here when either the outer or inner path is a UniquePath.
1307 merge_selec = approx_selectivity(root, mergeclauses,
1308 path->jpath.jointype);
1309 cost_qual_eval(&merge_qual_cost, mergeclauses, root);
1310 cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
1311 qp_qual_cost.startup -= merge_qual_cost.startup;
1312 qp_qual_cost.per_tuple -= merge_qual_cost.per_tuple;
1314 /* approx # tuples passing the merge quals */
1315 mergejointuples = clamp_row_est(merge_selec * outer_path_rows * inner_path_rows);
1318 * When there are equal merge keys in the outer relation, the mergejoin
1319 * must rescan any matching tuples in the inner relation. This means
1320 * re-fetching inner tuples. Our cost model for this is that a re-fetch
1321 * costs the same as an original fetch, which is probably an overestimate;
1322 * but on the other hand we ignore the bookkeeping costs of mark/restore.
1323 * Not clear if it's worth developing a more refined model.
1325 * The number of re-fetches can be estimated approximately as size of
1326 * merge join output minus size of inner relation. Assume that the
1327 * distinct key values are 1, 2, ..., and denote the number of values of
1328 * each key in the outer relation as m1, m2, ...; in the inner relation,
1329 * n1, n2, ... Then we have
1331 * size of join = m1 * n1 + m2 * n2 + ...
1333 * number of rescanned tuples = (m1 - 1) * n1 + (m2 - 1) * n2 + ... = m1 *
1334 * n1 + m2 * n2 + ... - (n1 + n2 + ...) = size of join - size of inner
1337 * This equation works correctly for outer tuples having no inner match
1338 * (nk = 0), but not for inner tuples having no outer match (mk = 0); we
1339 * are effectively subtracting those from the number of rescanned tuples,
1340 * when we should not. Can we do better without expensive selectivity
1343 if (IsA(outer_path, UniquePath))
1344 rescannedtuples = 0;
1347 rescannedtuples = mergejointuples - inner_path_rows;
1348 /* Must clamp because of possible underestimate */
1349 if (rescannedtuples < 0)
1350 rescannedtuples = 0;
1352 /* We'll inflate inner run cost this much to account for rescanning */
1353 rescanratio = 1.0 + (rescannedtuples / inner_path_rows);
1356 * A merge join will stop as soon as it exhausts either input stream
1357 * (unless it's an outer join, in which case the outer side has to be
1358 * scanned all the way anyway). Estimate fraction of the left and right
1359 * inputs that will actually need to be scanned. We use only the first
1360 * (most significant) merge clause for this purpose. Since
1361 * mergejoinscansel() is a fairly expensive computation, we cache the
1362 * results in the merge clause RestrictInfo.
1364 if (mergeclauses && path->jpath.jointype != JOIN_FULL)
1366 RestrictInfo *firstclause = (RestrictInfo *) linitial(mergeclauses);
1371 MergeScanSelCache *cache;
1373 /* Get the input pathkeys to determine the sort-order details */
1374 opathkeys = outersortkeys ? outersortkeys : outer_path->pathkeys;
1375 ipathkeys = innersortkeys ? innersortkeys : inner_path->pathkeys;
1378 opathkey = (PathKey *) linitial(opathkeys);
1379 ipathkey = (PathKey *) linitial(ipathkeys);
1380 /* debugging check */
1381 if (opathkey->pk_opfamily != ipathkey->pk_opfamily ||
1382 opathkey->pk_strategy != ipathkey->pk_strategy ||
1383 opathkey->pk_nulls_first != ipathkey->pk_nulls_first)
1384 elog(ERROR, "left and right pathkeys do not match in mergejoin");
1386 /* Get the selectivity with caching */
1387 cache = cached_scansel(root, firstclause, opathkey);
1389 if (bms_is_subset(firstclause->left_relids,
1390 outer_path->parent->relids))
1392 /* left side of clause is outer */
1393 outerscansel = cache->leftscansel;
1394 innerscansel = cache->rightscansel;
1398 /* left side of clause is inner */
1399 outerscansel = cache->rightscansel;
1400 innerscansel = cache->leftscansel;
1402 if (path->jpath.jointype == JOIN_LEFT)
1404 else if (path->jpath.jointype == JOIN_RIGHT)
1409 /* cope with clauseless or full mergejoin */
1410 outerscansel = innerscansel = 1.0;
1413 /* convert selectivity to row count; must scan at least one row */
1414 outer_rows = clamp_row_est(outer_path_rows * outerscansel);
1415 inner_rows = clamp_row_est(inner_path_rows * innerscansel);
1418 * Readjust scan selectivities to account for above rounding. This is
1419 * normally an insignificant effect, but when there are only a few rows in
1420 * the inputs, failing to do this makes for a large percentage error.
