RFC: Implement clean shutdown and in-memory document GC #233
Conversation
|
The latest updates on your projects. Learn more about Vercel for Git ↗︎
1 Skipped Deployment
|
|
@rbtying is attempting to deploy a commit to the drifting-corp Team on Vercel. A member of the Team first needs to authorize it. |
When the server receives a ctrl-c, stop taking new requests, immediately trigger a save on all open documents, wait for all the documents to persist, and then exit.
crates/y-sweet/src/server.rs
Outdated
| .layer(middleware::from_fn_with_state( | ||
| redact_errors, | ||
| Self::redact_error_middleware, | ||
| )); |
There was a problem hiding this comment.
I like the middleware approach here, but since the redact error middleware is a no-op if redact_errors = false, should we just conditionally not add the middleware in that case and avoid passing through the boolean state?
|
Unfortunately it looks like yrs is not thread-safe (y-crdt/y-crdt#373), so I might tinker a bit to see if there's a way to do the yrs subscription drops single-threaded before merging |
I think this is unblocked w/ the 0.19 update! |
Merged! Thanks for all your work on getting this across the finish line! |
I noticed in #228 that the y-sweet binary is currently missing at least best-effort persistence-before-shutdown, and GCing documents out of memory. I understand that y-sweet is currently mostly deployed as CloudFlare workers, so both of those are a bit moot - CloudFlare will kill the process entirely once there aren't any connections for long enough -- but I figured that it might be nice to have something for outside-CloudFlare as well.
This PR implements:
y-sweet-corewith implementation details of the server. The garbage collection task holds the cancellation token for the persistence worker, so in graceful shutdown scenarios it should always persist the current state.