[
https://issues.apache.org/jira/browse/HBASE-23679?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17013260#comment-17013260
]
Josh Elser commented on HBASE-23679:
------------------------------------
Dropping some more details in here, should anyone else ever find themselves
back here.
* Funny that this user's RegionServers never (rarely?) actually OOME'ed. They
were left in this crap state because the JVM just kept hanging on. GC verbose
logging showed us that heap was just sitting completely full. We just sat with
30-60s GC pauses, getting next to nothing done.
* A heapdump showed (at worst) 60k+ instances of DistributedFileSystem sitting
on the heap, nearly all of them referenced by {{FileSystem$CACHE}}. It takes
some digging to realize this.
* Digging into each DFS -> DFSClient -> UGI -> Tokens, you can see the
delegation tokens which let us confirm that Bulk Load was the culprit
creating/leaking these.
* For the user, it was an extremely slow growth. One RegionServer would get
into this state every ~20 days across about that many RegionServers. So, in
reality, we were probably looking at even longer than 20days for a single
RegionServer to leak enough of these to cause issues.
* UGI being a part of FileSystem caching is crap, and we should avoid it for
anything other than "just the HBase user". It's extremely obtuse and error
prone, likely to just give us more heartache (lots of issues around this in the
past).
> FileSystem instance leaks due to bulk loads with Kerberos enabled
> -----------------------------------------------------------------
>
> Key: HBASE-23679
> URL: https://issues.apache.org/jira/browse/HBASE-23679
> Project: HBase
> Issue Type: Bug
> Reporter: Josh Elser
> Assignee: Josh Elser
> Priority: Critical
>
> Spent the better part of a week chasing an issue on HBase 2.x where the
> number of DistributedFileSystem instances on the heap of a RegionServer would
> grow unbounded. Looking at multiple heap-dumps, it was obvious to see that we
> had an immense number of DFS instances cached (in FileSystem$Cache) for the
> same user, with the unique number of Tokens contained in that DFS's UGI
> member (one hbase delegation token, and two HDFS delegation tokens – we only
> do this for bulk loads). For the user's clusters, they eventually experienced
> 10x perf degradation as RegionServers spent all of their time in JVM GC (they
> were unlucky to not have RegionServers crash outright, as this would've,
> albeit temporarily, fixed the issue).
> The problem seems to be two-fold with changes by HBASE-15291 being largely
> the cause. This issue tried to close FileSystem instances which were being
> leaked – however, it did this by instrumenting the method
> {{SecureBulkLoadManager.cleanupBulkLoad(..)}}. Two big issues with this
> approach:
> # It relies on clients to call this method (client's hanging up will leak
> resources in RegionServers)
> # This method is only called on the RegionServer hosting the first Region of
> the table which was bulk-loaded into. For multiple RegionServers, they are
> left to leak resources.
> HBASE-21342 later tried to fix an issue where FS objects were now being
> closed prematurely via reference-counting (which appears to work fine), but
> does not address the other two issues above. Point #2 makes debugging this
> issue harder than normal because it doesn't manifest on a single node
> instance :)
> Through all of this, I (re)learned the dirty history of UGI and how its
> caching doesn't work so great HADOOP-6670. I see trying to continue to
> leverage the FileSystem$CACHE as a potentially dangerous thing (we've been
> back here multiple times already). My opinion at this point is that we should
> cleanly create a new FileSystem instance during the call to
> {{SecureBulkLoadManager#secureBulkLoadHFiles(..)}} and close it in a finally
> block in that same method. This both simplifies the lifecycle of a FileSystem
> instance in the bulk-load codepath but also helps us avoid future problems
> with UGI and FS caching. The one downside is that we pay the penalty to
> create a new FileSystem instance, but I'm of the opinion that we cross that
> bridge when we get there.
> Thanks for [~jdcryans] and [~busbey] for their help along the way.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
