| blob | ff689b65245fcd2605b79ec7746bc296d424c5f7 |
1 <!-- doc/src/sgml/perform.sgml -->
8 </indexterm>
10 <para>
11 Query performance can be affected by many things. Some of these can
12 be controlled by the user, while others are fundamental to the underlying
13 design of the system. This chapter provides some hints about understanding
15 </para>
22 </indexterm>
26 </indexterm>
28 <para>
30 plan</firstterm> for each query it receives. Choosing the right
31 plan to match the query structure and the properties of the data
32 is absolutely critical for good performance, so the system includes
35 to see what query plan the planner creates for any query.
36 Plan-reading is an art that requires some experience to master,
37 but this section attempts to cover the basics.
38 </para>
40 <para>
41 Examples in this section are drawn from the regression test database
43 You should be able to get similar results if you try the examples
44 yourself, but your estimated costs and row counts might vary slightly
46 than exact, and because costs are inherently somewhat platform-dependent.
47 </para>
49 <para>
51 format, which is compact and convenient for humans to read.
53 analysis, you should use one of its machine-readable output formats
54 (XML, JSON, or YAML) instead.
55 </para>
60 <para>
62 Nodes at the bottom level of the tree are scan nodes: they return raw rows
63 from a table. There are different types of scan nodes for different
64 table access methods: sequential scans, index scans, and bitmap index
67 own scan node types.
68 If the query requires joining, aggregation, sorting, or other
69 operations on the raw rows, then there will be additional nodes
70 above the scan nodes to perform these operations. Again,
71 there is usually more than one possible way to do these operations,
72 so different node types can appear here too. The output
74 tree, showing the basic node type plus the cost estimates that the planner
75 made for the execution of that plan node. Additional lines might appear,
76 indented from the node's summary line,
77 to show additional properties of the node.
78 The very first line (the summary line for the topmost
79 node) has the estimated total execution cost for the plan; it is this
80 number that the planner seeks to minimize.
81 </para>
83 <para>
84 Here is a trivial example, just to show what the output looks like:
86 <screen>
87 EXPLAIN SELECT * FROM tenk1;
89 QUERY PLAN
90 -------------------------------------------------------------
92 </screen>
93 </para>
95 <para>
97 rows of the table, so the planner has chosen to use a simple sequential
98 scan plan. The numbers that are quoted in parentheses are (left
99 to right):
101 <itemizedlist>
102 <listitem>
103 <para>
104 Estimated start-up cost. This is the time expended before the output
105 phase can begin, e.g., time to do the sorting in a sort node.
106 </para>
107 </listitem>
109 <listitem>
110 <para>
111 Estimated total cost. This is stated on the assumption that the plan
112 node is run to completion, i.e., all available rows are retrieved.
113 In practice a node's parent node might stop short of reading all
115 </para>
116 </listitem>
118 <listitem>
119 <para>
120 Estimated number of rows output by this plan node. Again, the node
121 is assumed to be run to completion.
122 </para>
123 </listitem>
125 <listitem>
126 <para>
127 Estimated average width of rows output by this plan node (in bytes).
128 </para>
129 </listitem>
130 </itemizedlist>
131 </para>
133 <para>
134 The costs are measured in arbitrary units determined by the planner's
136 Traditional practice is to measure the costs in units of disk page
139 to that. The examples in this section are run with the default cost
140 parameters.
141 </para>
143 <para>
144 It's important to understand that the cost of an upper-level node includes
145 the cost of all its child nodes. It's also important to realize that
146 the cost only reflects things that the planner cares about.
147 In particular, the cost does not consider the time spent to convert
148 output values to text form or to transmit them to the client, which
149 could be important factors in the real elapsed time; but the planner
150 ignores those costs because it cannot change them by altering the
151 plan. (Every correct plan will output the same row set, we trust.)
152 </para>
154 <para>
156 not the number of rows processed or scanned by the
157 plan node, but rather the number emitted by the node. This is often
158 less than the number scanned, as a result of filtering by any
160 Ideally the top-level rows estimate will approximate the number of rows
161 actually returned, updated, or deleted by the query.
162 </para>
164 <para>
165 Returning to our example:
167 <screen>
168 EXPLAIN SELECT * FROM tenk1;
170 QUERY PLAN
171 -------------------------------------------------------------
173 </screen>
174 </para>
176 <para>
177 These numbers are derived very straightforwardly. If you do:
179 <programlisting>
180 SELECT relpages, reltuples FROM pg_class WHERE relname = 'tenk1';
181 </programlisting>
184 pages and 10000 rows. The estimated cost is computed as (disk pages read *
189 </para>
191 <para>
194 <screen>
197 QUERY PLAN
198 ------------------------------------------------------------
201 </screen>
205 Scan plan node. This means that
206 the plan node checks the condition for each row it scans, and outputs
207 only the ones that pass the condition.
208 The estimate of output rows has been reduced because of the
210 However, the scan will still have to visit all 10000 rows, so the cost
214 </para>
216 <para>
218 estimate is only approximate. If you try to duplicate this experiment,
219 you may well get a slightly different estimate; moreover, it can
222 randomized sample of the table.
223 </para>
225 <para>
226 Now, let's make the condition more restrictive:
228 <screen>
231 QUERY PLAN
232 -------------------------------------------------------------------&zwsp;-----------
237 </screen>
239 Here the planner has decided to use a two-step plan: the child plan
240 node visits an index to find the locations of rows matching the index
241 condition, and then the upper plan node actually fetches those rows
242 from the table itself. Fetching rows separately is much more
243 expensive than reading them sequentially, but because not all the pages
244 of the table have to be visited, this is still cheaper than a sequential
245 scan. (The reason for using two plan levels is that the upper plan
246 node sorts the row locations identified by the index into physical order
247 before reading them, to minimize the cost of separate fetches.
249 does the sorting.)
250 </para>
252 <para>
255 <screen>
258 QUERY PLAN
259 -------------------------------------------------------------------&zwsp;-----------
262 Filter: (stringu1 = 'xxx'::name)
265 </screen>
268 output row count estimate, but not the cost because we still have to visit
270 cannot be applied as an index condition, since this index is only on
272 the rows retrieved using the index. Thus the cost has actually gone up
273 slightly to reflect this extra checking.
274 </para>
276 <para>
279 <screen>
280 EXPLAIN SELECT * FROM tenk1 WHERE unique1 = 42;
282 QUERY PLAN
283 -------------------------------------------------------------------&zwsp;----------
285 Index Cond: (unique1 = 42)
286 </screen>
288 In this type of plan the table rows are fetched in index order, which
289 makes them even more expensive to read, but there are so few that the
290 extra cost of sorting the row locations is not worth it. You'll most
291 often see this plan type for queries that fetch just a single row. It's
293 that matches the index order, because then no extra sorting step is needed
296 index already implicitly provides the requested ordering.
