[ 
https://issues.apache.org/jira/browse/HADOOP-1411?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Tom White updated HADOOP-1411:
------------------------------

    Attachment: createRetry-tw.patch

>I agree this interface is cleaner. But one disadvantage is that it is harder 
>for a client to use the framework to support RemoteException. A client needs 
>to build two exception2Policy maps and shouldRetry needs to be recusively 
>called one more time to search one more map.

It is slightly harder for clients to set up, but I would argue it is clearer. I 
wouldn't worry about the cost of an extra call to shouldRetry since it will be 
dwarfed by the time between retries.

>I agree that keying by Exception provides more type safety. But again calling 
>Class.forName is more costly than calling Class.getName. That's why I perfer 
>usign class name as the map's key. Alternatively we could enforce type safety 
>by providing an "add" handler to the map with Exception as a parameter and 
>internally implements the map using the class name as the key. But we could do 
>it in a different jira.

Agreed. I was imagining turning the Map<Exception, RetryPolicy> to a 
Map<String, RetryPolicy> on construction of RemoteExceptionDependentRetry 
rather then using Class.getName. I've attached a patch which does this. Does 
this work OK for you?

Finally, it would be nice to have a test for this new behaviour.


> AlreadyBeingCreatedException from task retries
> ----------------------------------------------
>
>                 Key: HADOOP-1411
>                 URL: https://issues.apache.org/jira/browse/HADOOP-1411
>             Project: Hadoop
>          Issue Type: Bug
>          Components: dfs
>    Affects Versions: 0.13.0
>            Reporter: Nigel Daley
>         Assigned To: Hairong Kuang
>            Priority: Blocker
>             Fix For: 0.13.0
>
>         Attachments: createRetry-tw.patch, createRetry.patch, 
> createRetry1.patch
>
>
> HADOOP-1407 indicates 2 bugs: a mapred bug which will be fixed as part of 
> 1407, and a DFSClient bug that will be fixed here.
> Note that the test run in 1407 was without speculation.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.

Reply via email to