Full-text Search Index Optimization — Alibaba Cloud RDS PostgreSQL Best Practices
Background
With built-in GIN indexes, PostgreSQL supports full-text search and searching multi-value data types including arrays.
Will indexes be used in full-text searches that do not contain a certain keyword?
GIN stands for Generalized Inverted Index. A query that does not contain a keyword is actually a scanning process that skips the token above the main tree.
It will be cost-effective if the skipped token contains a relatively large amount of data. PostgreSQL uses a CBO (cost-based optimizer) execution plan optimizer to automatically choose the best index.
Example 1. Full-text search — “Does Not Contain” query
1.Create a test table
postgres=# create table notcontain (id int, info tsvector);
CREATE TABLE2.Create a function to generate random strings
CREATE OR REPLACE FUNCTION
gen_rand_str(integer)
RETURNS text
LANGUAGE sql
STRICT
AS $function$
select string_agg(a[(random()*6)::int+1],'') from generate_series(1,$1), (select array['a','b','c','d','e','f',' ']) t(a);
$function$;3.Insert 1 million entries of test data
postgres=# insert into notcontain select generate_series(1,1000000), to_tsvector(gen_rand_str(256));
4.Create full-text index (a GIN index)
create index idx_notcontain_info on notcontain using gin (info);
5.Query one of the records
postgres=# select * from notcontain limit 1;
-[ RECORD 1 ]----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
id | 1
info | 'afbbeeccbf':3 'b':16 'bdcdfd':2 'bdcfbcecdeeaed':8 'bfedfecbfab':7 'cd':9 'cdcaefaccdccadeafadededddcbdecdaefbcfbdaefcec':14 'ceafecff':6 'd':17,18 'dbc':12 'dceabcdcbdca':10 'dddfdbffffeaca':13 'deafcccfbcdebdaecda':11 'dfbadcdebdedbfa':19 'eb':15 'ebe':1 'febdcbdaeaeabbdadacabdbbedfafcaeabbdcedaeca':5 'fedeecbcdfcdceabbabbfcdd':46.Test — Does Not Contain a keyword
The database automatically chooses Full Table Scan, and does not use the GIN index.
Why wasn’t an index used? As I have explained before, the number of data records containing the keyword is quite small. It is not cost-effective to use indexes for filtering in a query that does not contain a keyword.
(However, if the query contains a keyword, it will be very cost-effective to use the GIN indexes. “Contains” and “Does Not Contain” are inversions of each other, and so are the associated costs.)
select * from notcontain t1 where info @@ to_tsquery ('!eb');
postgres=# explain (analyze,verbose,timing,costs,buffers) select * from notcontain t1 where info @@ to_tsquery ('!eb');
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------
Seq Scan on postgres.notcontain t1 (cost=0.00..318054.51 rows=950820 width=412) (actual time=0.016..1087.463 rows=947911 loops=1)
Output: id, info
Filter: (t1.info @@ to_tsquery('!eb'::text))
Rows Removed by Filter: 52089
Buffers: shared hit=55549
Planning time: 0.131 ms
Execution time: 1134.571 ms
(7 rows)7.Force disable Full Table Scan, and allow the database to choose the indexes.
We can see that, searches that use index are really slow. Most of the time we can trust the database to find the most efficient method to complete our queries. (All of these can be calibrated as long as cost factors and environmental performance measuring are accurate enough.)
postgres=# set enable_seqscan=off;
SET
postgres=# explain (analyze,verbose,timing,costs,buffers) select * from notcontain t1 where info @@ to_tsquery ('!eb');
QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------------------
Bitmap Heap Scan on postgres.notcontain t1 (cost=82294981.00..82600120.25 rows=950820 width=412) (actual time=1325.587..1540.145 rows=947911 loops=1)
Output: id, info
Recheck Cond: (t1.info @@ to_tsquery('!eb'::text))
Heap Blocks: exact=55549
Buffers: shared hit=171948
-> Bitmap Index Scan on idx_notcontain_info (cost=0.00..82294743.30 rows=950820 width=0) (actual time=1315.663..1315.663 rows=947911 loops=1)
Index Cond: (t1.info @@ to_tsquery('!eb'::text))
Buffers: shared hit=116399
Planning time: 0.135 ms
Execution time: 1584.670 ms
(10 rows)Example 2. Full-text search — “Does Not Contain” query
In this example, I am going to create unevenly distributed data. This token contains a large number of repeated content, and I will filter them out with a “Does Not Contain” operator. Let’s see if the index will be used.
