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. This chapter provides some hints about understanding and tuning PostgreSQL performance.
PostgreSQL devises a query plan for each query it is given. Choosing the right plan to match the query structure and the properties of the data is absolutely critical for good performance. You can use the EXPLAIN command to see what query plan the system creates for any query. Plan-reading is an art that deserves an extensive tutorial, which this is not; but here is some basic information.
The numbers that are currently quoted by EXPLAIN are:
Estimated start-up cost (time expended before output scan can start, e.g., time to do the sorting in a SORT node).
Estimated total cost (if all tuples are retrieved, which they may not be --- a query with a LIMIT will stop short of paying the total cost, for example).
Estimated number of rows output by this plan node (again, without regard for any LIMIT).
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. If you want to experiment with these factors, see the list of run-time configuration parameters in the Administrator's Guide.)
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. Ideally the top-level rows estimate will approximate the number of rows actually returned, updated, or deleted by the query.
Here are some examples (using the regress test database after a vacuum analyze, and 7.2 development sources):
regression=# EXPLAIN SELECT * FROM tenk1; NOTICE: QUERY PLAN: 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 will find out that
has 233 disk pages and 10000 tuples. So the cost is estimated
at 233 page 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; NOTICE: QUERY PLAN: Seq Scan on tenk1 (cost=0.00..358.00 rows=1007 width=148)
The estimate of output rows has gone down because of the WHERE clause. 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.
The actual number of rows this query would select is 1000, but the estimate is only approximate. If you try to duplicate this experiment, you will probably get a slightly different estimate; moreover, it will change after each ANALYZE command, because the statistics produced by ANALYZE are taken from a randomized sample of the table.
Modify the query to restrict the qualification even more:
regression=# EXPLAIN SELECT * FROM tenk1 WHERE unique1 < 50; NOTICE: QUERY PLAN: Index Scan using tenk1_unique1 on tenk1 (cost=0.00..181.09 rows=49 width=148)
and you will see that if we make the WHERE condition selective enough, the planner will eventually decide that an index scan is cheaper than a sequential scan. This plan will only have to visit 50 tuples because of the index, so it wins despite the fact that each individual fetch is more expensive than reading a whole disk page sequentially.
Add another condition to the qualification:
regression=# EXPLAIN SELECT * FROM tenk1 WHERE unique1 < 50 AND regression-# stringu1 = 'xxx'; NOTICE: QUERY PLAN: Index Scan using tenk1_unique1 on tenk1 (cost=0.00..181.22 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 < 50 regression-# AND t1.unique2 = t2.unique2; NOTICE: QUERY PLAN: Nested Loop (cost=0.00..330.41 rows=49 width=296) -> Index Scan using tenk1_unique1 on tenk1 t1 (cost=0.00..181.09 rows=49 width=148) -> Index Scan using tenk2_unique2 on tenk2 t2 (cost=0.00..3.01 rows=1 width=148)
In this nested-loop join, the outer scan is the same index scan we had in the example before last, and so its cost and row count are the same because we are applying the unique1 < 50 WHERE clause at that node. The t1.unique2 = t2.unique2 clause is not relevant yet, so it doesn't affect row count of the outer scan. For the inner scan, the unique2 value of the current outer-scan tuple is plugged into the inner index scan to produce an index qualification 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 costs of the loop node are then set on the basis of the cost of the outer scan, plus one repetition of the inner scan for each outer tuple (49 * 3.01, 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 would decrease the output row count of the join node, but not change either input scan.
One way to look at variant plans is to force the planner to disregard whatever strategy it thought was the winner, using the enable/disable flags for each plan type. (This is a crude tool, but useful. See also Section 11.3.)
regression=# set enable_nestloop = off; SET VARIABLE regression=# EXPLAIN SELECT * FROM tenk1 t1, tenk2 t2 WHERE t1.unique1 < 50 regression-# AND t1.unique2 = t2.unique2; NOTICE: QUERY PLAN: Hash Join (cost=181.22..564.83 rows=49 width=296) -> Seq Scan on tenk2 t2 (cost=0.00..333.00 rows=10000 width=148) -> Hash (cost=181.09..181.09 rows=49 width=148) -> Index Scan using tenk1_unique1 on tenk1 t1 (cost=0.00..181.09 rows=49 width=148)
This plan proposes to extract the 50 interesting rows of
tenk1 using ye same olde index
scan, stash them into an in-memory hash table, and then do a
sequential scan of
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
start-up 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 hefty charge for CPU time to probe the hash
table 10000 times. Note, however, that we are NOT charging
10000 times 181.09; the hash table setup is only done once in
this plan type.
