Slow queries in Discourse

dev
1

From Database Log we find some queries with unnecessary ORDER BY in subquery produced by user.rb#L379.One of them is as follows:

SELECT 
  "group_users"."group_id" 
FROM 
  "group_users" 
WHERE 
  "group_users"."group_id" IN (
    SELECT 
      "groups"."id" 
    FROM 
      "groups" 
    WHERE 
      (groups.id > 0) 
    ORDER BY 
      name ASC
  ) 
  AND "group_users"."user_id" = 762

ORDER BY in subquery is meaningless and we can remove it to accelerate query. According to our test it can improve performance of query by 18%-90%(depend on the time wateing on sort in subquery)

1 Like
2

What are the actual benchmarks, in milliseconds?

1 Like
3

For this example, remove ORDER BY can decrease query time from 4711878 Nanosecond to 585849 Nanosecond.

4

We run specs in discourse, and from Database we find a slow query((6480259 nanosecond) compared with similar query) with unnecessary DISTINCT produced by
site_settings_controller.rb#L141as follows:

SELECT 
  DISTINCT users.id 
FROM 
  "users" CROSS 
  JOIN tags t 
  LEFT JOIN tag_users tu ON users.id = tu.user_id 
  AND t.id = tu.tag_id 
WHERE 
  (
    t.id IN (1825) 
    AND tu.notification_level IS NULL
  )

When tags.id and notification_level are specify value, because of UNIQUE (tag_id, user_id, notification_level) in tag_users and PRIMARY KEY(id) in tags, both cross join and left join will not create duplicate
records which means that we can remove DISTINCT to accelerate query.
This opt query take 4538891 nanosecond(improve 30%)

6

From Database Log we find a slow query(6064379 NanoSecond) with many subquery and union produced by group.rb#L112 as follows:

SELECT 
  "groups"."id", 
  "groups"."name" 
FROM 
  "groups" 
  INNER JOIN "group_users" ON "groups"."id" = "group_users"."group_id" 
WHERE 
  "group_users"."user_id" = 296 
  AND (groups.id > 0) 
  AND (
    groups.id IN (
      SELECT 
        id 
      FROM 
        groups 
      WHERE 
        visibility_level = 0 
      UNION ALL 
      SELECT 
        id 
      FROM 
        groups 
      WHERE 
        visibility_level = 1 
        AND 296 IS NOT NULL 
      UNION ALL 
      SELECT 
        g.id 
      FROM 
        groups g 
        JOIN group_users gu ON gu.group_id = g.id 
        AND gu.user_id = 296 
      WHERE 
        g.visibility_level = 2 
      UNION ALL 
      SELECT 
        g.id 
      FROM 
        groups g 
        LEFT JOIN group_users gu ON gu.group_id = g.id 
        AND gu.user_id = 296 
        AND gu.owner 
      WHERE 
        g.visibility_level = 3 
        AND (
          gu.id IS NOT NULL 
          OR FALSE
        ) 
      UNION ALL 
      SELECT 
        g.id 
      FROM 
        groups g 
        JOIN group_users gu ON gu.group_id = g.id 
        AND gu.user_id = 296 
        AND gu.owner 
      WHERE 
        g.visibility_level = 4
    )
  ) 
ORDER BY 
  name ASC

Obviously it is equal to query as below:

SELECT 
  "groups"."id", 
  "groups"."name" 
FROM 
  "groups" 
  INNER JOIN "group_users" ON "groups"."id" = "group_users"."group_id" 
WHERE 
  "group_users"."user_id" = 296 AND
  groups.visibility_level IN (0, 1, 2, 3, 4) 
ORDER BY 
  name ASC

This opt query takes 378062 Nanosecond(improve 93%).

7

How is it equal? you are looking for groups with no visibility level checks.

1 Like
8

I‘m sorry , the correct rewritten query should be

SELECT 
  "groups"."id", 
  "groups"."name" 
FROM 
  "groups" 
  INNER JOIN "group_users" gu ON "groups"."id" = gu."group_id" 
WHERE 
  "group_users"."user_id" = 296 
   AND (groups.id > 0) 
   AND (
    groups.visibility_level IN (0, 1, 2) 
    OR (groups.visibility_level IN (3,4) and gu.owner)
   ) 
ORDER BY 
  name ASC

subquery with predicate visibility_level = 0,1,2 can be replaced by groups.visibility_level IN (0, 1, 2) , and subquery with predicate visibility_level = 3,4 can be replaced by or (groups.visibility_level IN (3,4) and gu.owner)

9

From Database we find a slow query(1141257 Nanosecond) with unnecessary JOIN and DISTINCT produced by directory_items_controller.rb as follows

SELECT 
  COUNT(DISTINCT "directory_items"."id") 
FROM 
  "directory_items" 
  LEFT OUTER JOIN "users" ON "users"."id" = "directory_items"."user_id" 
  LEFT OUTER JOIN "group_users" ON "group_users"."user_id" = "users"."id" 
  LEFT OUTER JOIN "groups" ON "groups"."id" = "group_users"."group_id" 
  LEFT OUTER JOIN "user_stats" ON "user_stats"."user_id" = "directory_items"."user_id" 
WHERE 
  "directory_items"."period_type" = 1 
  AND "groups"."id" = 2898

It retrieved nothing from user_stats and groups.id = 2898 can be replcaed by group_users.group_id = 2898 which means we can remove table groups and user_stats from query. At the same time, because of UNIQUE (user_id, group_id) in group_users and PRIMARY KEY (id) in users, JOIN won’t create duplicate records, so we can also remove DISTINCT from query as shown below:

SELECT 
  COUNT("directory_items"."id") 
FROM 
  "directory_items" 
  LEFT OUTER JOIN "users" ON "users"."id" = "directory_items"."user_id" 
  LEFT OUTER JOIN "group_users" ON "group_users"."user_id" = "users"."id" 
WHERE 
  "directory_items"."period_type" = 1 
  AND "group_users"."group_id"  = 2898

This opt query take 941700 Nanosecond(improve 17.49%)

10

Are you running some kind of database scanner on Discourse?

1 Like
11

We run the specs in discourse, and from Database we find a slow query((331729 nanosecond)compared with similar query) with unnecessary JOIN produced by
ser_badge.rb#L18as follows:

SELECT 
  COUNT(*) 
FROM 
  "badges" 
  INNER JOIN "user_badges" ON "badges"."id" = "user_badges"."badge_id" 
WHERE 
  "user_badges"."user_id" = 2112 
  AND (
    user_badges.badge_id IN (
      SELECT 
        id 
      FROM 
        badges 
      WHERE 
        enabled
    )
  ) 
  AND "badges"."id" = 1

This subquery can be replaced by a simple prediacte badges.enabled as shown below:

SELECT 
  COUNT(*) 
FROM 
  "badges" 
  INNER JOIN "user_badges" ON "badges"."id" = "user_badges"."badge_id" 
WHERE 
  "user_badges"."user_id" = 2112 
  AND badges.enabled 
  AND "badges"."id" = 1

This opt query take 267383 nanosecond(improve 19%)