1.Generate test data
postgres=# truncate notcontain ;
TRUNCATE TABLE
postgres=# insert into notcontain select generate_series(1,1000000), to_tsvector('abc');
INSERT 0 10000002.Test search entries that do not contain ABC
The database automatically chooses indexed scanning and skips data blocks that do not need to be searched.
postgres=# explain (analyze,verbose,timing,costs,buffers) select * from notcontain t1 where info @@ to_tsquery ('!abc');
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------------
Bitmap Heap Scan on postgres.notcontain t1 (cost=220432.15..220433.71 rows=1 width=21) (actual time=107.936..107.936 rows=0 loops=1)
Output: id, info
Recheck Cond: (t1.info @@ to_tsquery('!abc'::text))
Buffers: shared hit=268
-> Bitmap Index Scan on idx_notcontain_info (cost=0.00..220432.15 rows=1 width=0) (actual time=107.933..107.933 rows=0 loops=1)
Index Cond: (t1.info @@ to_tsquery('!abc'::text))
Buffers: shared hit=268
Planning time: 0.183 ms
Execution time: 107.962 ms
(9 rows)3.Force enable Full Table Scan. We can see that the performance is truly no match for indexed scanning, which proves that PostgreSQL is a cost-based optimizer and can automatically choose the most optimal execution plan.
postgres=# set enable_bitmapscan =off;
SET
postgres=# explain (analyze,verbose,timing,costs,buffers) select * from notcontain t1 where info @@ to_tsquery ('!abc');
QUERY PLAN
----------------------------------------------------------------------------------------------------------------------------
Seq Scan on postgres.notcontain t1 (cost=0.00..268870.00 rows=1 width=21) (actual time=1065.436..1065.436 rows=0 loops=1)
Output: id, info
Filter: (t1.info @@ to_tsquery('!abc'::text))
Rows Removed by Filter: 1000000
Buffers: shared hit=6370
Planning time: 0.059 ms
Execution time: 1065.449 ms
(7 rows)Example 3. Ordinary B-tree indexes — “Not Equal To” searches
This example is an ordinary search that uses a B-tree index. Let’s see if PostgreSQL supports the “Not Equal To” indexed search.
The test method is similar to that of GIN test. In this test, I am going to use both unevenly and evenly distributed data.
1.For a query that uses the “Does Not Contain” operator on unevenly distributed data, the number of records filtered out with indexes is very small.
So far, “Does Not Contain” searches that uses the B-tree indexes are not supported at the kernel level. (Although this can be achieved by skipping Branch nodes that do not need to be scanned with INDEX SKIP SCAN.)
postgres=# truncate notcontain ;
TRUNCATE TABLE
postgres=# insert into notcontain select generate_series(1,1000000);
INSERT 0 1000000
postgres=# create index idx1 on notcontain (id);
CREATE INDEX
postgres=# set enable_bitmapscan =on;
SET
postgres=# explain (analyze,verbose,timing,costs,buffers) select * from notcontain t1 where id<>1;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------
Seq Scan on postgres.notcontain t1 (cost=0.00..16925.00 rows=999999 width=36) (actual time=0.011..110.592 rows=999999 loops=1)
Output: id, info
Filter: (t1.id <> 1)
Rows Removed by Filter: 1
Buffers: shared hit=4425
Planning time: 0.195 ms
Execution time: 156.013 ms
(7 rows)
postgres=# set enable_seqscan=off;
SET
postgres=# explain (analyze,verbose,timing,costs,buffers) select * from notcontain t1 where id<>1;
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------------------------------
Seq Scan on postgres.notcontain t1 (cost=10000000000.00..10000016925.00 rows=999999 width=36) (actual time=0.011..110.964 rows=999999 loops=1)
Output: id, info
Filter: (t1.id <> 1)
Rows Removed by Filter: 1
Buffers: shared hit=4425
Planning time: 0.062 ms
Execution time: 156.461 ms
(7 rows)2.Change the SQL write method to achieve indexed search. As INDEX SKIP SCAN is not used in this case, we need a JOIN process. The resulting performance is not very good.
