Great! 😁 I was just wondering because the memory graph showed sharp falls in memory usage every ~30 mins.
Exploring The Lemmy Archipelago 🏝️🍾
Great! 😁 I was just wondering because the memory graph showed sharp falls in memory usage every ~30 mins.
Is the memory leak still there?
I think the devs have been aware of the issue, theoretically, for a while. A proper solution requires some significant changes, so it was being postponed because this wasn’t considered urgent.
I think I found what eats the memory. DB iops isn’t the cause - looks like the server doesn’t reply before all the database operations are done. The problem is the unbounded queue in the activitypub_federation crate, spawned when creating the ActivityQueue struct. The point is, this queue holds all the “activities” - events to be sent to federated instances. If, for whatever reason, the events aren’t delivered to all the federated servers, they are retried with an exponential backoff for up to 2.5 days. If even a single federated instance is unreachable, all events remain in memory. For a large instance, this will eat up the memory for every upvote/downvote, post or comment.
Lemmy needs to figure out a scalable eventual consistency algorithm. Most importantly, to store the messages in the DB, not in memory.
I’m calling it - if there’s actually a memory leak in the Rust code, it’s gonna be the in memory queues because the DB’s iops can’t cope with the number of users.
But… but… Rust…
In terms of an optimal load spread, it’s best if the lemmiverse is split into multiple equally sized instances. If you use an instance just for yourself, it doesn’t actually decrease the load on the main servers in any way. The only thing you get is a guarantee that your instance won’t suddenly go down.
Does “Reject All” also object to legitimate interest?
As more and more people get direct fibre optic connections, this advantage will get smaller and smaller…
Rust has no garbage collector though. Memory is freed up as soon as the variable leaves the current scope.
I’m guessing the server was still set up to restart every 30 mins at the time this pic was taken. Then they tried disabling that and it was fine.