Unsupported versions: 7.0
This documentation is for an unsupported version of PostgreSQL.
You may want to view the same page for the current version, or one of the other supported versions listed above instead.

Chapter 17. Understanding Performance

Query performance can be affected by many things. Some of these can be manipulated by the user, while others are fundamental to the underlying design of the system.

Some performance issues, such as index creation and bulk data loading, are covered elsewhere. This chapter will discuss the EXPLAIN command, and will show how the details of a query can affect the query plan, and hence overall performance.


Author: Written by Tom Lane, from e-mail dated 2000-03-27.

Plan-reading is an art that deserves a tutorial, and I haven't had time to write one. Here is some quick & dirty explanation.

The numbers that are currently quoted by EXPLAIN are:

  • Estimated startup cost (time expended before output scan can start, eg, time to do the sorting in a SORT node).

  • Estimated total cost (if all tuples are retrieved, which they may not be --- LIMIT will stop short of paying the total cost, for example).

  • Estimated number of rows output by this plan node.

  • Estimated average width (in bytes) of rows output by this plan node.

The costs are measured in units of disk page fetches. (CPU effort estimates are converted into disk-page units using some fairly arbitrary fudge-factors. See the SET reference page if you want to experiment with these.) It's important to note that the cost of an upper-level node includes the cost of all its child nodes. It's also important to realize that the cost only reflects things that the planner/optimizer cares about. In particular, the cost does not consider the time spent transmitting result tuples to the frontend --- which could be a pretty dominant factor in the true elapsed time, but the planner ignores it because it cannot change it by altering the plan. (Every correct plan will output the same tuple set, we trust.)

Rows output is a little tricky because it is not the number of rows processed/scanned by the query --- it is usually less, reflecting the estimated selectivity of any WHERE-clause constraints that are being applied at this node.

Average width is pretty bogus because the thing really doesn't have any idea of the average length of variable-length columns. I'm thinking about improving that in the future, but it may not be worth the trouble, because the width isn't used for very much.

Here are some examples (using the regress test database after a vacuum analyze, and almost-7.0 sources):

regression=# explain select * from tenk1;

Seq Scan on tenk1  (cost=0.00..333.00 rows=10000 width=148)

This is about as straightforward as it gets. If you do

select * from pg_class where relname = 'tenk1';
you'll find out that tenk1 has 233 disk pages and 10000 tuples. So the cost is estimated at 233 block reads, defined as 1.0 apiece, plus 10000 * cpu_tuple_cost which is currently 0.01 (try show cpu_tuple_cost).

Now let's modify the query to add a qualification clause:

regression=# explain select * from tenk1 where unique1 < 1000;

Seq Scan on tenk1  (cost=0.00..358.00 rows=1000 width=148)
The estimate of output rows has gone down because of the WHERE clause. (The uncannily accurate estimate is just because tenk1 is a particularly simple case --- the unique1 column has 10000 distinct values ranging from 0 to 9999, so the estimator's linear interpolation between min and max column values is dead-on.) However, the scan will still have to visit all 10000 rows, so the cost hasn't decreased; in fact it has gone up a bit to reflect the extra CPU time spent checking the WHERE condition.

Modify the query to restrict the qualification even more:

regression=# explain select * from tenk1 where unique1 < 100;

Index Scan using tenk1_unique1 on tenk1  (cost=0.00..89.35 rows=100 width=148)
and you will see that if we make the WHERE condition selective enough, the planner will eventually decide that an indexscan is cheaper than a sequential scan. This plan will only have to visit 100 tuples because of the index, so it wins despite the fact that each individual fetch is expensive.

Add another condition to the qualification:

regression=# explain select * from tenk1 where unique1 < 100 and
regression-# stringu1 = 'xxx';

Index Scan using tenk1_unique1 on tenk1  (cost=0.00..89.60 rows=1 width=148)
The added clause "stringu1 = 'xxx'" reduces the output-rows estimate, but not the cost because we still have to visit the same set of tuples.

Let's try joining two tables, using the fields we have been discussing:

regression=# explain select * from tenk1 t1, tenk2 t2 where t1.unique1 < 100
regression-# and t1.unique2 = t2.unique2;

Nested Loop  (cost=0.00..144.07 rows=100 width=296)
  ->  Index Scan using tenk1_unique1 on tenk1 t1
             (cost=0.00..89.35 rows=100 width=148)
  ->  Index Scan using tenk2_unique2 on tenk2 t2
             (cost=0.00..0.53 rows=1 width=148)

In this nested-loop join, the outer scan is the same indexscan we had in the example before last, and so its cost and row count are the same because we are applying the "unique1 < 100" WHERE clause at that node. The "t1.unique2 = t2.unique2" clause isn't relevant yet, so it doesn't affect the outer scan's row count. For the inner scan, the current outer-scan tuple's unique2 value is plugged into the inner indexscan to produce an indexqual like "t2.unique2 = constant". So we get the same inner-scan plan and costs that we'd get from, say, "explain select * from tenk2 where unique2 = 42". The loop node's costs are then set on the basis of the outer scan's cost, plus one repetition of the inner scan for each outer tuple (100 * 0.53, here), plus a little CPU time for join processing.

In this example the loop's output row count is the same as the product of the two scans' row counts, but that's not true in general, because in general you can have WHERE clauses that mention both relations and so can only be applied at the join point, not to either input scan. For example, if we added "WHERE ... AND t1.hundred < t2.hundred", that'd decrease the output row count of the join node, but not change either input scan.

We can look at variant plans by forcing the planner to disregard whatever strategy it thought was the winner (a pretty crude tool, but it's what we've got at the moment):

regression=# set enable_nestloop = off;
regression=# explain select * from tenk1 t1, tenk2 t2 where t1.unique1 < 100
regression-# and t1.unique2 = t2.unique2;

Hash Join  (cost=89.60..574.10 rows=100 width=296)
  ->  Seq Scan on tenk2 t2
               (cost=0.00..333.00 rows=10000 width=148)
  ->  Hash  (cost=89.35..89.35 rows=100 width=148)
        ->  Index Scan using tenk1_unique1 on tenk1 t1
               (cost=0.00..89.35 rows=100 width=148)
This plan proposes to extract the 100 interesting rows of tenk1 using ye same olde indexscan, stash them into an in-memory hash table, and then do a sequential scan of tenk2, probing into the hash table for possible matches of "t1.unique2 = t2.unique2" at each tenk2 tuple. The cost to read tenk1 and set up the hash table is entirely startup cost for the hash join, since we won't get any tuples out until we can start reading tenk2. The total time estimate for the join also includes a pretty hefty charge for CPU time to probe the hash table 10000 times. Note, however, that we are NOT charging 10000 times 89.35; the hash table setup is only done once in this plan type.