postgres=# explain (analyze,verbose,timing,costs,buffers) select * from notcontain t1 where not exists (select 1 from notcontain t2 where t1.id=t2.id and t2.id=1);
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------------
Merge Anti Join (cost=0.85..25497.28 rows=999999 width=36) (actual time=0.023..277.639 rows=999999 loops=1)
Output: t1.id, t1.info
Merge Cond: (t1.id = t2.id)
Buffers: shared hit=7164
-> Index Scan using idx1 on postgres.notcontain t1 (cost=0.42..22994.22 rows=1000000 width=36) (actual time=0.009..148.520 rows=1000000 loops=1)
Output: t1.id, t1.info
Buffers: shared hit=7160
-> Index Only Scan using idx1 on postgres.notcontain t2 (cost=0.42..3.04 rows=1 width=4) (actual time=0.007..0.008 rows=1 loops=1)
Output: t2.id
Index Cond: (t2.id = 1)
Heap Fetches: 1
Buffers: shared hit=4
Planning time: 0.223 ms
Execution time: 322.798 ms
(14 rows)
postgres=# set enable_mergejoin=off;
SET
postgres=# explain (analyze,verbose,timing,costs,buffers) select * from notcontain t1 where not exists (select 1 from notcontain t2 where t1.id=t2.id and t2.id=1);
QUERY PLAN
----------------------------------------------------------------------------------------------------------------------------------------------
Hash Anti Join (cost=3.05..27053.05 rows=999999 width=36) (actual time=0.060..251.232 rows=999999 loops=1)
Output: t1.id, t1.info
Hash Cond: (t1.id = t2.id)
Buffers: shared hit=4432
-> Seq Scan on postgres.notcontain t1 (cost=0.00..14425.00 rows=1000000 width=36) (actual time=0.011..84.659 rows=1000000 loops=1)
Output: t1.id, t1.info
Buffers: shared hit=4425
-> Hash (cost=3.04..3.04 rows=1 width=4) (actual time=0.014..0.014 rows=1 loops=1)
Output: t2.id
Buckets: 1024 Batches: 1 Memory Usage: 9kB
Buffers: shared hit=4
-> Index Only Scan using idx1 on postgres.notcontain t2 (cost=0.42..3.04 rows=1 width=4) (actual time=0.010..0.011 rows=1 loops=1)
Output: t2.id
Index Cond: (t2.id = 1)
Heap Fetches: 1
Buffers: shared hit=4
Planning time: 0.147 ms
Execution time: 297.127 ms
(18 rows)
postgres=# set enable_seqscan=off;
SET
postgres=# explain (analyze,verbose,timing,costs,buffers) select * from notcontain t1 where not exists (select 1 from notcontain t2 where t1.id=t2.id and t2.id=1);
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------------
Hash Anti Join (cost=3.48..35622.27 rows=999999 width=36) (actual time=0.036..324.401 rows=999999 loops=1)
Output: t1.id, t1.info
Hash Cond: (t1.id = t2.id)
Buffers: shared hit=7164
-> Index Scan using idx1 on postgres.notcontain t1 (cost=0.42..22994.22 rows=1000000 width=36) (actual time=0.017..149.383 rows=1000000 loops=1)
Output: t1.id, t1.info
Buffers: shared hit=7160
-> Hash (cost=3.04..3.04 rows=1 width=4) (actual time=0.011..0.011 rows=1 loops=1)
Output: t2.id
Buckets: 1024 Batches: 1 Memory Usage: 9kB
Buffers: shared hit=4
-> Index Only Scan using idx1 on postgres.notcontain t2 (cost=0.42..3.04 rows=1 width=4) (actual time=0.008..0.009 rows=1 loops=1)
Output: t2.id
Index Cond: (t2.id = 1)
Heap Fetches: 1
Buffers: shared hit=4
Planning time: 0.141 ms
Execution time: 369.749 ms
(18 rows)3.PostgreSQL uses multi-core parallel computing to allow incredible performance improvements to Full Table Scan.
Parallel scanning can remarkably improve the performance in case with a significantly large number of records.
postgres=# create unlogged table ptbl(id int);
CREATE TABLE
postgres=# insert into ptbl select generate_series(1,100000000);
postgres=# alter table ptbl set (parallel_workers =32);
\timing
Nonparallel query:
postgres=# set max_parallel_workers_per_gather =0;
postgres=# select count(*) from ptbl where id<>1;
count
----------
99999999
(1 row)
Time: 11863.151 ms (00:11.863)
并行查询：
postgres=# set max_parallel_workers_per_gather =32;
postgres=# select count(*) from ptbl where id<>1;
count
----------
99999999
(1 row)
Time: 610.017 msPerformance see marked improvement when we use a parallel query.
Example 4. Ordinary partial B-tree indexes — “Not Equal To” searches
For fixed “Not Equal To” queries, we can use the partial index function of PostgreSQL.
create table tbl (id int, info text, crt_time timestamp, c1 int);select * from tbl where c1<>1;insert into tbl select generate_series(1,10000000), 'test', now(), 1;
insert into tbl values (1,'abc',now(),2);create index idx_tbl_1 on tbl(id) where c1<>1;
Cool! It takes only 0.03 ms to run a “Not Equal To” search over 10 million entries of data using partial index.
postgres=# explain (analyze,verbose,timing,costs,buffers) select * from tbl where c1<>1;
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------
Index Scan using idx_tbl_1 on postgres.tbl (cost=0.12..1.44 rows=1 width=21) (actual time=0.015..0.015 rows=1 loops=1)
Output: id, info, crt_time, c1
Buffers: shared hit=1 read=1
Planning time: 0.194 ms
Execution time: 0.030 ms
(5 rows)Summary
With built-in GIN indexes, PostgreSQL supports full-text search and searching multi-value data types including arrays.
PostgreSQL uses a cost-based execution plan optimizer. It automatically chooses the best execution plan. When running a “Does Not Contain” search, PostgreSQL automatically chooses whether to use indexed scanning.
With regards to B-tree indexes, technically it can be used to implement the “Not Equal To” searches (INDEX SKIP SCAN), but it is currently not supported at the kernel level. We can use indexed scanning by adjusting the current SQL write methods.
PostgreSQL uses multi-core parallel computing to allow incredible performance improvements to Full Table Scan. Parallel scanning can remarkably improve the performance in case with a significantly large number of records.
PostgreSQL supports partial index, which supports partitioned or partial indexes. Performance is outstanding for “Not Equal To” queries with fixed conditions.
Reference:
