| blob | a502a2aaba335cfc966387faa2444f8dedaef588 |
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 When using the enable/disable flags to disable plan node types, many of
583 the flags only discourage the use of the corresponding plan node and don't
584 outright disallow the planner's ability to use the plan node type. This
585 is by design so that the planner still maintains the ability to form a
586 plan for a given query. When the resulting plan contains a disabled node,
589 <screen>
590 SET enable_seqscan = off;
591 EXPLAIN SELECT * FROM unit;
593 QUERY PLAN
594 ---------------------------------------------------------
596 Disabled: true
597 </screen>
598 </para>
600 <para>
602 other means to read the table data, so the sequential scan is the only
603 option available to the query planner.
604 </para>
606 <para>
607 <indexterm>
609 </indexterm>
612 queries can sometimes be transformed into ordinary join plans, but
613 when they cannot be, we get plans like:
615 <screen>
616 EXPLAIN VERBOSE SELECT unique1
617 FROM tenk1 t
618 WHERE t.ten < ALL (SELECT o.ten FROM onek o WHERE o.four = t.four);
620 QUERY PLAN
621 -------------------------------------------------------------------&zwsp;------
623 Output: t.unique1
625 SubPlan 1
627 Output: o.ten
628 Filter: (o.four = t.four)
629 </screen>
631 This rather artificial example serves to illustrate a couple of
632 points: values from the outer plan level can be passed down into a
634 results of the sub-select are available to the outer plan. Those
636 like
640 </para>
642 <para>
643 <indexterm>
646 </indexterm>
648 subplan again for each row of the outer query (which accounts for the
649 high estimated cost). Some queries can use a <firstterm>hashed
650 subplan</firstterm> to avoid that:
652 <screen>
653 EXPLAIN SELECT *
654 FROM tenk1 t
655 WHERE t.unique1 NOT IN (SELECT o.unique1 FROM onek o);
657 QUERY PLAN
658 -------------------------------------------------------------------&zwsp;-------------------------
660 Filter: (NOT (ANY (unique1 = (hashed SubPlan 1).col1)))
661 SubPlan 1
663 (4 rows)
664 </screen>
666 Here, the subplan is run a single time and its output is loaded into
667 an in-memory hash table, which is then probed by the
671 amenable to hashing.
672 </para>
674 <para>
675 <indexterm>
677 </indexterm>
678 If, in addition to not referencing any variables of the outer query,
682 <screen>
683 EXPLAIN VERBOSE SELECT unique1
684 FROM tenk1 t1 WHERE t1.ten = (SELECT (random() * 10)::integer);
686 QUERY PLAN
687 ------------------------------------------------------------&zwsp;--------
689 Output: t1.unique1
690 Filter: (t1.ten = (InitPlan 1).col1)
691 InitPlan 1
693 Output: ((random() * '10'::double precision))::integer
694 </screen>
696 An initplan is run only once per execution of the outer plan, and its
697 results are saved for re-use in later rows of the outer plan. So in
700 randomly-chosen integer. That's quite different from what would
702 </para>
704 </sect2>
709 <para>
710 It is possible to check the accuracy of the planner's estimates
713 the true row counts and true run time accumulated within each plan node,
715 shows. For example, we might get a result like this:
717 <screen>
718 EXPLAIN ANALYZE SELECT *
719 FROM tenk1 t1, tenk2 t2
722 QUERY PLAN
723 -------------------------------------------------------------------&zwsp;--------------------------------------------------------------
726 -> 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)
728 Heap Blocks: exact=10
730 -> 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)
732 Buffers: shared hit=2
733 -> 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)
734 Index Cond: (unique2 = t1.unique2)
736 Planning:
738 Planning Time: 0.485 ms
739 Execution Time: 0.073 ms
740 </screen>
744 arbitrary units; so they are unlikely to match up.
745 The thing that's usually most important to look for is whether the
746 estimated row counts are reasonably close to reality. In this example
747 the estimates were all dead-on, but that's quite unusual in practice.
748 </para>
750 <para>
751 In some query plans, it is possible for a subplan node to be executed more
752 than once. For example, the inner index scan will be executed once per
753 outer row in the above nested-loop plan. In such cases, the
755 total number of executions of the node, and the actual time and rows
756 values shown are averages per-execution. This is done to make the numbers
757 comparable with the way that the cost estimates are shown. Multiply by
759 the node. In the above example, we spent a total of 0.030 milliseconds
761 </para>
763 <para>
765 statistics beyond the plan node execution times and row counts.
766 For example, Sort and Hash nodes provide extra information:
768 <screen>
769 EXPLAIN ANALYZE SELECT *
770 FROM tenk1 t1, tenk2 t2
773 QUERY PLAN
774 -------------------------------------------------------------------&zwsp;-------------------------------------------------------------------&zwsp;------
776 Sort Key: t1.fivethous
777 Sort Method: quicksort Memory: 74kB
778 Buffers: shared hit=440
779 -> Hash Join (cost=226.23..709.73 rows=100 width=488) (actual time=0.515..2.920 rows=100 loops=1)
780 Hash Cond: (t2.unique2 = t1.unique2)
781 Buffers: shared hit=437
782 -> Seq Scan on tenk2 t2 (cost=0.00..445.00 rows=10000 width=244) (actual time=0.026..1.790 rows=10000 loops=1)
783 Buffers: shared hit=345
786 Buffers: shared hit=92
787 -> 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)
789 Heap Blocks: exact=90
790 Buffers: shared hit=92
791 -> 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)
793 Buffers: shared hit=2
794 Planning:
795 Buffers: shared hit=12
796 Planning Time: 0.187 ms
797 Execution Time: 3.036 ms
798 </screen>
800 The Sort node shows the sort method used (in particular, whether the sort
801 was in-memory or on-disk) and the amount of memory or disk space needed.
