[
https://issues.apache.org/jira/browse/HBASE-16492?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15434645#comment-15434645
]
Phil Yang commented on HBASE-16492:
-----------------------------------
Sounds interesting :)
Indeed, "the suitable points" for checking if we can drop requests are usually
the points that we may block the current handler thread. Setting a timeout
provided by client to these blocking operations can free the handler much
earlier then now. And if we can make these operations non-blocking, we can make
a big progress to SEDA. And of course, not all check points need non-blocking,
I think the most suitable points for stages is the operation that relies on the
load of other nodes, for example, HDFS read/write operations. A slow datanode
may occupy all handler, which is very similar to the scene in HBASE-16388
> Drop timeout requests in read/write path
> ----------------------------------------
>
> Key: HBASE-16492
> URL: https://issues.apache.org/jira/browse/HBASE-16492
> Project: HBase
> Issue Type: Umbrella
> Reporter: Phil Yang
> Assignee: Phil Yang
>
> After HBASE-15593, we can use rpc timeout provided by client to prevent
> wasting time on requests that have been dropped by client. In
> request-handling path there are some points that are suitable to check if we
> can finish this request in time.
> We can do this work in several sub-tasks to make sure each one is simple and
> easy to maintain.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
