[
https://issues.apache.org/jira/browse/HBASE-14559?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14944604#comment-14944604
]
Mikhail Antonov commented on HBASE-14559:
-----------------------------------------
I think there're several things here (or maybe I'm missing something).
Conceptually it looks logical to me that on large cluster overall
responsiveness to admin commands would be improved by serving admin commands
from separate threadpool (is that actually right? if not, we can revert it
completely?). There were some deadlock-type bugs (these 2 referenced by
[~eclark]), but that should be fixed? In theory having yet one more threadpool
to separate admin commands from things like meta lookups might help prevent it,
but looks overkill.
For the tests, these features are revealed in corner case, as in minicluster
everything is running as admin, so admin threadpool is overloaded. On current
master as I'm seeing in HConstants:
bq. public static final int DEFAULT_REGION_SERVER_HIGH_PRIORITY_HANDLER_COUNT =
20;
as was set in HBASE-13351.
[~stack] so in these 2 oneliners, you're dropping down the number of high
priority handlers (to tighten up thread usage?) right?
> branch-1 test tweeks; disable assert explicit region lands post-restart and
> up a few handlers
> ---------------------------------------------------------------------------------------------
>
> Key: HBASE-14559
> URL: https://issues.apache.org/jira/browse/HBASE-14559
> Project: HBase
> Issue Type: Sub-task
> Components: test
> Reporter: stack
> Assignee: stack
> Fix For: 2.0.0, 1.2.0, 1.3.0
>
> Attachments: 14559.branch-1.txt, 14559.master.txt
>
>
> Running branch-1 on internal rig (trying to stabilize branch-1/branch-1.2).
> Small tweaks get tests to pass. Small one liners that up priority handler
> count and disable assert that seems wrong -- that we'll always get an explcit
> region to land on a newly started server.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
