We have no regions in transition.

As far as who's requesting the table descriptors:  We are calling 
HTableIterface.getTableDescriptor frequently.  We are going to cache the result 
to avoid the unnecessary network calls.

Thanks for the hint!

From: lars hofhansl [mailto:[email protected]]
Sent: Thursday, March 26, 2015 3:43 PM
To: [email protected]
Cc: Development
Subject: Re: master consumes large amount of CPU for days

Hi Ted,

Yes, it is safe to bounce the HMaster without taking the region servers down.

Are there any regions in transition (would be shown on the master's page)?

All of the threads involved with
- 
org.apache.hadoop.hdfs.DistributedFileSystem.listStatus(org.apache.hadoop.fs.Path)
 @bci=14, line=309 (Compiled frame)
  ...
- org.apache.hadoop.hbase.master.HMaster.getHTableDescriptors(java.util.List) 
@bci=50, line=1932 (Compiled frame)

are a bit suspect. Who's requesting all the table descriptors from the master?
-- Lars

________________________________
From: Ted Tuttle <[email protected]<mailto:[email protected]>>
To: "[email protected]<mailto:[email protected]>" 
<[email protected]<mailto:[email protected]>>
Cc: Development 
<[email protected]<mailto:[email protected]>>
Sent: Thursday, March 26, 2015 12:18 PM
Subject: master consumes large amount of CPU for days

Hello-

Our master process started consuming a large amount of CPU (75% of box) several 
days back and hasn't stopped.  I have 2 questions:

                1) what is it doing (stack dump and log below)
                2) is it safe to restart the master without taking the whole 
cluster down?

Master stack dump:

                http://pastebin.com/G0iNNEpC

Master log from last 15 mins or so:

                http://pastebin.com/WQNjhFGf

Thanks,
Ted

Reply via email to