802 The Hash node shows the number of hash buckets and batches as well as the
803 peak amount of memory used for the hash table. (If the number of batches
804 exceeds one, there will also be disk space usage involved, but that is not
805 shown.)
806 </para>
808 <para>
809 Another type of extra information is the number of rows removed by a
810 filter condition:
812 <screen>
815 QUERY PLAN
816 -------------------------------------------------------------------&zwsp;--------------------------------------
817 Seq Scan on tenk1 (cost=0.00..470.00 rows=7000 width=244) (actual time=0.030..1.995 rows=7000 loops=1)
819 Rows Removed by Filter: 3000
820 Buffers: shared hit=345
821 Planning Time: 0.102 ms
822 Execution Time: 2.145 ms
823 </screen>
825 These counts can be particularly valuable for filter conditions applied at
827 one scanned row, or potential join pair in the case of a join node,
828 is rejected by the filter condition.
829 </para>
831 <para>
833 index scans. For example, consider this search for polygons containing a
834 specific point:
836 <screen>
839 QUERY PLAN
840 -------------------------------------------------------------------&zwsp;-----------------------------------
841 Seq Scan on polygon_tbl (cost=0.00..1.09 rows=1 width=85) (actual time=0.023..0.023 rows=0 loops=1)
843 Rows Removed by Filter: 7
844 Buffers: shared hit=1
845 Planning Time: 0.039 ms
846 Execution Time: 0.033 ms
847 </screen>
849 The planner thinks (quite correctly) that this sample table is too small
850 to bother with an index scan, so we have a plain sequential scan in which
851 all the rows got rejected by the filter condition. But if we force an
852 index scan to be used, we see:
854 <screen>
855 SET enable_seqscan TO off;
859 QUERY PLAN
860 -------------------------------------------------------------------&zwsp;-------------------------------------------------------
861 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)
863 Rows Removed by Index Recheck: 1
864 Buffers: shared hit=1
865 Planning Time: 0.039 ms
866 Execution Time: 0.098 ms
867 </screen>
869 Here we can see that the index returned one candidate row, which was
870 then rejected by a recheck of the index condition. This happens because a
872 returns the rows with polygons that overlap the target, and then we have
873 to do the exact containment test on those rows.
874 </para>
876 <para>
878 provides additional detail about I/O operations performed during the
879 planning and execution of the given query. The buffer numbers displayed
880 show the count of the non-distinct buffers hit, read, dirtied, and written
881 for the given node and all of its child nodes. The
886 <screen>
887 EXPLAIN (ANALYZE, BUFFERS OFF) SELECT * FROM tenk1 WHERE unique1 < 100 AND unique2 > 9000;
889 QUERY PLAN
890 -------------------------------------------------------------------&zwsp;--------------------------------------------------------------
891 Bitmap Heap Scan on tenk1 (cost=25.07..60.11 rows=10 width=244) (actual time=0.105..0.114 rows=10 loops=1)
893 Heap Blocks: exact=10
895 -> 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)
897 -> 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)
899 Planning Time: 0.162 ms
900 Execution Time: 0.143 ms
901 </screen>
902 </para>
904 <para>
906 runs the query, any side-effects will happen as usual, even though
907 whatever results the query might output are discarded in favor of
909 data-modifying query without changing your tables, you can
910 roll the command back afterwards, for example:
912 <screen>
913 BEGIN;
917 QUERY PLAN
918 -------------------------------------------------------------------&zwsp;-------------------------------------------------------------
920 -> Bitmap Heap Scan on tenk1 (cost=5.06..225.23 rows=100 width=10) (actual time=0.065..0.141 rows=100 loops=1)
922 Heap Blocks: exact=90
924 -> 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)
926 Buffers: shared read=2
927 Planning Time: 0.151 ms
928 Execution Time: 1.856 ms
930 ROLLBACK;
931 </screen>
932 </para>
934 <para>
938 applying the table changes is done by a top-level Insert, Update,
939 Delete, or Merge plan node. The plan nodes underneath this node perform
940 the work of locating the old rows and/or computing the new data.
941 So above, we see the same sort of bitmap table scan we've seen already,
942 and its output is fed to an Update node that stores the updated rows.
943 It's worth noting that although the data-modifying node can take a
944 considerable amount of run time (here, it's consuming the lion's share
945 of the time), the planner does not currently add anything to the cost
946 estimates to account for that work. That's because the work to be done is
947 the same for every correct query plan, so it doesn't affect planning
948 decisions.