1422 outerscansel = outer_rows / outer_path_rows;
1423 innerscansel = inner_rows / inner_path_rows;
1425 /* cost of source data */
1427 if (outersortkeys) /* do we need to sort outer? */
1429 cost_sort(&sort_path,
1432 outer_path->total_cost,
1434 outer_path->parent->width);
1435 startup_cost += sort_path.startup_cost;
1436 run_cost += (sort_path.total_cost - sort_path.startup_cost)
1441 startup_cost += outer_path->startup_cost;
1442 run_cost += (outer_path->total_cost - outer_path->startup_cost)
1446 if (innersortkeys) /* do we need to sort inner? */
1448 cost_sort(&sort_path,
1451 inner_path->total_cost,
1453 inner_path->parent->width);
1454 startup_cost += sort_path.startup_cost;
1455 run_cost += (sort_path.total_cost - sort_path.startup_cost)
1456 * innerscansel * rescanratio;
1460 startup_cost += inner_path->startup_cost;
1461 run_cost += (inner_path->total_cost - inner_path->startup_cost)
1462 * innerscansel * rescanratio;
1468 * If we're doing JOIN_IN then we will stop outputting inner tuples for an
1469 * outer tuple as soon as we have one match. Account for the effects of
1470 * this by scaling down the cost estimates in proportion to the expected
1471 * output size. (This assumes that all the quals attached to the join are
1472 * IN quals, which should be true.)
1474 joininfactor = join_in_selectivity(&path->jpath, root);
1477 * The number of tuple comparisons needed is approximately number of outer
1478 * rows plus number of inner rows plus number of rescanned tuples (can we
1479 * refine this?). At each one, we need to evaluate the mergejoin quals.
1480 * NOTE: JOIN_IN mode does not save any work here, so do NOT include
1483 startup_cost += merge_qual_cost.startup;
1484 run_cost += merge_qual_cost.per_tuple *
1485 (outer_rows + inner_rows * rescanratio);
1488 * For each tuple that gets through the mergejoin proper, we charge
1489 * cpu_tuple_cost plus the cost of evaluating additional restriction
1490 * clauses that are to be applied at the join. (This is pessimistic since
1491 * not all of the quals may get evaluated at each tuple.) This work is
1492 * skipped in JOIN_IN mode, so apply the factor.
1494 startup_cost += qp_qual_cost.startup;
1495 cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
1496 run_cost += cpu_per_tuple * mergejointuples * joininfactor;
1498 path->jpath.path.startup_cost = startup_cost;
1499 path->jpath.path.total_cost = startup_cost + run_cost;
1503 * run mergejoinscansel() with caching
1505 static MergeScanSelCache *
1506 cached_scansel(PlannerInfo *root, RestrictInfo *rinfo, PathKey *pathkey)
1508 MergeScanSelCache *cache;
1510 Selectivity leftscansel,
1512 MemoryContext oldcontext;
1514 /* Do we have this result already? */
1515 foreach(lc, rinfo->scansel_cache)
1517 cache = (MergeScanSelCache *) lfirst(lc);
1518 if (cache->opfamily == pathkey->pk_opfamily &&
1519 cache->strategy == pathkey->pk_strategy &&
1520 cache->nulls_first == pathkey->pk_nulls_first)
1524 /* Nope, do the computation */
1525 mergejoinscansel(root,
1526 (Node *) rinfo->clause,
1527 pathkey->pk_opfamily,
1528 pathkey->pk_strategy,
1529 pathkey->pk_nulls_first,
1533 /* Cache the result in suitably long-lived workspace */
1534 oldcontext = MemoryContextSwitchTo(root->planner_cxt);
1536 cache = (MergeScanSelCache *) palloc(sizeof(MergeScanSelCache));
1537 cache->opfamily = pathkey->pk_opfamily;
1538 cache->strategy = pathkey->pk_strategy;
1539 cache->nulls_first = pathkey->pk_nulls_first;
1540 cache->leftscansel = leftscansel;
1541 cache->rightscansel = rightscansel;
1543 rinfo->scansel_cache = lappend(rinfo->scansel_cache, cache);
1545 MemoryContextSwitchTo(oldcontext);
1552 * Determines and returns the cost of joining two relations using the
1553 * hash join algorithm.
