Filter Duplicate Boosts

From neuromatch
Filter Duplicate Boosts
Description Prevent duplicated boosts in public timelines like in home timelines
Part Of Mastodon/Hacking
Contributors Jonny
Has Git Repository https://github.com/NeuromatchAcademy/mastodon
Completion Status Completed
Active Status Inactive


Problem

People like boosts in local timeline, but don't like seeing the same boost a zillion times.

Public feeds work differently than home feeds:

  • Public feeds use a public_feed model that is a scoped database query
  def get(limit, max_id = nil, since_id = nil, min_id = nil)
    scope = public_scope

    scope.merge!(without_local_only_scope) unless allow_local_only?
    scope.merge!(without_replies_scope) unless with_replies?
    scope.merge!(without_reblogs_scope) unless with_reblogs?
    scope.merge!(local_only_scope) if local_only?
    scope.merge!(remote_only_scope) if remote_only?
    scope.merge!(account_filters_scope) if account?
    scope.merge!(media_only_scope) if media_only?
    scope.merge!(language_scope) if account&.chosen_languages.present?

    scope.cache_ids.to_a_paginated_by_id(limit, max_id: max_id, since_id: since_id, min_id: min_id)
  end
  • Home feeds use a feed_manager class that inserts posts into a persistent list, which applies the boost filter at the time of that insertion:
  def add_to_feed(timeline_type, account_id, status, aggregate_reblogs: true)
    timeline_key = key(timeline_type, account_id)
    reblog_key   = key(timeline_type, account_id, 'reblogs')

    if status.reblog? && (aggregate_reblogs.nil? || aggregate_reblogs)
      # If the original status or a reblog of it is within
      # REBLOG_FALLOFF statuses from the top, do not re-insert it into
      # the feed
      rank = redis.zrevrank(timeline_key, status.reblog_of_id)

      return false if !rank.nil? && rank < FeedManager::REBLOG_FALLOFF

      # The ordered set at `reblog_key` holds statuses which have a reblog
      # in the top `REBLOG_FALLOFF` statuses of the timeline
      if redis.zadd(reblog_key, status.id, status.reblog_of_id, nx: true)
        # This is not something we've already seen reblogged, so we
        # can just add it to the feed (and note that we're reblogging it).
        redis.zadd(timeline_key, status.id, status.id)
      else
        # Another reblog of the same status was already in the
        # REBLOG_FALLOFF most recent statuses, so we note that this
        # is an "extra" reblog, by storing it in reblog_set_key.
        reblog_set_key = key(timeline_type, account_id, "reblogs:#{status.reblog_of_id}")
        redis.sadd(reblog_set_key, status.id)
        return false
      end
    else
      # A reblog may reach earlier than the original status because of the
      # delay of the worker delivering the original status, the late addition
      # by merging timelines, and other reasons.
      # If such a reblog already exists, just do not re-insert it into the feed.
      return false unless redis.zscore(reblog_key, status.id).nil?

      redis.zadd(timeline_key, status.id, status.id)
    end

    true
  end

Options

  • Model - Make some tricky scope that can filter repeated boosts in the existing public_feed model
  • Controller - Filter repeated boosts from the cached feed when serving the public_feed endpoint
  • Frontend - Filter repeated boosts in the web UI (but wouldn't work on apps, which is bad)

Implementation

Simpler Implementation (good)

Instead of doing all that garbage below, we can just restate the problem as "we want posts that aren't boosts, or posts that are boosts with the maximum ID for all boosts of a given post"

Like this: 

  def max_boost_id_scope
    Status.where(<<~SQL.squish)
      "statuses"."id" = (
        SELECT MAX(id)
        FROM "statuses" "s2"
        WHERE "s2"."reblog_of_id" = "statuses"."reblog_of_id"
          #{'AND ("s2"."local" = true OR "s2"."uri" IS NULL)' if local_only?}
          #{'AND "s2"."local" = false AND "s2"."uri" IS NOT NULL' if remote_only?}
        )
    SQL
  end

  def without_duplicate_reblogs
    Status.where(statuses: { reblog_of_id: nil })
          .or(max_boost_id_scope)
  end

Caveats

Note that we are using string interpolations of the Status.local and Status.remote scopes because there isn't a way, as far as Jonny can tell, to change the table alias of a scope. This shouldn't be that big of an issue (if, eg. the definition of those scopes changes) because the tests should catch most cases.

