Sylvain Wallez created ZOOKEEPER-3652:
-----------------------------------------
Summary: Improper synchronization in ClientCnxn
Key: ZOOKEEPER-3652
URL: https://issues.apache.org/jira/browse/ZOOKEEPER-3652
Project: ZooKeeper
Issue Type: Bug
Components: java client
Affects Versions: 3.5.6
Reporter: Sylvain Wallez
ZOOKEEPER-2111 introduced {{synchronized(state)}} statements in {{ClientCnxn}}
and {{ClientCnxn.SendThread}} to coordinate insertion in {{outgoingQueue}} and
draining it when the client connection isn't alive.
There are several issues with this approach:
- the value of the {{state}} field is not stable, meaning we don't always
synchronize on the same object.
- the {{state}} field is an enum value, which are global objects. So in an
application with several ZooKeeper clients connected to different servers, this
causes some contention between clients.
An easy fix is change those {{synchronized(state)}} statements to
{{synchronized(outgoingQueue)}} since it is local to each client and is what we
want to coordinate.
I'll be happy to prepare a PR with the above change if this is deemed to be the
correct way to fix it.
Another issue that makes contention worse is {{ClientCnxnSocketNIO.cleanup()}}
that is called from within the above synchronized block and contains
{{Thread.sleep(100)}}. Why is this sleep statement needed, and can we remove it?
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