1555 * 'path' is already filled in except for the cost fields
1557 * Note: path's hashclauses should be a subset of the joinrestrictinfo list
1560 cost_hashjoin(HashPath *path, PlannerInfo *root)
1562 Path *outer_path = path->jpath.outerjoinpath;
1563 Path *inner_path = path->jpath.innerjoinpath;
1564 List *hashclauses = path->path_hashclauses;
1565 Cost startup_cost = 0;
1568 Selectivity hash_selec;
1569 QualCost hash_qual_cost;
1570 QualCost qp_qual_cost;
1571 double hashjointuples;
1572 double outer_path_rows = PATH_ROWS(outer_path);
1573 double inner_path_rows = PATH_ROWS(inner_path);
1574 int num_hashclauses = list_length(hashclauses);
1577 double virtualbuckets;
1578 Selectivity innerbucketsize;
1579 Selectivity joininfactor;
1582 if (!enable_hashjoin)
1583 startup_cost += disable_cost;
1586 * Compute cost and selectivity of the hashquals and qpquals (other
1587 * restriction clauses) separately. We use approx_selectivity here for
1588 * speed --- in most cases, any errors won't affect the result much.
1590 * Note: it's probably bogus to use the normal selectivity calculation
1591 * here when either the outer or inner path is a UniquePath.
1593 hash_selec = approx_selectivity(root, hashclauses,
1594 path->jpath.jointype);
1595 cost_qual_eval(&hash_qual_cost, hashclauses, root);
1596 cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
1597 qp_qual_cost.startup -= hash_qual_cost.startup;
1598 qp_qual_cost.per_tuple -= hash_qual_cost.per_tuple;
1600 /* approx # tuples passing the hash quals */
1601 hashjointuples = clamp_row_est(hash_selec * outer_path_rows * inner_path_rows);
1603 /* cost of source data */
1604 startup_cost += outer_path->startup_cost;
1605 run_cost += outer_path->total_cost - outer_path->startup_cost;
1606 startup_cost += inner_path->total_cost;
1609 * Cost of computing hash function: must do it once per input tuple. We
1610 * charge one cpu_operator_cost for each column's hash function. Also,
1611 * tack on one cpu_tuple_cost per inner row, to model the costs of
1612 * inserting the row into the hashtable.
1614 * XXX when a hashclause is more complex than a single operator, we really
1615 * should charge the extra eval costs of the left or right side, as
1616 * appropriate, here. This seems more work than it's worth at the moment.
1618 startup_cost += (cpu_operator_cost * num_hashclauses + cpu_tuple_cost)
1620 run_cost += cpu_operator_cost * num_hashclauses * outer_path_rows;
1622 /* Get hash table size that executor would use for inner relation */
1623 ExecChooseHashTableSize(inner_path_rows,
1624 inner_path->parent->width,
1627 virtualbuckets = (double) numbuckets *(double) numbatches;
1630 * Determine bucketsize fraction for inner relation. We use the smallest
1631 * bucketsize estimated for any individual hashclause; this is undoubtedly
1634 * BUT: if inner relation has been unique-ified, we can assume it's good
1635 * for hashing. This is important both because it's the right answer, and
1636 * because we avoid contaminating the cache with a value that's wrong for
1637 * non-unique-ified paths.
1639 if (IsA(inner_path, UniquePath))
1640 innerbucketsize = 1.0 / virtualbuckets;
1643 innerbucketsize = 1.0;
1644 foreach(hcl, hashclauses)
1646 RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(hcl);
1647 Selectivity thisbucketsize;
1649 Assert(IsA(restrictinfo, RestrictInfo));
1652 * First we have to figure out which side of the hashjoin clause
1653 * is the inner side.
1655 * Since we tend to visit the same clauses over and over when
1656 * planning a large query, we cache the bucketsize estimate in the
1657 * RestrictInfo node to avoid repeated lookups of statistics.
1659 if (bms_is_subset(restrictinfo->right_relids,
1660 inner_path->parent->relids))
1662 /* righthand side is inner */
1663 thisbucketsize = restrictinfo->right_bucketsize;
1664 if (thisbucketsize < 0)
1666 /* not cached yet */
1668 estimate_hash_bucketsize(root,
1669 get_rightop(restrictinfo->clause),
1671 restrictinfo->right_bucketsize = thisbucketsize;
1676 Assert(bms_is_subset(restrictinfo->left_relids,
1677 inner_path->parent->relids));
1678 /* lefthand side is inner */
1679 thisbucketsize = restrictinfo->left_bucketsize;
1680 if (thisbucketsize < 0)
1682 /* not cached yet */
1684 estimate_hash_bucketsize(root,
1685 get_leftop(restrictinfo->clause),
1687 restrictinfo->left_bucketsize = thisbucketsize;
1691 if (innerbucketsize > thisbucketsize)
1692 innerbucketsize = thisbucketsize;
1697 * If inner relation is too big then we will need to "batch" the join,
1698 * which implies writing and reading most of the tuples to disk an extra
1699 * time. Charge seq_page_cost per page, since the I/O should be nice and
1700 * sequential. Writing the inner rel counts as startup cost,
1701 * all the rest as run cost.