949 </para>
951 <para>
954 inheritance hierarchy, the output might look like this:
956 <screen>
957 EXPLAIN UPDATE gtest_parent SET f1 = CURRENT_DATE WHERE f2 = 101;
959 QUERY PLAN
960 -------------------------------------------------------------------&zwsp;---------------------
962 Update on gtest_child gtest_parent_1
963 Update on gtest_child2 gtest_parent_2
964 Update on gtest_child3 gtest_parent_3
967 Filter: (f2 = 101)
969 Filter: (f2 = 101)
971 Filter: (f2 = 101)
972 </screen>
974 In this example the Update node needs to consider three child tables,
975 but not the originally-mentioned partitioned table (since that never
976 stores any data). So there are three input
977 scanning subplans, one per table. For clarity, the Update node is
978 annotated to show the specific target tables that will be updated, in the
979 same order as the corresponding subplans.
980 </para>
982 <para>
984 ANALYZE</command> is the time it took to generate the query plan from the
985 parsed query and optimize it. It does not include parsing or rewriting.
986 </para>
988 <para>
990 ANALYZE</command> includes executor start-up and shut-down time, as well
991 as the time to run any triggers that are fired, but it does not include
992 parsing, rewriting, or planning time.
994 the time for the related Insert, Update, or Delete node; but time
997 The total time spent in each trigger
999 Note that deferred constraint triggers will not be executed
1000 until end of transaction and are thus not considered at all by
1002 </para>
1004 <para>
1005 The time shown for the top-level node does not include any time needed
1006 to convert the query's output data into displayable form or to send it
1008 send the data to the client, it can be told to convert the query's
1009 output data to displayable form and measure the time needed for that,
1011 be shown separately, and it's also included in the
1013 </para>
1015 </sect2>
1020 <para>
1021 There are two significant ways in which run times measured by
1023 the same query. First, since no output rows are delivered to the client,
1024 network transmission costs are not included. I/O conversion costs are
1026 Second, the measurement overhead added by <command>EXPLAIN
1027 ANALYZE</command> can be significant, especially on machines with slow
1030 on your system.
1031 </para>
1033 <para>
1035 much different from the one you are actually testing; for example,
1036 results on a toy-sized table cannot be assumed to apply to large tables.
1037 The planner's cost estimates are not linear and so it might choose
1038 a different plan for a larger or smaller table. An extreme example
1039 is that on a table that only occupies one disk page, you'll nearly
1040 always get a sequential scan plan whether indexes are available or not.
1041 The planner realizes that it's going to take one disk page read to
1042 process the table in any case, so there's no value in expending additional
1043 page reads to look at an index. (We saw this happening in the
1045 </para>
1047 <para>
1048 There are cases in which the actual and estimated values won't match up
1049 well, but nothing is really wrong. One such case occurs when
1053 <screen>
1056 QUERY PLAN
1057 -------------------------------------------------------------------&zwsp;------------------------------------------------------------
1059 Buffers: shared hit=16
1060 -> 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)
1063 Rows Removed by Filter: 287
1064 Buffers: shared hit=16
1065 Planning Time: 0.077 ms
1066 Execution Time: 0.086 ms
1067 </screen>
1069 the estimated cost and row count for the Index Scan node are shown as
1070 though it were run to completion. But in reality the Limit node stopped
1071 requesting rows after it got two, so the actual row count is only 2 and
1072 the run time is less than the cost estimate would suggest. This is not
1073 an estimation error, only a discrepancy in the way the estimates and true
1074 values are displayed.
1075 </para>
1077 <para>
1078 Merge joins also have measurement artifacts that can confuse the unwary.
1079 A merge join will stop reading one input if it's exhausted the other input
1080 and the next key value in the one input is greater than the last key value
1081 of the other input; in such a case there can be no more matches and so no
1082 need to scan the rest of the first input. This results in not reading all
1084 Also, if the outer (first) child contains rows with duplicate key values,
1085 the inner (second) child is backed up and rescanned for the portion of its
1087 repeated emissions of the same inner rows as if they were real additional
1088 rows. When there are many outer duplicates, the reported actual row count
1089 for the inner child plan node can be significantly larger than the number
1090 of rows that are actually in the inner relation.
1091 </para>
1093 <para>
1094 BitmapAnd and BitmapOr nodes always report their actual row counts as zero,
1095 due to implementation limitations.
1096 </para>
1098 <para>
1100 created by the planner. However, there are cases where the executor
1101 can determine that certain nodes need not be executed because they
1102 cannot produce any rows, based on parameter values that were not
1103 available at planning time. (Currently this can only happen for child
1104 nodes of an Append or MergeAppend node that is scanning a partitioned
1105 table.) When this happens, those plan nodes are omitted from
1108 instead.
1109 </para>
1110 </sect2>
1112 </sect1>
1120 </indexterm>
1124 <para>
1125 As we saw in the previous section, the query planner needs to estimate
1126 the number of rows retrieved by a query in order to make good choices
1127 of query plans. This section provides a quick look at the statistics
1128 that the system uses for these estimates.