297 </para>
299 <para>
301 ways. The above example shows that such an ordering clause may be
302 implemented implicitly. The planner may also add an explicit
305 <screen>
306 EXPLAIN SELECT * FROM tenk1 ORDER BY unique1;
308 QUERY PLAN
309 -------------------------------------------------------------------
311 Sort Key: unique1
313 </screen>
315 If a part of the plan guarantees an ordering on a prefix of the
316 required sort keys, then the planner may instead decide to use an
319 <screen>
320 EXPLAIN SELECT * FROM tenk1 ORDER BY hundred, ten LIMIT 100;
322 QUERY PLAN
323 -------------------------------------------------------------------&zwsp;-----------------------------
326 Sort Key: hundred, ten
327 Presorted Key: hundred
329 </screen>
331 Compared to regular sorts, sorting incrementally allows returning tuples
332 before the entire result set has been sorted, which particularly enables
334 memory usage and the likelihood of spilling sorts to disk, but it comes at
335 the cost of the increased overhead of splitting the result set into multiple
336 sorting batches.
337 </para>
339 <para>
340 If there are separate indexes on several of the columns referenced
342 combination of the indexes:
344 <screen>
347 QUERY PLAN
348 -------------------------------------------------------------------&zwsp;------------------
356 </screen>
358 But this requires visiting both indexes, so it's not necessarily a win
359 compared to using just one index and treating the other condition as
360 a filter. If you vary the ranges involved you'll see the plan change
361 accordingly.
362 </para>
364 <para>
367 <screen>
370 QUERY PLAN
371 -------------------------------------------------------------------&zwsp;------------------
376 </screen>
377 </para>
379 <para>
381 not all the rows need be retrieved, and the planner changed its mind about
382 what to do. Notice that the total cost and row count of the Index Scan
383 node are shown as if it were run to completion. However, the Limit node
384 is expected to stop after retrieving only a fifth of those rows, so its
385 total cost is only a fifth as much, and that's the actual estimated cost
386 of the query. This plan is preferred over adding a Limit node to the
387 previous plan because the Limit could not avoid paying the startup cost
388 of the bitmap scan, so the total cost would be something over 25 units
389 with that approach.
390 </para>
392 <para>
393 Let's try joining two tables, using the columns we have been discussing:
395 <screen>
396 EXPLAIN SELECT *
397 FROM tenk1 t1, tenk2 t2
400 QUERY PLAN
401 -------------------------------------------------------------------&zwsp;-------------------
408 Index Cond: (unique2 = t1.unique2)
409 </screen>
410 </para>
412 <para>
413 In this plan, we have a nested-loop join node with two table scans as
414 inputs, or children. The indentation of the node summary lines reflects
416 is a bitmap scan similar to those we saw before. Its cost and row count
418 because we are
420 at that node.
422 so it doesn't affect the row count of the outer scan. The nested-loop
423 join node will run its second,
425 Column values from the current outer row can be plugged into the inner
427 so we get a plan and costs similar to what we saw above for a simple
429 (The estimated cost is actually a bit lower than what was seen above,
430 as a result of caching that's expected to occur during the repeated
432 costs of the loop node are then set on the basis of the cost of the outer
434 here), plus a little CPU time for join processing.
435 </para>
437 <para>
438 In this example the join's output row count is the same as the product
439 of the two scans' row counts, but that's not true in all cases because
441 and so can only be applied at the join point, not to either input scan.
442 Here's an example:
444 <screen>
445 EXPLAIN SELECT *
446 FROM tenk1 t1, tenk2 t2
449 QUERY PLAN
450 -------------------------------------------------------------------&zwsp;--------------------------
452 Join Filter: (t1.hundred < t2.hundred)
460 </screen>
464 join node. This reduces the estimated output row count of the join node,
465 but does not change either input scan.
466 </para>
468 <para>
470 relation of the join, by putting a Materialize plan node atop it. This
472 though the nested-loop join node needs to read that data ten times, once
473 for each row from the outer relation. The Materialize node saves the data
474 in memory as it's read, and then returns the data from memory on each
475 subsequent pass.
476 </para>
478 <para>
479 When dealing with outer joins, you might see join plan nodes with both
482 so a row that fails the Join Filter condition could still get emitted as
483 a null-extended row. But a plain Filter condition is applied after the
484 outer-join rules and so acts to remove rows unconditionally. In an inner
485 join there is no semantic difference between these types of filters.
486 </para>
488 <para>
489 If we change the query's selectivity a bit, we might get a very different
490 join plan:
492 <screen>
493 EXPLAIN SELECT *
494 FROM tenk1 t1, tenk2 t2
497 QUERY PLAN
498 -------------------------------------------------------------------&zwsp;-----------------------
500 Hash Cond: (t2.unique2 = t1.unique2)
507 </screen>
508 </para>
510 <para>
511 Here, the planner has chosen to use a hash join, in which rows of one
512 table are entered into an in-memory hash table, after which the other
513 table is scanned and the hash table is probed for matches to each row.
514 Again note how the indentation reflects the plan structure: the bitmap
516 the hash table. That's then returned to the Hash Join node, which reads
517 rows from its outer child plan and searches the hash table for each one.
518 </para>
520 <para>
521 Another possible type of join is a merge join, illustrated here:
523 <screen>
524 EXPLAIN SELECT *
525 FROM tenk1 t1, onek t2
528 QUERY PLAN
529 -------------------------------------------------------------------&zwsp;-----------------------
531 Merge Cond: (t1.unique2 = t2.unique2)
535 </screen>
536 </para>
538 <para>
539 Merge join requires its input data to be sorted on the join keys. In this
540 example each input is sorted by using an index scan to visit the rows
541 in the correct order; but a sequential scan and sort could also be used.
542 (Sequential-scan-and-sort frequently beats an index scan for sorting many rows,
543 because of the nonsequential disk access required by the index scan.)
544 </para>
546 <para>
547 One way to look at variant plans is to force the planner to disregard
548 whatever strategy it thought was the cheapest, using the enable/disable
550 (This is a crude tool, but useful. See
552 For example, if we're unconvinced that merge join is the best join
553 type for the previous example, we could try
555 <screen>
556 SET enable_mergejoin = off;
558 EXPLAIN SELECT *
559 FROM tenk1 t1, onek t2
562 QUERY PLAN
563 -------------------------------------------------------------------&zwsp;-----------------------
565 Hash Cond: (t2.unique2 = t1.unique2)
572 </screen>
574 which shows that the planner thinks that hash join would be nearly 50%
575 more expensive than merge join for this case.
