> On July 8, 2015, 5:37 p.m., Alexander Rukletsov wrote:
> > A high-level concern I would like to share with you guys, though it isn't 
> > directly related to this particular patch. I have a feeling that we don't 
> > really care about keeping the Allocator interface neat, brief, and concise, 
> > but rather abuse it and add random methods to it in order to satisfy our 
> > current needs for the built-in allocator. With several production-grade 
> > allocators on the market we won't be able to do this anymore. Shall we be 
> > more cautious about that? Maybe it is a good time to schedule allocator 
> > refactoring?

Alex, I totally agree with you that the current allocator interface needs a 
major refactor. Given the recent Quota design, I feel like we should do the 
refactor asap, otherwise the complexity will soon become unmanagable.

I think the key is to define a clear boundary between master and allocator. 
Currently, the boundary is not clear. For instance, the offer management is 
done by the master but we have some interfaces like reviveOffers in the 
allocator.

I would suggest that we open a ticket (or a doc) to list all the issues in the 
current allocator interfaces, and then we can think about solutions. Does that 
sound good to you?


- Jie


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/35947/#review90930
-----------------------------------------------------------


On June 26, 2015, 10:55 p.m., Michael Park wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/35947/
> -----------------------------------------------------------
> 
> (Updated June 26, 2015, 10:55 p.m.)
> 
> 
> Review request for mesos, Alexander Rukletsov, Benjamin Hindman, Ben Mahler, 
> and Jie Yu.
> 
> 
> Repository: mesos
> 
> 
> Description
> -------
> 
> Needed to implement the master HTTP endpoints: `/reserve`, `/unreserve`, 
> `/create` and `/destroy`.
> 
> This API is similar to `updateSlave` which is currently very specific to 
> oversubscription.
> I considered consolidating `updateAvailable` and `updateSlave` but it will 
> require making offers be generated within the allocator to enable this.
> 
> In specific, `updateAvailable` could fail if there aren't sufficient 
> available resources to update, whereas `updateSlave` avoids failing by 
> keeping the allocator in an over-allocated state. For `updateSlave`, leaving 
> the allocator in an over-allocated state is ok. This is because it does not 
> modify resources therefore `total - allocated` will work out to __empty__. 
> `updateAvailable` cannot leave the allocator in an over-allocated state, 
> because it modifies resources, and therefore `total - allocated` is not 
> guaranteed to yield __empty__.
> 
> 
> Diffs
> -----
> 
>   include/mesos/master/allocator.hpp 22992c0c77058af4fcd28aa8e4a1191693a16f44 
>   src/master/allocator/mesos/allocator.hpp 
> 72470ec7f56f84a9a9815c09adb88def90ef672f 
>   src/master/allocator/mesos/hierarchical.hpp 
> 3264d145d52b48852878abf7ab9be29ab98208cc 
>   src/tests/hierarchical_allocator_tests.cpp 
> 3258840135290cd008ca09235d18b7f093dafd2e 
>   src/tests/mesos.hpp 9157ac079808d2686592e54ea26a26e6a0825ed3 
> 
> Diff: https://reviews.apache.org/r/35947/diff/
> 
> 
> Testing
> -------
> 
> (1) Added `HierarchicalAllocatorTest.updateAvailableSuccess` and 
> `HierarchicalAllocatorTest.updateAvailableFail`
> (2) `make check`
> 
> 
> Thanks,
> 
> Michael Park
> 
>

Reply via email to