1129 </para>
1131 <para>
1132 One component of the statistics is the total number of entries in
1133 each table and index, as well as the number of disk blocks occupied
1134 by each table and index. This information is kept in the table
1138 queries similar to this one:
1140 <screen>
1141 SELECT relname, relkind, reltuples, relpages
1142 FROM pg_class
1143 WHERE relname LIKE 'tenk1%';
1145 relname | relkind | reltuples | relpages
1146 ----------------------+---------+-----------+----------
1152 (5 rows)
1153 </screen>
1156 rows, as do its indexes, but the indexes are (unsurprisingly) much
1157 smaller than the table.
1158 </para>
1160 <para>
1163 and so they usually contain somewhat out-of-date values.
1167 (which is commonly the case) will incrementally update the
1169 of the table it did scan, resulting in an approximate value.
1170 In any case, the planner
1172 to match the current physical table size, thus obtaining a closer
1173 approximation.
1174 </para>
1176 <indexterm>
1178 </indexterm>
1180 <para>
1181 Most queries retrieve only a fraction of the rows in a table, due
1183 examined. The planner thus needs to make an estimate of the
1185 the fraction of rows that match each condition in the
1187 stored in the
1191 ANALYZE</command> commands, and are always approximate even when freshly
1192 updated.
1193 </para>
1195 <indexterm>
1197 </indexterm>
1199 <para>
1201 it's better to look at its view
1204 is designed to be more easily readable. Furthermore,
1207 (This prevents unprivileged users from learning something about
1208 the contents of other people's tables from the statistics. The
1210 rows about tables that the current user can read.)
1211 For example, we might do:
1213 <screen>
1214 SELECT attname, inherited, n_distinct,
1215 array_to_string(most_common_vals, E'\n') as most_common_vals
1216 FROM pg_stats
1217 WHERE tablename = 'road';
1219 attname | inherited | n_distinct | most_common_vals
1220 ---------+-----------+------------+------------------------------------
1222 | | | I- 880 Ramp+
1223 | | | Sp Railroad +
1224 | | | I- 580 +
1225 | | | I- 680 Ramp+
1226 | | | I- 80 Ramp+
1227 | | | 14th St +
1228 | | | I- 880 +
1229 | | | Mac Arthur Blvd+
1230 | | | Mission Blvd+
1231 ...
1233 | | | I- 880 Ramp+
1234 | | | I- 580 +
1235 | | | I- 680 Ramp+
1236 | | | I- 80 Ramp+
1237 | | | Sp Railroad +
1238 | | | I- 880 +
1239 | | | State Hwy 13 Ramp+
1240 | | | I- 80 +
1241 | | | State Hwy 24 Ramp+
1242 ...
1243 thepath | f | 0 |
1244 thepath | t | 0 |
1245 (4 rows)
1246 </screen>
1248 Note that two rows are displayed for the same column, one corresponding
1249 to the complete inheritance hierarchy starting at the
1253 (For brevity, we have only shown the first ten most-common values for
1255 </para>
1257 <para>
1261 arrays for each column, can be set on a
1263 command, or globally by setting the
1265 The default limit is presently 100 entries. Raising the limit
1266 might allow more accurate planner estimates to be made, particularly for
1267 columns with irregular data distributions, at the price of consuming
1269 time to compute the estimates. Conversely, a lower limit might be
1270 sufficient for columns with simple data distributions.
1271 </para>
1273 <para>
1274 Further details about the planner's use of statistics can be found in
1276 </para>
1277 </sect2>
1285 </indexterm>
1287 <indexterm>
1290 </indexterm>
1292 <indexterm>
1294 </indexterm>
1296 <indexterm>
1298 </indexterm>
1300 <para>
1301 It is common to see slow queries running bad execution plans because
1302 multiple columns used in the query clauses are correlated.
1303 The planner normally assumes that multiple conditions
1304 are independent of each other,
1305 an assumption that does not hold when column values are correlated.
1306 Regular statistics, because of their per-individual-column nature,
1307 cannot capture any knowledge about cross-column correlation.
1310 such information.
1311 </para>
1313 <para>
1314 Because the number of possible column combinations is very large,
1315 it's impractical to compute multivariate statistics automatically.
1318 the server to obtain statistics across interesting sets of columns.
1319 </para>
1321 <para>
1322 Statistics objects are created using the
1324 Creation of such an object merely creates a catalog entry expressing
1325 interest in the statistics. Actual data collection is performed
1327 auto-analyze). The collected values can be examined in the
1328 <link linkend="catalog-pg-statistic-ext-data"><structname>pg_statistic_ext_data</structname></link>
1329 catalog.
1330 </para>
1332 <para>
1334 sample of table rows that it takes for computing regular single-column
1335 statistics. Since the sample size is increased by increasing the
1336 statistics target for the table or any of its columns (as described in
1337 the previous section), a larger statistics target will normally result in
1338 more accurate extended statistics, as well as more time spent calculating
1339 them.
1340 </para>
1342 <para>
1343 The following subsections describe the kinds of extended statistics
1344 that are currently supported.
1345 </para>
1350 <para>
1351 The simplest kind of extended statistics tracks <firstterm>functional
1352 dependencies</firstterm>, a concept used in definitions of database normal forms.
1358 In a fully normalized database, functional dependencies should exist
1359 only on primary keys and superkeys. However, in practice many data sets
1360 are not fully normalized for various reasons; intentional
1361 denormalization for performance reasons is a common example.
