mirror of
https://github.com/discourse/discourse.git
synced 2024-11-24 13:38:49 +08:00
2f5e66b6f8
In the query generated by `TopicTrackingState.report`, there are two
subqueies being executed. The first subquery fetches all the topics
that are new for a given user while the second subquery fetches all the topics with
unread posts for a given user. For the second subquery, there is a
filter `topics.updated_at >= user_stats.first_unread_at` which is used
as a performance optimisation to reduce the number of rows that PG has
to scan through the `topics` table.
However, we started to notice in production that the PG planner doesn't
always execute the filter first to reduce the number of rows that it has
to scan through. Running the following query in one of our production
instance,
```
EXPLAIN ANALYZE
SELECT
DISTINCT topics.id as topic_id,
u.id as user_id,
topics.created_at,
topics.updated_at,
topics.highest_staff_post_number AS highest_post_number,
last_read_post_number,
c.id as category_id,
c.topic_id AS category_topic_id,
tu.notification_level,
us.first_unread_at,
GREATEST(
CASE
WHEN COALESCE(uo.new_topic_duration_minutes, 2880) = -1 THEN u.created_at
WHEN COALESCE(uo.new_topic_duration_minutes, 2880) = -2 THEN COALESCE(
u.previous_visit_at,u.created_at
)
ELSE ('2023-07-31 03:29:45.737630'::timestamp - INTERVAL '1 MINUTE' * COALESCE(uo.new_topic_duration_minutes, 2880))
END, u.created_at, '2023-07-25 15:06:44'
) AS treat_as_new_topic_start_date
FROM topics
JOIN users u on u.id = 13455
JOIN user_stats AS us ON us.user_id = u.id
JOIN user_options AS uo ON uo.user_id = u.id
JOIN categories c ON c.id = topics.category_id
LEFT JOIN topic_users tu ON tu.topic_id = topics.id AND tu.user_id = u.id
WHERE u.id = 13455 AND
topics.updated_at >= us.first_unread_at AND
topics.archetype <> 'private_message' AND
(("topics"."deleted_at" IS NULL AND (tu.last_read_post_number < topics.highest_staff_post_number) AND (COALESCE(tu.notification_level, 1) >= 2)) OR (1=0)) AND
NOT (
COALESCE((select array_agg(tag_id) from topic_tags where topic_tags.topic_id = topics.id), ARRAY[]::int[]) && ARRAY[451,452,453]
) AND
topics.deleted_at IS NULL AND
NOT (
last_read_post_number IS NULL AND
(
topics.category_id IN (SELECT "categories"."id" FROM "categories" LEFT JOIN categories categories2 ON categories2.id = categories.parent_category_id LEFT JOIN category_users ON category_users.category_id = categories.id AND category_users.user_id = 13455 LEFT JOIN category_users category_users2 ON category_users2.category_id = categories2.id AND category_users2.user_id = 13455 WHERE ((category_users.id IS NULL AND COALESCE(category_users2.notification_level, 1) = 0) OR COALESCE(category_users.notification_level, 1) = 0))
AND tu.notification_level <= 1
)
)
```
we get the following
```
QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Unique (cost=201606.06..201608.15 rows=76 width=60) (actual time=91.279..91.294 rows=14 loops=1)
-> Sort (cost=201606.06..201606.25 rows=76 width=60) (actual time=91.278..91.284 rows=14 loops=1)
Sort Key: topics.id, topics.created_at, topics.updated_at, topics.highest_staff_post_number, tu.last_read_post_number, c.id, c.topic_id, tu.notification_level, us.first_unread_at, (GREATEST(CASE WHEN (COALESCE(uo.new_topic_duration_minutes, 2880) = '-1'::integer) THEN u.created_at WHEN (COALESCE(uo.new_topic_duration_minutes, 2880) = '-2'::integer) THEN COALESCE(u.previous_visit_at, u.created_at) ELSE ('2023-07-31 03:29:45.73763'::timestamp without time zone - ('00:01:00'::interval * (COALESCE(uo.new_topic_duration_minutes, 2880))::double precision)) END, u.created_at, '2023-07-25 15:06:44'::timestamp without time zone))
Sort Method: quicksort Memory: 26kB
-> Hash Join (cost=97519.51..201603.69 rows=76 width=60) (actual time=87.662..91.268 rows=14 loops=1)
Hash Cond: (topics.id = tu.topic_id)
Join Filter: ((tu.last_read_post_number < topics.highest_staff_post_number) AND ((tu.last_read_post_number IS NOT NULL) OR (NOT (hashed SubPlan 2)) OR (tu.notification_level > 1)))