1705 double outerpages = page_size(outer_path_rows,
1706 outer_path->parent->width);
1707 double innerpages = page_size(inner_path_rows,
1708 inner_path->parent->width);
1710 startup_cost += seq_page_cost * innerpages;
1711 run_cost += seq_page_cost * (innerpages + 2 * outerpages);
1717 * If we're doing JOIN_IN then we will stop comparing inner tuples to an
1718 * outer tuple as soon as we have one match. Account for the effects of
1719 * this by scaling down the cost estimates in proportion to the expected
1720 * output size. (This assumes that all the quals attached to the join are
1721 * IN quals, which should be true.)
1723 joininfactor = join_in_selectivity(&path->jpath, root);
1726 * The number of tuple comparisons needed is the number of outer tuples
1727 * times the typical number of tuples in a hash bucket, which is the inner
1728 * relation size times its bucketsize fraction. At each one, we need to
1729 * evaluate the hashjoin quals. But actually, charging the full qual eval
1730 * cost at each tuple is pessimistic, since we don't evaluate the quals
1731 * unless the hash values match exactly. For lack of a better idea, halve
1732 * the cost estimate to allow for that.
1734 startup_cost += hash_qual_cost.startup;
1735 run_cost += hash_qual_cost.per_tuple *
1736 outer_path_rows * clamp_row_est(inner_path_rows * innerbucketsize) *
1740 * For each tuple that gets through the hashjoin proper, we charge
1741 * cpu_tuple_cost plus the cost of evaluating additional restriction
1742 * clauses that are to be applied at the join. (This is pessimistic since
1743 * not all of the quals may get evaluated at each tuple.)
1745 startup_cost += qp_qual_cost.startup;
1746 cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
1747 run_cost += cpu_per_tuple * hashjointuples * joininfactor;
1749 path->jpath.path.startup_cost = startup_cost;
1750 path->jpath.path.total_cost = startup_cost + run_cost;
1756 * Estimate the CPU costs of evaluating a WHERE clause.
1757 * The input can be either an implicitly-ANDed list of boolean
1758 * expressions, or a list of RestrictInfo nodes. (The latter is
1759 * preferred since it allows caching of the results.)
1760 * The result includes both a one-time (startup) component,
1761 * and a per-evaluation component.
1764 cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
1766 cost_qual_eval_context context;
1769 context.root = root;
1770 context.total.startup = 0;
1771 context.total.per_tuple = 0;
1773 /* We don't charge any cost for the implicit ANDing at top level ... */
1777 Node *qual = (Node *) lfirst(l);
1779 cost_qual_eval_walker(qual, &context);
1782 *cost = context.total;
1786 * cost_qual_eval_node
1787 * As above, for a single RestrictInfo or expression.
1790 cost_qual_eval_node(QualCost *cost, Node *qual, PlannerInfo *root)
1792 cost_qual_eval_context context;
1794 context.root = root;
1795 context.total.startup = 0;
1796 context.total.per_tuple = 0;
1798 cost_qual_eval_walker(qual, &context);
1800 *cost = context.total;
1804 cost_qual_eval_walker(Node *node, cost_qual_eval_context *context)
1810 * RestrictInfo nodes contain an eval_cost field reserved for this
1811 * routine's use, so that it's not necessary to evaluate the qual
1812 * clause's cost more than once. If the clause's cost hasn't been
1813 * computed yet, the field's startup value will contain -1.
1815 if (IsA(node, RestrictInfo))
1817 RestrictInfo *rinfo = (RestrictInfo *) node;
1819 if (rinfo->eval_cost.startup < 0)
1821 cost_qual_eval_context locContext;
1823 locContext.root = context->root;
1824 locContext.total.startup = 0;
1825 locContext.total.per_tuple = 0;
1827 * For an OR clause, recurse into the marked-up tree so that
1828 * we set the eval_cost for contained RestrictInfos too.
