[ 
https://issues.apache.org/jira/browse/HBASE-4209?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13113859#comment-13113859
 ] 

Roman Shaposhnik commented on HBASE-4209:
-----------------------------------------

@stack

> At least log the exception.
> What is the exception we are suppressing?  Should we at least log it here too?

Basically, the only root cause of possible exceptions here is the code in 
suppressHdfsShutdownHook() (which is quite fragile and overly aggressive in
throwing exceptions if things don't look quite right). Now, if you look at 
how this code is executed in a non-standalone case -- HBase does bail on 
exceptions 
thrown from there. We can adopt the same approach in a standalone and unit 
testing
case, but I wasn't sure it was the right thing to do. After all, NOT bailing out
(and lets say simply logging these things) will be no worse than what we 
currently
have (not calling shutdown hooks at all). On the other hand -- if we do bail out
aggressively we get a better chance of catching incompatibilities between 
suppressHdfsShutdownHook() logic and future hadoop releases during unit tests.

So perhaps -- bailing out would make the most sense of all.

> The HBase hbase-daemon.sh SIGKILLs master when stopping it
> ----------------------------------------------------------
>
>                 Key: HBASE-4209
>                 URL: https://issues.apache.org/jira/browse/HBASE-4209
>             Project: HBase
>          Issue Type: Bug
>          Components: master
>            Reporter: Roman Shaposhnik
>            Assignee: Roman Shaposhnik
>         Attachments: HBASE-4209.patch.txt
>
>
> There's a bit of code in hbase-daemon.sh that makes HBase master being 
> SIGKILLed when stopping it rather than trying SIGTERM (like it does for other 
> daemons). When HBase is executed in a standalone mode (and the only daemon 
> you need to run is master) that causes newly created tables to go missing as 
> unflushed data is thrown out. If there was not a good reason to kill master 
> with SIGKILL perhaps we can take that special case out and rely on SIGTERM.

--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira

        

Reply via email to