This optimises searching for users when using queries consisting out of
one or two characters such as "ab". We optimise such cases by searching
for `LOWER(name)` and `LOWER(username)` instead of using `ILIKE`. Using
`LOWER` produces a _much_ better performing query.
For example, when searching for all users matching the term "a" we'd
produce the following plan:
Limit (cost=637.69..637.74 rows=20 width=805) (actual time=41.983..41.995 rows=20 loops=1)
Buffers: shared hit=8330
-> Sort (cost=637.69..638.61 rows=368 width=805) (actual time=41.982..41.990 rows=20 loops=1)
Sort Key: (CASE WHEN ((name)::text = 'a'::text) THEN 0 WHEN ((username)::text = 'a'::text) THEN 1 WHEN ((email)::text = 'a'::text) THEN 2 ELSE 3 END), name
Sort Method: top-N heapsort Memory: 35kB
Buffers: shared hit=8330
-> Bitmap Heap Scan on users (cost=75.47..627.89 rows=368 width=805) (actual time=9.452..41.305 rows=277 loops=1)
Recheck Cond: (((name)::text ~~* 'a'::text) OR ((username)::text ~~* 'a'::text) OR ((email)::text = 'a'::text))
Rows Removed by Index Recheck: 7601
Heap Blocks: exact=7636
Buffers: shared hit=8327
-> BitmapOr (cost=75.47..75.47 rows=368 width=0) (actual time=8.290..8.290 rows=0 loops=1)
Buffers: shared hit=691
-> Bitmap Index Scan on index_users_on_name_trigram (cost=0.00..38.85 rows=180 width=0) (actual time=4.369..4.369 rows=4071 loops=1)
Index Cond: ((name)::text ~~* 'a'::text)
Buffers: shared hit=360
-> Bitmap Index Scan on index_users_on_username_trigram (cost=0.00..34.41 rows=188 width=0) (actual time=3.896..3.896 rows=4140 loops=1)
Index Cond: ((username)::text ~~* 'a'::text)
Buffers: shared hit=328
-> Bitmap Index Scan on users_email_key (cost=0.00..1.94 rows=1 width=0) (actual time=0.022..0.022 rows=0 loops=1)
Index Cond: ((email)::text = 'a'::text)
Buffers: shared hit=3
Planning time: 3.912 ms
Execution time: 42.171 ms
With the changes in this commit we now produce the following plan
instead:
Limit (cost=13257.48..13257.53 rows=20 width=805) (actual time=1.567..1.579 rows=20 loops=1)
Buffers: shared hit=287
-> Sort (cost=13257.48..13280.93 rows=9379 width=805) (actual time=1.567..1.572 rows=20 loops=1)
Sort Key: (CASE WHEN ((name)::text = 'a'::text) THEN 0 WHEN ((username)::text = 'a'::text) THEN 1 WHEN ((email)::text = 'a'::text) THEN 2 ELSE 3 END), name
Sort Method: top-N heapsort Memory: 35kB
Buffers: shared hit=287
-> Bitmap Heap Scan on users (cost=135.66..13007.91 rows=9379 width=805) (actual time=0.194..1.107 rows=277 loops=1)
Recheck Cond: ((lower((name)::text) = 'a'::text) OR (lower((username)::text) = 'a'::text) OR ((email)::text = 'a'::text))
Heap Blocks: exact=277
Buffers: shared hit=287
-> BitmapOr (cost=135.66..135.66 rows=9379 width=0) (actual time=0.152..0.152 rows=0 loops=1)
Buffers: shared hit=10
-> Bitmap Index Scan on yorick_test_users (cost=0.00..124.75 rows=9377 width=0) (actual time=0.101..0.101 rows=277 loops=1)
Index Cond: (lower((name)::text) = 'a'::text)
Buffers: shared hit=4
-> Bitmap Index Scan on index_on_users_lower_username (cost=0.00..1.94 rows=1 width=0) (actual time=0.035..0.035 rows=1 loops=1)
Index Cond: (lower((username)::text) = 'a'::text)
Buffers: shared hit=3
-> Bitmap Index Scan on users_email_key (cost=0.00..1.94 rows=1 width=0) (actual time=0.014..0.014 rows=0 loops=1)
Index Cond: ((email)::text = 'a'::text)
Buffers: shared hit=3
Planning time: 0.303 ms
Execution time: 1.687 ms
Here we can see the new query is 25 times faster compared to the old
query.
Nested groups support uses queries along the lines of `path LIKE 'X/%'`.
For these queries to use an index on PostgreSQL we need to use either
the varchar_pattern_ops or text_pattern_ops operator class.
Fixes https://gitlab.com/gitlab-org/gitlab-ce/issues/29554
Performance is improved in two steps:
1. On PostgreSQL an expression index is used for checking lower(email)
and lower(username).
2. The check to determine if we're searching for a username or Email is
moved to Ruby. Thanks to @haynes for suggesting and writing the
initial implementation of this.
Moving the check to Ruby makes this method an additional 1.5 times
faster compared to doing the check in the SQL query.
With performance being improved I've now also tweaked the amount of
iterations required by the User.by_login benchmark. This method now runs
between 900 and 1000 iterations per second.
