Hello Esteban-
At the time of the ZK connection problems the client had an OOM event. However,
the client machine overall was in fine shape looking at ganglia reports; it
certainly wasn't swapping or spending significant cycles on I/O wait.
Similarly, our zookeeper server was real chilled as it always is.
Regarding client configuration:
<property>
<!--Loaded from hbase-default.xml-->
<name>hbase.client.pause</name>
<value>1000</value>
</property>
Thanks,
Ted
-----Original Message-----
From: Esteban Gutierrez [mailto:[email protected]]
Sent: Thursday, August 14, 2014 10:47 AM
To: [email protected]
Cc: [email protected]
Subject: Re: HBase client hangs after client-side OOM
Hi Ted,
I've see this kind of client "hangs" few times when the underlying environment
is under heavy swapping and with older versions of ZK as Rakesh mentioned, also
when hbase.client.pause is set to 0. Do you know if your environment is
experiencing a similar behavior with heavy IO due swapping ?
can you also share your client configuration too?
cheers,
esteban.
--
Cloudera, Inc.
On Thu, Aug 14, 2014 at 9:56 AM, Ted Tuttle <[email protected]> wrote:
> The client-side thread dump in here:
>
> http://pastebin.com/xU4MSq9k
>
> SendThread appears to be active.
>
> -----Original Message-----
> From: Rakesh R [mailto:[email protected]]
> Sent: Thursday, August 14, 2014 7:01 AM
> To: [email protected]; [email protected]
> Subject: RE: HBase client hangs after client-side OOM
>
> Hi,
>
> >> We are running ZK 3.3.4, Cloudera cdh3u3, HBase 0.94.16.
>
> ZK version is quite old. I could see ClientCnxn is only catching
> IOException and when there is OOME it will exit SendThread.
> I think, thats the reason for client hanging. Client side threaddump
> will help us to see the liveliness of SendThread.
>
> Client side exception handling has been modified in 3.4 & 3.5 branches.
> Can you check the possibility of upgrading to 3.4.6 latest release.
>
> Regards,
> Rakesh
>
> -----Original Message-----
> From: Qiang Tian [mailto:[email protected]]
> Sent: 14 August 2014 11:03
> To: [email protected]; [email protected]
> Subject: Re: HBase client hangs after client-side OOM
>
> the sendthread stacktrace looks not correct. Do you have the client log?
> (in case zk client code log sth there) from the zk code, it looks
> ClientCnxn$SendThread.run should have caught
> it(throwable) and done the cleanup work, e.g. notify the main thread,
> so that it can wake up from ClientCnxn.submitRequest..
>
> send to Zookeeper for help.
> thanks.
>
>
>
> On Thu, Aug 14, 2014 at 11:19 AM, Ted Tuttle <[email protected]> wrote:
>
> > Hi Lars-
> >
> > We are running ZK 3.3.4, Cloudera cdh3u3, HBase 0.94.16.
> >
> > Thanks,
> > Ted
> >
> > > On Aug 13, 2014, at 5:36 PM, "lars hofhansl" <[email protected]> wrote:
> > >
> > > Hey Ted,
> > >
> > > so this is a problem with the ZK client, it seems to not clean
> > > itself up
> > correctly upon receiving an exception at the wrong moment.
> > > Which version of ZK are you using?
> > >
> > >
> > > -- Lars
> > >
> > >
> > >
> > > ----- Original Message -----
> > > From: Ted Tuttle <[email protected]>
> > > To: "[email protected]" <[email protected]>
> > > Cc: Development <[email protected]>
> > > Sent: Wednesday, August 13, 2014 4:38 PM
> > > Subject: HBase client hangs after client-side OOM
> > >
> > > Hello-
> > >
> > > We are running HBase v0.94.16 on an 8 node cluster.
> > >
> > > We have a recurring problem w/ HBase clients hanging. In latest
> > occurrence, I observed the following sequence of events:
> > >
> > > 0) client plays w/ HBase for a long time w/o issue
> > > 1) client runs out of memory during HBase operation:
> > >
> > > http://pastebin.com/b5x44Lx7
> > >
> > > 3) Exception is thrown, memory is released
> > > 2) In some shutdown logic the client tries to access HBase again
> > > and
> > hangs:
> > >
> > > http://pastebin.com/xU4MSq9k
> > >
> > > Clearly I need to fix OOM. However, the fact that client hangs is
> > > not
> > nice. Any ideas why?
> > >
> > > BTW- I started by looking at zookeeper log. Not much there but
> > > here you
> > go:
> > >
> > > http://pastebin.com/wZvE0Fbv
> > >
> > > Thanks,
> > > Ted
> > >
> >
>
