this post was submitted on 23 Jun 2023
6 points (100.0% liked)

Lemmy Server Performance

420 readers
1 users here now

Lemmy Server Performance

lemmy_server uses the Diesel ORM that automatically generates SQL statements. There are serious performance problems in June and July 2023 preventing Lemmy from scaling. Topics include caching, PostgreSQL extensions for troubleshooting, Client/Server Code/SQL Data/server operator apps/sever operator API (performance and storage monitoring), etc.

founded 1 year ago
MODERATORS
[email protected]
[email protected]
6
Lemmy Benchmarking - Working on tooling (lemmy.ml)
submitted 1 year ago by [email protected] to c/[email protected]
23 comments fedilink hide all child comments
 

Heyho,

as a PostgreSQL guy, i'm currently working an tooling environment to simulate load on a lemmy instance and measure the database usage.

The tooling is written in Go (just because it is easy to write parallel load generators with it) and i'm using tools like PoWA to have a deep look at what happens on the database.

Currently, i have some troubles with lemmy itself, that make it hard to realy stress the database. For example the worker pool is far to small to bring the database near any real performance bootlenecks. Also, Ratelimiting per IP is an issue.

I though about ignoring the reverse proxy in front of the lemmy API and just spoof Forwarded-For headers to work around it.

Any ideas are welcome. Anyone willing to help is welcome.

Goals if this should be:

  • Get a good feeling for the current database usage of lemmy
  • Optimize Statements and DB Layout
  • Find possible improvements by caching

As your loved core devs for lemmy have large Issue Tracker backlog, some folks that talk rust fluently would be great, so we can work on this dedicated topic and provide finished PR's

Greatings, Loki (@tbe on github)

you are viewing a single comment's thread
view the rest of the comments
[–] [email protected] 1 points 1 year ago (2 children)

Also, Ratelimiting per IP is an issue.

I have concerns about Lemmy having a pattern of hiding these behaviors under the cover. In other words, people running servers having no kind of operator console to know that it is happening. Ideally to me, it would be a setting to adjust in a screen to set for an instance to disable/set threshold. If one doesn't exist, maybe we can identify where in the code the limit is enforced and hand-edit the code.

[–] [email protected] 1 points 1 year ago (1 children)

In the site settings, rate limiting can be configured.

But, yes there should be some way to see if limits where hit or not. Maybe this could be done via prometheus and just provide the option to gather this data outside of lemmy itself.

[–] [email protected] 0 points 1 year ago (2 children)

I'm assuming Federation backdoor, API for incoming server to server transactions, doesn't have a rate limit? But I haven't validated that assumption.

Maybe this could be done via prometheus and just provide the option to gather this data outside of lemmy itself.

Something, from what I've seen so far, Lemmy has no application specific logging and just dumps everything into the system log. I really think operators need some concept of how many signups, logins, post, comments, communities they are getting per hour/day/week - which external websites are out there publishing number of communities and users.

[–] [email protected] 2 points 1 year ago (1 children)

The docker-compose.yml file in the docker folder has a config for postgres logging. We use it to diagnose performance issues in prod. There's a DB tag on the lemmy issue tracker, I suggest using that to track performance issues.

[–] [email protected] 1 points 1 year ago (1 children)

The problem is we need to get some data out of the big sites, Beehaw, Lemmy.ml, Lemmy.world - so that we can see what it is like having far more comments, likes, federation activity, and interactive user loads.

Is someone with a big instance willing to publish their logs?

[–] [email protected] 1 points 1 year ago

These log settings are not very good for production. The DB would spend more time logging, then working.

But: We can simulate this kinds of load with local toolings. My first tests look quiet promising, if i would only find the bug in my old pg-exporter i wrote for mal last employer ;) (yes, it is open source).

[–] [email protected] 2 points 1 year ago* (last edited 1 year ago)

Maybe @[email protected] can shed some light on this. Not sure if the find time, but it would be great to get some more input.

[–] [email protected] 0 points 1 year ago (1 children)
[–] [email protected] 2 points 1 year ago

PoWA (PostgreSQL Workload Analyzer)... how much overhead do you think this adds to a server? run in production?