Reasonably efficient PostgreSQL live queries

The module implements query multiplexing, which is our implementation strategy for live queries (i.e. GraphQL subscriptions) made against Postgres. Fundamentally, our implementation is built around polling, which is never ideal, but it’s a lot easier to implement than trying to do something event-based. To minimize the resource cost of polling, we use multiplexing, which is essentially a two-tier batching strategy.

The high-level idea

The objective is to minimize the number of concurrent polling workers to reduce database load as much as possible. A very naïve strategy would be to group identical queries together so we only have one poller per unique active subscription. That’s a good start, but of course, in practice, most queries differ slightly. However, it happens that they very frequently /only differ in their variables/ (that is, GraphQL query variables and session variables), and in those cases, we try to generated parameterized SQL. This means that the same prepared SQL query can be reused, just with a different set of variables.

To give a concrete example, consider the following query:

subscription vote_count($post_id: Int!) {
  vote_count(where: {post_id: {_eq: $post_id}}) {
    votes
  }
}

No matter what the client provides for $post_id, we will always generate the same SQL:

SELECT votes FROM vote_count WHERE post_id = $1

If multiple clients subscribe to vote_count, we can certainly reuse the same prepared query. For example, imagine we had 10 concurrent subscribers, each listening on a distinct $post_id:

let postIds = [3, 11, 32, 56, 13, 97, 24, 43, 109, 48]

We could iterate over postIds in Haskell, executing the same prepared query 10 times:

for postIds $ \postId ->
  PG.withQE defaultTxErrorHandler preparedQuery (Identity postId) True

Sadly, that on its own isn’t good enough. The overhead of running each query is large enough that Postgres becomes overwhelmed if we have to serve lots of concurrent subscribers. Therefore, what we want to be able to do is somehow make one query instead of ten.

Multiplexing

This is where multiplexing comes in. By taking advantage of Postgres lateral joins, we can do the iteration in Postgres rather than in Haskell, allowing us to pay the query overhead just once for all ten subscribers. Essentially, lateral joins add map-like functionality to SQL, so we can run our query once per $post_id:

SELECT results.votes
FROM unnest($1::integer[]) query_variables (post_id)
LEFT JOIN LATERAL (
  SELECT coalesce(json_agg(votes), '[]')
  FROM vote_count WHERE vote_count.post_id = query_variables.post_id
) results ON true

If we generalize this approach just a little bit more, we can apply this transformation to arbitrary queries parameterized over arbitrary session and query variables!

Implementation overview

To support query multiplexing, we maintain a tree of the following types, where > should be read as “contains”:

SubscriptionsState > Poller > Cohort > Subscriber

Here’s a brief summary of each type’s role:

• A Subscriber is an actual client with an open websocket connection.
• A Cohort is a set of Subscribers that are all subscribed to the same query /with the exact same variables/. (By batching these together, we can do better than multiplexing, since we can just query the data once.)
• A Poller is a worker thread for a single, multiplexed query. It fetches data for a set of Cohorts that all use the same parameterized query, but have different sets of variables.
• Finally, the SubscriptionsState is the top-level container that holds all the active Pollers.

Additional details are provided by the documentation for individual bindings.