1830 if (rinfo->orclause)
1831 cost_qual_eval_walker((Node *) rinfo->orclause, &locContext);
1833 cost_qual_eval_walker((Node *) rinfo->clause, &locContext);
1835 * If the RestrictInfo is marked pseudoconstant, it will be tested
1836 * only once, so treat its cost as all startup cost.
1838 if (rinfo->pseudoconstant)
1840 /* count one execution during startup */
1841 locContext.total.startup += locContext.total.per_tuple;
1842 locContext.total.per_tuple = 0;
1844 rinfo->eval_cost = locContext.total;
1846 context->total.startup += rinfo->eval_cost.startup;
1847 context->total.per_tuple += rinfo->eval_cost.per_tuple;
1848 /* do NOT recurse into children */
1853 * For each operator or function node in the given tree, we charge the
1854 * estimated execution cost given by pg_proc.procost (remember to
1855 * multiply this by cpu_operator_cost).
1857 * Vars and Consts are charged zero, and so are boolean operators (AND,
1858 * OR, NOT). Simplistic, but a lot better than no model at all.
1860 * Should we try to account for the possibility of short-circuit
1861 * evaluation of AND/OR? Probably *not*, because that would make the
1862 * results depend on the clause ordering, and we are not in any position
1863 * to expect that the current ordering of the clauses is the one that's
1864 * going to end up being used. (Is it worth applying order_qual_clauses
1865 * much earlier in the planning process to fix this?)
1867 if (IsA(node, FuncExpr))
1869 context->total.per_tuple +=
1870 get_func_cost(((FuncExpr *) node)->funcid) * cpu_operator_cost;
1872 else if (IsA(node, OpExpr) ||
1873 IsA(node, DistinctExpr) ||
1874 IsA(node, NullIfExpr))
1876 /* rely on struct equivalence to treat these all alike */
1877 set_opfuncid((OpExpr *) node);
1878 context->total.per_tuple +=
1879 get_func_cost(((OpExpr *) node)->opfuncid) * cpu_operator_cost;
1881 else if (IsA(node, ScalarArrayOpExpr))
1884 * Estimate that the operator will be applied to about half of the
1885 * array elements before the answer is determined.
1887 ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) node;
1888 Node *arraynode = (Node *) lsecond(saop->args);
1890 set_sa_opfuncid(saop);
1891 context->total.per_tuple += get_func_cost(saop->opfuncid) *
1892 cpu_operator_cost * estimate_array_length(arraynode) * 0.5;
1894 else if (IsA(node, RowCompareExpr))
1896 /* Conservatively assume we will check all the columns */
1897 RowCompareExpr *rcexpr = (RowCompareExpr *) node;
1900 foreach(lc, rcexpr->opnos)
1902 Oid opid = lfirst_oid(lc);
1904 context->total.per_tuple += get_func_cost(get_opcode(opid)) *
1908 else if (IsA(node, SubLink))
1910 /* This routine should not be applied to un-planned expressions */
1911 elog(ERROR, "cannot handle unplanned sub-select");
1913 else if (IsA(node, SubPlan))
1916 * A subplan node in an expression typically indicates that the
1917 * subplan will be executed on each evaluation, so charge accordingly.
1918 * (Sub-selects that can be executed as InitPlans have already been
1919 * removed from the expression.)
1921 * An exception occurs when we have decided we can implement the
1922 * subplan by hashing.
1924 SubPlan *subplan = (SubPlan *) node;
1925 Plan *plan = planner_subplan_get_plan(context->root, subplan);
1927 if (subplan->useHashTable)
1930 * If we are using a hash table for the subquery outputs, then the
1931 * cost of evaluating the query is a one-time cost. We charge one
1932 * cpu_operator_cost per tuple for the work of loading the
1935 context->total.startup += plan->total_cost +
1936 cpu_operator_cost * plan->plan_rows;
1939 * The per-tuple costs include the cost of evaluating the lefthand
1940 * expressions, plus the cost of probing the hashtable. Recursion
1941 * into the testexpr will handle the lefthand expressions
1942 * properly, and will count one cpu_operator_cost for each
1943 * comparison operator. That is probably too low for the probing
1944 * cost, but it's hard to make a better estimate, so live with it
1951 * Otherwise we will be rescanning the subplan output on each
1952 * evaluation. We need to estimate how much of the output we will
1953 * actually need to scan. NOTE: this logic should agree with the
1954 * estimates used by make_subplan() in plan/subselect.c.