576 Of course, the next question is whether it's right about that.
579 </para>
581 <para>
582 <indexterm>
584 </indexterm>
587 queries can sometimes be transformed into ordinary join plans, but
588 when they cannot be, we get plans like:
590 <screen>
591 EXPLAIN VERBOSE SELECT unique1
592 FROM tenk1 t
593 WHERE t.ten < ALL (SELECT o.ten FROM onek o WHERE o.four = t.four);
595 QUERY PLAN
596 -------------------------------------------------------------------&zwsp;------
598 Output: t.unique1
600 SubPlan 1
602 Output: o.ten
603 Filter: (o.four = t.four)
604 </screen>
606 This rather artificial example serves to illustrate a couple of
607 points: values from the outer plan level can be passed down into a
609 results of the sub-select are available to the outer plan. Those
611 like
615 </para>
617 <para>
618 <indexterm>
621 </indexterm>
623 subplan again for each row of the outer query (which accounts for the
624 high estimated cost). Some queries can use a <firstterm>hashed
625 subplan</firstterm> to avoid that:
627 <screen>
628 EXPLAIN SELECT *
629 FROM tenk1 t
630 WHERE t.unique1 NOT IN (SELECT o.unique1 FROM onek o);
632 QUERY PLAN
633 -------------------------------------------------------------------&zwsp;-------------------------
635 Filter: (NOT (ANY (unique1 = (hashed SubPlan 1).col1)))
636 SubPlan 1
638 (4 rows)
639 </screen>
641 Here, the subplan is run a single time and its output is loaded into
642 an in-memory hash table, which is then probed by the
646 amenable to hashing.
647 </para>
649 <para>
650 <indexterm>
652 </indexterm>
653 If, in addition to not referencing any variables of the outer query,
657 <screen>
658 EXPLAIN VERBOSE SELECT unique1
659 FROM tenk1 t1 WHERE t1.ten = (SELECT (random() * 10)::integer);
661 QUERY PLAN
662 ------------------------------------------------------------&zwsp;--------
664 Output: t1.unique1
665 Filter: (t1.ten = (InitPlan 1).col1)
666 InitPlan 1
668 Output: ((random() * '10'::double precision))::integer
669 </screen>
671 An initplan is run only once per execution of the outer plan, and its
672 results are saved for re-use in later rows of the outer plan. So in
675 randomly-chosen integer. That's quite different from what would
677 </para>
679 </sect2>
684 <para>
685 It is possible to check the accuracy of the planner's estimates
688 the true row counts and true run time accumulated within each plan node,
690 shows. For example, we might get a result like this:
692 <screen>
693 EXPLAIN ANALYZE SELECT *
694 FROM tenk1 t1, tenk2 t2
697 QUERY PLAN
698 -------------------------------------------------------------------&zwsp;--------------------------------------------------------------
700 -> Bitmap Heap Scan on tenk1 t1 (cost=4.36..39.38 rows=10 width=244) (actual time=0.009..0.017 rows=10 loops=1)
702 Heap Blocks: exact=10
703 -> Bitmap Index Scan on tenk1_unique1 (cost=0.00..4.36 rows=10 width=0) (actual time=0.004..0.004 rows=10 loops=1)
705 -> Index Scan using tenk2_unique2 on tenk2 t2 (cost=0.29..7.90 rows=1 width=244) (actual time=0.003..0.003 rows=1 loops=10)
706 Index Cond: (unique2 = t1.unique2)
707 Planning Time: 0.485 ms
708 Execution Time: 0.073 ms
709 </screen>
713 arbitrary units; so they are unlikely to match up.
714 The thing that's usually most important to look for is whether the
715 estimated row counts are reasonably close to reality. In this example
716 the estimates were all dead-on, but that's quite unusual in practice.
717 </para>
719 <para>
720 In some query plans, it is possible for a subplan node to be executed more
721 than once. For example, the inner index scan will be executed once per
722 outer row in the above nested-loop plan. In such cases, the
724 total number of executions of the node, and the actual time and rows
725 values shown are averages per-execution. This is done to make the numbers
726 comparable with the way that the cost estimates are shown. Multiply by
728 the node. In the above example, we spent a total of 0.030 milliseconds
730 </para>
732 <para>
734 statistics beyond the plan node execution times and row counts.
735 For example, Sort and Hash nodes provide extra information:
737 <screen>
738 EXPLAIN ANALYZE SELECT *
739 FROM tenk1 t1, tenk2 t2
742 QUERY PLAN
743 -------------------------------------------------------------------&zwsp;-------------------------------------------------------------------&zwsp;------
745 Sort Key: t1.fivethous
746 Sort Method: quicksort Memory: 74kB
747 -> Hash Join (cost=226.23..709.73 rows=100 width=488) (actual time=0.515..2.920 rows=100 loops=1)
748 Hash Cond: (t2.unique2 = t1.unique2)
749 -> Seq Scan on tenk2 t2 (cost=0.00..445.00 rows=10000 width=244) (actual time=0.026..1.790 rows=10000 loops=1)
752 -> Bitmap Heap Scan on tenk1 t1 (cost=5.06..224.98 rows=100 width=244) (actual time=0.030..0.450 rows=100 loops=1)
754 Heap Blocks: exact=90
755 -> Bitmap Index Scan on tenk1_unique1 (cost=0.00..5.04 rows=100 width=0) (actual time=0.013..0.013 rows=100 loops=1)
757 Planning Time: 0.187 ms
758 Execution Time: 3.036 ms
759 </screen>
761 The Sort node shows the sort method used (in particular, whether the sort
762 was in-memory or on-disk) and the amount of memory or disk space needed.
763 The Hash node shows the number of hash buckets and batches as well as the
764 peak amount of memory used for the hash table. (If the number of batches
765 exceeds one, there will also be disk space usage involved, but that is not
766 shown.)
767 </para>
769 <para>
770 Another type of extra information is the number of rows removed by a
771 filter condition:
773 <screen>
776 QUERY PLAN
777 -------------------------------------------------------------------&zwsp;--------------------------------------
778 Seq Scan on tenk1 (cost=0.00..470.00 rows=7000 width=244) (actual time=0.030..1.995 rows=7000 loops=1)
780 Rows Removed by Filter: 3000
781 Planning Time: 0.102 ms
782 Execution Time: 2.145 ms
783 </screen>
785 These counts can be particularly valuable for filter conditions applied at
787 one scanned row, or potential join pair in the case of a join node,
788 is rejected by the filter condition.
789 </para>
791 <para>
793 index scans. For example, consider this search for polygons containing a
794 specific point:
796 <screen>
799 QUERY PLAN
800 -------------------------------------------------------------------&zwsp;-----------------------------------
801 Seq Scan on polygon_tbl (cost=0.00..1.09 rows=1 width=85) (actual time=0.023..0.023 rows=0 loops=1)
803 Rows Removed by Filter: 7
804 Planning Time: 0.039 ms
805 Execution Time: 0.033 ms
806 </screen>
808 The planner thinks (quite correctly) that this sample table is too small
809 to bother with an index scan, so we have a plain sequential scan in which
810 all the rows got rejected by the filter condition. But if we force an
811 index scan to be used, we see:
813 <screen>
814 SET enable_seqscan TO off;
818 QUERY PLAN
819 -------------------------------------------------------------------&zwsp;-------------------------------------------------------
820 Index Scan using gpolygonind on polygon_tbl (cost=0.13..8.15 rows=1 width=85) (actual time=0.074..0.074 rows=0 loops=1)
822 Rows Removed by Index Recheck: 1
823 Planning Time: 0.039 ms
824 Execution Time: 0.098 ms
825 </screen>
827 Here we can see that the index returned one candidate row, which was
828 then rejected by a recheck of the index condition. This happens because a
830 returns the rows with polygons that overlap the target, and then we have
831 to do the exact containment test on those rows.
