[
https://issues.apache.org/jira/browse/ACCUMULO-4391?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15474860#comment-15474860
]
Ivan Bella commented on ACCUMULO-4391:
--------------------------------------
[~kturner] There was a slew of different exceptions that would occur, mostly
stemming from the fact that the same decompressor was being used by multiple
threads:
1) Most of them stem from using the same decompressor across multiple threads.
The root cause of this was that the same decompressor was being returned to the
codec pool multiple times and then later the same decompressor was being
returned to multiple threads resulting is mass chaos. This is basically solved
by ensuring that the decompressor is returned only once in the BCFile finish
call.
2) The second set of issues stemmed from the close being called at the same
time other threads were reading from the FSDataInputStream. This resulted in
the decompressor being returned underneath and subsequently reused by some
other read all while some read was occurring. This is solved by the
synchronization added in the BoundedRangeInputFileStream. The synchronization
within the CacheableBlockFile was added more by examination of the code to
ensure we were not closing the same FSDataInputStream concurrently while
reading/closing within the BoundedRangeInputFileStream.
3) The third problem was the available call being called concurrently on the
FSDataInputStream while it was being closed in the BoundedRangeInputFileStream.
I initially added synchronization there as well, however after our initial
discussions in the pull request it was determined (and verified on our system)
that the available call is really not being used except by the initialization
of the CompressionInputStream the results of which are used in the getPos call
which is not used. Hence simply avoiding the underlying available call seemed
the best course of action.
> Source deepcopies cannot be used safely in separate threads in tserver
> ----------------------------------------------------------------------
>
> Key: ACCUMULO-4391
> URL: https://issues.apache.org/jira/browse/ACCUMULO-4391
> Project: Accumulo
> Issue Type: Bug
> Components: core
> Affects Versions: 1.6.5
> Reporter: Ivan Bella
> Assignee: Ivan Bella
> Fix For: 1.6.6, 1.7.3, 1.8.1, 2.0.0
>
> Original Estimate: 24h
> Time Spent: 11h 40m
> Remaining Estimate: 12h 20m
>
> We have iterators that create deep copies of the source and use them in
> separate threads. As it turns out this is not safe and we end up with many
> exceptions, mostly down in the ZlibDecompressor library. Curiously if you
> turn on the data cache for the table being scanned then the errors disappear.
> After much hunting it turns out that the real bug is in the
> BoundedRangeFileInputStream. The read() method therein appropriately
> synchronizes on the underlying FSDataInputStream, however the available()
> method does not. Adding similar synchronization on that stream fixes the
> issues. On a side note, the available() call is only invoked within the
> hadoop CompressionInputStream for use in the getPos() call. That call does
> not appear to actually be used at least in this context.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
