hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Dmitriy Lyubimov <dlie...@gmail.com>
Subject Re: Hbase Master Failover Issue
Date Sun, 15 May 2011 19:50:04 GMT
The problem was multinic configuration at master nodes.

i saw that the processes starts listening on a wrong NIC

I read the source code and saw that with default settings it would use
whatever ip is reported by canonical hostname, i.e. whatever retruned
by something like

ping `hostname`,


our canonical hostname was resolving of course  the wrong nic.

i kind of did not want to edit /etc/hostsnames (i guessed our admins
had a reason to point hostname to that nic), so i forcefully set
'eth0' as hbase.master.dns.interface (if i remember that property name
correctly).

it started listening on what was pointed by eth0:0 isntead of eth0
which solved the problem anyway.

(funny thing though i still couldn't make it listen on eth0 ip, but
rather on eth0:0 only although both had reverse dns. apparently
whatever native code is used, lists both ips for that interface and
then the first one that has reverse dns is used, so there's no way to
force it to listen on other ones).

Bottom line, with multinic configurations your hostname better points
to the ip you want it to listen on in /etc/hosts. If it's different,
one cannot use the default configuration.

-d

On Sat, May 14, 2011 at 3:04 PM, Stack <stack@duboce.net> wrote:
> What did you do to solve it?
> Thanks,
> St.Ack
>
> On Fri, May 13, 2011 at 6:17 PM, Dmitriy Lyubimov <dlieu.7@gmail.com> wrote:
>> Ok i think the issue is largely solved. Thanks for your help, guys.
>>
>> -d
>>
>> On Fri, May 13, 2011 at 5:32 PM, Dmitriy Lyubimov <dlieu.7@gmail.com> wrote:
>>> ok the problem seems to be multi-nic hosting on masters. the hbase
>>> master starts up and uses canonical hostname to listen on which points
>>> to a wrong nic. I am not sure why so i am not changign this but i am
>>> struggling to override this at the moment as nothing seems to work
>>> (master.dns.interface=eth2, master.dns.server=ip2 ... tried all
>>> possible combinatiosn... it probably has something to do with reverse
>>> lookup so i added entry to hosts files to no avail so far. i will have
>>> to talk to our admins to see why we can't switch the canonical host
>>> name to ip that all the nodes are supposed to use it with .
>>>
>>> thanks.
>>> -d
>>>
>>> On Fri, May 13, 2011 at 3:39 PM, Dmitriy Lyubimov <dlieu.7@gmail.com> wrote:
>>>> Thanks, Jean-Daniel.
>>>>
>>>> Logs don't show anything abnormal (not even warnings). How soon you
>>>> think the region servers should join?
>>>>
>>>> I am guessing the sequence should be something along the lines --
>>>>  zookeeper needs to timeout old master session first (2 mins or so ) ,
>>>> then hot spare should wean next master election (we probably should
>>>> see that happening if we can tail its log, right?)
>>>> and then the rest of the crowd should join in something like what
>>>> seems to be governed by hbase.regionserver.msginterval property , if i
>>>> read the code correctly?
>>>>
>>>> So all -in -all probably something like 3 minutes should warrant
>>>> everybody has found the new master one way or another , right? if not,
>>>> we have a problem, right?
>>>>
>>>> Thanks.
>>>> -Dmitriy
>>>>
>>>> On Fri, May 13, 2011 at 12:34 PM, Jean-Daniel Cryans
