Steven McDonald created ZOOKEEPER-3315:
------------------------------------------
Summary: Exceptions in callbacks should be handlable by the
application
Key: ZOOKEEPER-3315
URL: https://issues.apache.org/jira/browse/ZOOKEEPER-3315
Project: ZooKeeper
Issue Type: Improvement
Reporter: Steven McDonald
Attachments: ExceptionTest.java
Hi,
In [KAFKA-7898|https://issues.apache.org/jira/browse/KAFKA-7898], a
{{NullPointerException}} in a {{MultiCallback}} caused a Kafka cluster to
become unhealthy in such a way that manual intervention was needed to recover.
The cause of this particular {{NullPointerException}} is fixed in Kafka 2.2.x
(with a proposed documentation update in
[ZOOKEEPER-3314|https://issues.apache.org/jira/projects/ZOOKEEPER/issues/ZOOKEEPER-3314]),
but I am interested in improving the resiliency of Kafka (and by extension the
Zookeeper client library) against such bugs.
Quoting the stack trace from KAFKA-7898:
{code}
[2019-02-05 14:28:12,525] ERROR Caught unexpected throwable
(org.apache.zookeeper.ClientCnxn)
java.lang.NullPointerException
at
kafka.zookeeper.ZooKeeperClient$$anon$8.processResult(ZooKeeperClient.scala:217)
at org.apache.zookeeper.ClientCnxn$EventThread.processEvent(ClientCnxn.java:633)
at org.apache.zookeeper.ClientCnxn$EventThread.run(ClientCnxn.java:508)
{code}
The "caught unexpected throwable" message comes from [within the Zookeeper
client
library|https://github.com/apache/zookeeper/blob/release-3.4.13/src/java/main/org/apache/zookeeper/ClientCnxn.java#L641].
I think that try/catch is pointless, because removing it causes the message to
instead be logged
[here|https://github.com/apache/zookeeper/blob/release-3.4.13/src/java/main/org/apache/zookeeper/server/ZooKeeperThread.java#L60],
with no discernable change in behaviour otherwise. Explicitly exiting the
{{EventThread}} when this happens does not help (I don't think it gets
restarted).
I suspect the best approach here might be to allow the application to register
a callback to notify it of unhandlable exceptions within the Zookeeper library,
since Zookeeper has no way of knowing what approach makes sense for the
application. Of course, this is already technically possible in this case by
having the application catch all exceptions in every callback, but that doesn't
seem very practical.
This is especially problematic with distributed applications, since they are
generally designed to tolerate the loss of a node, so it is preferable to have
the application be allowed to terminate itself rather than risk inconsistent
state.
I am attaching a simple Zookeeper client which does nothing except throw a
{{NullPointerException}} as soon as it receives a callback, to illustrate the
problem. Running this results in:
{code}
232 [main-EventThread] ERROR org.apache.zookeeper.ClientCnxn - Error while
calling watcher
java.lang.NullPointerException
at ExceptionTest.process(ExceptionTest.java:31)
at
org.apache.zookeeper.ClientCnxn$EventThread.processEvent(ClientCnxn.java:539)
at org.apache.zookeeper.ClientCnxn$EventThread.run(ClientCnxn.java:514)
{code}
This comes from
[here|https://github.com/apache/zookeeper/blob/7256d01a26412cd35a46edab6de9ac8c5adf5bb3/zookeeper-server/src/main/java/org/apache/zookeeper/ClientCnxn.java#L541],
which simply logs the occurrence but provides no way for my application to
handle the failure.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