Rows Removed by Join Filter: 10
-> Nested Loop (cost=1.54..104075.36 rows=3511 width=68) (actual time=0.055..3.609 rows=548 loops=1)
-> Nested Loop (cost=1.13..25.20 rows=1 width=32) (actual time=0.027..0.033 rows=1 loops=1)
-> Nested Loop (cost=0.71..16.76 rows=1 width=28) (actual time=0.020..0.023 rows=1 loops=1)
-> Index Scan using users_pkey on users u (cost=0.42..8.44 rows=1 width=20) (actual time=0.010..0.012 rows=1 loops=1)
Index Cond: (id = 13455)
-> Index Scan using user_stats_pkey on user_stats us (cost=0.29..8.31 rows=1 width=12) (actual time=0.008..0.010 rows=1 loops=1)
Index Cond: (user_id = 13455)
-> Index Scan using index_user_options_on_user_id_and_default_calendar on user_options uo (cost=0.42..8.44 rows=1 width=8) (actual time=0.007..0.008 rows=1 loops=1)
Index Cond: (user_id = 13455)
-> Nested Loop (cost=0.41..104015.12 rows=3504 width=36) (actual time=0.026..3.503 rows=548 loops=1)
-> Seq Scan on categories c (cost=0.00..13.73 rows=73 width=8) (actual time=0.003..0.039 rows=73 loops=1)
-> Index Only Scan using index_topics_on_updated_at_public on topics (cost=0.41..1424.20 rows=48 width=28) (actual time=0.012..0.046 rows=8 loops=73)
Index Cond: ((updated_at >= us.first_unread_at) AND (category_id = c.id))
Filter: (NOT (COALESCE((SubPlan 1), '{}'::integer[]) && '{451,452,453}'::integer[]))
Heap Fetches: 553
SubPlan 1
-> Aggregate (cost=4.31..4.32 rows=1 width=32) (actual time=0.002..0.002 rows=1 loops=548)
-> Index Only Scan using index_topic_tags_on_topic_id_and_tag_id on topic_tags (cost=0.29..4.31 rows=1 width=4) (actual time=0.002..0.002 rows=1 loops=548)
Index Cond: (topic_id = topics.id)
Heap Fetches: 178
-> Hash (cost=97222.14..97222.14 rows=19914 width=16) (actual time=87.545..87.546 rows=42884 loops=1)
Buckets: 65536 (originally 32768) Batches: 1 (originally 1) Memory Usage: 2387kB
-> Bitmap Heap Scan on topic_users tu (cost=1217.47..97222.14 rows=19914 width=16) (actual time=14.419..78.286 rows=42884 loops=1)
Recheck Cond: (user_id = 13455)
Filter: (COALESCE(notification_level, 1) >= 2)
Rows Removed by Filter: 15839
Heap Blocks: exact=45285
-> Bitmap Index Scan on index_topic_users_on_user_id_and_topic_id (cost=0.00..1212.49 rows=59741 width=0) (actual time=6.448..6.448 rows=58723 loops=1)
Index Cond: (user_id = 13455)
SubPlan 2
-> Nested Loop Left Join (cost=0.74..46.90 rows=1 width=4) (never executed)
Join Filter: (category_users2.category_id = categories2.id)
Filter: (((category_users.id IS NULL) AND (COALESCE(category_users2.notification_level, 1) = 0)) OR (COALESCE(category_users.notification_level, 1) = 0))
-> Nested Loop Left Join (cost=0.45..32.31 rows=73 width=16) (never executed)
Join Filter: (category_users.category_id = categories.id)
-> Nested Loop Left Join (cost=0.15..18.45 rows=73 width=8) (never executed)
-> Seq Scan on categories (cost=0.00..13.73 rows=73 width=8) (never executed)
-> Memoize (cost=0.15..0.28 rows=1 width=4) (never executed)
Cache Key: categories.parent_category_id
Cache Mode: logical
-> Index Only Scan using categories_pkey on categories categories2 (cost=0.14..0.27 rows=1 width=4) (never executed)
Index Cond: (id = categories.parent_category_id)
Heap Fetches: 0
-> Materialize (cost=0.29..11.69 rows=2 width=12) (never executed)
-> Index Scan using idx_category_users_user_id_category_id on category_users (cost=0.29..11.68 rows=2 width=12) (never executed)
Index Cond: (user_id = 13455)
-> Materialize (cost=0.29..11.69 rows=2 width=8) (never executed)
-> Index Scan using idx_category_users_user_id_category_id on category_users category_users2 (cost=0.29..11.68 rows=2 width=8) (never executed)
Index Cond: (user_id = 13455)
Planning Time: 1.740 ms
Execution Time: 91.414 ms
(59 rows)
```
From the execution plan, we can see the most of the time is spent
joining about 42888 rows in the `topics` table to the `topic_users` table.
