[jira] [Commented] (CASSANDRA-5412) Lots of deleted rows came back to life after upgrade from 1.1.6 to 1.1.10

["Robert Coli (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/CASSANDRA-5412?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13893848#comment-13893848 ] 

Robert Coli commented on CASSANDRA-5412:
----------------------------------------

In retrospect, this seems likely to have been CASSANDRA-6503, I'm going to link them for google searchers.

> Lots of deleted rows came back to life after upgrade from 1.1.6 to 1.1.10
> -------------------------------------------------------------------------
>
>                 Key: CASSANDRA-5412
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-5412
>             Project: Cassandra
>          Issue Type: Bug
>          Components: Core
>    Affects Versions: 1.1.10
>         Environment: Ubuntu 10.04 LTS
> Sun Java 6 u39
> 1.1.6 => 1.1.10
>            Reporter: Arya Goudarzi
>
> Also per discussion here:  http://www.mail-archive.com/user@cassandra.apache.org/msg28905.html
> I was not able to find any answers as to why a simple upgrade process could bring back a lot of (millions) of deleted rows to life. We have successful repairs running on our cluster every night. Unless repair is not doing its job, it is not possible to the best of my knowledge that the deleted rows come back unless there is a bug. I have previously experienced this issue when I upgraded our sandbox cluster. I failed at every single attempt to reproduce the issue by restoring a fresh cluster from snapshot, and performing the upgrade from 1.1.6 to 1.1.10. I even exercised this with the snapshot of our production cluster before upgrading and was not successful. So, I finally made the decision to upgrade, and guess what?! Millions of deleted rows came back after the upgrade. 
> This time I confirmed the timestamps of the deleted rows that came back; they were actually before the time there were deleted. So, this is just like when tombstones get purged before they get propagated. We use nanosecond precision timestamps (19 digits).
> My discussion on the mailing list did not lead anywhere, though Aaron helped me find one another possible way of this happening by Hinted Handoff which I filed a separate ticket for. I don't believe this is an issue for us as we don't have nodes down for a long period of time. 



--
This message was sent by Atlassian JIRA
(v6.1.5#6160)