1956 Cost plan_run_cost = plan->total_cost - plan->startup_cost;
1958 if (subplan->subLinkType == EXISTS_SUBLINK)
1960 /* we only need to fetch 1 tuple */
1961 context->total.per_tuple += plan_run_cost / plan->plan_rows;
1963 else if (subplan->subLinkType == ALL_SUBLINK ||
1964 subplan->subLinkType == ANY_SUBLINK)
1966 /* assume we need 50% of the tuples */
1967 context->total.per_tuple += 0.50 * plan_run_cost;
1968 /* also charge a cpu_operator_cost per row examined */
1969 context->total.per_tuple +=
1970 0.50 * plan->plan_rows * cpu_operator_cost;
1974 /* assume we need all tuples */
1975 context->total.per_tuple += plan_run_cost;
1979 * Also account for subplan's startup cost. If the subplan is
1980 * uncorrelated or undirect correlated, AND its topmost node is a
1981 * Sort or Material node, assume that we'll only need to pay its
1982 * startup cost once; otherwise assume we pay the startup cost
1985 if (subplan->parParam == NIL &&
1987 IsA(plan, Material)))
1988 context->total.startup += plan->startup_cost;
1990 context->total.per_tuple += plan->startup_cost;
1994 /* recurse into children */
1995 return expression_tree_walker(node, cost_qual_eval_walker,
2001 * approx_selectivity
2002 * Quick-and-dirty estimation of clause selectivities.
2003 * The input can be either an implicitly-ANDed list of boolean
2004 * expressions, or a list of RestrictInfo nodes (typically the latter).
2006 * This is quick-and-dirty because we bypass clauselist_selectivity, and
2007 * simply multiply the independent clause selectivities together. Now
2008 * clauselist_selectivity often can't do any better than that anyhow, but
2009 * for some situations (such as range constraints) it is smarter. However,
2010 * we can't effectively cache the results of clauselist_selectivity, whereas
2011 * the individual clause selectivities can be and are cached.
2013 * Since we are only using the results to estimate how many potential
2014 * output tuples are generated and passed through qpqual checking, it
2015 * seems OK to live with the approximation.
2018 approx_selectivity(PlannerInfo *root, List *quals, JoinType jointype)
2020 Selectivity total = 1.0;
2025 Node *qual = (Node *) lfirst(l);
2027 /* Note that clause_selectivity will be able to cache its result */
2028 total *= clause_selectivity(root, qual, 0, jointype);
2035 * set_baserel_size_estimates
2036 * Set the size estimates for the given base relation.
2038 * The rel's targetlist and restrictinfo list must have been constructed
2041 * We set the following fields of the rel node:
2042 * rows: the estimated number of output tuples (after applying
2043 * restriction clauses).
2044 * width: the estimated average output tuple width in bytes.
2045 * baserestrictcost: estimated cost of evaluating baserestrictinfo clauses.
2048 set_baserel_size_estimates(PlannerInfo *root, RelOptInfo *rel)
2052 /* Should only be applied to base relations */
2053 Assert(rel->relid > 0);
2055 nrows = rel->tuples *
2056 clauselist_selectivity(root,
2057 rel->baserestrictinfo,
2061 rel->rows = clamp_row_est(nrows);
2063 cost_qual_eval(&rel->baserestrictcost, rel->baserestrictinfo, root);
2065 set_rel_width(root, rel);
2069 * set_joinrel_size_estimates
2070 * Set the size estimates for the given join relation.
2072 * The rel's targetlist must have been constructed already, and a
2073 * restriction clause list that matches the given component rels must
2076 * Since there is more than one way to make a joinrel for more than two
2077 * base relations, the results we get here could depend on which component
2078 * rel pair is provided. In theory we should get the same answers no matter
2079 * which pair is provided; in practice, since the selectivity estimation
2080 * routines don't handle all cases equally well, we might not. But there's
2081 * not much to be done about it. (Would it make sense to repeat the
2082 * calculations for each pair of input rels that's encountered, and somehow
2083 * average the results? Probably way more trouble than it's worth.)
2085 * It's important that the results for symmetric JoinTypes be symmetric,
2086 * eg, (rel1, rel2, JOIN_LEFT) should produce the same result as (rel2,
2087 * rel1, JOIN_RIGHT). Also, JOIN_IN should produce the same result as
2088 * JOIN_UNIQUE_INNER, likewise JOIN_REVERSE_IN == JOIN_UNIQUE_OUTER.
2090 * We set only the rows field here. The width field was already set by
2091 * build_joinrel_tlist, and baserestrictcost is not used for join rels.
