Tuesday, April 12, 2016

100x faster Postgres performance by changing 1 line

A simple but non-obvious one-line change (ANY(ARRAY[...]) toANY(VALUES(...))) in a (bad) PostgreSQL 9.0 query cuts query time from 20s to 0.2s. Starting with low-level metrics we make our way to your best friend: EXPLAIN ANALYZE. The amount of time invested will pay off a hundred times over. The Postgres community is your second best friend.

Monitoring slow Postgres queries with Postgres

Earlier this week the performance of one of our (many) databases was plagued by a few pathologically large, primary-key queries in a smallish table (10 GB, 15 million rows) used to feed our graph editor.
In 99.9% of accounts these queries would be zippy. In a few cases where the number of tags used to annotate metrics is large, these queries would take up to 20 seconds. That much time spent in a database meant that someone was waiting in front of their browser for the graph editor to respond. This is obviously not the best experience for the unlucky 0.1%.
Here is the surprisingly simple culprit:
SELECT c.key,
c.x_key,
c.tags,
x.name
FROM context c
JOIN x
ON c.x_key = x.key
WHERE c.key = ANY (ARRAY[15368196, -- 11,000 other keys --)])
AND c.x_key = 1
AND c.tags @> ARRAY[E'blah'];
view rawslow-query.sql hosted with ❤ by GitHub
Table X contains a few thousand rows, Table C contains 15 millions rows. Both “key” columns are primary keys with proper indexing. Pretty straightforward, it’s a simple primary key lookup. Where things get interesting is when you increase the number of keys. At 11,000 keys we get the following plan, by prefixing EXPLAIN (ANALYZE, BUFFERS) to the query:
Nested Loop (cost=6923.33..11770.59 rows=1 width=362) (actual time=17128.188..22109.283 rows=10858 loops=1)
Buffers: shared hit=83494
-> Bitmap Heap Scan on context c (cost=6923.33..11762.31 rows=1 width=329) (actual time=17128.121..22031.783 rows=10858 loops=1)
Recheck Cond: ((tags @> '{blah}'::text[]) AND (x_key = 1))
Filter: (key = ANY ('{15368196,(a lot more keys here)}'::integer[]))
Buffers: shared hit=50919
-> BitmapAnd (cost=6923.33..6923.33 rows=269 width=0) (actual time=132.910..132.910 rows=0 loops=1)
Buffers: shared hit=1342
-> Bitmap Index Scan on context_tags_idx (cost=0.00..1149.61 rows=15891 width=0) (actual time=64.614..64.614 rows=264777 loops=1)
Index Cond: (tags @> '{blah}'::text[])
Buffers: shared hit=401
-> Bitmap Index Scan on context_x_id_source_type_id_idx (cost=0.00..5773.47 rows=268667 width=0) (actual time=54.648..54.648 rows=267659 loops=1)
Index Cond: (x_id = 1)
Buffers: shared hit=941
-> Index Scan using x_pkey on x (cost=0.00..8.27 rows=1 width=37) (actual time=0.003..0.004 rows=1 loops=10858)
Index Cond: (x.key = 1)
Buffers: shared hit=32575
Total runtime: 22117.417 ms
view rawgistfile1.sql hosted with ❤ by GitHub
As you can see at the bottom of the plan, the query took 22 seconds to execute. These 22 seconds can be visualized on the following graph as pure CPU execution split 90/10 between Postgres and the OS; very little disk I/O.
Postgres
At the lowest level these queries would look like these spikes of CPU utilization. CPU graphs are seldom useful but in this case it confirmed a crucial point: the database was not waiting for the disks to read data. It was doing things like sorting, hashing and comparing rows.
The second interesting metric that tracks the spikes very closely is the number of rows that are “fetched” by Postgres (in this case not returned, just looked at and discarded).
Postgres row stats
Clearly something is regularly and methodically going through a lot of rows: our query.

