[
https://issues.apache.org/jira/browse/HBASE-12282?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14180217#comment-14180217
]
Anoop Sam John commented on HBASE-12282:
----------------------------------------
bq. BBs may have more index and position management internally. BR may come
handy here.
What we really need is to have a new impl of BB so that we can avoid overhead
in read APIs. Also the changes can make sure that the inlining will be
possible. So change to BR alone won't really help us. And it will add extra
overhead of BR to BB. Andy had done some testing for the inlining and we are
not getting inlining for some of the APIs.
bq.If we want to deal with BBs one issue in case of DBB is that max size of
integer is the max size of a DBB that can be allocated. How can that be
addressed. Can our HFileBlock be bigger than that?
You mean the ByteBuffer#allocateDirect(int capacity) and int limitation? Any
way that is there with byte[] also no? I think int max size is fair enough
bq.Some of the steps followed in the patch could be hacky but if we could find
ways to address them then we can move towards the goal of BB
Yes looks hacky. We can think to better it.
We discussed abt having APIs in Cell to do the compares (like row compare) but
the issue is within one Cell impl it knows abt itself but not about the
incoming. If we continue to do this compares based on getxxxArray() APIs, we
will end up in many times byte[] creation and data copy from buffer to this
array. This is applicable in the Heap compares in Scanner as well as some of
the Filters (like SCVF)
> Ensure Cells and its implementations work with Buffers also
> -----------------------------------------------------------
>
> Key: HBASE-12282
> URL: https://issues.apache.org/jira/browse/HBASE-12282
> Project: HBase
> Issue Type: Sub-task
> Components: regionserver, Scanners
> Affects Versions: 0.99.1
> Reporter: ramkrishna.s.vasudevan
> Assignee: ramkrishna.s.vasudevan
> Fix For: 2.0.0, 0.99.2
>
> Attachments: HBASE-12224_2.patch
>
>
> This issue can be used to brainstorm and then do the necessary changes for
> the offheap work. All impl of cells deal with byte[] but when we change the
> Hfileblocks/Readers to work purely with Buffers then the byte[] usage would
> mean that always the data is copied to the onheap. Cell may need some
> interface change to implement this.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
