[
https://issues.apache.org/jira/browse/HBASE-27355?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17600347#comment-17600347
]
ruanhui commented on HBASE-27355:
---------------------------------
{quote}
So the solution here is to have a special handler group for handling meta
requests from master only?
{quote}
Yes. we plan to dispatch master meta requests to scan handlers and dispatch
client requests to read handlers.
{quote}
Is it possible to not expose another QOS number in HConstants?
{quote}
Yes, It is really problematic to define MASTER_READ_QUESTS in HConstants. The
content in HConstants should be public.
{quote}
Could we just check whether the requests are coming from master and give it a
high prority automatically?
{quote}
We have also considered this solution. Then we found that it is not very easy
to distinguish the request from the master and the request from the client. If
using the source IP, we might get confused whether the request is coming from
the master or from a client deployed on the master node. Finally we choose the
first one. But the latter situation is not common, so I think the second
solution works too. Let me address it.
> Separate meta read requests from master and client
> ---------------------------------------------------
>
> Key: HBASE-27355
> URL: https://issues.apache.org/jira/browse/HBASE-27355
> Project: HBase
> Issue Type: Improvement
> Components: IPC/RPC
> Affects Versions: 3.0.0-alpha-3
> Reporter: ruanhui
> Assignee: ruanhui
> Priority: Minor
> Fix For: 3.0.0-alpha-4
>
>
> If we have a large number of store files in a single region or the response
> from hdfs is slow, the region transition can be slow, the client may put a
> lot of pressure on the meta server when retrying. This may block the master
> system read requests. Maybe we can set a special priority for the master
> request to isolate read requests from master and client.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