832 </para>
834 <para>
838 <screen>
841 QUERY PLAN
842 -------------------------------------------------------------------&zwsp;--------------------------------------------------------------
843 Bitmap Heap Scan on tenk1 (cost=25.07..60.11 rows=10 width=244) (actual time=0.105..0.114 rows=10 loops=1)
845 Heap Blocks: exact=10
849 -> Bitmap Index Scan on tenk1_unique1 (cost=0.00..5.04 rows=100 width=0) (actual time=0.027..0.027 rows=100 loops=1)
851 Buffers: shared hit=2
852 -> Bitmap Index Scan on tenk1_unique2 (cost=0.00..19.78 rows=999 width=0) (actual time=0.070..0.070 rows=999 loops=1)
855 Planning:
856 Buffers: shared hit=3
857 Planning Time: 0.162 ms
858 Execution Time: 0.143 ms
859 </screen>
862 of the query are the most I/O-intensive.
863 </para>
865 <para>
867 runs the query, any side-effects will happen as usual, even though
868 whatever results the query might output are discarded in favor of
870 data-modifying query without changing your tables, you can
871 roll the command back afterwards, for example:
873 <screen>
874 BEGIN;
878 QUERY PLAN
879 -------------------------------------------------------------------&zwsp;-------------------------------------------------------------
881 -> Bitmap Heap Scan on tenk1 (cost=5.06..225.23 rows=100 width=10) (actual time=0.065..0.141 rows=100 loops=1)
883 Heap Blocks: exact=90
884 -> Bitmap Index Scan on tenk1_unique1 (cost=0.00..5.04 rows=100 width=0) (actual time=0.031..0.031 rows=100 loops=1)
886 Planning Time: 0.151 ms
887 Execution Time: 1.856 ms
889 ROLLBACK;
890 </screen>
891 </para>
893 <para>
897 applying the table changes is done by a top-level Insert, Update,
898 Delete, or Merge plan node. The plan nodes underneath this node perform
899 the work of locating the old rows and/or computing the new data.
900 So above, we see the same sort of bitmap table scan we've seen already,
901 and its output is fed to an Update node that stores the updated rows.
902 It's worth noting that although the data-modifying node can take a
903 considerable amount of run time (here, it's consuming the lion's share
904 of the time), the planner does not currently add anything to the cost
905 estimates to account for that work. That's because the work to be done is
906 the same for every correct query plan, so it doesn't affect planning
907 decisions.
908 </para>
910 <para>
913 inheritance hierarchy, the output might look like this:
915 <screen>
916 EXPLAIN UPDATE gtest_parent SET f1 = CURRENT_DATE WHERE f2 = 101;
918 QUERY PLAN
919 -------------------------------------------------------------------&zwsp;---------------------
921 Update on gtest_child gtest_parent_1
922 Update on gtest_child2 gtest_parent_2
923 Update on gtest_child3 gtest_parent_3
926 Filter: (f2 = 101)
928 Filter: (f2 = 101)
930 Filter: (f2 = 101)
931 </screen>
933 In this example the Update node needs to consider three child tables,
934 but not the originally-mentioned partitioned table (since that never
935 stores any data). So there are three input
936 scanning subplans, one per table. For clarity, the Update node is
937 annotated to show the specific target tables that will be updated, in the
938 same order as the corresponding subplans.
939 </para>
941 <para>
943 ANALYZE</command> is the time it took to generate the query plan from the
944 parsed query and optimize it. It does not include parsing or rewriting.
945 </para>
947 <para>
949 ANALYZE</command> includes executor start-up and shut-down time, as well
950 as the time to run any triggers that are fired, but it does not include
951 parsing, rewriting, or planning time.
953 the time for the related Insert, Update, or Delete node; but time
956 The total time spent in each trigger
958 Note that deferred constraint triggers will not be executed
959 until end of transaction and are thus not considered at all by
961 </para>
963 <para>
964 The time shown for the top-level node does not include any time needed
965 to convert the query's output data into displayable form or to send it
967 send the data to the client, it can be told to convert the query's
968 output data to displayable form and measure the time needed for that,
970 be shown separately, and it's also included in the
972 </para>
974 </sect2>
979 <para>
980 There are two significant ways in which run times measured by
982 the same query. First, since no output rows are delivered to the client,
983 network transmission costs are not included. I/O conversion costs are
985 Second, the measurement overhead added by <command>EXPLAIN
986 ANALYZE</command> can be significant, especially on machines with slow
989 on your system.
990 </para>
992 <para>
994 much different from the one you are actually testing; for example,
995 results on a toy-sized table cannot be assumed to apply to large tables.
996 The planner's cost estimates are not linear and so it might choose
997 a different plan for a larger or smaller table. An extreme example
998 is that on a table that only occupies one disk page, you'll nearly
999 always get a sequential scan plan whether indexes are available or not.
1000 The planner realizes that it's going to take one disk page read to
1001 process the table in any case, so there's no value in expending additional
1002 page reads to look at an index. (We saw this happening in the
1004 </para>
1006 <para>
1007 There are cases in which the actual and estimated values won't match up
1008 well, but nothing is really wrong. One such case occurs when
1012 <screen>
1015 QUERY PLAN
1016 -------------------------------------------------------------------&zwsp;------------------------------------------------------------
1018 -> Index Scan using tenk1_unique2 on tenk1 (cost=0.29..70.50 rows=10 width=244) (actual time=0.051..0.070 rows=2 loops=1)
1021 Rows Removed by Filter: 287
1022 Planning Time: 0.077 ms
1023 Execution Time: 0.086 ms
1024 </screen>
1026 the estimated cost and row count for the Index Scan node are shown as
1027 though it were run to completion. But in reality the Limit node stopped
1028 requesting rows after it got two, so the actual row count is only 2 and
1029 the run time is less than the cost estimate would suggest. This is not
1030 an estimation error, only a discrepancy in the way the estimates and true
1031 values are displayed.
