[
https://issues.apache.org/jira/browse/IGNITE-11574?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ivan Bessonov updated IGNITE-11574:
-----------------------------------
Description:
The problem is in this code (GridCachePartitionExchangeManager#discoLsnr) :
{code:java}
if (cache.state().transition()) {
if (log.isDebugEnabled())
log.debug("Adding pending event: " + evt);
pendingEvts.add(new PendingDiscoveryEvent(evt, cache));
}{code}
Problem occurs when "setBaseline" and "nodeLeft" events happen simultaneously
(+ some undetermined conditions).
"nodeLeft" provokes exchange, and while that exchange isn't finished
"setBaseline" is invoked. This moves cluster into a transition state and
"CacheAffinityChangeMessage" from the exchange cannot be processed properly. At
the same time "setBaseline" cannot be completed before the exchange, so we have
a deadlock.
Reproducer attached.
was:
The problem is in this code (GridCachePartitionExchangeManager#start0) :
{code:java}
if (cache.state().transition()) {
if (log.isDebugEnabled())
log.debug("Adding pending event: " + evt);
pendingEvts.add(new PendingDiscoveryEvent(evt, cache));
}{code}
Problem occurs when "setBaseline" and "nodeLeft" events happen simultaneously
(+ some undetermined conditions).
"nodeLeft" provokes exchange, and while that exchange isn't finished
"setBaseline" is invoked. This moves cluster into a transition state and
"CacheAffinityChangeMessage" from the exchange cannot be processed properly. At
the same time "setBaseline" cannot be completed before the exchange, so we have
a deadlock.
Reproducer attached.
> Exchange on NodeLeft event hangs when cluster is in transition state
> --------------------------------------------------------------------
>
> Key: IGNITE-11574
> URL: https://issues.apache.org/jira/browse/IGNITE-11574
> Project: Ignite
> Issue Type: Bug
> Reporter: Ivan Bessonov
> Assignee: Ivan Bessonov
> Priority: Major
> Attachments: ExchangeDeadlockTest.java
>
>
> The problem is in this code (GridCachePartitionExchangeManager#discoLsnr) :
> {code:java}
> if (cache.state().transition()) {
> if (log.isDebugEnabled())
> log.debug("Adding pending event: " + evt);
> pendingEvts.add(new PendingDiscoveryEvent(evt, cache));
> }{code}
> Problem occurs when "setBaseline" and "nodeLeft" events happen simultaneously
> (+ some undetermined conditions).
> "nodeLeft" provokes exchange, and while that exchange isn't finished
> "setBaseline" is invoked. This moves cluster into a transition state and
> "CacheAffinityChangeMessage" from the exchange cannot be processed properly.
> At the same time "setBaseline" cannot be completed before the exchange, so we
> have a deadlock.
> Reproducer attached.
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