1362 Even in a fully normalized database, there may be partial correlation
1363 between some columns, which can be expressed as partial functional
1364 dependency.
1365 </para>
1367 <para>
1368 The existence of functional dependencies directly affects the accuracy
1369 of estimates in certain queries. If a query contains conditions on
1370 both the independent and the dependent column(s), the
1371 conditions on the dependent columns do not further reduce the result
1372 size; but without knowledge of the functional dependency, the query
1373 planner will assume that the conditions are independent, resulting
1374 in underestimating the result size.
1375 </para>
1377 <para>
1379 can collect measurements of cross-column dependency. Assessing the
1380 degree of dependency between all sets of columns would be prohibitively
1381 expensive, so data collection is limited to those groups of columns
1382 appearing together in a statistics object defined with
1385 strongly correlated, to avoid unnecessary overhead in both
1387 </para>
1389 <para>
1390 Here is an example of collecting functional-dependency statistics:
1391 <programlisting>
1392 CREATE STATISTICS stts (dependencies) ON city, zip FROM zipcodes;
1394 ANALYZE zipcodes;
1396 SELECT stxname, stxkeys, stxddependencies
1397 FROM pg_statistic_ext join pg_statistic_ext_data on (oid = stxoid)
1398 WHERE stxname = 'stts';
1399 stxname | stxkeys | stxddependencies
1400 ---------+---------+------------------------------------------
1402 (1 row)
1403 </programlisting>
1404 Here it can be seen that column 1 (zip code) fully determines column
1407 represented by more than a single ZIP code.
1408 </para>
1410 <para>
1411 When computing the selectivity for a query involving functionally
1412 dependent columns, the planner adjusts the per-condition selectivity
1413 estimates using the dependency coefficients so as not to produce
1414 an underestimate.
1415 </para>
1420 <para>
1421 Functional dependencies are currently only applied when considering
1422 simple equality conditions that compare columns to constant values,
1424 They are not used to improve estimates for equality conditions
1425 comparing two columns or comparing a column to an expression, nor for
1427 </para>
1429 <para>
1430 When estimating with functional dependencies, the planner assumes that
1431 conditions on the involved columns are compatible and hence redundant.
1432 If they are incompatible, the correct estimate would be zero rows, but
1433 that possibility is not considered. For example, given a query like
1434 <programlisting>
1435 SELECT * FROM zipcodes WHERE city = 'San Francisco' AND zip = '94105';
1436 </programlisting>
1438 changing the selectivity, which is correct. However, it will make
1439 the same assumption about
1440 <programlisting>
1441 SELECT * FROM zipcodes WHERE city = 'San Francisco' AND zip = '90210';
1442 </programlisting>
1443 even though there will really be zero rows satisfying this query.
1444 Functional dependency statistics do not provide enough information
1445 to conclude that, however.
1446 </para>
1448 <para>
1449 In many practical situations, this assumption is usually satisfied;
1450 for example, there might be a GUI in the application that only allows
1451 selecting compatible city and ZIP code values to use in a query.
1452 But if that's not the case, functional dependencies may not be a viable
1453 option.
1454 </para>
1455 </sect4>
1456 </sect3>
1461 <para>
1462 Single-column statistics store the number of distinct values in each
1463 column. Estimates of the number of distinct values when combining more
1465 frequently wrong when the planner only has single-column statistical
1466 data, causing it to select bad plans.
1467 </para>
1469 <para>
1471 statistics for groups of columns. As before, it's impractical to do
1472 this for every possible column grouping, so data is collected only for
1473 those groups of columns appearing together in a statistics object
1475 for each possible combination of two or more columns from the set of
1476 listed columns.
1477 </para>
1479 <para>
1480 Continuing the previous example, the n-distinct counts in a
1481 table of ZIP codes might look like the following:
1482 <programlisting>
1483 CREATE STATISTICS stts2 (ndistinct) ON city, state, zip FROM zipcodes;
1485 ANALYZE zipcodes;
1487 SELECT stxkeys AS k, stxdndistinct AS nd
1488 FROM pg_statistic_ext join pg_statistic_ext_data on (oid = stxoid)
1489 WHERE stxname = 'stts2';
1491 k | 1 2 5
1493 (1 row)
1494 </programlisting>
1495 This indicates that there are three combinations of columns that
1496 have 33178 distinct values: ZIP code and state; ZIP code and city;
1497 and ZIP code, city and state (the fact that they are all equal is
1498 expected given that ZIP code alone is unique in this table). On the
1499 other hand, the combination of city and state has only 27435 distinct
1500 values.
1501 </para>
1503 <para>
1505 on combinations of columns that are actually used for grouping, and
1506 for which misestimation of the number of groups is resulting in bad
1508 </para>
1509 </sect3>
1514 <para>
1515 Another type of statistic stored for each column are most-common value
1516 lists. This allows very accurate estimates for individual columns, but
1517 may result in significant misestimates for queries with conditions on
1518 multiple columns.
