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.
Minor thing but over night both wefwef and Memmy clients are showing the wrong comment score (karma) against my profile, and given they are showing the same amount I assume itβs related to API fed data. Value was correct yesterday. Easy for me to confirm given I have only 2 dozen posts and the value has dropped to single digits.
Not a biggie, but figured Iβd report it in case there was some issue causing that. Might be some optimisation around indexing or something has intentionally or unintentionally impacted that.
Otherwise the service feels much more stable currently. No timeouts today where itβs been very frequent the past few days. Nice job. π
I notified the wefwef dev of this.
Cheers. Seems to be incrementing ok with my posts today, but total seems like it was reset or is incrementing only from the past 24-48 hours of comments, so is short 50 or so.
Servers still performing well since my last post but 5 hours earlier.