Jie Huang created ZOOKEEPER-3774:
------------------------------------
Summary: Close quorum socket asynchronously on the leader to avoid
ping being blocked by long socket closing time
Key: ZOOKEEPER-3774
URL: https://issues.apache.org/jira/browse/ZOOKEEPER-3774
Project: ZooKeeper
Issue Type: Sub-task
Components: server
Reporter: Jie Huang
Fix For: 3.7.0
In ZOOKEEPER-3574 we close the quorum sockets on followers asynchronously when
a leader is partitioned away so the shutdown process will not be stalled by
long socket closing time and the followers can quickly establish a new quorum
to serve client requests.
We've found that the long socket closing time can cause trouble on the leader
too when a follower is partitioned away if the partition is detected by
PingLaggingDetector. When the ping thread detects partition, it tries to
disconnect the follower. If the socket closing time is long, the ping thread
will be blocked and no ping is sent to any follower--even the ones still
connected to the leader--since the ping thread is responsible for sending pings
to all followers. When followers don't receive pings, they don't send ping
response. When the leader don't receive ping response, the sessions expire.
To prevent good sessions from expiring, we need to close the socket
asynchronously on the leader too.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
