[ 
https://issues.apache.org/jira/browse/CASSANDRA-6943?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13950765#comment-13950765
 ] 

Benedict commented on CASSANDRA-6943:
-------------------------------------

For anyone looking for a little more information: this problem would be hit 
when inserting a larger number of records into an existing record than were 
currently present in that record (excluding empty records) causing a over-spill 
into a new parent node, after which there was a competing update to the 
underlying record, and the update that overspilled aborted before finishing the 
construction of its replacement btree. In this case it did not clear the state 
of the builder for the spilled-over parent, as it only visited all those from 
its original root/parent down to reset the state (it would ascend to the real 
parent on finishing construction), so that when the next update finished it 
included this old partially constructed parent, resulting in a malformed btree.

The simple solution is to always go up resetting all parents on completion, as 
well as down

> UpdateFunction.abortEarly can cause BTree.update to leave its Builder in a 
> bad state, which affects future operations
> ---------------------------------------------------------------------------------------------------------------------
>
>                 Key: CASSANDRA-6943
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-6943
>             Project: Cassandra
>          Issue Type: Bug
>            Reporter: Russ Hatch
>            Assignee: Benedict
>             Fix For: 2.1 beta2
>
>         Attachments: 6943.2.txt, 6943.txt, node4.log, node4_jstack.log, 
> node5.log, node5_jstack.log, node6.log, node6_jstack.log, node7.log, 
> node7_jstack.log, screenshot.png, screenshot2.png, stress_jstack.log
>
>
> Running performance scenarios I have seen this characteristic drop in 
> performance happen several times. A similar effect was reproduced in another 
> test cluster. Operations eventually come to a standstill.
> !screenshot.png!
> !screenshot2.png!



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Reply via email to