However, we know that we only have to scan through a
subset of the `topics` table because the user's last unread at is '2023-07-20 11:33:05'.
If we filter the `topics` table with `topics.updated_at >= '2023-07-20 11:33:05'`, this would only
return about 1500 rows.
From our testing in production, the PG planner is able to execute a
better query plan when we avoid the unnecessary joins on `user_stats` just to be
able to get the user's `UserStat#first_unread_at`. Instead, we can just
pass the value of `UserStat#first_unread_at` directly as a query
parameter.
```
EXPLAIN ANALYZE
SELECT
DISTINCT topics.id as topic_id,
u.id as user_id,
topics.created_at,
topics.updated_at,
topics.highest_staff_post_number AS highest_post_number,
last_read_post_number,
c.id as category_id,
c.topic_id AS category_topic_id,
tu.notification_level,
GREATEST(
CASE
WHEN COALESCE(uo.new_topic_duration_minutes, 2880) = -1 THEN u.created_at
WHEN COALESCE(uo.new_topic_duration_minutes, 2880) = -2 THEN COALESCE(
u.previous_visit_at,u.created_at
)
ELSE ('2023-07-31 03:29:45.737630'::timestamp - INTERVAL '1 MINUTE' * COALESCE(uo.new_topic_duration_minutes, 2880))
END, u.created_at, '2023-07-25 15:06:44'
) AS treat_as_new_topic_start_date
FROM topics
JOIN users u on u.id = 13455
JOIN user_options AS uo ON uo.user_id = u.id
JOIN categories c ON c.id = topics.category_id
LEFT JOIN topic_users tu ON tu.topic_id = topics.id AND tu.user_id = u.id
WHERE u.id = 13455 AND
topics.updated_at >= '2023-07-20 11:33:05' AND
topics.archetype <> 'private_message' AND
(("topics"."deleted_at" IS NULL AND (tu.last_read_post_number < topics.highest_staff_post_number) AND (COALESCE(tu.notification_level, 1) >= 2)) OR (1=0)) AND
NOT (
COALESCE((select array_agg(tag_id) from topic_tags where topic_tags.topic_id = topics.id), ARRAY[]::int[]) && ARRAY[451,452,453]
) AND
topics.deleted_at IS NULL AND
NOT (
last_read_post_number IS NULL AND
(
topics.category_id IN (SELECT "categories"."id" FROM "categories" LEFT JOIN categories categories2 ON categories2.id = categories.parent_category_id LEFT JOIN category_users ON category_users.category_id = categories.id AND category_users.user_id = 13455 LEFT JOIN category_users category_users2 ON category_users2.category_id = categories2.id AND category_users2.user_id = 13455 WHERE ((category_users.id IS NULL AND COALESCE(category_users2.notification_level, 1) = 0) OR COALESCE(category_users.notification_level, 1) = 0))
AND tu.notification_level <= 1
)
);
```
Note how the filter is now `topics.updated_at >= '2023-07-20 11:33:05'`
instead of `topics.updated_at >= us.first_unread_at`. The modified query
above generates the following execution plan.
