[
https://issues.apache.org/jira/browse/HDFS-2492?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13135781#comment-13135781
]
Steve Loughran commented on HDFS-2492:
--------------------------------------
I'm not sure it's quite so serious, as for the normal script based mapper it
works; it looks for the script file and assumes the presence of the file means
cross-rack.
It's only for people who have written their own DNS mapper that the problem
will arise, and the problem isn't topology mapping, its that the replication
logic that tries to ensure that replicas are spread across multiple racks.
The risk then is: if you have a custom DNS mapper, all your blocks may end up
on the same rack, making you less resilient to rack failures. Your data is
still spread across machines, and the actual placement code itself will work to
place the blocks in multiple racks. The only issue is that at a switch or rack
failure is slightly more likely to take all replicas of a block offline until
the switch comes back up.
It's easy to fix though, I will try and do something today.
Before then: who is running a custom DNS Mapper? Do you expect it to work
unchanged on 0.23+, or is adding another method to an interface something you
can handle?
> BlockManager cross-rack replication checks only work for ScriptBasedMapping
> ---------------------------------------------------------------------------
>
> Key: HDFS-2492
> URL: https://issues.apache.org/jira/browse/HDFS-2492
> Project: Hadoop HDFS
> Issue Type: Bug
> Affects Versions: 0.23.0, 0.24.0
> Reporter: Steve Loughran
> Priority: Blocker
>
> The BlockManager cross-rack replication checks only works if script files are
> used for replication, not if alternate plugins provide the topology
> information.
> This is because the BlockManager sets its rack checking flag if there is a
> filename key
> {code}
> shouldCheckForEnoughRacks =
> conf.get(DFSConfigKeys.NET_TOPOLOGY_SCRIPT_FILE_NAME_KEY) != null;
> {code}
> yet this filename key is only used if the topology mapper defined by
> {code}
> DFSConfigKeys.NET_TOPOLOGY_NODE_SWITCH_MAPPING_IMPL_KEY
> {code}
> is an instance of {{ScriptBasedMapping}}
> If any other mapper is used, the system may be multi rack, but the Block
> Manager will not be aware of this fact unless the filename key is set to
> something non-null
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
