> We could add a metric that did an iteration of handlers in HBaseServer > emitting how many were in progress. There is no provision for doing this > currently. Would have to add accessors, etc.
If already going into the code and changing things... Having a bounded thread pool is important, so availability can degrade gracefully (more or less), as opposed to the whole regionserver becoming livelocked. But, we don't need for the pool to be preallocated as is done currently. How about changing the RPC server thread pool such that the user can specify a minimum and maximum number of handler threads? The pool would start with the minimum, allocate more up to the max to handle additional concurrency, then terminate unused threads after some time back down to the minimum. Then we can do things like set a maximum of 100 handlers or such without taking on the overhead of 100 threads until it is needed. - Andy ----- Original Message ---- > From: stack <st...@duboce.net> > To: hbase-dev@hadoop.apache.org > Sent: Sun, January 17, 2010 11:56:45 AM > Subject: Re: [jira] Resolved: (HBASE-2133) Increase default number of client > handlers > > We could add a metric that did an iteration of handlers in HBaseServer > emitting how many were in progress. There is no provision for doing this > currently. Would have to add accessors, etc. Its a good idea. As a > metric, we might miss a burst of requests filling all slots but sustained > high request numbers would show. > > St.Ack > > On Sun, Jan 17, 2010 at 10:17 AM, Lars George wrote: > > > Hi Andrew, > > > > I have it at 30. What I would like to know is how to detect this sort > > of shortage? I also have the feeling we need this sort of indicators > > as metrics to the various servers so that the max and current can be > > graphed. What do you think? > > > > Thanks, > > Lars > > > > On Sat, Jan 16, 2010 at 2:02 AM, Andrew Purtell (JIRA) > > wrote: > > > > > > [ > > > https://issues.apache.org/jira/browse/HBASE-2133?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel] > > > > > > Andrew Purtell resolved HBASE-2133. > > > ----------------------------------- > > > > > > Resolution: Fixed > > > Hadoop Flags: [Reviewed] > > > > > > Committed following change to trunk and 0.20 branch: > > > \\ > > > {noformat} > > > --- conf/hbase-default.xml (revision 899849) > > > +++ conf/hbase-default.xml (working copy) > > > @@ -153,10 +153,10 @@ > > > > > > > > > hbase.regionserver.handler.count > > > - 10 > > > + 25 > > > Count of RPC Server instances spun up on RegionServers > > > Same property is used by the HMaster for count of master handlers. > > > - Default is 10. > > > + Default is 25. > > > > > > > > > > > > {noformat} > > > > > > We might want to increase this again depending on user feedback. I know I > > needed 100 to avoid trouble with high read/write load once above ~200 > > regions/server. > > > > > >> Increase default number of client handlers > > >> ------------------------------------------ > > >> > > >> Key: HBASE-2133 > > >> URL: https://issues.apache.org/jira/browse/HBASE-2133 > > >> Project: Hadoop HBase > > >> Issue Type: Improvement > > >> Reporter: Andrew Purtell > > >> Assignee: Andrew Purtell > > >> Fix For: 0.20.3, 0.21.0 > > >> > > >> > > >> Any reason not to just go ahead and change hbase-default.xml to include: > > >> {noformat} > > >> > > >> hbase.regionserver.handler.count > > >> 100 > > >> > > >> > > >> hbase.zookeeper.property.maxClientCnxns > > >> 100 > > >> > > >> {noformat} > > >> ? > > >> The current default for both, 10, is anemic. > > > > > > -- > > > This message is automatically generated by JIRA. > > > - > > > You can reply to this email to add a comment to the issue online. > > > > > > > >
