cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Tyler Hobbs (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (CASSANDRA-11087) Queries on compact storage tables in mixed version clusters can return incorrect results
Date Thu, 28 Jan 2016 17:01:40 GMT

    [ https://issues.apache.org/jira/browse/CASSANDRA-11087?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15121889#comment-15121889
] 

Tyler Hobbs commented on CASSANDRA-11087:
-----------------------------------------

Should we add a flag to the upgrade tests to control flushing before upgrading?  It seems
like we want to test both behaviors (to ensure commitlog replay is also covered), but there
could easily be other legacy sstable reading bugs out there.

> Queries on compact storage tables in mixed version clusters can return incorrect results
> ----------------------------------------------------------------------------------------
>
>                 Key: CASSANDRA-11087
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-11087
>             Project: Cassandra
>          Issue Type: Bug
>          Components: Local Write-Read Paths
>            Reporter: Sam Tunnicliffe
>            Assignee: Sam Tunnicliffe
>             Fix For: 3.0.x, 3.x
>
>
> Whilst writing a dtest for CASSANDRA-11045, it becomes apparent that queries on compact
storage tables are broken during the 3.0 upgrade (and this has probably been the case since
day 1). 
> tl;dr In a cluster with a mix of < 3.0 and 3.0 nodes, reads on COMPACT STORAGE tables
may not include all results. 
> To repro: tables are created and data written before any nodes are upgraded to 3.0+,
some nodes are then upgraded putting the cluster into a mixed state.
> Now, when a query is run where the coordinator is a < 3.0 node, any 3.0+ replica which
has not yet run upgradesstables always returns 0 results.  Once upgradesstables is run, the
replica returns the correct results. Likewise, if the data is inserted after the node is upgraded,
the results are correct. If the 3.0 node acts as the coordinator, the results are also correct
and so once all nodes are upgraded, the problem goes away.
> The behaviour can be seen for both single partition and range requests as [this dtest|https://github.com/beobal/cassandra-dtest/commit/91bb9ffd8fb761ad3454187d2f05f05a6e7af930]
demonstrates.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message