[
https://issues.apache.org/jira/browse/PHOENIX-5482?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17049642#comment-17049642
]
Chinmay Kulkarni commented on PHOENIX-5482:
-------------------------------------------
[~mihir6692] your PR changes look good as far as this Jira is considered,
however I have more of a "are we sure we want to do this" kind of question.
[~gjacoby] as per
[HBASE-15470|https://issues.apache.org/jira/browse/HBASE-15470]:
"Meta can have very different request rates than any other region. We shouldn't
use the same call queue length.
For example:
If a normal request takes 100ms ( long scan or large get )
but a call to meta is all in memory so it takes < 1ms
So a call queue length of 100 represents multiple seconds of work for normal
requests, but less than a second for meta."
This sort of indicates that hbase.ipc.server.priority.max.callqueue.length
would be set to a larger value than ipc.server.max.callqueue.length. The same
assumption about a request to meta being quick does not necessarily apply to a
request for a Phoenix index or SYSTEM table though, right?
Please correct me if my understanding is wrong here.
> PhoenixRpcScheduler should use HBase priority max queue length
> --------------------------------------------------------------
>
> Key: PHOENIX-5482
> URL: https://issues.apache.org/jira/browse/PHOENIX-5482
> Project: Phoenix
> Issue Type: Improvement
> Reporter: Geoffrey Jacoby
> Assignee: Mihir Monani
> Priority: Major
> Attachments: PHOENIX-5482.patch
>
> Time Spent: 20m
> Remaining Estimate: 0h
>
> HBase 1.3 and 2.0 introduced the
> hbase.ipc.server.priority.max.callqueue.length setting in HBASE-15470, to
> allow operators to vary the max queue length size for hbase:meta requests
> independently of normal regions.
> The PhoenixRpcScheduler doesn't check this setting, and uses the existing
> ipc.server.max.callqueue.length for all of its queues (normal, index, and
> metadata). While we should still default to this setting if the newer
> priority callqueue setting isn't present, Phoenix should use the new setting
> for index and metadata queues if it's present.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
