[
https://issues.apache.org/jira/browse/HBASE-9440?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13767365#comment-13767365
]
Lars Hofhansl commented on HBASE-9440:
--------------------------------------
bq. You mean going via front door?
Sorry, yes.
bq. What should take-away be? What we need to dig in on? To go faster, we need
to do the prefixtreeblocks and pull blocks up out of hfile?
Not entirely sure... The data suggest that with 50 cols the best we can do is a
~5x improvement (and that is if we can pass the KVs up with *no* overhead).
For tall tables, we might want to check what the per row overhead is (is it the
creation of the Result object for example?)
Yes, to go faster we need to be able to scan encoded KVs and pass the up
unchanged to the various heaps, to avoid all that baggage of the key for every
column (0.94 and trunk still do that for the prefix encoders). We need to be
able to pass KVs around that are not backed by a continuous byte[].
> Pass blocks of KVs from HFile scanner to the StoreFileScanner and up
> --------------------------------------------------------------------
>
> Key: HBASE-9440
> URL: https://issues.apache.org/jira/browse/HBASE-9440
> Project: HBase
> Issue Type: Bug
> Reporter: Lars Hofhansl
>
> Currently we read KVs from an HFileScanner one-by-one and pass them up the
> scanner/heap tree. Many time the ranges of KVs retrieved from
> StoreFileScanner (by StoreScanners) and HFileScanner (by StoreFileScanner)
> will be non-overlapping. If chunks of KVs do not overlap we can sort entire
> chunks just by comparing the start/end key of the chunk. Only if chunks are
> overlapping do we need to sort KV by KV as we do now.
> I have no patch, but I wanted to float this idea.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
