hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Nicolas Liochon <nkey...@gmail.com>
Subject Re: Handling regionserver crashes in production cluster
Date Thu, 13 Jun 2013 07:00:46 GMT
Hum... So even a simple get shows the issue?
It would be a (surprising) critical bug. Could you please try the 95.1 or
the 94.8? Or write an unit test?

Thanks,

Nicolas


On Thu, Jun 13, 2013 at 5:43 AM, kiran <kiran.sarvabhotla@gmail.com> wrote:

> Its a simple kill...
> Scan is used using startrow and stoprow
> Scan scan = new Scan(Bytes.toBytes("adidas"), Bytes.toBytes("adidas1"));
>
>
> Our cluster size is 15. The load average when I see in master is 78%...It
> is not that overloaded. but writes are happening in the cluster...
>
> Thanks
> Kiran
>
>
>
> On Wed, Jun 12, 2013 at 10:49 PM, Nicolas Liochon <nkeywal@gmail.com>
> wrote:
>
> > Yeah, it should not block the other regions.
> >
> > For the region server, was it a kill -9 or in simple kill (the former
> > triggers a recovery, the later will close the region before stopping the
> > process)?
> >
> > How do you select the scan scope? With stop/start rows?
> > Can you share the client code you're using?
> > What's the cluster size? Was it already very loaded before you killed the
> > region server?
> >
> > Nicolas
> >
> >
> >
> > On Wed, Jun 12, 2013 at 6:11 PM, kiran <kiran.sarvabhotla@gmail.com>
> > wrote:
> >
> > > Yes we killed the region server but datanode is still running on the
> > > node...
> > >
> > > Sample Test scenario: Assume, I have table with pre-splits a upto z
> > (about
> > > 26 regions). I brought down region server purposefully with regions
> > having
> > > prefixes c and d. Then I used client API to scan data from regions with
> > > prefixes other than c and d. The response was very slow and sometimes
> not
> > > coming at all.
> > >
> > > My doubt was if only regions with prefix c and d are getting relocated
> or
> > > in transition. Why is it affecting the regions with other prefixes....
> > But
> > > once the region transition is over, the response is very fast as
> > expected.
> > >
> > >
> > >
> > > On Wed, Jun 12, 2013 at 8:50 PM, rajesh babu chintaguntla <
> > > chrajeshbabu32@gmail.com> wrote:
> > >
> > > > You can configure below to more value to close more regions at a
> time.
> > > >
> > > >  <property>
> > > >     <name>hbase.regionserver.executor.closeregion.threads</name>
> > > >     <value>3</value>
> > > >   </property>
> > > >
> > > >
> > > > On Wed, Jun 12, 2013 at 7:38 PM, Nicolas Liochon <nkeywal@gmail.com>
> > > > wrote:
> > > >
> > > > > What was your test exactly? You killed -9 a region server but kept
> > the
> > > > > datanode alive?
> > > > > Could you detail the queries you were doing?
> > > > >
> > > > >
> > > > > On Wed, Jun 12, 2013 at 2:10 PM, kiran <
> kiran.sarvabhotla@gmail.com>
> > > > > wrote:
> > > > >
> > > > > > It is not possible for us to migrate to new version immediately.
> > > > > >
> > > > > > @Anoop we purposefully brought down one regionserver, then we
> > > observed
> > > > > the
> > > > > > website is taking too much time to respond. We observed the
> pattern
> > > for
> > > > > > about 5 min till the regions are relocated.
> > > > > > Also we issued queries in our website taking care that the
> queries
> > > did
> > > > > n't
> > > > > > come under the regions in the regionserver we brought down.
> > > > > >
> > > > > > Is there any configuration workaround to mitigate it??
> > > > > >
> > > > > > Thanks
> > > > > > Kiran
> > > > > >
> > > > > >
> > > > > >
> > > > > > On Thu, Jun 6, 2013 at 8:27 PM, Jean-Marc Spaggiari <
> > > > > > jean-marc@spaggiari.org
> > > > > > > wrote:
> > > > > >
> > > > > > > Hi Kiran,
> > > > > > >
> > > > > > > Also, any chance for you to migrate to 0.94.8? There have
been
> > > > > > > hundreds of fixes since 0.94.1...
> > > > > > >
> > > > > > > JM
> > > > > > >
> > > > > > > 2013/6/6 Anoop John <anoop.hbase@gmail.com>:
> > > > > > > > How many total RS in the cluster?  You mean u can
not do any
> > > > > operation
> > > > > > on
> > > > > > > > other regions in the live clusters?  It should not
happen..
>  Is
> > > it
> > > > so
> > > > > > > > happening that the client ops are targetted at the
regions
> > which
> > > > were
> > > > > > in
> > > > > > > > the dead RS( and in transition now)?   Can u have
a closer
> look
> > > and
> > > > > > see?
> > > > > > > > If not pls check the RS threads were they are getting
> blocked.
> > > > > > > >
> > > > > > > > -Anoop-
> > > > > > > >
> > > > > > > > On Wed, Jun 5, 2013 at 10:50 PM, kiran <
> > > > kiran.sarvabhotla@gmail.com>
> > > > > > > wrote:
> > > > > > > >
> > > > > > > >> Dear All,
> > > > > > > >>
> > > > > > > >> We have production cluster that runs on hbase
0.94.1. The
> > issue
> > > we
> > > > > are
> > > > > > > >> facing is whenever one regionserver goes down,
the cluster
> > > becomes
> > > > > > > >> unresponsive until all the regions are allocated
to another
> > > > > > > >> regionserver(s). The transition is taking about
3-5 mins and
> > > > during
> > > > > > this
> > > > > > > >> time we are unable to any do client operation
on the
> cluster.
> > > > > > > >>
> > > > > > > >> Is there any way we can make the transition to
run in
> > > background ?
> > > > > > > >>
> > > > > > > >> Also, it is acceptable for us if the client operations
such
> as
> > > > scan
> > > > > or
> > > > > > > get
> > > > > > > >> does not work on the rowkeys of regions in transition.
But,
> > they
> > > > are
> > > > > > not
> > > > > > > >> working on the entire cluster until all the regions
are
> moved
> > > out
> > > > of
> > > > > > > >> transition. We can't afford 3-5 minutes of downtime.
> > > > > > > >>
> > > > > > > >> --
> > > > > > > >> Thank you
> > > > > > > >> Kiran Sarvabhotla
> > > > > > > >>
> > > > > > > >> -----Even a correct decision is wrong when it
is taken late
> > > > > > > >>
> > > > > > >
> > > > > >
> > > > > >
> > > > > >
> > > > > > --
> > > > > > Thank you
> > > > > > Kiran Sarvabhotla
> > > > > >
> > > > > > -----Even a correct decision is wrong when it is taken late
> > > > > >
> > > > >
> > > >
> > >
> > >
> > >
> > > --
> > > Thank you
> > > Kiran Sarvabhotla
> > >
> > > -----Even a correct decision is wrong when it is taken late
> > >
> >
>
>
>
> --
> Thank you
> Kiran Sarvabhotla
>
> -----Even a correct decision is wrong when it is taken late
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message