1032 </para>
1034 <para>
1035 Merge joins also have measurement artifacts that can confuse the unwary.
1036 A merge join will stop reading one input if it's exhausted the other input
1037 and the next key value in the one input is greater than the last key value
1038 of the other input; in such a case there can be no more matches and so no
1039 need to scan the rest of the first input. This results in not reading all
1041 Also, if the outer (first) child contains rows with duplicate key values,
1042 the inner (second) child is backed up and rescanned for the portion of its
1044 repeated emissions of the same inner rows as if they were real additional
1045 rows. When there are many outer duplicates, the reported actual row count
1046 for the inner child plan node can be significantly larger than the number
1047 of rows that are actually in the inner relation.
1048 </para>
1050 <para>
1051 BitmapAnd and BitmapOr nodes always report their actual row counts as zero,
1052 due to implementation limitations.
1053 </para>
1055 <para>
1057 created by the planner. However, there are cases where the executor
1058 can determine that certain nodes need not be executed because they
1059 cannot produce any rows, based on parameter values that were not
1060 available at planning time. (Currently this can only happen for child
1061 nodes of an Append or MergeAppend node that is scanning a partitioned
1062 table.) When this happens, those plan nodes are omitted from
1065 instead.
1066 </para>
1067 </sect2>
1069 </sect1>
1077 </indexterm>
1081 <para>
1082 As we saw in the previous section, the query planner needs to estimate
1083 the number of rows retrieved by a query in order to make good choices
1084 of query plans. This section provides a quick look at the statistics
1085 that the system uses for these estimates.
1086 </para>
1088 <para>
1089 One component of the statistics is the total number of entries in
1090 each table and index, as well as the number of disk blocks occupied
1091 by each table and index. This information is kept in the table
1095 queries similar to this one:
1097 <screen>
1098 SELECT relname, relkind, reltuples, relpages
1099 FROM pg_class
1100 WHERE relname LIKE 'tenk1%';
1102 relname | relkind | reltuples | relpages
1103 ----------------------+---------+-----------+----------
1109 (5 rows)
1110 </screen>
1113 rows, as do its indexes, but the indexes are (unsurprisingly) much
1114 smaller than the table.
1115 </para>
1117 <para>
1120 and so they usually contain somewhat out-of-date values.
1124 (which is commonly the case) will incrementally update the
1126 of the table it did scan, resulting in an approximate value.
1127 In any case, the planner
1129 to match the current physical table size, thus obtaining a closer
1130 approximation.
1131 </para>
1133 <indexterm>
1135 </indexterm>
1137 <para>
1138 Most queries retrieve only a fraction of the rows in a table, due
1140 examined. The planner thus needs to make an estimate of the
1142 the fraction of rows that match each condition in the
1144 stored in the
1148 ANALYZE</command> commands, and are always approximate even when freshly
1149 updated.
1150 </para>
1152 <indexterm>
1154 </indexterm>
1156 <para>
1158 it's better to look at its view
1161 is designed to be more easily readable. Furthermore,
1164 (This prevents unprivileged users from learning something about
1165 the contents of other people's tables from the statistics. The
1167 rows about tables that the current user can read.)
1168 For example, we might do:
1170 <screen>
1171 SELECT attname, inherited, n_distinct,
1172 array_to_string(most_common_vals, E'\n') as most_common_vals
1173 FROM pg_stats
1174 WHERE tablename = 'road';
1176 attname | inherited | n_distinct | most_common_vals
1177 ---------+-----------+------------+------------------------------------
1179 | | | I- 880 Ramp+
1180 | | | Sp Railroad +
1181 | | | I- 580 +
1182 | | | I- 680 Ramp+
1183 | | | I- 80 Ramp+
1184 | | | 14th St +
1185 | | | I- 880 +
1186 | | | Mac Arthur Blvd+
1187 | | | Mission Blvd+
1188 ...
1190 | | | I- 880 Ramp+
1191 | | | I- 580 +
1192 | | | I- 680 Ramp+
1193 | | | I- 80 Ramp+
1194 | | | Sp Railroad +
1195 | | | I- 880 +
1196 | | | State Hwy 13 Ramp+
1197 | | | I- 80 +
1198 | | | State Hwy 24 Ramp+
1199 ...
1200 thepath | f | 0 |
1201 thepath | t | 0 |
1202 (4 rows)
1203 </screen>
1205 Note that two rows are displayed for the same column, one corresponding
1206 to the complete inheritance hierarchy starting at the
1210 (For brevity, we have only shown the first ten most-common values for
1212 </para>
1214 <para>
1218 arrays for each column, can be set on a
1220 command, or globally by setting the
1222 The default limit is presently 100 entries. Raising the limit
1223 might allow more accurate planner estimates to be made, particularly for
1224 columns with irregular data distributions, at the price of consuming
1226 time to compute the estimates. Conversely, a lower limit might be
1227 sufficient for columns with simple data distributions.
1228 </para>
1230 <para>
1231 Further details about the planner's use of statistics can be found in
1233 </para>
1234 </sect2>
1242 </indexterm>
1244 <indexterm>
1247 </indexterm>
1249 <indexterm>
1251 </indexterm>
1253 <indexterm>
1255 </indexterm>
1257 <para>
1258 It is common to see slow queries running bad execution plans because
1259 multiple columns used in the query clauses are correlated.
1260 The planner normally assumes that multiple conditions
1261 are independent of each other,
1262 an assumption that does not hold when column values are correlated.
1263 Regular statistics, because of their per-individual-column nature,
1264 cannot capture any knowledge about cross-column correlation.
1267 such information.
1268 </para>
1270 <para>
1271 Because the number of possible column combinations is very large,
1272 it's impractical to compute multivariate statistics automatically.
1275 the server to obtain statistics across interesting sets of columns.
1276 </para>
1278 <para>
1279 Statistics objects are created using the
1281 Creation of such an object merely creates a catalog entry expressing
1282 interest in the statistics. Actual data collection is performed
1284 auto-analyze). The collected values can be examined in the
1285 <link linkend="catalog-pg-statistic-ext-data"><structname>pg_statistic_ext_data</structname></link>
1286 catalog.
1287 </para>
1289 <para>
1291 sample of table rows that it takes for computing regular single-column
1292 statistics. Since the sample size is increased by increasing the
1293 statistics target for the table or any of its columns (as described in
1294 the previous section), a larger statistics target will normally result in
1295 more accurate extended statistics, as well as more time spent calculating
1296 them.
1297 </para>
1299 <para>
1300 The following subsections describe the kinds of extended statistics
1301 that are currently supported.
1302 </para>
1307 <para>
1308 The simplest kind of extended statistics tracks <firstterm>functional
1309 dependencies</firstterm>, a concept used in definitions of database normal forms.
1315 In a fully normalized database, functional dependencies should exist
1316 only on primary keys and superkeys. However, in practice many data sets
1317 are not fully normalized for various reasons; intentional
1318 denormalization for performance reasons is a common example.
1319 Even in a fully normalized database, there may be partial correlation
1320 between some columns, which can be expressed as partial functional
1321 dependency.
