core: Fix shutdown failing accepted RPCs #12539
Conversation
This fixes a race where RPCs could fail with "UNAVAILABLE: Channel shutdown invoked" even though they were created before channel.shutdown(). This basically adopts the internalStart() logic from DelayedStream, although the stream is a bit different because it has APIs that can be called before start() and doesn't need to handle cancel() without start(). The ManagedChannelImpltest had the number of due tasks increase because start() running earlier creates a DelayedStream. Previously the stream wasn't created until runDueTasks() so the mockPicker had already been installed and it could use a real stream from the beginning. But that's specific to the test; in practice it'd be a delayed stream before and after this change. See grpc#12536
|
Read your comment:
If this PR solves the root cause for race, then why are we not reverting prev PR #12536, which was just fixing the test to make it green? |
Because the test was broken in other ways. I had written earlier in that PR:
The flake that I had reproduced is solved by this PR. But the test still had issues that are fixed by the other PR. |
3 participants
This fixes a race where RPCs could fail with "UNAVAILABLE: Channel shutdown invoked" even though they were created before channel.shutdown().
This basically adopts the internalStart() logic from DelayedStream, although the stream is a bit different because it has APIs that can be called before start() and doesn't need to handle cancel() without start().
The ManagedChannelImpltest had the number of due tasks increase because start() running earlier creates a DelayedStream. Previously the stream wasn't created until runDueTasks() so the mockPicker had already been installed and it could use a real stream from the beginning. But that's specific to the test; in practice it'd be a delayed stream before and after this change.
See #12536