lemmy_server PostgreSQL table for comment does not keep parent comment id directly, it uses a path field of ltree type.

by default, every comment has a path of it’s own primary key id.

comment id 101, path = “0.101”
comment id 102, path = “0.102”
comment id 103, path = “0.101.103”
comment id 104, path = “0.101.103.104”

comment 103 is a reply to comment 101, 104 is a reply to 103.

A second table named comment_aggregates has a count field with comment_id column linking to comment table id key. On each new comment reply, lemmy_server issues an update statement to update the counts on every parent in the tree. Rust code issues this to PostgreSQL:

        if let Some(parent_id) = parent_id {
          let top_parent = format!("0.{}", parent_id);
          let update_child_count_stmt = format!(
            "
update comment_aggregates ca set child_count = c.child_count
from (
  select c.id, c.path, count(c2.id) as child_count from comment c
  join comment c2 on c2.path <@ c.path and c2.path != c.path
  and c.path <@ '{top_parent}'
  group by c.id
) as c
where ca.comment_id = c.id"
          );
      sql_query(update_child_count_stmt).execute(conn).await?;
    }

I’ve been playing with doing bulk INSERT of thousands of comments at once to test SELECT query performance.

So far, this is the only SQL statement I have found that does a mass UPDATE of child_count from path for the entire comment table:

UPDATE
    comment_aggregates ca
SET
    child_count = c2.child_count
FROM (
    SELECT
        c.id,
        c.path,
        count(c2.id) AS child_count
    FROM
        comment c
    LEFT JOIN comment c2 ON c2.path <@ c.path
        AND c2.path != c.path
GROUP BY
    c.id) AS c2
WHERE
    ca.comment_id = c2.id;

There are 1 to 2 millions comments stored on lemmy.ml and lemmy.world - this rebuild of child_count can take hours, and may not complete at all. Even on 100,000 rows in a test system, it’s a harsh UPDATE statement to execute. EDIT: I found my API connection to production server was timing out and the run-time on the total rebuild isn’t as bad as I thought. With my testing system I’m also finding it is taking under 19 seconds with 312684 comments. The query does seem to execute and run normal, not stuck.

Anyone have suggestions on how to improve this and help make Lemmy PostgreSQL servers more efficient?

EDIT: lemmy 0.18.3 and 0.18.4 are munging the less-than and greater-than signs in these code blocks.

  • RoundSparrowOP
    link
    fedilink
    arrow-up
    1
    ·
    1 year ago

    I agree there is potential to reuse the child_count from child/grandchild rows. But there has to be some sense to the order they are updated in so that the deepest child gets count updated first?

    • bahmanm
      link
      fedilink
      English
      arrow-up
      1
      ·
      1 year ago

      potential to reuse

      I have a feeling that it’s going to make a noticeable difference; it’s way cheaper than a JOIN ... GROUP BY query.


      order they are updated in so that the deepest child gets count updated first

      Given the declarative nature of SQL, I’m afraid that’s not possible - at least to my knowledge.

      But worry not! That’s why there are stored procedures in almost every RDBMS; to add an imperative flare to the engine.

      In purely technical terms, Implementing what you’re thinking about is rather straight-forward in a stored procedure using a CURSOR. This could be possibly the quickest win (plus the idea of COUNT(*) if applicable.)


      Now, I’d like to suggest a possibly longer route which I think may be more scalable. The idea is based around the fact that comments themselves are utterly more important than the number of child comments.

      1. The first priority should be to ensure INSERT/UPDATE/SELECT are super quick on comment and post.
      2. The second priority should be to ensure child_count is eventually correctly updated when (1) happens.
      3. The last priority should be to try to make (2) as fast as we can while making sure (3) doesn’t interfere w/ (1) and (2) performance.

      Before rambling on, I’d like to ask if you think the priorities make sense? If they do, I can elaborate on the implementation.

        • RoundSparrowOP
          link
          fedilink
          arrow-up
          1
          ·
          1 year ago

          I found the total table update wasn’t as bad performing as I thought and the API gateway was timing out. I’m still generating larger amounts of test data to see how it performs in edge worst-case situations.

          • bahmanm
            link
            fedilink
            English
            arrow-up
            2
            ·
            1 year ago

            Can you keep this thread posted please? Or you can share a PR link so I can follow up the progress there. Am very interested.