hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From psm...@aconex.com
Subject Re: JMX bean registrations for Master & RegionServers
Date Fri, 09 Apr 2010 07:40:42 GMT

On 09/04/2010, at 17:15, Ryan Rawson <ryanobjc@gmail.com> wrote:

> In the HMaster, there is this:
>  private void startServiceThreads() {
>    // Do after main thread name has been set
>    this.metrics = new MasterMetrics();
> which calls the MBeanUtil registration (a hadoop utility) which
> registers the MasterMetrics which contains all those metrics you are
> interested in.  The HMaster constructor will block until the ZK master
> node is available.
> does this help?

Thanks for replying. I'm not sure I get it though (it's Friday and  
I've had a beer), for a Master I can block and wait for that to  
complete but then if the current process is a regionserver it would  
wait forever.  How can I differentiate a Master from a RegionServer  

One way would be to change the JMX namespace for the RPC bean in the  
master so it doesn't look like it's an RS?  Be nice if my requirement  
didn't require HBase code changes though.

This sniffing works ok for straight Hadoop nodes because the NN, JT,  
TT and DN all have distinct JMX names to sniff for.

> On Thu, Apr 8, 2010 at 11:48 PM, Paul Smith <psmith@aconex.com> wrote:
>> Hi gents, related to the Parfait/PCP discussion recently, I'm  
>> hitting a problem that I need your help.
>> Right now a MetricContext implementation is initialised fairly  
>> early within the Hbase/Hadoop lifecycle, and is given a chance to  
>> initialise which is good.  However what I've noticed is that other  
>> parts of the JMX namespace do not appear for, say, the HMaster  
>> process until quite late; well after the MetricContext impl is  
>> initialized.
>> My PCPContext implementation I'm working on is going to want to  
>> 'sniff' and identify what type of process it is, mainly so as to be  
>> able to route metrics to a locally named file identifying it as  
>> metrics from that service.  There is no way that I can spot at the  
>> moment where a MetricContext can be told that the process own  
>> initialization is complete.  I'd like to have this so that a proper  
>> exploration of the JMX namespace can happen to export things like  
>> JVM metrics etc.
>> Can someone point me in the right direction here?  For example,  
>> once the HMaster is started up, there exists a JMX bean with the  
>> name pattern 'hadoop:service=Master' namespace.  Unfortunately the  
>> RegionServer process has a pattern that is 'hadoop:service=HBase',  
>> but that ALSO appears in the Master process (and early) so right  
>> now my naive topology sniffing is seeing this and thinking the  
>> Master is a RegionServer.
>> If I can delay my PCPContext initialization (even by just waiting  
>> until I can detect something has occurred) until the rest of the  
>> HBase process is properly up and running, then I think I'm off and  
>> racing.
>> Any hooks or ideas on this would be appreciated.
>> Paul

View raw message