Another day, another update.
More troubleshooting was done today. What did we do:
- Yesterday evening @phiresky@phiresky@lemmy.world did some SQL troubleshooting with some of the lemmy.world admins. After that, phiresky submitted some PRs to github.
- @cetra3@lemmy.ml created a docker image containing 3PR's: Disable retry queue, Get follower Inbox Fix, Admin Index Fix
- We started using this image, and saw a big drop in CPU usage and disk load.
- We saw thousands of errors per minute in the nginx log for old clients trying to access the websockets (which were removed in 0.18), so we added a
return 404in nginx conf for/api/v3/ws. - We updated lemmy-ui from RC7 to RC10 which fixed a lot, among which the issue with replying to DMs
- We found that the many 502-errors were caused by an issue in Lemmy/markdown-it.actix or whatever, causing nginx to temporarily mark an upstream to be dead. As a workaround we can either 1.) Only use 1 container or 2.) set ~~
proxy_next_upstream timeout;~~max_fails=5in nginx.
Currently we're running with 1 lemmy container, so the 502-errors are completely gone so far, and because of the fixes in the Lemmy code everything seems to be running smooth. If needed we could spin up a second lemmy container using the ~~proxy_next_upstream timeout;~~ max_fails=5 workaround but for now it seems to hold with 1.
Thanks to @phiresky@lemmy.world , @cetra3@lemmy.ml , @stanford@discuss.as200950.com, @db0@lemmy.dbzer0.com , @jelloeater85@lemmy.world , @TragicNotCute@lemmy.world for their help!
And not to forget, thanks to @nutomic@lemmy.ml and @dessalines@lemmy.ml for their continuing hard work on Lemmy!
And thank you all for your patience, we'll keep working on it!
Oh, and as bonus, an image (thanks Phiresky!) of the change in bandwidth after implementing the new Lemmy docker image with the PRs.
Edit So as soon as the US folks wake up (hi!) we seem to need the second Lemmy container for performance. So that's now started, and I noticed the proxy_next_upstream timeout setting didn't work (or I didn't set it properly) so I used max_fails=5 for each upstream, that does actually work.
This is why having a big popular instance isn't all bad. It helps detect and fix the scaling problems and inefficiencies for all the other 1000s of instances out there!
This, if everyone kept just spreading out to smaller instances as suggested in the beginning, while still a sensible thing to do, no one would have noticed these performance issues. We need to think a few years out, assuming Lemmy succeeds and Reddit dies, and expect that "small instance" will mean 50k users.
I sincerely doubt reddit will die anytime soon, it'll just exist as its own thing that it's new target audience gets bored with and moves on from in a few years when something new and flashy catches their eye in the app store. Just like they do all the other apps designed in exactly the same fashion that reddit is currently morphing into.
Meanwhile Lemmy will be slowly building it's communities up to be what reddit used to be.
I'm actually kinda waiting a few releases to start promoting my instance anywhere, letting some other brave instance admins work the kinks out a bit first.
Agreed. I decided to keep my community on lemmy.world specifically because of the community investment I see being put into it.
If this project is to stay for the long haul, we gotta load test it and stabilize it. These folks are doing the important work here. Large instances are more or less inevitable if Lemmy sticks.