[
https://issues.apache.org/jira/browse/HBASE-17507?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15833446#comment-15833446
]
Yu Li commented on HBASE-17507:
-------------------------------
bq. Yes, A switch can be added, but will increase the complicity of the patch.
WAL is not only appended in put/delete/Append/Increment, but also when writing
open/close/compaction/flush markers
The mvcc preassign feature introduced a per-region level lock, and this is some
additional cost for RS with few regions or cluster with light workload. And as
I mentioned, we also observed some limitations in other case. I'm glad that the
design is well recognized, but I'd say be cautious to make it non-configurable.
[~allan163]
bq. Have you compared with multiwal?
Yes, in our online product cluster we're using 4 wals. With more wals could
lower the contention caused by CountDownLatch, but still cannot resolve our
problem, which forced us to make this fix. [~Apache9]
P.S. I'm wondering why the "Reply" function is removed from JIRA, I think
"Reply" makes discussion looks more compacted... [~stack] [~busbey]
> Add document for the mvcc preassign feature introduced in HBASE-16698
> ---------------------------------------------------------------------
>
> Key: HBASE-17507
> URL: https://issues.apache.org/jira/browse/HBASE-17507
> Project: HBase
> Issue Type: Task
> Reporter: Yu Li
> Assignee: Yu Li
>
> In this JIRA we will supply a design doc of the mvcc preassign mechanism
> introduced in HBASE-16698, as well as more detailed performance data of it
> for better reference.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
