Are there any additional steps I can take to speed up query execution?
I have a table with more than 100m rows and I need to do search for matching strings. For that I checked two options:
- Compare text with to_tsvector
@@
(to_tsquery or plainto_tsquery)
This works very fast (under 1s on all data) but it has some problems with finding text similarity - Compare text with pg_trgm similarity
This works fine on text comparison but works bad on large amount of data.
I found that I can use indexes to improve performance.
For my GiST index I tried to increase siglen
from small number to 2024, but for some reason Postgres uses 512
and not higher.
CREATE INDEX trgm_idx_512_gg ON table USING GIST (name gist_trgm_ops(siglen=512));
Query:
SELECT name, similarity(name, 'ноутбук MSI GF63 Thin 10SC 086XKR 9S7 16R512 086') as sm
FROM table
WHERE name % 'ноутбук MSI GF63 Thin 10SC 086XKR 9S7 16R512 086'
EXPLAIN
output:
Bitmap Heap Scan on table (cost=1632.01..40051.57 rows=9737 width=126)
Recheck Cond: ((name)::text % 'ноутбук MSI GF63 Thin 10SC 086XKR 9S7 16R512 086'::text)
-> Bitmap Index Scan on trgm_idx_512_gg (cost=0.00..1629.57 rows=9737 width=0)
Index Cond: ((name)::text % 'ноутбук MSI GF63 Thin 10SC 086XKR 9S7 16R512 086'::text)
Execution time was about 120 sec.
Question
How can I improve or speed up query? Maybe I need to use a different approach or just add something else?
Output for EXPLAIN (ANALYZE, BUFFERS)
(searching for a different name so that the search is completely new and not from the cache):
Bitmap Heap Scan on table (cost=1632.01..40051.57 rows=9737 width=126) (actual time=159119.258..159960.251 rows=5645 loops=1)
Recheck Cond: ((name)::text % 'Чехол на realme C25s / Реалми Ц25с c рисунком / прозрачный с принтом, Andy&Paul'::text)
Heap Blocks: exact=3795
Buffers: shared read=1289378
-> Bitmap Index Scan on trgm_idx_512_gg (cost=0.00..1629.57 rows=9737 width=0) (actual time=159118.616..159118.616 rows=5645 loops=1)
Index Cond: ((name)::text % 'Чехол на realme C25s / Реалми Ц25с c рисунком / прозрачный с принтом, Andy&Paul'::text)
Buffers: shared read=1285583
Planning:
Buffers: shared read=5
Planning Time: 4.063 ms
Execution Time: 159961.121 ms
I also created a GIN index (but Postgres kept using the GiST):
CREATE INDEX gin_gg ON table USING GIN (name gin_trgm_ops);
Size: 12 GB.
GIST index: 31GB
2
Answers
Your indexes are correct, besides that, here are some strategies in order of importance according to this specific case to tune your query performance:
In your execution plan can be noticed that buffers are being read
instead of hit (means that data is missing from prostgres buffer
cache and is read from disk) see Using buffers for query
optimization.
Issue a
SELECT setting, unit FROM pg_settings WHERE name = 'shared_buffers';
to see the size of your database buffers (thatnumber must be mutiplied by unit, 80k usually to see actual size in
kbs).
It’s recommended 25% of server’s RAM for buffers, and usually
is left on it’s defult value, that is 16384 * 8kb = 128Mb, but
should be adapt to the situation. In tests i have seen a slight
improvement (not much when storage is in SSDs) in query performance, you can change this parameter in postgresql.conf file and a database restart is
required.
Peform a
vacuum verbose analyze <table>;
to check if there is datablocked by pending transactions (dead rows blocked) and at the same
time execute a vacuum analyze operation, the PostgreSQL query planner
relies on statistical information about the contents of tables in
order to generate good plans for queries. These statistics are
gathered by the ANALYZE command, which can be invoked by itself or as
an optional step in VACUUM. It is important to have reasonably
accurate statistics, otherwise poor choices of plans might degrade
database performance, see
vaccuming.
For many installations, it is sufficient to let vacuuming be
performed by the autovacuum daemon, but some database administrators
will want to supplement or replace the daemon’s activities with
manually-managed VACUUM commands, which typically are executed
according to a schedule by cron or Task Scheduler scripts. This
updates the visibility map, which speeds up index-only scans (which
in your case doesn’t fully apply, because mainly, data is obtained by
heap scan).
Could execute a
cluster <table> using <index>;
(when database isout of production because its a blocking method) to reorganize data
for faster access, however for this use case i don’t see performance
improvement, see sql-cluster.
A trigram GiST index with
siglen=512
on 100m rows is very large, and will probably never be cached efficiently. (Default issiglen=12
i.e. 12 bytes.) What makes you think this large signature would be a good choice? The manual:Looks like you went overboard with the size.
I have better experience with a trigram GIN indexes, especially in current versions of Postgres. If the query planner is confused by the existence of an additional GiST index, you’ll have to remove that one, to test results with the GIN index.
But first, to get a size comparison, look at the output of:
(Ideally, add the result to the questions.)
Your query plan shows vast amounts of
Buffers: shared read
for index and main relation (heap). So nothing was found in cache. The key to better performance will be to read fewer data pages to satisfy your queries, and more of them from cache:hit
instead ofread
in the query plan.Reducing the size of table and indexes helps in this reagard.
The selectivity of the trigram similarity operator
%
is set by the customized optionpg_trgm.similarity_threshold
. The default0.3
is rather lax and allows many hits. A higher similarity threshold will filter fewer (better matching) result rows. What do you do withrows=5645
result rows anyway? Try:Then retry your query.
See:
The latest version or Postgres, better server configuration and more RAM can also help in this regard. You disclosed no information about either of these.