[
https://issues.apache.org/jira/browse/TAJO-704?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14074258#comment-14074258
]
ASF GitHub Bot commented on TAJO-704:
-------------------------------------
Github user blrunner commented on the pull request:
https://github.com/apache/tajo/pull/77#issuecomment-50131715
Hi, @jihoonson
Thanks for your detailed review. I replied for your descriptions as follows:
* I've just fixed NPE in the web UI via Hyunsik. Because my desktop has
broken down.
* DerbyStore is not appropriate for TajoMaster HA. As you know, just one
tajo master invoke the store. If another master try to invoke it, it never run
itself. Also, if we set another catalog uri for backup master, it is a
incorrect configuration. Because they are unequal in every way.
* If you run two masters on same host and you set same directory for pid,
you must meet pid error. But I think that it is trivial problem. Because we can
avoid easily by documents.
* If backup master become active, existing running query will finish
successfully, and you will see success message in tsql. But you will never find
executed query history in the web UI. Because TajoMaster contains its query
history in memory, thus backup master can't know previous active master
information. I already know this problem. And I'll save query simple history in
share storage (like HDFS).
For reference, I'll write documents for TajoMaster HA. And I'll update
start-tajo.sh to start backup masters.
Cheers
> TajoMaster HA
> -------------
>
> Key: TAJO-704
> URL: https://issues.apache.org/jira/browse/TAJO-704
> Project: Tajo
> Issue Type: New Feature
> Components: tajo master
> Affects Versions: 0.9.0
> Reporter: Jaehwa Jung
> Assignee: Jaehwa Jung
> Fix For: 0.9.0
>
> Attachments: TAJO-704.Henrick.01.patch.txt, TajoMasterHAdraft.pdf
>
>
> TajoMaster is a Single Point of Failure in a Tajo Cluster because TajoMaster
> is the central controlling entity for all components of the Tajo system.
> TajoMaster failure prevents clients from submitting new queries to the
> cluster, and results in the disruption of the ability to run insert overwrite
> queries because the TajoWorker can’t apply its statistical information to
> CatalogStore. Therefore, the high-availability (HA) of TajoMaster is
> essential for the high-availability of Tajo generally.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
