[
https://issues.apache.org/jira/browse/HBASE-17744?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15899735#comment-15899735
]
Yu Li commented on HBASE-17744:
-------------------------------
I'd like to list the issues/improvements we found through this simple embedded
mode:
1. It's better to use soft instead of weak reference for IdReadWriteLock, see
HBASE-17747
2. There's correctness issue for increment and append under high concurrency,
as revealed by the newly add test case in this JIRA. Refer to the "FIXME" part
of {{TestEmbeddedHBase}}
What's more, we found the performance of master code base is better than our
customized 1.1.2 under embedded mode, but the advantage disappeared in the
end-2-end case with YCSB, which IMO shows there're some perf regression on the
wire, and we could give it some further investigation after this simple
embedded mode is in.
We could also contribute the changes in YCSB to test embedded mode, and
actually it's easy to implement one for an embedded library.
> Implement simple embedded mode
> ------------------------------
>
> Key: HBASE-17744
> URL: https://issues.apache.org/jira/browse/HBASE-17744
> Project: HBase
> Issue Type: Sub-task
> Reporter: Yu Li
> Assignee: Yu Li
> Attachments: HBASE-17744.patch
>
>
> In simple embedded mode, we will
> 1. Start only one HMaster thread, which also serves as regionserver, and
> could serve tables other than meta
> 2. Implement direct access w/o RPC
> 3. Keep using the existing region assignment, procedure and recovery mechanism
> 4. Keep using zookeeper (mini-cluster)
> More details please refer to the patch.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