1519 </para>
1521 <para>
1523 lists on combinations of columns. Similarly to functional dependencies
1524 and n-distinct coefficients, it's impractical to do this for every
1525 possible column grouping. Even more so in this case, as the MCV list
1526 (unlike functional dependencies and n-distinct coefficients) does store
1527 the common column values. So data is collected only for those groups
1528 of columns appearing together in a statistics object defined with the
1530 </para>
1532 <para>
1533 Continuing the previous example, the MCV list for a table of ZIP codes
1534 might look like the following (unlike for simpler types of statistics,
1535 a function is required for inspection of MCV contents):
1537 <programlisting>
1538 CREATE STATISTICS stts3 (mcv) ON city, state FROM zipcodes;
1540 ANALYZE zipcodes;
1542 SELECT m.* FROM pg_statistic_ext join pg_statistic_ext_data on (oid = stxoid),
1543 pg_mcv_list_items(stxdmcv) m WHERE stxname = 'stts3';
1545 index | values | nulls | frequency | base_frequency
1546 -------+------------------------+-------+-----------+----------------
1557 ...
1558 (99 rows)
1559 </programlisting>
1560 This indicates that the most common combination of city and state is
1561 Washington in DC, with actual frequency (in the sample) about 0.35%.
1562 The base frequency of the combination (as computed from the simple
1563 per-column frequencies) is only 0.0027%, resulting in two orders of
1564 magnitude under-estimates.
1565 </para>
1567 <para>
1569 on combinations of columns that are actually used in conditions together,
1570 and for which misestimation of the number of groups is resulting in bad
1572 are just wasted.
1573 </para>
1574 </sect3>
1576 </sect2>
1577 </sect1>
1585 </indexterm>
1587 <para>
1588 It is possible
1590 syntax. To see why this matters, we first need some background.
1591 </para>
1593 <para>
1594 In a simple join query, such as:
1595 <programlisting>
1596 SELECT * FROM a, b, c WHERE a.id = b.id AND b.ref = c.id;
1597 </programlisting>
1598 the planner is free to join the given tables in any order. For
1599 example, it could generate a query plan that joins A to B, using
1602 condition. Or it could join B to C and then join A to that result.
1603 Or it could join A to C and then join them with B — but that
1604 would be inefficient, since the full Cartesian product of A and C
1605 would have to be formed, there being no applicable condition in the
1608 between two input tables, so it's necessary to build up the result
1609 in one or another of these fashions.) The important point is that
1610 these different join possibilities give semantically equivalent
1611 results but might have hugely different execution costs. Therefore,
1612 the planner will explore all of them to try to find the most
1613 efficient query plan.
1614 </para>
1616 <para>
1617 When a query only involves two or three tables, there aren't many join
1618 orders to worry about. But the number of possible join orders grows
1619 exponentially as the number of tables expands. Beyond ten or so input
1620 tables it's no longer practical to do an exhaustive search of all the
1621 possibilities, and even for six or seven tables planning might take an
1622 annoyingly long time. When there are too many input tables, the
1625 through a limited number of possibilities. (The switch-over threshold is
1627 parameter.)
1628 The genetic search takes less time, but it won't
1629 necessarily find the best possible plan.
1630 </para>
1632 <para>
1633 When the query involves outer joins, the planner has less freedom
1634 than it does for plain (inner) joins. For example, consider:
1635 <programlisting>
1636 SELECT * FROM a LEFT JOIN (b JOIN c ON (b.ref = c.id)) ON (a.id = b.id);
1637 </programlisting>
1638 Although this query's restrictions are superficially similar to the
1639 previous example, the semantics are different because a row must be
1640 emitted for each row of A that has no matching row in the join of B and C.
1641 Therefore the planner has no choice of join order here: it must join
1642 B to C and then join A to that result. Accordingly, this query takes
1643 less time to plan than the previous query. In other cases, the planner
1644 might be able to determine that more than one join order is safe.
1645 For example, given:
1646 <programlisting>
1647 SELECT * FROM a LEFT JOIN b ON (a.bid = b.id) LEFT JOIN c ON (a.cid = c.id);
1648 </programlisting>
1649 it is valid to join A to either B or C first. Currently, only
1652 can be rearranged to some extent.
1653 </para>
1655 <para>
1659 constrain the join order.
1660 </para>
1662 <para>
1664 the join order, it is possible to instruct the
1667 For example, these three queries are logically equivalent:
1668 <programlisting>
1669 SELECT * FROM a, b, c WHERE a.id = b.id AND b.ref = c.id;
1670 SELECT * FROM a CROSS JOIN b CROSS JOIN c WHERE a.id = b.id AND b.ref = c.id;
1671 SELECT * FROM a JOIN (b JOIN c ON (b.ref = c.id)) ON (a.id = b.id);
1672 </programlisting>
1674 the second and third take less time to plan than the first. This effect
1675 is not worth worrying about for only three tables, but it can be a
1676 lifesaver with many tables.
1677 </para>
1679 <para>
1680 To force the planner to follow the join order laid out by explicit
1683 (Other possible values are discussed below.)
1684 </para>
1686 <para>
1687 You do not need to constrain the join order completely in order to
1690 <programlisting>
1691 SELECT * FROM a CROSS JOIN b, c, d, e WHERE ...;
1692 </programlisting>
1694 forces the planner to join A to B before joining them to other tables,
1695 but doesn't constrain its choices otherwise. In this example, the
1696 number of possible join orders is reduced by a factor of 5.