Example Query

SELECT "statuses"."id",
       "statuses"."updated_at"
FROM "statuses"
INNER JOIN "accounts" ON "accounts"."id" = "statuses"."account_id"
WHERE "statuses"."visibility" = $1
  AND "accounts"."suspended_at" IS NULL
  AND "accounts"."silenced_at" IS NULL
  AND (statuses.reply = FALSE
       OR statuses.in_reply_to_account_id = statuses.account_id)
  AND ("statuses"."reblog_of_id" IS NULL
       OR "statuses"."id" =
         (SELECT MAX(id)
          FROM statuses s2
          WHERE s2.reblog_of_id = statuses.reblog_of_id
            AND ("s2"."local" = true OR "s2"."uri" IS NULL)
       ))
  AND ("statuses"."local" = $2
       OR "statuses"."uri" IS NULL)
  AND "statuses"."deleted_at" IS NULL
  AND 1=1
  AND "statuses"."id" < 111813463418866657
ORDER BY "statuses"."id" DESC LIMIT $3 [["visibility", 0], ["local", true], ["LIMIT", 20]]

Original Implementation (bad)

Actually pretty damn simple. Add an additional scope in public_feed.rb

The core of the scope is this: 

  def without_duplicate_reblogs(limit, max_id, since_id, min_id)
    inner_query = Status.select('DISTINCT ON (reblog_of_id) statuses.id')
                        .reorder(reblog_of_id: :desc, id: :desc)
    Status.where(statuses: { reblog_of_id: nil })
          .or(Status.where(id: inner_query))

which selects either

  • Statuses that aren't a boost (reblog_of_id == nil)
  • Statuses that are a boost
    • sorted by the boosted status ID and the ID of the boost itself,
    • using postgres' [DISTINCT ON https://www.postgresql.org/docs/current/sql-select.html] clause to select only the first matching row
    • since id is a snowflake ID, and thus chronological, this will be the most recent boost.

There are some problems with this naive implementation though:

  • The inner query will select all boosts from all time, every time, because the outer pagination parameters passed to PublicFeed.get don't propagate to the inner query
  • The inner query necessarily needs to sort by reblog_of_id first, so just applying a LIMIT will filter on the recency of the original post rather than the boost, meaning we will miss most boosts most of the time

So we add the pagination information from PublicFeed.get, mimicking Paginable.to_a_paginated_by_id's use of the parameters to make a WHERE filter selecting whatever statuses would also be included in the given page that is being fetched 

  def without_duplicate_reblogs(limit, max_id, since_id, min_id)
    inner_query = Status.select('DISTINCT ON (reblog_of_id) statuses.id').reorder(reblog_of_id: :desc, id: :desc)
    if min_id.present?
      inner_query = inner_query.where(min_id < :id)
    elsif since_id.present?
      inner_query = inner_query.where(since_id < :id)
    end
    inner_query = inner_query.where(max_id > :id) if max_id.present?
    inner_query = inner_query.limit(limit) if limit.present?

    Status.where(statuses: { reblog_of_id: nil })
          .or(Status.where(id: inner_query))
  end

But that still doesn't quite get us there, since the first page of a local feed load doesn't have any pagination information in it except for the limit, so we still have the same problem as before. So instead we create a set of n candidate statuses where n is some arbitrary multiple of limit such that we assume that whatever is being filtered out by the other scopes is less than n - ie. we are considering the same set of candidate statuses as the outer scope. This should be relatively fast because we're just taking a slice of the index, rather than applying some complex SQL shit.

