You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@ignite.apache.org by "Pavel Kovalenko (JIRA)" <ji...@apache.org> on 2018/03/01 12:05:00 UTC

[jira] [Assigned] (IGNITE-7467) Verify partition update counters and sizes on partition map exchange

     [ https://issues.apache.org/jira/browse/IGNITE-7467?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Pavel Kovalenko reassigned IGNITE-7467:
---------------------------------------

    Assignee: Alexey Goncharuk  (was: Pavel Kovalenko)

Ready to review.
PR: https://github.com/apache/ignite/pull/3579
TC Run: https://ci.ignite.apache.org/project.html?projectId=IgniteTests24Java8&branch_IgniteTests24Java8=pull%2F3579%2Fhead

Changes show problem with update counters inconsistency in a couple of tests. List of the affected tests will be published later.

> Verify partition update counters and sizes on partition map exchange
> --------------------------------------------------------------------
>
>                 Key: IGNITE-7467
>                 URL: https://issues.apache.org/jira/browse/IGNITE-7467
>             Project: Ignite
>          Issue Type: Improvement
>          Components: persistence
>    Affects Versions: 2.1
>            Reporter: Alexey Goncharuk
>            Assignee: Alexey Goncharuk
>            Priority: Major
>
> In Ignite we heavily rely on an invariant that under no load owning partitions will have equal sizes and, more importantly, equal partition counters. This invariant becomes even more important when persistence is enabled.
> However, due to a possible bug in the code, this invariant can be violated which in a long run may lead to an undetected data loss. We need to take best effort to detect such a situation as soon as possible.
> Currently, we already send partition update counters during partition map exchange. We can also send partition sizes and verify that corresponding partitions in OWNING state have equal partition update counters and sizes.
> If a divergence detected, we can:
> 1) Always print out an error message to the log
> 2) Move the corresponding caches to the read-only state to prevent further corruption or operating on invalid data
> Also, we can introduce a ./control.sh command which will trigger an empty exchange to verify the partition states.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)