1697 </para>
1699 <para>
1700 Constraining the planner's search in this way is a useful technique
1701 both for reducing planning time and for directing the planner to a
1702 good query plan. If the planner chooses a bad join order by default,
1704 — assuming that you know of a better order, that is. Experimentation
1705 is recommended.
1706 </para>
1708 <para>
1709 A closely related issue that affects planning time is collapsing of
1710 subqueries into their parent query. For example, consider:
1711 <programlisting>
1712 SELECT *
1713 FROM x, y,
1714 (SELECT * FROM a, b, c WHERE something) AS ss
1715 WHERE somethingelse;
1716 </programlisting>
1717 This situation might arise from use of a view that contains a join;
1719 reference, yielding a query much like the above. Normally, the planner
1720 will try to collapse the subquery into the parent, yielding:
1721 <programlisting>
1722 SELECT * FROM x, y, a, b, c WHERE something AND somethingelse;
1723 </programlisting>
1724 This usually results in a better plan than planning the subquery
1726 joining X to A first eliminates many rows of A, thus avoiding the need to
1727 form the full logical output of the subquery.) But at the same time,
1728 we have increased the planning time; here, we have a five-way join
1729 problem replacing two separate three-way join problems. Because of the
1730 exponential growth of the number of possibilities, this makes a big
1731 difference. The planner tries to avoid getting stuck in huge join search
1734 query. You can trade off planning time against quality of plan by
1735 adjusting this run-time parameter up or down.
1736 </para>
1738 <para>
1741 are similarly named because they do almost the same thing: one controls
1743 other controls when it will flatten out explicit joins. Typically
1747 to control join order with explicit joins). But you might set them
1748 differently if you are trying to fine-tune the trade-off between planning
1749 time and run time.
1750 </para>
1751 </sect1>
1756 <para>
1757 One might need to insert a large amount of data when first populating
1758 a database. This section contains some suggestions on how to make
1759 this process as efficient as possible.
1760 </para>
1765 <indexterm>
1768 </indexterm>
1770 <para>
1772 one commit at the end. (In plain
1775 do this behind your back, in which case you need to make sure the
1776 library does it when you want it done.) If you allow each
1777 insertion to be committed separately,
1779 each row that is added. An additional benefit of doing all
1780 insertions in one transaction is that if the insertion of one row
1781 were to fail then the insertion of all rows inserted up to that
1782 point would be rolled back, so you won't be stuck with partially
1783 loaded data.
1784 </para>
1785 </sect2>
1790 <para>
1792 all the rows in one command, instead of using a series of
1794 command is optimized for loading large numbers of rows; it is less
1797 is a single command, there is no need to disable autocommit if you
1798 use this method to populate a table.
1799 </para>
1801 <para>
1806 some of the overhead of repeatedly parsing and planning
1809 documentation.
1810 </para>
1812 <para>
1813 Note that loading a large number of rows using
1816 multiple insertions are batched into a single transaction.
1817 </para>
1819 <para>
1823 needs to be written, because in case of an error, the files
1824 containing the newly loaded data will be removed anyway.
1825 However, this consideration only applies when
1827 as all commands must write WAL otherwise.
1828 </para>
1830 </sect2>
1835 <para>
1836 If you are loading a freshly created table, the fastest method is to
1837 create the table, bulk load the table's data using
1839 table. Creating an index on pre-existing data is quicker than
1840 updating it incrementally as each row is loaded.
1841 </para>
1843 <para>
1844 If you are adding large amounts of data to an existing table,
1845 it might be a win to drop the indexes,
1846 load the table, and then recreate the indexes. Of course, the
1847 database performance for other users might suffer
1848 during the time the indexes are missing. One should also think
1849 twice before dropping a unique index, since the error checking
1850 afforded by the unique constraint will be lost while the index is
1851 missing.
1852 </para>
1853 </sect2>
1858 <para>
1859 Just as with indexes, a foreign key constraint can be checked
1861 useful to drop foreign key constraints, load data, and re-create
1862 the constraints. Again, there is a trade-off between data load
1863 speed and loss of error checking while the constraint is missing.
1864 </para>
1866 <para>
1867 What's more, when you load data into a table with existing foreign key
1868 constraints, each new row requires an entry in the server's list of
1869 pending trigger events (since it is the firing of a trigger that checks
1870 the row's foreign key constraint). Loading many millions of rows can
1871 cause the trigger event queue to overflow available memory, leading to
1872 intolerable swapping or even outright failure of the command. Therefore
1874 foreign keys when loading large amounts of data. If temporarily removing
1875 the constraint isn't acceptable, the only other recourse may be to split
1876 up the load operation into smaller transactions.
1877 </para>
1878 </sect2>
1883 <para>
1885 configuration variable when loading large amounts of data can
1886 lead to improved performance. This will help to speed up <command>CREATE
1889 only useful when you are using one or both of the above techniques.
1890 </para>
1891 </sect2>
1896 <para>
1898 configuration variable can also
1899 make large data loads faster. This is because loading a large
1901 cause checkpoints to occur more often than the normal checkpoint
1903 configuration variable). Whenever a checkpoint occurs, all dirty
1904 pages must be flushed to disk. By increasing
1906 data loads, the number of checkpoints that are required can be
1907 reduced.
