does the etherpad-proxy itself scale?

[Loki-Afro]

i took a look into the code

• a local db is used (sqlite)
• i saw several places with locking mechanisms
• didn't see any code related to sync info with multiple proxies since a db is not used

thus i have the suspicion that the proxy itself does not scale

is that correct?

[SamTV12345]

Hi @Loki-Afro Actually the proxy is very performant. It uses minimal footprint. You can easily put hundreds of requests against it. The mutexes are only in places where it is absolutely necessary so that things that change data in the map don't overwrite each other concurrently and thus corrupt data.
SQLite is actually pretty performant. So you can just mount a volume. Do you require other databases?

[Loki-Afro]

@SamTV12345 Thanks for pointing this out! From what I’ve seen, the proxy itself doesn’t seem to be scalable. In other words, you can already set up an Etherpad cluster, but the proxy component becomes a single point of failure since it can’t scale horizontally.

To run multiple proxies in parallel, they would need to synchronize their state. As far as I know, SQLite isn’t really suitable for that—especially since, as you mentioned, mounting a volume isn’t a safe or reliable approach.

In a containerized setup, database state isn’t typically shared through a volume mount anyway. That becomes even more complex if your cluster spans multiple datacenters, because you’d need an underlying system that replicates the volume across them reliably and in near-real time.

Something lightweight like Redis or Valkey could already cover this use case, and if more complexity is required down the line, you could even extend it with a pub/sub mechanism for state distribution.

That said, I might be missing something here

Also take a look here: https://medium.com/lumen-engineering-blog/how-to-implement-a-distributed-and-auto-scalable-websocket-server-architecture-on-kubernetes-4cc32e1dfa45 maybe it gets clearer where i'm coming from, websockets are a pain to scale ...

i know you are not using yjs, but also an approach which works well with tldraw for example :)

[SamTV12345]

You can now use postgres. I'll have a look at the other options.