[
https://issues.apache.org/jira/browse/HADOOP-6834?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12881362#action_12881362
]
Ahad Rana commented on HADOOP-6834:
-----------------------------------
Hi Hong,
Yes I would be happy to provide a patch. As far as triggering the failure, if
you append a LongWritable with a negative value as the first key value and use
the LongWritable's default Comparator, this will trigger the bug. I would be
happy to provide a code snippet if required. In the case of LongWritable's
comparator, it interprets the uninitialized zero bytes of the lastKey as a zero
(since it does not check length), and thus this causes the out-of-order
insertion exception. Either LongWritable's comparator is implemented
improperly, or the contract implied by RawComparator is not clear on what
happens in the case where you compare a non-zero byte array to a zero byte
array. Probably both LongWritable's RawComparator should be fixed, but either
way, passing in a uninitialized lastKey is going to cause problems for someone
else down the road, so probably best to fix this as well ?
> TFile.append compares initial key against null lastKey
> --------------------------------------------------------
>
> Key: HADOOP-6834
> URL: https://issues.apache.org/jira/browse/HADOOP-6834
> Project: Hadoop Common
> Issue Type: Bug
> Components: io
> Affects Versions: 0.20.1, 0.20.2
> Reporter: Ahad Rana
>
> The following code in TFile.KeyReigster.close:
> byte[] lastKey = lastKeyBufferOS.getBuffer();
> int lastLen = lastKeyBufferOS.size();
> if (tfileMeta.getComparator().compare(key, 0, len, lastKey, 0,
> lastLen) < 0) {
> throw new IOException("Keys are not added in sorted order");
> }
> compares the initial key (passed in via TFile.Writer.append) against a
> technically NULL lastKey. lastKey is not initialized until after the first
> call to TFile.Writer.append. The underlying RawComparator interface used for
> comparisons does not stipulate the proper behavior when either length 1 or
> length 2 is zero. In the case of LongWritable, its WritableComparator
> implementation does an unsafe read on the passed in byte arrays b1 and b2.
> Since TFile pre-allocates the buffer used for storing lastKey, this passes a
> valid buffer with zero count to LongWritable's comparator, which ignores
> length and thus produces incorrect results.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
