Pierre Salagnac created SOLR-17405:
--------------------------------------
Summary: Zookeeper session can be re-established by multiple
threads concurrently
Key: SOLR-17405
URL: https://issues.apache.org/jira/browse/SOLR-17405
Project: Solr
Issue Type: Bug
Security Level: Public (Default Security Level. Issues are Public)
Affects Versions: 9.6, 8.11
Reporter: Pierre Salagnac
Attachments: stack.png
Because of a bug in SolrCloud, the Zookeeper session can be re-established by
multiple threads concurrently when an expiration occurs.
This portion of the code assumes it is mono-threaded. Because of the bug, the
last thread re-establishing the session can waif for 30 seconds per core,
waiting for it to be marked {{DOWN}} while it was previously marked {{ACTIVE}}
by another thread. With a high number of cores, the Solr server can hang for
hours before taking traffic again.
Following exception shows two threads were reestablishing the session
concurrently. {{ZkController.createEphemeralLiveNode()}} should never be
invoked twice for the same Zookeeper session.
!stack.png!
h2. Root cause
This bug occurs because several threads can re-establish the session
concurrently.
It cannot happen at the first expiration of the session, thanks to a thread
pool with a single thread to execute the zookeeper Watcher.
Bellow is a code snippet from class {{SolrZkClient.ProcessWatchWithExecutor}}
{code:java}
if (watcher instanceof ConnectionManager) {
zkConnManagerCallbackExecutor.submit(() -> watcher.process(event));
} else {
.......
}
{code}
Using this dedicated thread pool (with a single thread) is supposed to ensure
we don’t handle watches for connection related events with multiple threads.
This works well for the first session expiration.
Now, when we re-establish the session after the first expiration, we don’t use
this wrapper to register the watch.
It is done directly in {{ConnectionManager}} without wrapping the ZK watch. In
the following snippet, _“this”_ is the ZK watcher instance, but it is not
wrapper to use a {{{}ProcessWatchWithExecutor{}}}. This means the next events
will directly be handled by any ZK callback thread.
{code:java}
connectionStrategy.reconnect(zkServerAddress,client.getZkClientTimeout(), this,
{code}
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