1322 </para>
1324 <para>
1325 The existence of functional dependencies directly affects the accuracy
1326 of estimates in certain queries. If a query contains conditions on
1327 both the independent and the dependent column(s), the
1328 conditions on the dependent columns do not further reduce the result
1329 size; but without knowledge of the functional dependency, the query
1330 planner will assume that the conditions are independent, resulting
1331 in underestimating the result size.
1332 </para>
1334 <para>
1336 can collect measurements of cross-column dependency. Assessing the
1337 degree of dependency between all sets of columns would be prohibitively
1338 expensive, so data collection is limited to those groups of columns
1339 appearing together in a statistics object defined with
1342 strongly correlated, to avoid unnecessary overhead in both
1344 </para>
1346 <para>
1347 Here is an example of collecting functional-dependency statistics:
1348 <programlisting>
1349 CREATE STATISTICS stts (dependencies) ON city, zip FROM zipcodes;
1351 ANALYZE zipcodes;
1353 SELECT stxname, stxkeys, stxddependencies
1354 FROM pg_statistic_ext join pg_statistic_ext_data on (oid = stxoid)
1355 WHERE stxname = 'stts';
1356 stxname | stxkeys | stxddependencies
1357 ---------+---------+------------------------------------------
1359 (1 row)
1360 </programlisting>
1361 Here it can be seen that column 1 (zip code) fully determines column
1364 represented by more than a single ZIP code.
1365 </para>
1367 <para>
1368 When computing the selectivity for a query involving functionally
1369 dependent columns, the planner adjusts the per-condition selectivity
1370 estimates using the dependency coefficients so as not to produce
1371 an underestimate.
1372 </para>
1377 <para>
1378 Functional dependencies are currently only applied when considering
1379 simple equality conditions that compare columns to constant values,
1381 They are not used to improve estimates for equality conditions
1382 comparing two columns or comparing a column to an expression, nor for
1384 </para>
1386 <para>
1387 When estimating with functional dependencies, the planner assumes that
1388 conditions on the involved columns are compatible and hence redundant.
1389 If they are incompatible, the correct estimate would be zero rows, but
1390 that possibility is not considered. For example, given a query like
1391 <programlisting>
1392 SELECT * FROM zipcodes WHERE city = 'San Francisco' AND zip = '94105';
1393 </programlisting>
1395 changing the selectivity, which is correct. However, it will make
1396 the same assumption about
1397 <programlisting>
1398 SELECT * FROM zipcodes WHERE city = 'San Francisco' AND zip = '90210';
1399 </programlisting>
1400 even though there will really be zero rows satisfying this query.
1401 Functional dependency statistics do not provide enough information
1402 to conclude that, however.
1403 </para>
1405 <para>
1406 In many practical situations, this assumption is usually satisfied;
1407 for example, there might be a GUI in the application that only allows
1408 selecting compatible city and ZIP code values to use in a query.
1409 But if that's not the case, functional dependencies may not be a viable
1410 option.
1411 </para>
1412 </sect4>
1413 </sect3>
1418 <para>
1419 Single-column statistics store the number of distinct values in each
1420 column. Estimates of the number of distinct values when combining more
1422 frequently wrong when the planner only has single-column statistical
1423 data, causing it to select bad plans.
1424 </para>
1426 <para>
1428 statistics for groups of columns. As before, it's impractical to do
1429 this for every possible column grouping, so data is collected only for
1430 those groups of columns appearing together in a statistics object
1432 for each possible combination of two or more columns from the set of
1433 listed columns.
1434 </para>
1436 <para>
1437 Continuing the previous example, the n-distinct counts in a
1438 table of ZIP codes might look like the following:
1439 <programlisting>
1440 CREATE STATISTICS stts2 (ndistinct) ON city, state, zip FROM zipcodes;
1442 ANALYZE zipcodes;
1444 SELECT stxkeys AS k, stxdndistinct AS nd
1445 FROM pg_statistic_ext join pg_statistic_ext_data on (oid = stxoid)
1446 WHERE stxname = 'stts2';
1448 k | 1 2 5
1450 (1 row)
1451 </programlisting>
1452 This indicates that there are three combinations of columns that
1453 have 33178 distinct values: ZIP code and state; ZIP code and city;
1454 and ZIP code, city and state (the fact that they are all equal is
1455 expected given that ZIP code alone is unique in this table). On the
1456 other hand, the combination of city and state has only 27435 distinct
1457 values.
1458 </para>
1460 <para>
1462 on combinations of columns that are actually used for grouping, and
1463 for which misestimation of the number of groups is resulting in bad
1465 </para>
1466 </sect3>
1471 <para>
1472 Another type of statistic stored for each column are most-common value
1473 lists. This allows very accurate estimates for individual columns, but
1474 may result in significant misestimates for queries with conditions on
1475 multiple columns.
1476 </para>
1478 <para>
1480 lists on combinations of columns. Similarly to functional dependencies
1481 and n-distinct coefficients, it's impractical to do this for every
1482 possible column grouping. Even more so in this case, as the MCV list
1483 (unlike functional dependencies and n-distinct coefficients) does store
1484 the common column values. So data is collected only for those groups
1485 of columns appearing together in a statistics object defined with the
1487 </para>
1489 <para>
1490 Continuing the previous example, the MCV list for a table of ZIP codes
1491 might look like the following (unlike for simpler types of statistics,
1492 a function is required for inspection of MCV contents):
1494 <programlisting>
1495 CREATE STATISTICS stts3 (mcv) ON city, state FROM zipcodes;
1497 ANALYZE zipcodes;
1499 SELECT m.* FROM pg_statistic_ext join pg_statistic_ext_data on (oid = stxoid),
1500 pg_mcv_list_items(stxdmcv) m WHERE stxname = 'stts3';
1502 index | values | nulls | frequency | base_frequency
1503 -------+------------------------+-------+-----------+----------------
1514 ...
1515 (99 rows)
1516 </programlisting>
1517 This indicates that the most common combination of city and state is
1518 Washington in DC, with actual frequency (in the sample) about 0.35%.
1519 The base frequency of the combination (as computed from the simple
1520 per-column frequencies) is only 0.0027%, resulting in two orders of
1521 magnitude under-estimates.
1522 </para>
1524 <para>
1526 on combinations of columns that are actually used in conditions together,
1527 and for which misestimation of the number of groups is resulting in bad
1529 are just wasted.
1530 </para>
1531 </sect3>
1533 </sect2>
1534 </sect1>
1542 </indexterm>
1544 <para>
1545 It is possible
1547 syntax. To see why this matters, we first need some background.
1548 </para>
1550 <para>
1551 In a simple join query, such as:
1552 <programlisting>
1553 SELECT * FROM a, b, c WHERE a.id = b.id AND b.ref = c.id;
1554 </programlisting>
1555 the planner is free to join the given tables in any order. For
1556 example, it could generate a query plan that joins A to B, using
1559 condition. Or it could join B to C and then join A to that result.
