kvcoord: Release catchup reservation before re-acquire attempt #105083
Conversation
|
This change is |
Release catchup scan reservation prior to attemt to re-acquire it. Failure to do so could result in a stuck mux rangefeed when enough ranges encounter an error at the same time (which can happen if e.g. a node gets restarted). Fixes cockroachdb#105058 Release note (bug fix): Fix a bug in mux rangefeed implementation that may cause mux rangefeed to become stuck if enough ranges enounter certain error concurrently.
There was a problem hiding this comment.
I can understand how current change makes reservations aligned with activeMuxRangeFeed lifecycle.
BuI can only see us having double reservations if we had to do a errInfo.resolveSpan branch and have new feeds being blocked by feed they are trying to replace. This doesn't look like the case of nodes being restarted, but having many concurrent splits when table grows.
If that's the case, can you update the description?
|
bors r+ |
|
Build succeeded: |
|
How about v22.2 @miretskiy? Or should we simply say mux is not recommended on v22.2? |
not applicable. this is rewritten code that didn't exist (in its current shape) in 22.2 |
Release catchup scan reservation prior to attemt to re-acquire it. Failure to do so could result in a stuck mux rangefeed when enough ranges encounter an error, such as range split, prior to receiving the first checkpoint event, that would cause additional attempts to acquire catchup scan quota.
Fixes #105058
Release note (bug fix): Fix a bug in mux rangefeed implementation that may cause mux rangefeed to become stuck if enough ranges encounter certain error concurrently.