Change the range/block index scheme from [start,end) to (start, end], and index
range/block by endKey, specially in HFile
-------------------------------------------------------------------------------------------------------------------------
Key: HBASE-1978
URL: https://issues.apache.org/jira/browse/HBASE-1978
Project: Hadoop HBase
Issue Type: New Feature
Components: io, master, regionserver
Reporter: Schubert Zhang
>From the code review of HFile (HBASE-1818), we found the HFile allows
>duplicated key. But the old implementation would lead to missing of duplicated
>key when seek and scan, when the duplicated key span multiple blocks.
We provide a patch (HBASE-1841 is't step1) to resolve above issue. This patch
modified HFile.Writer to avoid generating a problem hfile with above
cross-block duplicated key. It only start a new block when current appending
key is different from the last appended key. But it still has a rish when the
user of HFile.Writer append many same duplicated key which lead to a very large
block and need much memory or Out-of-memory.
The current HFile's block-index use startKey to index a block, i.e. the
range/block index scheme is [startKey,endKey).
As refering to the section 5.1 of the Google Bigtable paper.
"The METADATA table stores the location of a tablet under a row key that is an
encoding of the tablet's table identifer and its end row."
The theory of Bigtable's METADATA is same as the BlockIndex in a SSTable or
HFile, so we should use EndKey in HFile's BlockIndex. In my experiences of
Hypertable, the METADATA is also "tableID:endRow".
We would change the index scheme in HFile, from [startKey,endKey) to
(startKey,endKey]. And change the binary search method to meet this index
scheme.
This change can resolve above duplicated-key issue.
Note:
The totally fix need to modify many modules in HBase, seems include HFile, META
schema, some internal code, etc.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