1560 Or it could join A to C and then join them with B — but that
1561 would be inefficient, since the full Cartesian product of A and C
1562 would have to be formed, there being no applicable condition in the
1565 between two input tables, so it's necessary to build up the result
1566 in one or another of these fashions.) The important point is that
1567 these different join possibilities give semantically equivalent
1568 results but might have hugely different execution costs. Therefore,
1569 the planner will explore all of them to try to find the most
1570 efficient query plan.
1571 </para>
1573 <para>
1574 When a query only involves two or three tables, there aren't many join
1575 orders to worry about. But the number of possible join orders grows
1576 exponentially as the number of tables expands. Beyond ten or so input
1577 tables it's no longer practical to do an exhaustive search of all the
1578 possibilities, and even for six or seven tables planning might take an
1579 annoyingly long time. When there are too many input tables, the
1582 through a limited number of possibilities. (The switch-over threshold is
1584 parameter.)
1585 The genetic search takes less time, but it won't
1586 necessarily find the best possible plan.
1587 </para>
1589 <para>
1590 When the query involves outer joins, the planner has less freedom
1591 than it does for plain (inner) joins. For example, consider:
1592 <programlisting>
1593 SELECT * FROM a LEFT JOIN (b JOIN c ON (b.ref = c.id)) ON (a.id = b.id);
1594 </programlisting>
1595 Although this query's restrictions are superficially similar to the
1596 previous example, the semantics are different because a row must be
1597 emitted for each row of A that has no matching row in the join of B and C.
1598 Therefore the planner has no choice of join order here: it must join
1599 B to C and then join A to that result. Accordingly, this query takes
1600 less time to plan than the previous query. In other cases, the planner
1601 might be able to determine that more than one join order is safe.
1602 For example, given:
1603 <programlisting>
1604 SELECT * FROM a LEFT JOIN b ON (a.bid = b.id) LEFT JOIN c ON (a.cid = c.id);
1605 </programlisting>
1606 it is valid to join A to either B or C first. Currently, only
1609 can be rearranged to some extent.
1610 </para>
1612 <para>
1616 constrain the join order.
1617 </para>
1619 <para>
1621 the join order, it is possible to instruct the
1624 For example, these three queries are logically equivalent:
1625 <programlisting>
1626 SELECT * FROM a, b, c WHERE a.id = b.id AND b.ref = c.id;
1627 SELECT * FROM a CROSS JOIN b CROSS JOIN c WHERE a.id = b.id AND b.ref = c.id;
1628 SELECT * FROM a JOIN (b JOIN c ON (b.ref = c.id)) ON (a.id = b.id);
1629 </programlisting>
1631 the second and third take less time to plan than the first. This effect
1632 is not worth worrying about for only three tables, but it can be a
1633 lifesaver with many tables.
1634 </para>
1636 <para>
1637 To force the planner to follow the join order laid out by explicit
1640 (Other possible values are discussed below.)
1641 </para>
1643 <para>
1644 You do not need to constrain the join order completely in order to
1647 <programlisting>
1648 SELECT * FROM a CROSS JOIN b, c, d, e WHERE ...;
1649 </programlisting>
1651 forces the planner to join A to B before joining them to other tables,
1652 but doesn't constrain its choices otherwise. In this example, the
1653 number of possible join orders is reduced by a factor of 5.
1654 </para>
1656 <para>
1657 Constraining the planner's search in this way is a useful technique
1658 both for reducing planning time and for directing the planner to a
1659 good query plan. If the planner chooses a bad join order by default,
1661 — assuming that you know of a better order, that is. Experimentation
1662 is recommended.
1663 </para>
1665 <para>
1666 A closely related issue that affects planning time is collapsing of
1667 subqueries into their parent query. For example, consider:
1668 <programlisting>
1669 SELECT *
1670 FROM x, y,
1671 (SELECT * FROM a, b, c WHERE something) AS ss
1672 WHERE somethingelse;
1673 </programlisting>
1674 This situation might arise from use of a view that contains a join;
1676 reference, yielding a query much like the above. Normally, the planner
1677 will try to collapse the subquery into the parent, yielding:
1678 <programlisting>
1679 SELECT * FROM x, y, a, b, c WHERE something AND somethingelse;
1680 </programlisting>
1681 This usually results in a better plan than planning the subquery
1683 joining X to A first eliminates many rows of A, thus avoiding the need to
1684 form the full logical output of the subquery.) But at the same time,
1685 we have increased the planning time; here, we have a five-way join
1686 problem replacing two separate three-way join problems. Because of the
1687 exponential growth of the number of possibilities, this makes a big
1688 difference. The planner tries to avoid getting stuck in huge join search
1691 query. You can trade off planning time against quality of plan by
1692 adjusting this run-time parameter up or down.
1693 </para>
1695 <para>
1698 are similarly named because they do almost the same thing: one controls
1700 other controls when it will flatten out explicit joins. Typically
1704 to control join order with explicit joins). But you might set them
1705 differently if you are trying to fine-tune the trade-off between planning
1706 time and run time.
1707 </para>
1708 </sect1>
1713 <para>
1714 One might need to insert a large amount of data when first populating
1715 a database. This section contains some suggestions on how to make
1716 this process as efficient as possible.
1717 </para>
1722 <indexterm>
1725 </indexterm>
1727 <para>
1729 one commit at the end. (In plain
1732 do this behind your back, in which case you need to make sure the
1733 library does it when you want it done.) If you allow each
1734 insertion to be committed separately,
1736 each row that is added. An additional benefit of doing all
1737 insertions in one transaction is that if the insertion of one row
1738 were to fail then the insertion of all rows inserted up to that
1739 point would be rolled back, so you won't be stuck with partially
1740 loaded data.
1741 </para>
1742 </sect2>
1747 <para>
1749 all the rows in one command, instead of using a series of
1751 command is optimized for loading large numbers of rows; it is less
1754 is a single command, there is no need to disable autocommit if you
1755 use this method to populate a table.
1756 </para>
1758 <para>
1763 some of the overhead of repeatedly parsing and planning
1766 documentation.
1767 </para>
1769 <para>
1770 Note that loading a large number of rows using
1773 multiple insertions are batched into a single transaction.
1774 </para>
1776 <para>
1780 needs to be written, because in case of an error, the files
1781 containing the newly loaded data will be removed anyway.
1782 However, this consideration only applies when
1784 as all commands must write WAL otherwise.
1785 </para>
1787 </sect2>
1792 <para>
1793 If you are loading a freshly created table, the fastest method is to
1794 create the table, bulk load the table's data using
1796 table. Creating an index on pre-existing data is quicker than
1797 updating it incrementally as each row is loaded.
