Hello everyone!
I’ll be working on migrating our image uploads from local files over to object store this Sunday starting at about 11am PT.
This may require extended lemmy.ca downtime as pict-rs has to be stopped while the migration happens, but I’m hoping I can keep things running as read-only and run the migration off a second instance.
Updates will be posted here - https://status.lemmy.ca/maintenance/257501
As someone who’s gone through that recently - brace yourself.
I believe you can actually avoid downtime altogether if you’re willing to lose post thumbnails that would be generated during the migration. And even those might be possible to get regenerated, although I haven’t looked into it.
Issue 1 - lemmy-ui will crash and burn if it can’t load the site logo. Specifically UI will simply say Server error and dev tools will show 500. Nginx will also only show 500 with no more detail and lemmy-ui will spam you something about having received an empty array. Apps will continue to work fine as they use API rather than the UI. Because I found out about it after having started the migration, I had to resort to setting site logo to NULL in the database. You might get away with just unsetting the logo in /admin, but prepare the query just in case. On mobile now, can’t give you the query itself.
Issue 2 - pitcrs is not actually stateless. You MUST save that weird sled database and keep using it after the migration. Otherwise you’ll find out the Server error gremlins coming out more and more often.
Issue 3 - pictrs migration code is a bit on the shit side. It tries to handle missing files (why are they missing in the first place?! I haven’t deleted anything ffs!), but eventually gives up and stops the migration. I had to restart it enough times to lose count. Luckily enough it does resume, but keeps retrying all missing files, not ever discarding them. Output becomes unreadable.
Issue 4 - migration is dogshit slow. Took me nearly 4 hours to migrate ~20GB.
There might be something else in your case as you’re running a much larger instance, but this is a prime example of how much alpha lemmy+pictrs really is.
lol yeah I discovered this one a while ago. Thanks for the tip on setting it to null.
Yeah I was assuming so, didn’t think it was stateless.
Awesome.
Even more awesome. I do know there have been performance improvements in recent patches, so if you did your migration awhile ago then we should hopefully be faster. We do have 350gb to migrate though.
My current thought is:
I did the migration last week. London DigitalOcean to Amsterdam Backblaze. Sticking to “stable” pictrs, so 0.4.2. I think the biggest issue for the migration is that it’s single threaded. If there’s a way to tell it to do many items at once - that would help a great deal. Might be worth looking into 0.5.x alpha releases.
The problem I see with your plan is running pictrs during migration - you will still get thumbnails generated during that time. All this will lead to split brain. Not too big of a deal as you simply won’t have thumbnails for recent sites, but, technically speaking, you will still have data loss in terms of imaging. Might as well avoid the overhead of making a copy of the db.
$0.02
Yep but I was thinking that losing some thumbnails was better than having image delivery completely off for hours…
Fair point 👍
I got a good laugh at this. What a fucking shit show my god. Absolute insanity lol. Good on you for sharing