hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ravi Prakash (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HDFS-6075) Introducing "non-replication mode"
Date Wed, 12 Mar 2014 20:24:43 GMT

    [ https://issues.apache.org/jira/browse/HDFS-6075?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13932300#comment-13932300

Ravi Prakash commented on HDFS-6075:

I feel another option to do this would be to disable replication on a set of *nodes* temporarily
(not cluster wide). i.e. the list of nodes, and a timeout after which replications should
be done.

> Introducing "non-replication mode"
> ----------------------------------
>                 Key: HDFS-6075
>                 URL: https://issues.apache.org/jira/browse/HDFS-6075
>             Project: Hadoop HDFS
>          Issue Type: New Feature
>          Components: datanode, namenode
>            Reporter: Adam Kawa
>            Priority: Minor
> Afaik, HDFS does not provide an easy way to temporarily disable the replication of missing
> If you would like to temporarily disable the replication, you would have to
> * set dfs.namenode.replication.interval (_The periodicity in seconds with which the namenode
computes repliaction work for datanodes_ Default 3) to something very high. *Disadvantage*:
you have to restart the NN
> * go into the safe-mode. *Disadvantage*: all write operations will fail
> We have the situation that we need to replace our top-of-rack switches for each rack.
Replacing a switch should take around 30 minutes. Each rack has around 0.6 PB of data. We
would like to avoid an expensive replication, since we know that we will put this rack online
quickly. To avoid any downtime, or excessive network transfer, we think that temporarily disabling
the replication could fit us.
> The default block placement policy puts blocks into two racks, so when one rack temporarily
goes offline, we still have an access to at least replica of each block. Of course, if we
lose this replica, then we would have to wait until the rack goes back online. This is what
the administrator should be aware of.
> This feature could disable the replication
> * globally - for a whole cluster
> * partially - e.g. only for missing blocks that come from a specified set of DataNodes.
So a file like "we_will_be_back_soon" :) could be introduced, similar to include and exclude.

This message was sent by Atlassian JIRA

View raw message