Recall that this is still one of several scopes ANDed together, so we will not return any boosts that are filtered by the other scopes, even if they are present in this scope. 

  def without_duplicate_reblogs(limit, max_id, since_id, min_id)
    candidate_statuses = Status.select(:id).reorder(id: :desc)

    if min_id.present?
      candidate_statuses = candidate_statuses.where(min_id < :id)
    elsif since_id.present?
      candidate_statuses = candidate_statuses.where(since_id < :id)
    end
    candidate_statuses = candidate_statuses.where(max_id > :id) if max_id.present?

    if limit.present?
      limit *= 5
      candidate_statuses = candidate_statuses.limit(limit)
    end

    inner_query = Status
                  .where(id: candidate_statuses)
                  .select('DISTINCT ON (reblog_of_id) statuses.id')
                  .reorder(reblog_of_id: :desc, id: :desc)

    Status.where(statuses: { reblog_of_id: nil })
          .or(Status.where(id: inner_query))
  end

But one MORE problem - since we're just considering one page at a time, we will still get duplicated boosts across pages. So...

  • When no minimum ID is provided, we use the "multiply the limit" strategy to avoid querying all statuses from all time
  • When a minimum ID is provided, we
    • Use no limit
    • If a maximum ID is also provided, we add 1 day worth of time to the ID so we also don't query arbitrarily into the future when fetching past pages
  def without_duplicate_reblogs(limit, max_id, since_id, min_id)
    candidate_statuses = Status.select(:id).reorder(id: :desc)

    if min_id.present?
      candidate_statuses = candidate_statuses.where(Status.arel_table[:id].gt(min_id))
    elsif since_id.present?
      candidate_statuses = candidate_statuses.where(Status.arel_table[:id].gt(since_id))
    elsif limit.present?
      limit *= 5
      candidate_statuses = candidate_statuses.limit(limit)
    end

    if max_id.present?
      max_time = Mastodon::Snowflake.to_time(id)
      max_time += 1.day
      max_id = Mastodon::Snowflake.id_at(max_time)
      candidate_statuses = candidate_statuses.where(Status.arel_table[:id].lt(max_id))
    end

    inner_query = Status
                  .where(id: candidate_statuses)
                  .select('DISTINCT ON (reblog_of_id) statuses.id')
                  .reorder(reblog_of_id: :desc, id: :desc)

    Status.where(statuses: { reblog_of_id: nil })
          .or(Status.where(id: inner_query))
  end

Example Query

SELECT "statuses"."id", "statuses"."updated_at" 
FROM "statuses" 
INNER JOIN "accounts" ON "accounts"."id" = "statuses"."account_id" 
WHERE "statuses"."visibility" = $1 
  AND "accounts"."suspended_at" IS NULL 
  AND "accounts"."silenced_at" IS NULL 
  AND (
    statuses.reply = FALSE 
    OR statuses.in_reply_to_account_id = statuses.account_id
  ) 
  AND (
    "statuses"."reblog_of_id" IS NULL 
    OR "statuses"."id" IN (
      SELECT DISTINCT ON (reblog_of_id) statuses.id 
      FROM "statuses" 
      WHERE "statuses"."deleted_at" IS NULL 
        AND "statuses"."id" IN (
          SELECT "statuses"."id" FROM "statuses" 
          WHERE "statuses"."deleted_at" IS NULL 
            AND "statuses"."id" < 111819125737828654 
          ORDER BY "statuses"."id" DESC 
          LIMIT $2
        ) 
      ORDER BY "statuses"."reblog_of_id" DESC, "statuses"."id" DESC
    )
  ) 
  AND (
    "statuses"."local" = $3 
    OR "statuses"."uri" IS NULL
  ) 
  AND "statuses"."deleted_at" IS NULL 
  AND 1=1 
  AND "statuses"."id" < 111813463418866657 
ORDER BY "statuses"."id" DESC LIMIT $4

[["visibility", 0], ["LIMIT", 100], ["local", true], ["LIMIT", 20]]

References

Retrieved from "https://wiki.neuromatch.social/index.php?title=Filter_Duplicate_Boosts&oldid=1204"