Mailing-List: contact issues-help@hbase.apache.org; run by ezmlm
Precedence: bulk
Date: Thu, 30 Jun 2016 19:14:10 +0000 (UTC)
From: "stack (JIRA)" <jira@apache.org>
To: issues@hbase.apache.org
Message-ID: <JIRA.12984774.1467178103000.6990.1467314050745@Atlassian.JIRA>
In-Reply-To: <JIRA.12984774.1467178103000@Atlassian.JIRA>
References: <JIRA.12984774.1467178103000@Atlassian.JIRA> <JIRA.12984774.1467178103983@arcas>
Subject: [jira] [Commented] (HBASE-16142) Trigger JFR session when under
 duress -- e.g. backed-up request queue count -- and dump the recording to
 log dir
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit
archived-at: Thu, 30 Jun 2016 19:14:12 -0000


    [ https://issues.apache.org/jira/browse/HBASE-16142?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15357676#comment-15357676 ] 

stack commented on HBASE-16142:
-------------------------------

Yes, off by default though I think I'd like to enable these flags by default -XX:+UnlockCommercialFeatures -XX:+FlightRecorder (as long as they are ignored by other JVMs and don't pull down perf (they don't in my experience).

Yes [~mantonov], this would not be tracing. It would be something that flipped on programmatically when struggling. It would bring down the perf by a few percent but the upside would be good detail on what was going on in JVM during the struggle.

> Trigger JFR session when under duress -- e.g. backed-up request queue count -- and dump the recording to log dir
> ----------------------------------------------------------------------------------------------------------------
>
>                 Key: HBASE-16142
>                 URL: https://issues.apache.org/jira/browse/HBASE-16142
>             Project: HBase
>          Issue Type: Task
>          Components: Operability
>            Reporter: stack
>            Priority: Minor
>              Labels: beginner
>
> Chatting today w/ a mighty hbase operator on how to figure what is happening during transitory latency spike or any other transitory 'weirdness' in a server, the idea came up that a java flight recording during a spike would include a pretty good picture of what is going on during the time of duress (more ideal would be a trace of the explicit slow queries showing call stack with timings dumped to a sink for later review; i.e. trigger an htrace when a query is slow...).
> Taking a look, programmatically triggering a JFR recording seems doable, if awkward (MBean invocations). There is even a means of specifying 'triggers' based off any published mbean emission -- e.g. a query queue count threshold -- which looks nice. See https://community.oracle.com/thread/3676275?start=0&tstart=0 and https://docs.oracle.com/javacomponents/jmc-5-4/jfr-runtime-guide/run.htm#JFRUH184
> This feature could start out as a blog post describing how to do it for one server. A plugin on Canary that looks at mbean values and if over a configured threshold, triggers a recording remotely could be next. Finally could integrate a couple of triggers that fire when issue via the trigger mechanism.
> Marking as beginner feature.


--
This message was sent by Atlassian JIRA
(v6.3.4#6332)