Optimizing a Postgres query with a large IN

Posted on

Question :

This query gets a list of posts created by people you follow. You can follow an unlimited number of people, but most people follow < 1000 others.

With this style of query, the obvious optimization would be to cache the "Post" ids, but unfortunately I do not have the time for that right now.

EXPLAIN ANALYZE SELECT
    "Post"."id",
    "Post"."actionId",
    "Post"."commentCount",
    ...
FROM
    "Posts" AS "Post"
INNER JOIN "Users" AS "user" ON "Post"."userId" = "user"."id"
LEFT OUTER JOIN "ActivityLogs" AS "activityLog" ON "Post"."activityLogId" = "activityLog"."id"
LEFT OUTER JOIN "WeightLogs" AS "weightLog" ON "Post"."weightLogId" = "weightLog"."id"
LEFT OUTER JOIN "Workouts" AS "workout" ON "Post"."workoutId" = "workout"."id"
LEFT OUTER JOIN "WorkoutLogs" AS "workoutLog" ON "Post"."workoutLogId" = "workoutLog"."id"
LEFT OUTER JOIN "Workouts" AS "workoutLog.workout" ON "workoutLog"."workoutId" = "workoutLog.workout"."id"
WHERE
"Post"."userId" IN (
    201486,
    1825186,
    998608,
    340844,
    271909,
    308218,
    341986,
    216893,
    1917226,
    ...  -- many more
)
AND "Post"."private" IS NULL
ORDER BY
    "Post"."createdAt" DESC
LIMIT 10;

Yields:

Limit  (cost=3.01..4555.20 rows=10 width=2601) (actual time=7923.011..7973.138 rows=10 loops=1)
  ->  Nested Loop Left Join  (cost=3.01..9019264.02 rows=19813 width=2601) (actual time=7923.010..7973.133 rows=10 loops=1)
        ->  Nested Loop Left Join  (cost=2.58..8935617.96 rows=19813 width=2376) (actual time=7922.995..7973.063 rows=10 loops=1)
              ->  Nested Loop Left Join  (cost=2.15..8821537.89 rows=19813 width=2315) (actual time=7922.984..7961.868 rows=10 loops=1)
                    ->  Nested Loop Left Join  (cost=1.71..8700662.11 rows=19813 width=2090) (actual time=7922.981..7961.846 rows=10 loops=1)
                          ->  Nested Loop Left Join  (cost=1.29..8610743.68 rows=19813 width=2021) (actual time=7922.977..7961.816 rows=10 loops=1)
                                ->  Nested Loop  (cost=0.86..8498351.81 rows=19813 width=1964) (actual time=7922.972..7960.723 rows=10 loops=1)
                                      ->  Index Scan using posts_createdat_public_index on "Posts" "Post"  (cost=0.43..8366309.39 rows=20327 width=261) (actual time=7922.869..7960.509 rows=10 loops=1)
                                            Filter: ("userId" = ANY ('{201486,1825186,998608,340844,271909,308218,341986,216893,1917226, ... many more ...}'::integer[]))
                                            Rows Removed by Filter: 218360
                                      ->  Index Scan using "Users_pkey" on "Users" "user"  (cost=0.43..6.49 rows=1 width=1703) (actual time=0.005..0.006 rows=1 loops=10)
                                            Index Cond: (id = "Post"."userId")
                                ->  Index Scan using "ActivityLogs_pkey" on "ActivityLogs" "activityLog"  (cost=0.43..5.66 rows=1 width=57) (actual time=0.107..0.107 rows=0 loops=10)
                                      Index Cond: ("Post"."activityLogId" = id)
                          ->  Index Scan using "WeightLogs_pkey" on "WeightLogs" "weightLog"  (cost=0.42..4.53 rows=1 width=69) (actual time=0.001..0.001 rows=0 loops=10)
                                Index Cond: ("Post"."weightLogId" = id)
                    ->  Index Scan using "Workouts_pkey" on "Workouts" workout  (cost=0.43..6.09 rows=1 width=225) (actual time=0.001..0.001 rows=0 loops=10)
                          Index Cond: ("Post"."workoutId" = id)
              ->  Index Scan using "WorkoutLogs_pkey" on "WorkoutLogs" "workoutLog"  (cost=0.43..5.75 rows=1 width=61) (actual time=1.118..1.118 rows=0 loops=10)
                    Index Cond: ("Post"."workoutLogId" = id)
        ->  Index Scan using "Workouts_pkey" on "Workouts" "workoutLog.workout"  (cost=0.43..4.21 rows=1 width=225) (actual time=0.004..0.004 rows=0 loops=10)
              Index Cond: ("workoutLog"."workoutId" = id)
Total runtime: 7974.524 ms

How can this be optimized for the time being?

I have the following relevant indexes:

-- Gets used
CREATE INDEX  "posts_createdat_public_index" ON "public"."Posts" USING btree("createdAt" DESC) WHERE "private" IS null;
-- Don't get used
CREATE INDEX  "posts_userid_fk_index" ON "public"."Posts" USING btree("userId");
CREATE INDEX  "posts_following_index" ON "public"."Posts" USING btree("userId", "createdAt" DESC) WHERE "private" IS null;

Perhaps this requires a large partial composite index with createdAt and userId where private IS NULL?

Answer :

Instead of using a huge IN-list, join on a VALUES expression, or if the list is large enough, use a temp table, index it, then join on it.

It’d be nice if PostgreSQL could do this internally & automatically but at this point the planner doesn’t know how.

Similar topics:

There are actually two different variants of the IN construct in Postgres. One works with a subquery expression (returning a set), the other one with a list of values, which is just shorthand for

expression = value1
OR
expression = value2
OR
...

You are using the second form, which is fine for a short list, but much slower for long lists. Provide your list of values as subquery expression instead. I was recently made aware of this variant:

WHERE "Post"."userId" IN (VALUES (201486), (1825186), (998608), ... )

I like to pass an array, unnest and join to it. Similar performance, but the syntax is shorter:

...
FROM   unnest('{201486,1825186,998608, ...}'::int[]) "userId"
JOIN   "Posts" "Post" USING ("userId")

Equivalent as long as there are no duplicates in the provided set / array. Else the second form with a JOIN returns duplicate rows, while the first with IN only returns a single instance. This subtle difference causes different query plans, too.

Obviously, you need an index on "Posts"."userId".
For very long lists (thousands), go with an indexed temp table like @Craig suggested. This allows combined bitmap index scans over both tables, which is typically faster as soon as there are multiple tuples per data page to fetch from disk.

Related:

Aside: your naming convention is not very helpful, makes your code verbose and hard to read. Rather use legal, lower-case, unquoted identifiers.

You can disable nested loops to force a hash join indexing the values. similar to what Craig Ringer said

set enable_nestloop to off;

This setting only applies to your current db session, if you need to turn it back on after that query, just execute

set enable_nestloop to on;

at the end.

You can expect improvement in several orders of magnitude in execution time.

Leave a Reply

Your email address will not be published. Required fields are marked *