[
https://issues.apache.org/jira/browse/HBASE-10719?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13931425#comment-13931425
]
Matt Corgan commented on HBASE-10719:
-------------------------------------
I debated that when making the Cell interface but thought it might be too big
of a step. I'm excited to see things are moving to where that's a possibility!
Some thoughts:
Downsides:
* ByteBuffers are heavyweight compared to array
** for HeapBB it would be an additional layer of indirection between Cell and
the byte[] that backs HeapByteBuffer
** looking at source there's more method calls to get to the backing array
including big/little-endian logic
* more complex to work with
* not reusable - have to spawn a new BB with .wrap(..) to move to a new backing
array
* encoders copy bytes to assemble the key anyway, so simpler to stick with
array for now
* current KeyValue is backed by array, so would be a big step to move
everything to BB
Upsides:
* i've seen sometimes DirectBB outperforms array (especially if you don't have
to copy it to byte[] first)
* HeapBB calls probably get inlined enough to match byte[] read performance
* probably better longer term solution if off-heap becomes the norm
I thought a first step could be making only the value field a ByteBuffer (or
ByteRange) since even encoders don't (or shouldn't) need to copy that.
As Nick mentions, maybe ByteRanges with varying implementations are the way to
go since they have a friendlier interface (comparable unsigned bytes). They
could have a ByteBuffer implementation, albeit with an additional level of
indirection over the underlying on-heap byte[] (inlining to the rescue?).
> Cell interface may need a ByteBuffer rather than a byte[]
> ---------------------------------------------------------
>
> Key: HBASE-10719
> URL: https://issues.apache.org/jira/browse/HBASE-10719
> Project: HBase
> Issue Type: Sub-task
> Reporter: ramkrishna.s.vasudevan
> Assignee: ramkrishna.s.vasudevan
>
> See discussion in HBASE-10531.
> The byte[] backing a cell may be off heap also. In such a case having a
> byte[] may require a copy offheap to on heap. So better to have ByteBuffer
> backing a Cell.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
