Nous avons récemment mis à jour à nouveau pour intégrer cette fonctionnalité FEATURE: New and Unread messages for user personal messages. by tgxworld · Pull Request #13603 · discourse/discourse · GitHub (c’est-à-dire que nous sommes actuellement sur la version 2.8.0beta6).
Il semble que l’opération « obtenir tous les MP pour un utilisateur », qui aboutit finalement à l’endroit indiqué ici FEATURE: New and Unread messages for user personal messages. by tgxworld · Pull Request #13603 · discourse/discourse · GitHub, génère une requête de ce type :
SELECT "topics"."id", "topics"."title", "topics"."last_posted_at", "topics"."created_at", "topics"."updated_at", "topics"."views", "topics"."posts_count", "topics"."user_id", "topics"."last_post_user_id", "topics"."reply_count", "topics"."featured_user1_id", "topics"."featured_user2_id", "topics"."featured_user3_id", "topics"."deleted_at", "topics"."highest_post_number", "topics"."like_count", "topics"."incoming_link_count", "topics"."category_id", "topics"."visible", "topics"."moderator_posts_count", "topics"."closed", "topics"."archived", "topics"."bumped_at", "topics"."has_summary", "topics"."archetype", "topics"."featured_user4_id", "topics"."notify_moderators_count", "topics"."spam_count", "topics"."pinned_at", "topics"."score", "topics"."percent_rank", "topics"."subtype", "topics"."slug", "topics"."deleted_by_id", "topics"."participant_count", "topics"."word_count", "topics"."excerpt", "topics"."pinned_globally", "topics"."pinned_until", "topics"."fancy_title", "topics"."highest_staff_post_number", "topics"."featured_link", "topics"."reviewable_score", "topics"."image_upload_id", "topics"."slow_mode_seconds" FROM "topics" LEFT OUTER JOIN topic_users AS tu ON (topics.id = tu.topic_id AND tu.user_id = 1234) LEFT JOIN group_archived_messages gm ON gm.topic_id = topics.id
LEFT JOIN user_archived_messages um
ON um.user_id = 1234
AND um.topic_id = topics.id WHERE "topics"."deleted_at" IS NULL AND (topics.id IN (
SELECT topic_id
FROM topic_allowed_users
WHERE user_id = 1234
UNION ALL
SELECT topic_id FROM topic_allowed_groups
WHERE group_id IN (
SELECT group_id FROM group_users WHERE user_id = 1234
)
)) AND "topics"."archetype" = 'private_message' AND "topics"."visible" = TRUE AND (um.user_id IS NULL AND gm.topic_id IS NULL) ORDER BY topics.bumped_at DESC LIMIT 30;
Cette requête fait planter notre serveur RDS. Un membre de notre équipe a suggéré de filtrer les données en utilisant les index disponibles avant de procéder aux jointures. Cette requête s’exécute en moins de 100 ms :
SELECT "topics_filter"."id", "topics_filter"."title", "topics_filter"."last_posted_at", "topics_filter"."created_at", "topics_filter"."updated_at",
"topics_filter"."views", "topics_filter"."posts_count", "topics_filter"."user_id", "topics_filter"."last_post_user_id", "topics_filter"."reply_count", "topics_filter"."featured_user1_id",
"topics_filter"."featured_user2_id", "topics_filter"."featured_user3_id", "topics_filter"."deleted_at", "topics_filter"."highest_post_number", "topics_filter"."like_count",
"topics_filter"."incoming_link_count", "topics_filter"."category_id", "topics_filter"."visible", "topics_filter"."moderator_posts_count", "topics_filter"."closed",
"topics_filter"."archived", "topics_filter"."bumped_at", "topics_filter"."has_summary", "topics_filter"."archetype", "topics_filter"."featured_user4_id", "topics_filter"."notify_moderators_count",
"topics_filter"."spam_count", "topics_filter"."pinned_at", "topics_filter"."score", "topics_filter"."percent_rank", "topics_filter"."subtype", "topics_filter"."slug",
"topics_filter"."deleted_by_id", "topics_filter"."participant_count", "topics_filter"."word_count", "topics_filter"."excerpt", "topics_filter"."pinned_globally",
"topics_filter"."pinned_until", "topics_filter"."fancy_title", "topics_filter"."highest_staff_post_number", "topics_filter"."featured_link", "topics_filter"."reviewable_score",
"topics_filter"."image_upload_id", "topics_filter"."slow_mode_seconds"
FROM (select * from "topics"
where "topics"."deleted_at" IS NULL
and (topics.id IN ( SELECT topic_id FROM topic_allowed_users WHERE user_id = 1234
UNION ALL
SELECT topic_id FROM topic_allowed_groups WHERE group_id IN ( SELECT group_id FROM group_users WHERE user_id = 1234 )
))
AND "topics"."archetype" = 'private_message'
AND "topics"."visible" = TRUE
) as topics_filter
LEFT OUTER JOIN (select topic_id from topic_users where user_id = 1234) AS tu ON (topics_filter.id = tu.topic_id )
LEFT JOIN group_archived_messages gm ON gm.topic_id = topics_filter.id
LEFT JOIN ( select user_id, topic_id from user_archived_messages where user_id = 1234) um
ON um.topic_id = topics_filter.id
WHERE (um.user_id IS NULL AND gm.topic_id IS NULL) ORDER BY topics_filter.bumped_at DESC LIMIT 30;
Pour référence : topics > 10 millions de lignes, topic_users > 50 millions de lignes.
Est-il possible de restructurer la requête pour résoudre ces problèmes d’une manière ou d’une autre ?
Édition :
Pour préciser, lorsque nous examinons nos requêtes les plus longues (> 2 minutes, voire 5 minutes), ce sont les seules qui apparaissent. Avec le dernier changement, le CPU de notre Postgres est constamment à 99 %. Lorsque je termine ces requêtes, nous constatons que l’utilisation du CPU revient à la normale.
) : 
Toujours plus que ce à quoi je m’attendais !