[
https://issues.apache.org/jira/browse/CASSANDRA-919?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12850045#action_12850045
]
Johan Oskarsson commented on CASSANDRA-919:
-------------------------------------------
There's quite a big difference between retrying one rpc request and retrying a
whole map task. A map task has a big overhead, in the common case a whole new
JVM has to be started per task. It would also have to refetch a significant
chunk of data while one rpc retry is a only a few thousand rows.
So if there's a short period where a few rpcs fail (due to gc:ing or similar)
retrying them instead of the task will speed up the overall job by a noticeable
amount of time. It's also worth noting that the hdfs client used by the
standard MapReduce jobs has a retry mechanism.
> Add thrift request retries to Hadoop input format
> -------------------------------------------------
>
> Key: CASSANDRA-919
> URL: https://issues.apache.org/jira/browse/CASSANDRA-919
> Project: Cassandra
> Issue Type: Improvement
> Components: Contrib
> Reporter: Johan Oskarsson
> Assignee: Johan Oskarsson
> Priority: Trivial
> Fix For: 0.7
>
> Attachments: CASSANDRA-919.patch
>
>
> In order to decrease overhead of restarting a map task and increase
> reliability of the record reader we should retry the get_range_slices
> requests if they fail.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