>>>> <jdcryans@apache.org> wrote:
>>>>> Maybe there is something else in there, would be useful to see logs
>>>>> from the region servers when you are shutting down master 1 and
>>>>> bringing up master2.
>>>>>
>>>>> About "I have no failover for a critical component of my
>>>>> infrastructure.", so is the Namenode, and for the moment you can't do
>>>>> much about it. What's usually recommended is to put both the master
>>>>> and the NN together on a more reliable machine. And the master ain't
>>>>> that critical, almost everything works without it.
>>>>>
>>>>> J-D
>>>>>
>>>>> On Fri, May 13, 2011 at 12:08 PM, sean barden <sbarden@gmail.com>
wrote:
>>>>>> So I updated one of my clusters from CDHb1 to u0 with no issues(in
the
>>>>>> upgrade).  Hbase failed over to it's "backup" master server just
find
>>>>>> in the older version.  As 0.90.1+15.18, I had hoped the fix would
be
>>>>>> in u0 for the failover issue.  However, I'm having the same issue.
>>>>>> master1 fails or I shut it down,  master2 waits for RS'es to check
in
>>>>>> forever.  Restarting the services for master2 and all RS's does
>>>>>> nothing until I start up master1.  So, essentially, I have no failover
>>>>>> for a critical component of my infrastructure.  Needless to say
I'm
>>>>>> exceptionally frustrated.  Any ideas to a fix or workaround would
be
>>>>>> greatly appreciated.
>>>>>>
>>>>>> Regards,
>>>>>>
>>>>>> Sean
>>>>>>
>>>>>> On Thu, May 5, 2011 at 11:59 AM, Jean-Daniel Cryans <jdcryans@apache.org>
wrote:
>>>>>>> Upgrade to CDH3u0 which as far as I can tell has it:
>>>>>>> http://archive.cloudera.com/cdh/3/hbase-0.90.1+15.18.CHANGES.txt
>>>>>>>
>>>>>>> J-D
>>>>>>>
>>>>>>> On Thu, May 5, 2011 at 9:55 AM, sean barden <sbarden@gmail.com>
wrote:
>>>>>>>> Looks like my issue.  We're using 0.90.1-CDH3B4 .  Looks
like an
>>>>>>>> upgrade is in order.  Can you suggest a workaround?
>>>>>>>>
>>>>>>>> thx,
>>>>>>>>
>>>>>>>> Sean
>>>>>>>>
>>>>>>>> On Thu, May 5, 2011 at 11:49 AM, Jean-Daniel Cryans <jdcryans@apache.org>
wrote:
>>>>>>>>> This sounds like https://issues.apache.org/jira/browse/HBASE-3545
>>>>>>>>> which was fix in 0.90.2, which version are you testing?
>>>>>>>>>
>>>>>>>>> J-D
>>>>>>>>>
>>>>>>>>> On Thu, May 5, 2011 at 9:23 AM, sean barden <sbarden@gmail.com>
wrote:
>>>>>>>>>> Hi,
>>>>>>>>>>
>>>>>>>>>> I'm testing failing over from one master to another
by stopping
>>>>>>>>>> master1(master2 is always running).  Master2 web
i/f kicks in and I can
>>>>>>>>>> zk_dump but the region servers never show up.  Logs
on master2 show repeated
>>>>>>>>>> entries below:
>>>>>>>>>>
>>>>>>>>>> 2011-05-05 09:10:05,938 INFO org.apache.hadoop.hbase.master.ServerManager:
>>>>>>>>>> Waiting on regionserver(s) to checkin
>>>>>>>>>> 2011-05-05 09:10:07,440 INFO org.apache.hadoop.hbase.master.ServerManager:
>>>>>>>>>> Waiting on regionserver(s) to checkin
>>>>>>>>>>
>>>>>>>>>> Obviously the RS are not checking in.  Not sure
why.
>>>>>>>>>>
>>>>>>>>>> Any ideas?
>>>>>>>>>>
>>>>>>>>>> thx,
>>>>>>>>>>
>>>>>>>>>> --
>>>>>>>>>> Sean Barden
>>>>>>>>>> sbarden@gmail.com
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> --
>>>>>>>> Sean Barden
>>>>>>>> sbarden@gmail.com
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> --
>>>>>> Sean Barden
>>>>>> sbarden@gmail.com
>>>>>>
>>>>>
>>>>
>>>
>>
>

Mime
View raw message