FIX: Ensure that aggregating search shows the post with the higest rank. (PR #10177)

Previously, we would only take either the MIN or MAX for post_number during aggregation meaning that the ranking is not considered.

require 'benchmark/ips'

Benchmark.ips do |x|
  x.config(time: 10, warmup: 2)

  x.report("current aggregate search query") do
    DB.exec <<~SQL
    SELECT "posts"."id", "posts"."user_id", "posts"."topic_id", "posts"."post_number", "posts"."raw", "posts"."cooked", "posts"."created_at", "posts"."updated_at", "posts"."reply_to_post_number", "posts"."reply_count", "posts"."quote_count", "posts"."deleted_at", "posts"."off_topic_count", "posts"."like_count", "posts"."incoming_link_count", "posts"."bookmark_count", "posts"."score", "posts"."reads", "posts"."post_type", "posts"."sort_order", "posts"."last_editor_id", "posts"."hidden", "posts"."hidden_reason_id", "posts"."notify_moderators_count", "posts"."spam_count", "posts"."illegal_count", "posts"."inappropriate_count", "posts"."last_version_at", "posts"."user_deleted", "posts"."reply_to_user_id", "posts"."percent_rank", "posts"."notify_user_count", "posts"."like_score", "posts"."deleted_by_id", "posts"."edit_reason", "posts"."word_count", "posts"."version", "posts"."cook_method", "posts"."wiki", "posts"."baked_at", "posts"."baked_version", "posts"."hidden_at", "posts"."self_edits", "posts"."reply_quoted", "posts"."via_email", "posts"."raw_email", "posts"."public_version", "posts"."action_code", "posts"."locked_by_id", "posts"."image_upload_id" FROM "posts" JOIN (SELECT *, row_number() over() row_number FROM (SELECT topics.id, min(posts.post_number) post_number FROM "posts" INNER JOIN "post_search_data" ON "post_search_data"."post_id" = "posts"."id" INNER JOIN "topics" ON "topics"."id" = "posts"."topic_id" AND ("topics"."deleted_at" IS NULL) LEFT JOIN categories ON categories.id = topics.category_id WHERE ("posts"."deleted_at" IS NULL) AND "posts"."post_type" IN (1, 2, 3, 4) AND (topics.visible) AND (topics.archetype <> 'private_message') AND (post_search_data.search_data @@ TO_TSQUERY('english', '''postgres'':*ABCD')) AND (categories.id NOT IN (
      SELECT categories.id WHERE categories.search_priority = 1
    )
    ) AND ((categories.id IS NULL) OR (NOT categories.read_restricted)) GROUP BY topics.id ORDER BY MAX((
      TS_RANK_CD(
        post_search_data.search_data,
        TO_TSQUERY('english', '''postgres'':*ABCD'),
        1|32
      ) *
      (
        CASE categories.search_priority
        WHEN 2
        THEN 0.6
        WHEN 3
        THEN 0.8
        WHEN 4
        THEN 1.2
        WHEN 5
        THEN 1.4
        ELSE
          CASE WHEN topics.closed
          THEN 0.9
          ELSE 1
          END
        END
      )
    )
    ) DESC, topics.bumped_at DESC LIMIT 51 OFFSET 0) xxx) x ON x.id = posts.topic_id AND x.post_number = posts.post_number WHERE ("posts"."deleted_at" IS NULL) ORDER BY row_number;
    SQL
  end

