[
https://issues.apache.org/jira/browse/HBASE-8367?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13644677#comment-13644677
]
Brian Dougan commented on HBASE-8367:
-------------------------------------
Yes, that's correct, the finally block is hit by that test. However, when that
test runs, the code simply return from inside the finally block. Therefore,
the if statement I'm adding after the try/finally to make sure a failure code
is returned if one or more items remains in the queue is not hit as once the
finally block executes, the method exits.
The only time I can hit the new block I'm adding is if the try/finally exits
without error and there are still items in the queue. I haven't investigated
every case that can make this occur, but the one I ran into is if one or more
files times out or fails server-side (assuming it fails enough to exhaust the
retries). As the tests are currently written, I'm struggling to come up with a
way replicate this. We'd either need to be able to intercept the server call
(to force it to fail) or access to the server side to make it fail after the
call is successful.
> LoadIncrementalHFiles does not return an error code or throw Exception when
> failures occur due to timeouts.
> -----------------------------------------------------------------------------------------------------------
>
> Key: HBASE-8367
> URL: https://issues.apache.org/jira/browse/HBASE-8367
> Project: HBase
> Issue Type: Improvement
> Components: mapreduce
> Affects Versions: 0.92.1, 0.92.2
> Environment: Red Hat 6.2
> Java 1.6.0_26
> Hadoop 2.0.0-mr1-cdh4.1.1
> HBase 0.92.1-cdh4.1.1
> Reporter: Brian Dougan
> Priority: Minor
> Fix For: 0.94.8
>
> Attachments: LoadIncrementalHFiles-HBASE-8367.patch
>
>
> The LoadIncrementalHFiles (completebulkload) command will exit with a success
> code (or lack of Exception) when one or more of the HFiles fail to be
> imported through a few ways (mainly when timeouts occur). Instead, it simply
> logs error messages to the log which makes it difficult to automate the
> import of HFiles programmatically.
> The heart of the LoadIncrementalHFiles class (doBulkLoad) returns void and
> has essentially the following structure.
> {code:title=LoadIncrementalHFiles.java}
> try {
> ...
>
> } finally {
> pool.shutdown();
> if (queue != null && !queue.isEmpty()) {
> StringBuilder err = new StringBuilder();
> err.append("-------------------------------------------------\n");
> err.append("Bulk load aborted with some files not yet loaded:\n");
> err.append("-------------------------------------------------\n");
> for (LoadQueueItem q : queue) {
> err.append(" ").append(q.hfilePath).append('\n');
> }
> LOG.error(err);
> }
> }
> {code}
> As you can see, instead of returning an error code, a success indicator, or
> simply throwing an Exception, an error message is sent to the log. This
> results in something like the following in the logs.
> {quote}
> Bulk load aborted with some files not yet loaded:
> -------------------------------------------------
>
> hdfs://prmdprod/user/userxxx/hfile/TABLE-1365721885510/record/_tmp/TABLE,2.bottom
>
> hdfs://prmdprod/user/userxxx/hfile/TABLE-1365721885510/record/_tmp/TABLE,2.top
>
> hdfs://prmdprod/user/userxxx/hfile/TABLE-1365721885510/record/_tmp/TABLE,1.bottom
>
> hdfs://prmdprod/user/userxxx/hfile/TABLE-1365721885510/record/_tmp/TABLE,1.top
> {quote}
> Without some sort of indication, it's not currently possible to chain this
> command to another or to programmatically consume this class and be certain
> of a successful import.
> This class should be enhanced to return non-success in whatever way makes
> sense to the community. I don't really have a strong preference, but one of
> the following should work fine (at least for my needs).
> * boolean return value on doBulkLoad (non-zero on run method)
> * Response object on doBulkLoad detailing the files that failed (non-zero on
> run method)
> * throw Exception in the finally block when files failed after the error is
> written to the log (should automatically cause non-zero on run method)
> It would also be nice to get this to the 0.94.x stream so it get included in
> the next Cloudera release. Thanks!
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
