[
https://issues.apache.org/jira/browse/PHOENIX-3950?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16638895#comment-16638895
]
Vincent Poon commented on PHOENIX-3950:
---------------------------------------
+1 to [~tdsilva]'s idea.
I think we should still try to handle the case I described above. We have to
keep the IndexTool and MetadataRegionObserver task from interfering with each
other.
An easy "brute force" approach for now would be to permanently disable an index
(index_disable_timestamp=0) if the transition is from 'b' -> 'x'.
In other words, if a write failure happens during an initial index build, then
operator intervention is required (likely to rebuild the index), or if we can
detect it, then we can do it programmatically as well.
> IndexTool marks index active even when index building fails
> -----------------------------------------------------------
>
> Key: PHOENIX-3950
> URL: https://issues.apache.org/jira/browse/PHOENIX-3950
> Project: Phoenix
> Issue Type: Bug
> Reporter: James Taylor
> Priority: Major
>
> We've seen cases in which the MR-based index building tool (IndexTool) marks
> an index as active when the index building clearly did not succeed. It might
> be related to a timeout occurring and then not being propagated to the client.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