1908 </para>
1909 </sect2>
1914 <para>
1915 When loading large amounts of data into an installation that uses
1916 WAL archiving or streaming replication, it might be faster to take a
1917 new base backup after the load has completed than to process a large
1918 amount of incremental WAL data. To prevent incremental WAL logging
1919 while loading, disable archiving and streaming replication, by setting
1923 But note that changing these settings requires a server restart,
1924 and makes any base backups taken before unavailable for archive
1925 recovery and standby server, which may lead to data loss.
1926 </para>
1928 <para>
1929 Aside from avoiding the time for the archiver or WAL sender to process the
1930 WAL data, doing this will actually make certain commands faster, because
1933 transaction) created or truncated the table or index they change. (They
1934 can guarantee crash safety more cheaply by doing
1936 </para>
1937 </sect2>
1942 <para>
1943 Whenever you have significantly altered the distribution of data
1944 within a table, running <link linkend="sql-analyze"><command>ANALYZE</command></link> is strongly recommended. This
1945 includes bulk loading large amounts of data into the table. Running
1947 ensures that the planner has up-to-date statistics about the
1948 table. With no statistics or obsolete statistics, the planner might
1949 make poor decisions during query planning, leading to poor
1950 performance on any tables with inaccurate or nonexistent
1951 statistics. Note that if the autovacuum daemon is enabled, it might
1955 </para>
1956 </sect2>
1961 <para>
1963 several, but not all, of the above guidelines. To restore a
1965 do a few extra things manually. (Note that these points apply while
1967 The same points apply whether loading a text dump with
1970 </para>
1972 <para>
1974 it is generating a complete schema-and-data dump, it is careful to
1975 load data before creating indexes and foreign keys. So in this case
1976 several guidelines are handled automatically. What is left
1977 for you to do is to:
1978 <itemizedlist>
1979 <listitem>
1980 <para>
1981 Set appropriate (i.e., larger than normal) values for
1984 </para>
1985 </listitem>
1986 <listitem>
1987 <para>
1988 If using WAL archiving or streaming replication, consider disabling
1993 Afterwards, set them back to the right values and take a fresh
1994 base backup.
1995 </para>
1996 </listitem>
1997 <listitem>
1998 <para>
1999 Experiment with the parallel dump and restore modes of both
2001 optimal number of concurrent jobs to use. Dumping and restoring in
2003 significantly higher performance over the serial mode.
2004 </para>
2005 </listitem>
2006 <listitem>
2007 <para>
2008 Consider whether the whole dump should be restored as a single
2012 mode, even the smallest of errors will rollback the entire restore,
2013 possibly discarding many hours of processing. Depending on how
2014 interrelated the data is, that might seem preferable to manual cleanup,
2016 transaction and have WAL archiving turned off.
2017 </para>
2018 </listitem>
2019 <listitem>
2020 <para>
2021 If multiple CPUs are available in the database server, consider using
2023 allows concurrent data loading and index creation.
2024 </para>
2025 </listitem>
2026 <listitem>
2027 <para>
2029 </para>
2030 </listitem>
2031 </itemizedlist>
2032 </para>
2034 <para>
2036 drop or recreate indexes, and it does not normally touch foreign
2037 keys.
2039 <footnote>
2040 <para>
2041 You can get the effect of disabling foreign keys by using
2043 that eliminates, rather than just postpones, foreign key
2044 validation, and so it is possible to insert bad data if you use it.
2045 </para>
2046 </footnote>
2048 So when loading a data-only dump, it is up to you to drop and recreate
2049 indexes and foreign keys if you wish to use those techniques.
2051 while loading the data, but don't bother increasing
2053 manually recreating indexes and foreign keys afterwards.
2057 </para>
2058 </sect2>
2059 </sect1>
2066 </indexterm>
2068 <para>
2069 Durability is a database feature that guarantees the recording of
2070 committed transactions even if the server crashes or loses
2071 power. However, durability adds significant database overhead,
2072 so if your site does not require such a guarantee,
2074 much faster. The following are configuration changes you can make
2075 to improve performance in such cases. Except as noted below, durability
2076 is still guaranteed in case of a crash of the database software;
2077 only an abrupt operating system crash creates a risk of data loss
2078 or corruption when these settings are used.
2080 <itemizedlist>
2081 <listitem>
2082 <para>
2083 Place the database cluster's data directory in a memory-backed
2085 database disk I/O, but limits data storage to the amount of
2086 available memory (and perhaps swap).
2087 </para>
2088 </listitem>
2090 <listitem>
2091 <para>
2093 data to disk.
2094 </para>
2095 </listitem>
2097 <listitem>
2098 <para>
2101 commit. This setting does risk transaction loss (though not data
2103 </para>
2104 </listitem>
2106 <listitem>
2107 <para>
2109 to guard against partial page writes.
2110 </para>
2111 </listitem>
2113 <listitem>
2114 <para>
2117 of checkpoints, but increases the storage requirements of
2119 </para>
2120 </listitem>
2122 <listitem>
2123 <para>
2126 makes the tables non-crash-safe.
2127 </para>
2128 </listitem>
2130 </itemizedlist>
2131 </para>
2132 </sect1>
2134 </chapter>