[
https://issues.apache.org/jira/browse/HBASE-14283?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14715147#comment-14715147
]
Ben Lau commented on HBASE-14283:
---------------------------------
Hi Anoop. The problem isn't that we read a previous block and see that the
block is not the expected type. prevBlockOffset guarantees that we can seek to
the previous block of the same type as the current one. See the comments on
HFileBlock.getPrevBlockOffset(). We are always seeking to the previous data
block, we are simply not calculating how much to read correctly once we have
seeked to that previous data block because our prev data block size calculation
can include other blocks because of the layout of scannable section in
HFileV2+. We need a way of knowing apriori what the size of the previous data
block is. The method you describe is used in
HFileReaderImpl.readNextDataBlock(). Note that the reason this method works is
because this method can use the method
curBlock.getNextBlockOnDiskSizeWithHeader(). We need something similar to that
when seeking backwards in order to achieve optimal performance. Let me know if
I misunderstood what you meant.
> Reverse scan doesn’t work with HFile inline index/bloom blocks
> --------------------------------------------------------------
>
> Key: HBASE-14283
> URL: https://issues.apache.org/jira/browse/HBASE-14283
> Project: HBase
> Issue Type: Bug
> Reporter: Ben Lau
> Assignee: Ben Lau
> Attachments: HBASE-14283.patch, hfile-seek-before.patch
>
>
> Reverse scans do not work if an HFile contains inline bloom blocks or leaf
> level index blocks. The reason is because the seekBefore() call calculates
> the previous data block’s size by assuming data blocks are contiguous which
> is not the case in HFile V2 and beyond.
> Attached is a first cut patch (targeting
> bcef28eefaf192b0ad48c8011f98b8e944340da5 on trunk) which includes:
> (1) a unit test which exposes the bug and demonstrates failures for both
> inline bloom blocks and inline index blocks
> (2) a proposed fix for inline index blocks that does not require a new HFile
> version change, but is only performant for 1 and 2-level indexes and not 3+.
> 3+ requires an HFile format update for optimal performance.
> This patch does not fix the bloom filter blocks bug. But the fix should be
> similar to the case of inline index blocks. The reason I haven’t made the
> change yet is I want to confirm that you guys would be fine with me revising
> the HFile.Reader interface.
> Specifically, these 2 functions (getGeneralBloomFilterMetadata and
> getDeleteBloomFilterMetadata) need to return the BloomFilter. Right now the
> HFileReader class doesn’t have a reference to the bloom filters (and hence
> their indices) and only constructs the IO streams and hence has no way to
> know where the bloom blocks are in the HFile. It seems that the HFile.Reader
> bloom method comments state that they “know nothing about how that metadata
> is structured” but I do not know if that is a requirement of the abstraction
> (why?) or just an incidental current property.
> We would like to do 3 things with community approval:
> (1) Update the HFile.Reader interface and implementation to contain and
> return BloomFilters directly rather than unstructured IO streams
> (2) Merge the fixes for index blocks and bloom blocks into open source
> (3) Create a new Jira ticket for open source HBase to add a ‘prevBlockSize’
> field in the block header in the next HFile version, so that seekBefore()
> calls can not only be correct but performant in all cases.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