  x.report("current aggregate search query with proper ranking") do
    DB.exec <<~SQL
    SELECT "posts"."id", "posts"."user_id", "posts"."topic_id", "posts"."post_number", "posts"."raw", "posts"."cooked", "posts"."created_at", "posts"."updated_at", "posts"."reply_to_post_number", "posts"."reply_count", "posts"."quote_count", "posts"."deleted_at", "posts"."off_topic_count", "posts"."like_count", "posts"."incoming_link_count", "posts"."bookmark_count", "posts"."score", "posts"."reads", "posts"."post_type", "posts"."sort_order", "posts"."last_editor_id", "posts"."hidden", "posts"."hidden_reason_id", "posts"."notify_moderators_count", "posts"."spam_count", "posts"."illegal_count", "posts"."inappropriate_count", "posts"."last_version_at", "posts"."user_deleted", "posts"."reply_to_user_id", "posts"."percent_rank", "posts"."notify_user_count", "posts"."like_score", "posts"."deleted_by_id", "posts"."edit_reason", "posts"."word_count", "posts"."version", "posts"."cook_method", "posts"."wiki", "posts"."baked_at", "posts"."baked_version", "posts"."hidden_at", "posts"."self_edits", "posts"."reply_quoted", "posts"."via_email", "posts"."raw_email", "posts"."public_version", "posts"."action_code", "posts"."locked_by_id", "posts"."image_upload_id" FROM "posts" JOIN (SELECT *, row_number() over() row_number FROM (SELECT subquery.topic_id id, (ARRAY_AGG(subquery.post_number))[1] post_number, MAX(subquery.rank) rank, MAX(subquery.bumped_at) bumped_at FROM (SELECT "posts"."id", "posts"."user_id", "posts"."topic_id", "posts"."post_number", "posts"."raw", "posts"."cooked", "posts"."created_at", "posts"."updated_at", "posts"."reply_to_post_number", "posts"."reply_count", "posts"."quote_count", "posts"."deleted_at", "posts"."off_topic_count", "posts"."like_count", "posts"."incoming_link_count", "posts"."bookmark_count", "posts"."score", "posts"."reads", "posts"."post_type", "posts"."sort_order", "posts"."last_editor_id", "posts"."hidden", "posts"."hidden_reason_id", "posts"."notify_moderators_count", "posts"."spam_count", "posts"."illegal_count", "posts"."inappropriate_count", "posts"."last_version_at", "posts"."user_deleted", "posts"."reply_to_user_id", "posts"."percent_rank", "posts"."notify_user_count", "posts"."like_score", "posts"."deleted_by_id", "posts"."edit_reason", "posts"."word_count", "posts"."version", "posts"."cook_method", "posts"."wiki", "posts"."baked_at", "posts"."baked_version", "posts"."hidden_at", "posts"."self_edits", "posts"."reply_quoted", "posts"."via_email", "posts"."raw_email", "posts"."public_version", "posts"."action_code", "posts"."locked_by_id", "posts"."image_upload_id", (
      TS_RANK_CD(
        post_search_data.search_data,
        TO_TSQUERY('english', '''postgres'':*ABCD'),
        1|32
      ) *
      (
        CASE categories.search_priority
        WHEN 2
        THEN 0.6
        WHEN 3
        THEN 0.8
        WHEN 4
        THEN 1.2
        WHEN 5
        THEN 1.4
        ELSE
          CASE WHEN topics.closed
          THEN 0.9
          ELSE 1
          END
        END
      )
    )
     rank, topics.bumped_at bumped_at FROM "posts" INNER JOIN "post_search_data" ON "post_search_data"."post_id" = "posts"."id" INNER JOIN "topics" ON "topics"."id" = "posts"."topic_id" AND ("topics"."deleted_at" IS NULL) LEFT JOIN categories ON categories.id = topics.category_id WHERE ("posts"."deleted_at" IS NULL) AND "posts"."post_type" IN (1, 2, 3, 4) AND (topics.visible) AND (topics.archetype <> 'private_message') AND (post_search_data.search_data @@ TO_TSQUERY('english', '''postgres'':*ABCD')) AND (categories.id NOT IN (
      SELECT categories.id WHERE categories.search_priority = 1
    )
    ) AND ((categories.id IS NULL) OR (NOT categories.read_restricted))) subquery GROUP BY subquery.topic_id ORDER BY rank DESC, bumped_at DESC LIMIT 51 OFFSET 0) xxx) x ON x.id = posts.topic_id AND x.post_number = posts.post_number WHERE ("posts"."deleted_at" IS NULL) ORDER BY row_number;
    SQL
  end

  x.compare!
end
Warming up --------------------------------------
current aggregate search query
                         1.000  i/100ms
current aggregate search query with proper ranking
                         1.000  i/100ms
Calculating -------------------------------------
current aggregate search query
                         17.726  (± 0.0%) i/s -    178.000  in  10.045107s
current aggregate search query with proper ranking
                         17.802  (± 0.0%) i/s -    178.000  in  10.002230s

Comparison:
current aggregate search query with proper ranking:       17.8 i/s
current aggregate search query:       17.7 i/s - 1.00x  (± 0.00) slower

GitHub

The title of this pull request changed from “FIX: Ensure that aggregating search shows the right posts.” to "FIX: Ensure that aggregating search shows the post with the higest rank.

I would add a comment explaining what “correctly” means (eg. shows the highest ranked post per topic)

Done :slight_smile:

1 Like

To be honest this change makes me nervous, a lot of stuff is shuffled, that said overall risk (at least conceptually) is not too high given we already had a MIN in the query to start with.

Go ahead and merge, watch it carefully please.

@SamSaffron One thing I can do is to introduce this behind a hidden site setting and enable it on dev/meta first.

1 Like

Merged in https://github.com/discourse/discourse/commit/d8c796bc44d84a9e177737153793e5acc276adec

1 Like