> On March 8, 2016, 12:37 a.m., Ian Downes wrote:
> > src/master/master.cpp, line 6461
> > <https://reviews.apache.org/r/44474/diff/1/?file=1283503#file1283503line6461>
> >
> >     I'm not familiar with this code but it appears to be changing the 
> > behavior substantially. 
> >     
> >     The original code sums across all slaves that are registered at the 
> > time of the metric call. The new code tracks counters for the various 
> > states when a task on a registered slave changes. IIUC, this is a 
> > fundamental change in how the metrics are determined because the slave 
> > registration state is considered at different times. For example, a task 
> > may change state on a registered slave and it gets counted, then the slave 
> > becomes unregistered, then the metric endpoint is queried. Old and new code 
> > will give different numbers?
> >     
> >     If my understanding is correct, then perhaps (somewhat) more performant 
> > code could be achieved by maintaining counters at the slave level, and then 
> > aggregating the counters from *registered* slaves?

So are you saying I miss the slave unregistration event? But when we unregister 
a slave, we remove all the tasks from that slave too, and the counter is 
updated in removeTask(). Or I misunderstand your point?

As the slaves/tasks states are complicated, I am not surprised at all I miss 
some case, please point it out explicitly so that I can fix it. ;)


> On March 8, 2016, 12:37 a.m., Ian Downes wrote:
> > src/master/metrics.cpp, lines 38-46
> > <https://reviews.apache.org/r/44474/diff/1/?file=1283505#file1283505line38>
> >
> >     Please keep the original order which more closely matches the 
> > progression of states, i.e., staging before starting.

Nope, we use TASK_* as an index for this array, therefore it has to be in this 
order.


- Cong


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/44474/#review122423
-----------------------------------------------------------


On March 7, 2016, 10:51 p.m., Cong Wang wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/44474/
> -----------------------------------------------------------
> 
> (Updated March 7, 2016, 10:51 p.m.)
> 
> 
> Review request for mesos, Ian Downes and Vinod Kone.
> 
> 
> Bugs: MESOS-4740
>     https://issues.apache.org/jira/browse/MESOS-4740
> 
> 
> Repository: mesos
> 
> 
> Description
> -------
> 
> Avoid iterate the list of slaves, instead just maintain some counters.
> 
> 
> Diffs
> -----
> 
>   src/master/master.hpp ea26670e6c6c67314406fded510e8fdd46053dc8 
>   src/master/master.cpp 57ff4a39039f573b8586bc03f873f97826b97f6f 
>   src/master/metrics.hpp 9d201fcce1c46a890c86a889ab31029f9a061561 
>   src/master/metrics.cpp 30c091198a8fdd6d6a957a351dc37d3dae7788e4 
> 
> Diff: https://reviews.apache.org/r/44474/diff/
> 
> 
> Testing
> -------
> 
> Manual check
> 
> 
> Thanks,
> 
> Cong Wang
> 
>

Reply via email to