[
https://issues.apache.org/jira/browse/HBASE-19543?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16296060#comment-16296060
]
Duo Zhang commented on HBASE-19543:
-----------------------------------
After a deep thought, I think we need to redesign the stack of replication
management architecture. I plan to split the stack into two levels, the lower
level is the storage level, and we may have zk based or table based
implementation. And most logic are in the upper layer, such as peer management
and queue management.
> Implement a replication peer storage for master use only
> --------------------------------------------------------
>
> Key: HBASE-19543
> URL: https://issues.apache.org/jira/browse/HBASE-19543
> Project: HBase
> Issue Type: Sub-task
> Components: proc-v2, Replication
> Reporter: Duo Zhang
> Assignee: Duo Zhang
>
> For now, we will do sanity checks at the same time when updating replication
> peer. But this is not a safe way for procedure based replication peer
> modification.
> For the old zk watcher way, the only thing is updating the data on zk, so if
> the data is updated and then we crashes, there is no problem.
> For the new procedure way, we need to trigger refresh by ourselves after
> updating zk. If we crashes after the updating and before we record the state
> change of the procedure, we may fail with IllegalArgumentException when we
> execute the procedure next time since the data on zk has already been updated.
> So the current way is to do sanity checks in PRE_PEER_MODIFICATION state, and
> in UPDATE_STORAGE state we will not do sanity checks any more, just
> update(overwrite) the peer storage.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
