[
https://issues.apache.org/jira/browse/HBASE-17484?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15829324#comment-15829324
]
ramkrishna.s.vasudevan commented on HBASE-17484:
------------------------------------------------
bq.Any evidence to present to justify why we should have these new types? We've
gone this route a bunch of times in our history adding caching and then undoing
it. The argument is that reading these lengths by parsing offheap bytes is slow?
Previously this offheapKV was getting used only in read path and we found that
caching it did not have a big impact because they are all short lived objects
that gets created in case of random reads. And during a block seek caching
these values helped us.
But in case of write path we add them to the memstore and we see that if we
have more caching then the flush related calculations related to blocking the
incoming write requests (due to higher limit breach) are impacted more by this
caching. If you see the KeyValue case we don't do these caching and so we try
to do the same here for the write path so that we have the same data and heap
overhead for both KeyValues and OffheapKeyValues.
bq.After this patch, how many versions of KeyValue do we have?
We thought about this and this would be a question. I think after that
ExtendedCell was added we have considerably reduced the variations. Let me
check on that on exact number.
bq.Do we have to have a OffheapKeyValue and Cached.
With 'Cached' I meant the an OffheapKV with more of its states cached.
ByteBufferCell for now is purely only for direct. Anything with non-direct we
create KV out of it.
bq.OffheapKeyValue.FIXED_OVERHEAD + Bytes.SIZEOF_INT + Bytes.SIZEOF_SHORT; ....
is that a double ''?
I think it is ok as we cache the keyLen (an int) and rowLen (a short) here.
> Add non cached version of OffheapKV for write path
> --------------------------------------------------
>
> Key: HBASE-17484
> URL: https://issues.apache.org/jira/browse/HBASE-17484
> Project: HBase
> Issue Type: Sub-task
> Components: regionserver
> Affects Versions: 2.0.0
> Reporter: ramkrishna.s.vasudevan
> Assignee: ramkrishna.s.vasudevan
> Fix For: 2.0.0
>
> Attachments: HBASE-17484.patch
>
>
> After running lot of different performance tests for various scenarios and
> with multi threads we thought that is better to have a version of OffheapKV
> in write path that does not cache anything and its fixed_overhead is equal to
> that in KeyValue.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