1798 </para>
1800 <para>
1801 If you are adding large amounts of data to an existing table,
1802 it might be a win to drop the indexes,
1803 load the table, and then recreate the indexes. Of course, the
1804 database performance for other users might suffer
1805 during the time the indexes are missing. One should also think
1806 twice before dropping a unique index, since the error checking
1807 afforded by the unique constraint will be lost while the index is
1808 missing.
1809 </para>
1810 </sect2>
1815 <para>
1816 Just as with indexes, a foreign key constraint can be checked
1818 useful to drop foreign key constraints, load data, and re-create
1819 the constraints. Again, there is a trade-off between data load
1820 speed and loss of error checking while the constraint is missing.
1821 </para>
1823 <para>
1824 What's more, when you load data into a table with existing foreign key
1825 constraints, each new row requires an entry in the server's list of
1826 pending trigger events (since it is the firing of a trigger that checks
1827 the row's foreign key constraint). Loading many millions of rows can
1828 cause the trigger event queue to overflow available memory, leading to
1829 intolerable swapping or even outright failure of the command. Therefore
1831 foreign keys when loading large amounts of data. If temporarily removing
1832 the constraint isn't acceptable, the only other recourse may be to split
1833 up the load operation into smaller transactions.
1834 </para>
1835 </sect2>
1840 <para>
1842 configuration variable when loading large amounts of data can
1843 lead to improved performance. This will help to speed up <command>CREATE
1846 only useful when you are using one or both of the above techniques.
1847 </para>
1848 </sect2>
1853 <para>
1855 configuration variable can also
1856 make large data loads faster. This is because loading a large
1858 cause checkpoints to occur more often than the normal checkpoint
1860 configuration variable). Whenever a checkpoint occurs, all dirty
1861 pages must be flushed to disk. By increasing
1863 data loads, the number of checkpoints that are required can be
1864 reduced.
1865 </para>
1866 </sect2>
1871 <para>
1872 When loading large amounts of data into an installation that uses
1873 WAL archiving or streaming replication, it might be faster to take a
1874 new base backup after the load has completed than to process a large
1875 amount of incremental WAL data. To prevent incremental WAL logging
1876 while loading, disable archiving and streaming replication, by setting
1880 But note that changing these settings requires a server restart,
1881 and makes any base backups taken before unavailable for archive
1882 recovery and standby server, which may lead to data loss.
1883 </para>
1885 <para>
1886 Aside from avoiding the time for the archiver or WAL sender to process the
1887 WAL data, doing this will actually make certain commands faster, because
1890 transaction) created or truncated the table or index they change. (They
1891 can guarantee crash safety more cheaply by doing
1893 </para>
1894 </sect2>
1899 <para>
1900 Whenever you have significantly altered the distribution of data
1901 within a table, running <link linkend="sql-analyze"><command>ANALYZE</command></link> is strongly recommended. This
1902 includes bulk loading large amounts of data into the table. Running
1904 ensures that the planner has up-to-date statistics about the
1905 table. With no statistics or obsolete statistics, the planner might
1906 make poor decisions during query planning, leading to poor
1907 performance on any tables with inaccurate or nonexistent
1908 statistics. Note that if the autovacuum daemon is enabled, it might
1912 </para>
1913 </sect2>
1918 <para>
1920 several, but not all, of the above guidelines. To restore a
1922 do a few extra things manually. (Note that these points apply while
1924 The same points apply whether loading a text dump with
1927 </para>
1929 <para>
1931 it is generating a complete schema-and-data dump, it is careful to
1932 load data before creating indexes and foreign keys. So in this case
1933 several guidelines are handled automatically. What is left
1934 for you to do is to:
1935 <itemizedlist>
1936 <listitem>
1937 <para>
1938 Set appropriate (i.e., larger than normal) values for
1941 </para>
1942 </listitem>
1943 <listitem>
1944 <para>
1945 If using WAL archiving or streaming replication, consider disabling
1950 Afterwards, set them back to the right values and take a fresh
1951 base backup.
1952 </para>
1953 </listitem>
1954 <listitem>
1955 <para>
1956 Experiment with the parallel dump and restore modes of both
1958 optimal number of concurrent jobs to use. Dumping and restoring in
1960 significantly higher performance over the serial mode.
1961 </para>
1962 </listitem>
1963 <listitem>
1964 <para>
1965 Consider whether the whole dump should be restored as a single
1969 mode, even the smallest of errors will rollback the entire restore,
1970 possibly discarding many hours of processing. Depending on how
1971 interrelated the data is, that might seem preferable to manual cleanup,
1973 transaction and have WAL archiving turned off.
1974 </para>
1975 </listitem>
1976 <listitem>
1977 <para>
1978 If multiple CPUs are available in the database server, consider using
1980 allows concurrent data loading and index creation.
1981 </para>
1982 </listitem>
1983 <listitem>
1984 <para>
1986 </para>
1987 </listitem>
1988 </itemizedlist>
1989 </para>
1991 <para>
1993 drop or recreate indexes, and it does not normally touch foreign
1994 keys.
1996 <footnote>
1997 <para>
1998 You can get the effect of disabling foreign keys by using
2000 that eliminates, rather than just postpones, foreign key
2001 validation, and so it is possible to insert bad data if you use it.
2002 </para>
2003 </footnote>
2005 So when loading a data-only dump, it is up to you to drop and recreate
2006 indexes and foreign keys if you wish to use those techniques.
2008 while loading the data, but don't bother increasing
2010 manually recreating indexes and foreign keys afterwards.
2014 </para>
2015 </sect2>
2016 </sect1>
2023 </indexterm>
2025 <para>
2026 Durability is a database feature that guarantees the recording of
2027 committed transactions even if the server crashes or loses
2028 power. However, durability adds significant database overhead,
2029 so if your site does not require such a guarantee,
2031 much faster. The following are configuration changes you can make
2032 to improve performance in such cases. Except as noted below, durability
2033 is still guaranteed in case of a crash of the database software;
2034 only an abrupt operating system crash creates a risk of data loss
2035 or corruption when these settings are used.
2037 <itemizedlist>
2038 <listitem>
2039 <para>
2040 Place the database cluster's data directory in a memory-backed
2042 database disk I/O, but limits data storage to the amount of
2043 available memory (and perhaps swap).
2044 </para>
2045 </listitem>
2047 <listitem>
2048 <para>
2050 data to disk.
2051 </para>
2052 </listitem>
2054 <listitem>
2055 <para>
2058 commit. This setting does risk transaction loss (though not data
2060 </para>
2061 </listitem>
2063 <listitem>
2064 <para>
2066 to guard against partial page writes.
2067 </para>
2068 </listitem>
2070 <listitem>
2071 <para>
2074 of checkpoints, but increases the storage requirements of
2076 </para>
2077 </listitem>
2079 <listitem>
2080 <para>
2083 makes the tables non-crash-safe.
2084 </para>
2085 </listitem>
2087 </itemizedlist>
2088 </para>
2089 </sect1>
2091 </chapter>