[
https://issues.apache.org/jira/browse/SOLR-11932?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16353357#comment-16353357
]
John Gallagher commented on SOLR-11932:
---------------------------------------
[~shalinmangar] very good point! I think that actually applies to the
ConnectionLoss exception as well? Those SolrZooKeeper instances were probably
not valid after connection loss. I understand your recommendation - but I took
a slightly different approach to reduce the amount of changed code, and to
bring back the connection-loss retry that I think was also affected by the
problem you pointed out:
[https://github.com/apache/lucene-solr/pull/321]
After some initial test failures that I fixed in the shutdown hook, I am down
to just one repeatable error:
org.apache.solr.schema.ManagedSchemaRoundRobinCloudTest I'm not sure if it is
due to a leaking retry thread or not. I believe it is just the thread leak
detection, because adding a sleep in the @AfterClass method of that class
(allowing the retry to drain) seems to pass the test, but nevertheless, here is
the failed test result:
[^org.apache.solr.schema.ManagedSchemaRoundRobinCloudTest.error.txt]
> ZkCmdExecutor: Retry ZkOperation on SessionExpired
> ---------------------------------------------------
>
> Key: SOLR-11932
> URL: https://issues.apache.org/jira/browse/SOLR-11932
> Project: Solr
> Issue Type: Bug
> Security Level: Public(Default Security Level. Issues are Public)
> Affects Versions: 7.2
> Reporter: John Gallagher
> Assignee: Ishan Chattopadhyaya
> Priority: Major
> Attachments: SessionExpiredLog.txt,
> org.apache.solr.schema.ManagedSchemaRoundRobinCloudTest.error.txt,
> zk_retry.patch
>
> Time Spent: 50m
> Remaining Estimate: 0h
>
> We are seeing situations where an operation, such as changing a replica's
> state to active after a recovery, fails because the zk session has expired.
> However, these operations seem like they are retryable, because the
> ZookeeperConnect receives an event that the session expired and tries to
> reconnect.
> That makes the SessionExpired handling scenario seem very similar to the
> ConnectionLoss handling scenario, so the ZkCmdExecutor seems like it could
> handle them in the same way.
>
> Here's an example stack trace with some slight redactions:
> [^SessionExpiredLog.txt] In this case, a zk operation (a read) failed with a
> SessionExpired event, which seems retriable. The exception kicked off a
> reconnection, but seems like the subsequent operation, (publishing as active)
> failed (perhaps it was using a stale connection handle at that point?)
>
> Regardless, the watch mechanism that reestablishes connection on
> SessionExpired seems sufficient to allow the ZkCmdExecutor to retry that
> operation at a later time and have hope of succeeding.
>
> I have included a simple patch we are trying that catches both exceptions
> instead of just ConnectionLossException: [^zk_retry.patch]
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
