[infinispan-issues] [JBoss JIRA] (ISPN-3835) Index Update command is processed before the registry listener is triggered

[Sanne Grinovero (JIRA)]

    [ https://issues.jboss.org/browse/ISPN-3835?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13032073#comment-13032073 ] 

Sanne Grinovero commented on ISPN-3835:
---------------------------------------

It might still be relevant but as far as I remember we heavily cleaned up this code area this summer, so maybe not.
By "command" I mean the RPC: the problem was that a node receiving an indexing command from a different node on the backend, would not have the same safeguarding methods like we have in the QueryInterceptor to verify if the the is already registered.

It's hard to reproduce as in theory the sender of the RPC did the registration first, but I witnessed situations in which you could still have a race condition. The race condition might be gone, as I think the registry is meant to be updated in a blocking way on all nodes? Still I think the receiver of the RPC should better be safe and check for the type to be registered, so to compensate for eventuality inconsistencies in the registry state itself.

> Index Update command is processed before the registry listener is triggered
> ---------------------------------------------------------------------------
>
>                 Key: ISPN-3835
>                 URL: https://issues.jboss.org/browse/ISPN-3835
>             Project: Infinispan
>          Issue Type: Bug
>          Components: Embedded Querying
>    Affects Versions: 6.0.0.Final
>            Reporter: Sanne Grinovero
>            Assignee: Gustavo Fernandes
>            Priority: Critical
>              Labels: 64QueryBlockers
>             Fix For: 7.1.0.CR1
>
>
> When using the InfinispanIndexManager backend the master node might receive an index update command about an index which it hasn't defined yet.
> Index definitions are triggered by the type registry, which in turn is driven by the ClusterRegistry and an event listener on the ClusterRegistry. It looks like slaves are sending update requests before the master has processed the configuration event.
> This leads to index update commands to be lost (with a stacktrace logged)



--
This message was sent by Atlassian JIRA
(v6.3.11#6341)