2094 set_joinrel_size_estimates(PlannerInfo *root, RelOptInfo *rel,
2095 RelOptInfo *outer_rel,
2096 RelOptInfo *inner_rel,
2106 * Compute joinclause selectivity. Note that we are only considering
2107 * clauses that become restriction clauses at this join level; we are not
2108 * double-counting them because they were not considered in estimating the
2109 * sizes of the component rels.
2111 * For an outer join, we have to distinguish the selectivity of the
2112 * join's own clauses (JOIN/ON conditions) from any clauses that were
2113 * "pushed down". For inner joins we just count them all as joinclauses.
2115 if (IS_OUTER_JOIN(jointype))
2117 List *joinquals = NIL;
2118 List *pushedquals = NIL;
2121 /* Grovel through the clauses to separate into two lists */
2122 foreach(l, restrictlist)
2124 RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
2126 Assert(IsA(rinfo, RestrictInfo));
2127 if (rinfo->is_pushed_down)
2128 pushedquals = lappend(pushedquals, rinfo);
2130 joinquals = lappend(joinquals, rinfo);
2133 /* Get the separate selectivities */
2134 jselec = clauselist_selectivity(root,
2138 pselec = clauselist_selectivity(root,
2143 /* Avoid leaking a lot of ListCells */
2144 list_free(joinquals);
2145 list_free(pushedquals);
2149 jselec = clauselist_selectivity(root,
2153 pselec = 0.0; /* not used, keep compiler quiet */
2157 * Basically, we multiply size of Cartesian product by selectivity.
2159 * If we are doing an outer join, take that into account: the joinqual
2160 * selectivity has to be clamped using the knowledge that the output must
2161 * be at least as large as the non-nullable input. However, any
2162 * pushed-down quals are applied after the outer join, so their
2163 * selectivity applies fully.
2165 * For JOIN_IN and variants, the Cartesian product is figured with respect
2166 * to a unique-ified input, and then we can clamp to the size of the other
2172 nrows = outer_rel->rows * inner_rel->rows * jselec;
2175 nrows = outer_rel->rows * inner_rel->rows * jselec;
2176 if (nrows < outer_rel->rows)
2177 nrows = outer_rel->rows;
2181 nrows = outer_rel->rows * inner_rel->rows * jselec;
2182 if (nrows < inner_rel->rows)
2183 nrows = inner_rel->rows;
2187 nrows = outer_rel->rows * inner_rel->rows * jselec;
2188 if (nrows < outer_rel->rows)
2189 nrows = outer_rel->rows;
2190 if (nrows < inner_rel->rows)
2191 nrows = inner_rel->rows;
2195 case JOIN_UNIQUE_INNER:
2196 upath = create_unique_path(root, inner_rel,
2197 inner_rel->cheapest_total_path);
2198 nrows = outer_rel->rows * upath->rows * jselec;
2199 if (nrows > outer_rel->rows)
2200 nrows = outer_rel->rows;
2202 case JOIN_REVERSE_IN:
2203 case JOIN_UNIQUE_OUTER:
2204 upath = create_unique_path(root, outer_rel,
2205 outer_rel->cheapest_total_path);
2206 nrows = upath->rows * inner_rel->rows * jselec;
2207 if (nrows > inner_rel->rows)
2208 nrows = inner_rel->rows;
2211 elog(ERROR, "unrecognized join type: %d", (int) jointype);
2212 nrows = 0; /* keep compiler quiet */
2216 rel->rows = clamp_row_est(nrows);
2220 * join_in_selectivity
2221 * Determines the factor by which a JOIN_IN join's result is expected
2222 * to be smaller than an ordinary inner join.
2224 * 'path' is already filled in except for the cost fields
2227 join_in_selectivity(JoinPath *path, PlannerInfo *root)
2229 RelOptInfo *innerrel;
2230 UniquePath *innerunique;
2234 /* Return 1.0 whenever it's not JOIN_IN */
2235 if (path->jointype != JOIN_IN)
2239 * Return 1.0 if the inner side is already known unique. The case where
2240 * the inner path is already a UniquePath probably cannot happen in
2241 * current usage, but check it anyway for completeness. The interesting
2242 * case is where we've determined the inner relation itself is unique,
2243 * which we can check by looking at the rows estimate for its UniquePath.
2245 if (IsA(path->innerjoinpath, UniquePath))
2247 innerrel = path->innerjoinpath->parent;
2248 innerunique = create_unique_path(root,
2250 innerrel->cheapest_total_path);
2251 if (innerunique->rows >= innerrel->rows)
2255 * Compute same result set_joinrel_size_estimates would compute for
2256 * JOIN_INNER. Note that we use the input rels' absolute size estimates,
2257 * not PATH_ROWS() which might be less; if we used PATH_ROWS() we'd be
2258 * double-counting the effects of any join clauses used in input scans.
