hadoop-hdfs-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Chris Nauroth (JIRA)" <j...@apache.org>
Subject [jira] [Resolved] (HDFS-350) DFSClient more robust if the namenode is busy doing GC
Date Tue, 22 Mar 2016 21:30:25 GMT

     [ https://issues.apache.org/jira/browse/HDFS-350?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel

Chris Nauroth resolved HDFS-350.
    Resolution: Not A Problem

I'm resolving this issue.  In current versions, the client is more robust to this kind of
failure.  The RPC layer implements retry policies.  Retried operations are handled gracefully
using either an inherently idempotent implementation of the RPC or the retry cache for at-most-once
execution.  In the event of an extremely long GC, the client would either retry and succeed
after completion of the GC, or in more extreme cases it would trigger an HA failover and the
client would successfully issue its call to the the new active NameNode.

> DFSClient more robust if the namenode is busy doing GC
> ------------------------------------------------------
>                 Key: HDFS-350
>                 URL: https://issues.apache.org/jira/browse/HDFS-350
>             Project: Hadoop HDFS
>          Issue Type: Improvement
>            Reporter: dhruba borthakur
>            Assignee: dhruba borthakur
> In the current code, if the client (writer) encounters an RPC error while fetching a
new block id from the namenode, it does not retry. It throws an exception to the application.
This becomes especially bad if the namenode is in the middle of a GC and does not respond
in time. The reason the client throws an exception is because it does not know whether the
namenode successfully allocated a block for this file.
> One possible enhancement would be to make the client retry the addBlock RPC if needed.
The client can send the block list that it currently has. The namenode can match the block
list send by the client with what it has in its own metadata and then send back a new blockid
(or a previously allocated blockid that the client had not yet received because the earlier
RPC timedout). This will make the client more robust!
> This works even when we support Appends because the namenode will *always* verify that
the client has the lease for the file in question.

This message was sent by Atlassian JIRA

View raw message