[
https://issues.apache.org/jira/browse/HBASE-10355?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13901036#comment-13901036
]
Devaraj Das commented on HBASE-10355:
-------------------------------------
Looks good overall. One question -
Consider a case where a table is created with three replicas. All the replicas
are not assigned (hence, the meta is not updated with the locations yet). Now a
client comes in, and the information that he sees would be partial (let's say
location_for_replica_0 and location_for_replica_2 are known). The location for
replica_1 would be null. It seems like the client wouldn't know about the
missing location until he refetches the row from the meta (which would not
happen until there is activity like a region move). I am talking about the
ReplicaRegionServerCallable class. Maybe if the location is null, the client
should hit the meta always?
When we support online schema changes for #replicas, in the case where we
reduce the count, we would have to cancel the RPCs that are trying to hit the
no-longer-existent replica. Similarly, when we increase the count, we should
discover these new replicas.
One nit is that all the methods/fields that you made public for tests should be
annotated with VisibleForTesting.
> Failover RPC's from client using region replicas
> ------------------------------------------------
>
> Key: HBASE-10355
> URL: https://issues.apache.org/jira/browse/HBASE-10355
> Project: HBase
> Issue Type: Sub-task
> Components: Client
> Reporter: Enis Soztutar
> Assignee: Nicolas Liochon
> Fix For: 0.99.0
>
> Attachments: 10355.v1.patch
>
>
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