It is possible to check on the accuracy of the planner's estimated costs by using EXPLAIN ANALYZE. This command actually executes the query, and then displays the true runtime accumulated within each plan node along with the same estimated costs that a plain EXPLAIN shows. For example, we might get a result like this:
regression=# EXPLAIN ANALYZE regression-# SELECT * FROM tenk1 t1, tenk2 t2 regression-# WHERE t1.unique1 < 50 AND t1.unique2 = t2.unique2; NOTICE: QUERY PLAN: Nested Loop (cost=0.00..330.41 rows=49 width=296) (actual time=1.31..28.90 rows=50 loops=1) -> Index Scan using tenk1_unique1 on tenk1 t1 (cost=0.00..181.09 rows=49 width=148) (actual time=0.69..8.84 rows=50 loops=1) -> Index Scan using tenk2_unique2 on tenk2 t2 (cost=0.00..3.01 rows=1 width=148) (actual time=0.28..0.31 rows=1 loops=50) Total runtime: 30.67 msec
Note that the "actual time" values are in milliseconds of real time, whereas the "cost" estimates are expressed in arbitrary units of disk fetches; so they are unlikely to match up. The thing to pay attention to is the ratios.
In some query plans, it is possible for a subplan node to be executed more than once. For example, the inner index scan is executed once per outer tuple in the above nested-loop plan. In such cases, the "loops" value reports the total number of executions of the node, and the actual time and rows values shown are averages per-execution. This is done to make the numbers comparable with the way that the cost estimates are shown. Multiply by the "loops" value to get the total time actually spent in the node.
The "total runtime" shown by EXPLAIN ANALYZE includes executor startup and shutdown time, as well as time spent processing the result tuples. It does not include parsing, rewriting, or planning time. For a SELECT query, the total runtime will normally be just a little larger than the total time reported for the top-level plan node. For INSERT, UPDATE, and DELETE queries, the total runtime may be considerably larger, because it includes the time spent processing the output tuples. In these queries, the time for the top plan node essentially is the time spent computing the new tuples and/or locating the old ones, but it doesn't include the time spent making the changes.
It is worth noting that EXPLAIN results should not be extrapolated to situations other than the one you are actually testing; for example, results on a toy-sized table can't be assumed to apply to large tables. The planner's cost estimates are not linear and so it may well choose a different plan for a larger or smaller table. An extreme example is that on a table that only occupies one disk page, you'll nearly always get a sequential scan plan whether indexes are available or not. The planner realizes that it's going to take one disk page read to process the table in any case, so there's no value in expending additional page reads to look at an index.
Also see the parameters taken by postmaster on startup.
Note the -B BUFFERS and -f (fsync() options).
# postmaster --help
postmaster is the PostgreSQL server.
-B NBUFFERS number of shared buffers (default 64)
-c NAME=VALUE set run-time parameter
-d 1-5 debugging level
-D DATADIR database directory
-F turn fsync off
-h HOSTNAME host name or IP address to listen on
-i enable TCP/IP connections
-k DIRECTORY Unix-domain socket location
-l enable SSL connections
-N MAX-CONNECT maximum number of allowed connections (default 32)
-o OPTIONS pass 'OPTIONS' to each backend server
-p PORT port number to listen on (default 5432)
-S silent mode (start in background without logging output)
Since postgresql have this (imho) nasty habit of torturing some filesystems by creating very many files, it would also be nice to see documentation (and maybe even benchmarks) of postgresql performance on different filesystems.
More performance tips regarding the postgresql.conf file would also be nice to have.
The performance of different types of database drivers such as DBI/DBD, ODBC etc. would also be nice to gain some more knowledge about.