-Re: decommission tablet servers
Aaron Cordova 2012-01-30, 14:39
Beautiful .. I'll try that. 1.4 is awesome.
On Jan 30, 2012, at 9:27 AM, John W Vines wrote:
> I believe the ./accumulo admin stop <node> safely brings down all services in a node, forcing flushes for all tablets who utilize that logger. I believe that this behavior exists in 1.4+
> ----- Original Message -----
> | From: "Aaron Cordova" <[EMAIL PROTECTED]>
> | To: [EMAIL PROTECTED]
> | Sent: Monday, January 30, 2012 3:35:42 AM
> | Subject: Re: decommission tablet servers
> | Also - it would be good to be able to do this programmatically, and
> | through the shell.
> | I would assign this ticket to myself, should it need to become a
> | ticket, and probably assign it to the 1.5 release. I saw a ticket a
> | while back to make the monitor page more of a controller, and I think
> | a ticket to add security to the monitor page so not just anyone could
> | restart the cluster, etc. ... how's that going?
> | This 'decommission tablet server' feature I'm talking about could be
> | implemented and added to the API and shell before the monitor is
> | ready.
> | Aaron
> | On Jan 30, 2012, at 3:33 AM, Aaron Cordova wrote:
> | > Is there a way for a user to request that the master decommission
> | > one or more tabletservers, causing the master to migrate all the
> | > tablets away without triggering a recovery and finally shutting down
> | > the process or at least refusing to assign future tablets to it? Of
> | > course there would need to be a way to un-decommission a server too
> | > in that case.
> | >
> | > HDFS does this for dataNodes through a list of machines in a file
> | > and a command to refreshNodes. I think it'd be better to be able to
> | > decommission servers through the monitor page, and perhaps the
> | > behaviors is: once a server is requested to be decommissioned, the
> | > master moves away tablets and the tabletserver process kills itself
> | > or the master kills it, and if a tabletserver process is started on
> | > that machine thereafter, it joins the cluster like any new server,
> | > meaning, we don't refuse that machine rejoining the cluster...
> | >
> | > Thoughts?
> | >
> | > Aaron