> On May 17, 2013, 7:12 p.m., Vinod Kone wrote:
> > src/master/master.cpp, lines 1266-1272
> > <https://reviews.apache.org/r/11121/diff/2/?file=291857#file291857line1266>
> >
> >     If we do decide to do this (I'm not completely convinced yet), we 
> > should do it in the allocator.
> 
> Brenden Matthews wrote:
>     Care to elaborate? It looks like in 
> HierarchicalAllocatorProcess::allocate() the offers are in a hash map, so I'm 
> not sure how you could shuffle them there.

Yes. You are right. The master gets a hashmap.

That said, I'm still not convinced that master is the right place to do. For 
example, can some frameworks benefit from getting the offers in the same order? 
It seems to me that its pretty easy for frameworks to randomize, if they choose 
to, but it would be hard for them to de-randomize randomized offers. Makes 
sense? I was ok with doing in this in allocator because allocators are 
pluggable.


- Vinod


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/11121/#review20716
-----------------------------------------------------------


On June 11, 2013, 9:54 p.m., Brenden Matthews wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/11121/
> -----------------------------------------------------------
> 
> (Updated June 11, 2013, 9:54 p.m.)
> 
> 
> Review request for mesos.
> 
> 
> Description
> -------
> 
> Shuffle offers before sending to frameworks.
> 
> This helps ensure fair allocation of resources.
> 
> Review: https://reviews.apache.org/r/11121
> 
> 
> Diffs
> -----
> 
>   src/master/master.cpp 60c6d4f88f8024796c5e495d7cc6ddde2b754887 
> 
> Diff: https://reviews.apache.org/r/11121/diff/
> 
> 
> Testing
> -------
> 
> Used in production at airbnb.
> 
> make -j10 check && cd hadoop && make hadoop-2.0.0-mr1-cdh4.2.1 && make 
> hadoop-0.20.205.0 && make hadoop-0.20.2-cdh3u3
> 
> 
> Thanks,
> 
> Brenden Matthews
> 
>

Reply via email to