hadoop-yarn-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Arinto Murdopo <ari...@gmail.com>
Subject TestRMRestart - app1 and appState refer to the same instance
Date Thu, 15 Nov 2012 00:17:43 GMT
Dear all,

Refer to YARN-128.full-code-4.patch on Yarn-128 issue, I have these
following observations:

   1. In *TestRMRestart.java* Line 78, app1 and appState refer to the same
   instance because we are using memory to store the states (
   MemoryRMStateStore). Therefore, the assert result will always be True.
   2. ApplicationState is stored when we invoke MockRM's submitApp method.
   More precisely, it is in ClientRMService class, line 266. Interestingly,
   the state that we store contains the resource request from client. In this
   case, the value of resource request is 200. However, if we wait for some
   time, the value will be 1024 (which is the normalized value given by the
   Scheduler).
   3. Currently our school project is trying to persist the state in
   persistent storage, and the assert statement in our modified test class
   returns error. Since our storage stores the resource value before updated
   by the scheduler.

Based on above observations, should we update the persisted memory value
with the new value assigned by scheduler?
Since we are going to restart both ApplicationMaster and NodeManager when
there is failure in ResourceManager, I think the answer is no, we can use
the original value requested by user. But I'm not really sure with my own
reasoning.. soo.. please comment on it. :) . If the answer is yes, then we
should wait until Scheduler updates the resource value before persisting it
into the storage.

best regards,

Arinto Murdopo
European Master in Distributed Computing (EMDC)
Universitat Politècnica de Catalunya · BarcelonaTech, Barcelona, Spain
KTH Royal Institute of Technology, Stockholm, Sweden
Phone: +46 725 548 759

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