2260 selec = clauselist_selectivity(root,
2261 path->joinrestrictinfo,
2264 nrows = path->outerjoinpath->parent->rows * innerrel->rows * selec;
2266 nrows = clamp_row_est(nrows);
2268 /* See if it's larger than the actual JOIN_IN size estimate */
2269 if (nrows > path->path.parent->rows)
2270 return path->path.parent->rows / nrows;
2276 * set_function_size_estimates
2277 * Set the size estimates for a base relation that is a function call.
2279 * The rel's targetlist and restrictinfo list must have been constructed
2282 * We set the same fields as set_baserel_size_estimates.
2285 set_function_size_estimates(PlannerInfo *root, RelOptInfo *rel)
2289 /* Should only be applied to base relations that are functions */
2290 Assert(rel->relid > 0);
2291 rte = rt_fetch(rel->relid, root->parse->rtable);
2292 Assert(rte->rtekind == RTE_FUNCTION);
2294 /* Estimate number of rows the function itself will return */
2295 rel->tuples = clamp_row_est(expression_returns_set_rows(rte->funcexpr));
2297 /* Now estimate number of output rows, etc */
2298 set_baserel_size_estimates(root, rel);
2302 * set_values_size_estimates
2303 * Set the size estimates for a base relation that is a values list.
2305 * The rel's targetlist and restrictinfo list must have been constructed
2308 * We set the same fields as set_baserel_size_estimates.
2311 set_values_size_estimates(PlannerInfo *root, RelOptInfo *rel)
2315 /* Should only be applied to base relations that are values lists */
2316 Assert(rel->relid > 0);
2317 rte = rt_fetch(rel->relid, root->parse->rtable);
2318 Assert(rte->rtekind == RTE_VALUES);
2321 * Estimate number of rows the values list will return. We know this
2322 * precisely based on the list length (well, barring set-returning
2323 * functions in list items, but that's a refinement not catered for
2324 * anywhere else either).
2326 rel->tuples = list_length(rte->values_lists);
2328 /* Now estimate number of output rows, etc */
2329 set_baserel_size_estimates(root, rel);
2335 * Set the estimated output width of a base relation.
2337 * NB: this works best on plain relations because it prefers to look at
2338 * real Vars. It will fail to make use of pg_statistic info when applied
2339 * to a subquery relation, even if the subquery outputs are simple vars
2340 * that we could have gotten info for. Is it worth trying to be smarter
2343 * The per-attribute width estimates are cached for possible re-use while
2344 * building join relations.
2347 set_rel_width(PlannerInfo *root, RelOptInfo *rel)
2349 int32 tuple_width = 0;
2352 foreach(tllist, rel->reltargetlist)
2354 Var *var = (Var *) lfirst(tllist);
2359 /* For now, punt on whole-row child Vars */
2362 tuple_width += 32; /* arbitrary */
2366 ndx = var->varattno - rel->min_attr;
2369 * The width probably hasn't been cached yet, but may as well check
2371 if (rel->attr_widths[ndx] > 0)
2373 tuple_width += rel->attr_widths[ndx];
2377 relid = getrelid(var->varno, root->parse->rtable);
2378 if (relid != InvalidOid)
2380 item_width = get_attavgwidth(relid, var->varattno);
2383 rel->attr_widths[ndx] = item_width;
2384 tuple_width += item_width;
2390 * Not a plain relation, or can't find statistics for it. Estimate
2391 * using just the type info.
2393 item_width = get_typavgwidth(var->vartype, var->vartypmod);
2394 Assert(item_width > 0);
2395 rel->attr_widths[ndx] = item_width;
2396 tuple_width += item_width;
2398 Assert(tuple_width >= 0);
2399 rel->width = tuple_width;
2403 * relation_byte_size
2404 * Estimate the storage space in bytes for a given number of tuples
2405 * of a given width (size in bytes).
2408 relation_byte_size(double tuples, int width)
2410 return tuples * (MAXALIGN(width) + MAXALIGN(sizeof(HeapTupleHeaderData)));
2415 * Returns an estimate of the number of pages covered by a given
2416 * number of tuples of a given width (size in bytes).
2419 page_size(double tuples, int width)
2421 return ceil(relation_byte_size(tuples, width) / BLCKSZ);