[
https://issues.apache.org/jira/browse/HDDS-400?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16604137#comment-16604137
]
Elek, Marton commented on HDDS-400:
-----------------------------------
Yes. That's an other option to check replication state only in case of node
failures. This was more safe to me (will work, even if datanode failure or
problems in the retry logic) and the only downside is the performance
implication but I don't think it has big overhead. It also handles the case
when a new node is introduced which has some more replicas.
I am not sure which one is the best option this is just a just good enough
implementation for me.
Maybe we can modify to Mapping/ContainerStateManager to check the replication
numbers in case of any replication information update. But it would be part of
a bigger refactor of the container report handling.
> Check global replication state of the reported containers on SCM
> ----------------------------------------------------------------
>
> Key: HDDS-400
> URL: https://issues.apache.org/jira/browse/HDDS-400
> Project: Hadoop Distributed Data Store
> Issue Type: Bug
> Components: SCM
> Reporter: Elek, Marton
> Assignee: Elek, Marton
> Priority: Blocker
> Fix For: 0.2.1
>
> Attachments: HDDS-400.001.patch, HDDS-400.002.patch
>
>
> Current container replication handler compare the reported containers with
> the previous report. It handles over an under replicated state.
> But there is no logic to check the cluster-wide replication count. If a node
> is went down it won't be detected.
> For the sake of simplicity I would add this check to the
> ContainerReportHandler (as of now). So all the reported container should have
> enough replicas.
> We can check the performance implication with genesis, but as a first
> implementation I think it could be good enough.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