```
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Unique (cost=5189.86..5189.88 rows=1 width=52) (actual time=4.991..5.002 rows=14 loops=1)
-> Sort (cost=5189.86..5189.86 rows=1 width=52) (actual time=4.990..4.994 rows=14 loops=1)
Sort Key: topics.id, topics.created_at, topics.updated_at, topics.highest_staff_post_number, tu.last_read_post_number, c.id, c.topic_id, tu.notification_level, (GREATEST(CASE WHEN (COALESCE(uo.new_topic_duration_minutes, 2880) = '-1'::integer) THEN u.created_at WHEN (COALESCE(uo.new_topic_duration_minutes, 2880) = '-2'::integer) THEN COALESCE(u.previous_visit_at, u.created_at) ELSE ('2023-07-31 03:29:45.73763'::timestamp without time zone - ('00:01:00'::interval * (COALESCE(uo.new_topic_duration_minutes, 2880))::double precision)) END, u.created_at, '2023-07-25 15:06:44'::timestamp without time zone))
Sort Method: quicksort Memory: 26kB
-> Nested Loop (cost=52.11..5189.85 rows=1 width=52) (actual time=0.093..4.974 rows=14 loops=1)
-> Nested Loop (cost=51.70..5181.39 rows=1 width=60) (actual time=0.084..4.931 rows=14 loops=1)
-> Nested Loop (cost=51.28..5172.94 rows=1 width=44) (actual time=0.076..4.887 rows=14 loops=1)
-> Nested Loop (cost=0.41..1698.46 rows=59 width=36) (actual time=0.029..3.537 rows=548 loops=1)
-> Seq Scan on categories c (cost=0.00..13.73 rows=73 width=8) (actual time=0.005..0.039 rows=73 loops=1)
-> Index Only Scan using index_topics_on_updated_at_public on topics (cost=0.41..23.07 rows=1 width=28) (actual time=0.012..0.047 rows=8 loops=73)
Index Cond: ((updated_at >= '2023-07-20 11:33:05'::timestamp without time zone) AND (category_id = c.id))
Filter: (NOT (COALESCE((SubPlan 1), '{}'::integer[]) && '{451,452,453}'::integer[]))
Heap Fetches: 552
SubPlan 1
-> Aggregate (cost=4.31..4.32 rows=1 width=32) (actual time=0.002..0.002 rows=1 loops=548)
-> Index Only Scan using index_topic_tags_on_topic_id_and_tag_id on topic_tags (cost=0.29..4.31 rows=1 width=4) (actual time=0.002..0.002 rows=1 loops=548)
Index Cond: (topic_id = topics.id)
Heap Fetches: 178
-> Index Scan using index_topic_users_on_user_id_and_topic_id on topic_users tu (cost=50.86..58.88 rows=1 width=16) (actual time=0.002..0.002 rows=0 loops=548)
Index Cond: ((user_id = 13455) AND (topic_id = topics.id))
Filter: ((COALESCE(notification_level, 1) >= 2) AND (last_read_post_number < topics.highest_staff_post_number) AND ((last_read_post_number IS NOT NULL) OR (NOT (hashed SubPlan 2)) OR (notification_level > 1)))
Rows Removed by Filter: 0
SubPlan 2
-> Nested Loop Left Join (cost=0.74..50.43 rows=1 width=4) (never executed)
Join Filter: (category_users2.category_id = categories2.id)
Filter: (((category_users.id IS NULL) AND (COALESCE(category_users2.notification_level, 1) = 0)) OR (COALESCE(category_users.notification_level, 1) = 0))
-> Nested Loop Left Join (cost=0.45..35.84 rows=73 width=16) (never executed)
Join Filter: (category_users.category_id = categories.id)
-> Nested Loop Left Join (cost=0.15..21.97 rows=73 width=8) (never executed)
-> Seq Scan on categories (cost=0.00..13.73 rows=73 width=8) (never executed)
-> Memoize (cost=0.15..0.61 rows=1 width=4) (never executed)
Cache Key: categories.parent_category_id
Cache Mode: logical
-> Index Only Scan using categories_pkey on categories categories2 (cost=0.14..0.60 rows=1 width=4) (never executed)
Index Cond: (id = categories.parent_category_id)
Heap Fetches: 0
-> Materialize (cost=0.29..11.69 rows=2 width=12) (never executed)
-> Index Scan using idx_category_users_user_id_category_id on category_users (cost=0.29..11.68 rows=2 width=12) (never executed)
Index Cond: (user_id = 13455)
-> Materialize (cost=0.29..11.69 rows=2 width=8) (never executed)
-> Index Scan using idx_category_users_user_id_category_id on category_users category_users2 (cost=0.29..11.68 rows=2 width=8) (never executed)
Index Cond: (user_id = 13455)
-> Index Scan using users_pkey on users u (cost=0.42..8.44 rows=1 width=20) (actual time=0.003..0.003 rows=1 loops=14)
Index Cond: (id = 13455)
-> Index Scan using index_user_options_on_user_id_and_default_calendar on user_options uo (cost=0.42..8.44 rows=1 width=8) (actual time=0.002..0.002 rows=1 loops=14)
Index Cond: (user_id = 13455)
Planning Time: 1.281 ms
Execution Time: 5.092 ms
(48 rows)
```
With the new query, PG first does an index scan using the `index_topics_on_updated_at_public` index to filter away most of the topics making the subsequent joins much cheaper. Total query time has been reduced from ~90ms to ~5ms.
This optimisation will mostly affect users with very few/recent unread topics since a large `UserStat#firsts_unread_at` value will still mean scanning through a large portion of the `topics` table.
|
||
|---|---|---|
| .. | ||
| assets | ||
| controllers | ||
| helpers | ||
| jobs | ||
| mailers | ||
| models | ||
| serializers | ||
| services | ||
| views | ||