The Postgres Performance Problem: Bitmap Heap Scan

The rows_fetched metric is consistent with the following part of the plan:
Buffers: shared hit=83494
-> Bitmap Heap Scan on context c (cost=6923.33..11762.31 rows=1 width=329) (actual time=17128.121..22031.783 rows=10858 loops=1)
Recheck Cond: ((tags @> '{blah}'::text[]) AND (x_key = 1))
Filter: (key = ANY ('{15368196,(a lot more keys here)}'::integer[]))
Buffers: shared hit=50919
view rawplan-detail-blog.sql hosted with ❤ by GitHub
Postgres is reading Table C using a Bitmap Heap Scan. When the number of keys to check stays small, it can efficiently use the index to build the bitmap in memory. If the bitmap gets too large, the query optimizer changes the way it looks up data. In our case it has a large number of keys to check so it uses the more approximative way to retrieve the candidate rows and checks each row individually for a match on x_key and tags. All this “loading in memory” and “checking individual row” takes time (the Recheck Cond in the plan). Luckily for us the table is 30% loaded in RAM so it is not as bad as retrieving the rows from disk. It still has a very noticeable impact on performance. Remember that the query is quite simple. It’s a primary key lookup so there aren’t many obvious ways to fix it without dramatically re-architecting the database or the application. We turned to the community for help via the PGSQL-Performance mailing list.

The solution

This is yet another reason why we love open-source and its culture of helping users. Tom Lane, one of the most prolific open-source authors around, suggested to try this instead:
SELECT c.key,
c.x_key,
c.tags,
x.name
FROM context c
JOIN x
ON c.x_key = x.key
WHERE c.key = ANY (VALUES (15368196), -- 11,000 other keys --)
AND c.x_key = 1
AND c.tags @> ARRAY[E'blah'];
Can you spot the difference? ARRAY has been changed to VALUES.
The query optimizer is fooled by the use of ARRAY[...] to list all the primary keys to look up. VALUES (...), (...) lets the optimizer make full use of the primary key index instead. It is literally a one-line change, which makes no semantic difference.
Here is the plan of the new query. The 2 major differences are on lines 3 and 14.
Nested Loop (cost=168.22..2116.29 rows=148 width=362) (actual time=22.134..256.531 rows=10858 loops=1)
Buffers: shared hit=44967
-> Index Scan using x_pkey on x (cost=0.00..8.27 rows=1 width=37) (actual time=0.071..0.073 rows=1 loops=1)
Index Cond: (id = 1)
Buffers: shared hit=4
-> Nested Loop (cost=168.22..2106.54 rows=148 width=329) (actual time=22.060..242.406 rows=10858 loops=1)
Buffers: shared hit=44963
-> HashAggregate (cost=168.22..170.22 rows=200 width=4) (actual time=21.529..32.820 rows=11215 loops=1)
-> Values Scan on "*VALUES*" (cost=0.00..140.19 rows=11215 width=4) (actual time=0.005..9.527 rows=11215 loops=1)
-> Index Scan using context_pkey on context c (cost=0.00..9.67 rows=1 width=329) (actual time=0.015..0.016 rows=1 loops=11215)
Index Cond: (c.key = "*VALUES*".column1)
Filter: ((c.tags @> '{blah}'::text[]) AND (c.x_id = 1))
Buffers: shared hit=44963
Total runtime: 263.639 ms
From 22,000ms to 200ms. That’s a 100x speedup for a simple one-line change.

The new query in production

A short code deploy later:
Hotfix to put in Postgres performance change
And the database looks much happier.
Postgres cpu release marker
Postgres row stats with marker

The takeaway

The slow Postgres query is gone. The 0.1% unlucky few who would have been affected by the issue are happy too. Seeing the impact of the change using Datadog allowed us to instantly validate that altering that part of the query was the right thing to do.
Update: this change need only be applied on 9.0. As of Postgres 9.3,ANY(ARRAY[...]) is performing properly again.
You can get that visibility set up in minutes by signing up for a free Datadog account if you’re looking to try out this change to your Postgres queries.

No comments:

Post a Comment