Abhishek Pal created RATIS-2611:
-----------------------------------
Summary: Some gRPC install-snapshot replies don't clear the
pending queue
Key: RATIS-2611
URL: https://issues.apache.org/jira/browse/RATIS-2611
Project: Ratis
Issue Type: Bug
Reporter: Abhishek Pal
Assignee: Abhishek Pal
When calling the onNext method, in the various result cases
[here|https://github.com/apache/ratis/blob/f4c3781468e756220836e88d953fd5598177afdf/ratis-grpc/src/main/java/org/apache/ratis/grpc/server/GrpcLogAppender.java#L676]
some cases do not call `removePending()`.
- NOT_LEADER
- CONF_MISMATCH
- UNRECOGNIZED
- SNAPSHOT_EXPIRED
these cases do not clear the pending queue and this may possibly fail on the
Preconditions.assertSame() check.
`pending.poll()` removes the oldest enqueued request index. The handler assumes
that for each reply we will have one addPending() and then one removePending()
call in order.
So if we take an example of chunk install with `SNAPSHOT_EXPIRED`
1. Let us assume we send chunk [0,1,2] -> three `addPending()` calls -> pending
queue [0, 1, 2]
2. Reply with chunk 0 as success -> one `removePending()` call ->
pending.poll() returns 0, condition passed
3. Reply with chunk 1 as SNAPSHOT_EXPIRED -> pending queue still has [1, 2]
This is not a critical bug as onError() a new handler is generated which causes
stream teardown and that abandons the broken handler, but it would be good to
fix this issue.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
