Tanuj Khurana created PHOENIX-8029:
--------------------------------------
Summary: Make the replication mode-switch notification non-blocking
Key: PHOENIX-8029
URL: https://issues.apache.org/jira/browse/PHOENIX-8029
Project: Phoenix
Issue Type: Sub-task
Reporter: Tanuj Khurana
Assignee: Tanuj Khurana
_*ReplicationLogGroup.checkAndSetModeAndNotify*_ wakes the Disruptor event
handler by calling the blocking sync(). It runs on the ZooKeeper cache-event
thread (the ACTIVE_IN_SYNC and ACTIVE_NOT_IN_SYNC state listeners) and on the
forwarder thread. That sync() can block for up to the replication sync timeout
(WAL sync timeout plus ZK session timeout), which stalls HA state notifications
for the group, and it an abort the RegionServer from a ZK or forwarder thread.
Fix: publish a non-blocking MODE_CHECK marker with RingBuffer.tryNext(), as the
rotation SWAP marker does. The handler applies the switch on the marker when no
syncs are pending and the current batch is empty; otherwise it applies at the
next sync point. A closed group doesn't apply switches requested by other
threads during the close() drain.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
