> On July 13, 2015, 4:46 p.m., Alexander Rukletsov wrote:
> > src/master/http.cpp, lines 515-516
> > <https://reviews.apache.org/r/35702/diff/9/?file=994080#file994080line515>
> >
> >     It looks like we actually have the role, but it's buried in resources. 
> > Do you envision having resources collection with various roles in one 
> > request? Maybe it makes sense to add a validation step which ensures there 
> > is just one role per request and use it here, also avoiding changes in the 
> > `validate()`function.
> Michael Park wrote:
>     I didn't see a good reason to require a "one role per request" condition. 
> The current interface accurately models the fact that an operator does not 
> have a role associated to it like a framework does, and I don't think 
> "avoiding changes in the `validate()` function" should have any influence in 
> deciding how an interface behaves.
>     If we required such a condition, the per-request atomicity guarantee 
> comes with a limitation that it can only be for a single role. While I'm not 
> sure of its value, I'm also not sure what we gain by requiring it from the 
> user's perspective?
> Alexander Rukletsov wrote:
>     I think I'm missing something, my understanding is that each dynamic 
> reservation is associated with a role, regardless, who issues a reservation 
> request. I don't think limiting users to one role per request gives them any 
> benefit, but it looks like we can be closer to framework-issued request if we 
> do so. What am I missing?
> Michael Park wrote:
>     Your understanding is correct. Aside from the resources being associated 
> with a role, frameworks are also associated with a role. We check that every 
> resource is being reserved for the framework's role because a framework is 
> associated with a role and it wouldn't make sense to allow a framework to 
> reserve resources for a role that does not match its role. On the contrary, 
> the same rule doesn't apply for an operator since there's no such thing as an 
> "operator's role".

That's right. Let me try to reformulate my proposal. If we require an operator 
to reserve resources for one role per request, it can be interpreted as an 
"operator role". An advantage here is that `validate()` method doesn't need to 
be changed, while a disadvantage is that this approach is a bit artificial and 
can lead to confusion. What do you think?

- Alexander

This is an automatically generated e-mail. To reply, visit:

On July 28, 2015, 9:03 p.m., Michael Park wrote:
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/35702/
> -----------------------------------------------------------
> (Updated July 28, 2015, 9:03 p.m.)
> Review request for mesos, Adam B, Benjamin Hindman, Ben Mahler, Jie Yu, Joris 
> Van Remoortere, and Vinod Kone.
> Bugs: MESOS-2600
>     https://issues.apache.org/jira/browse/MESOS-2600
> Repository: mesos
> Description
> -------
> This involved a lot more challenges than I anticipated, I've captured the 
> various approaches and limitations and deal-breakers of those approaches 
> here: [Master Endpoint Implementation 
> Challenges](https://docs.google.com/document/d/1cwVz4aKiCYP9Y4MOwHYZkyaiuEv7fArCye-vPvB2lAI/edit#)
> Key points:
> * This is a stop-gap solution until we shift the offer creation/management 
> logic from the master to the allocator.
> * `updateAvailable` and `updateSlave` are kept separate because
>   (1) `updateAvailable` is allowed to fail whereas `updateSlave` must not.
>   (2) `updateAvailable` returns a `Future` whereas `updateSlave` does not.
>   (3) `updateAvailable` never leaves the allocator in an over-allocated state 
> and must not, whereas `updateSlave` does, and can.
> * The algorithm:
>     * Initially, the master pessimistically assume that what seems like 
> "available" resources will be gone.
>       This is due to the race between the allocator scheduling an `allocate` 
> call to itself vs master's `allocator->updateAvailable` invocation.
>       As such, we first try to satisfy the request only with the offered 
> resources.
>     * We greedily rescind one offer at a time until we've rescinded 
> sufficiently many offers.
>       IMPORTANT: We perform `recoverResources(..., Filters())` rather than 
> `recoverResources(..., None())` so that we can pretty much always win the 
> race against `allocate`.
>                  In the case that we lose, no disaster occurs. We simply fail 
> to satisfy the request.
>     * If we still don't have enough resources after resciding all offers, be 
> optimistic and forward the request to the allocator since there may be 
> available resources to satisfy the request.
>     * If the allocator returns a failure, report the error to the user with 
> `PreconditionFailed`. This could be updated to be `Forbidden`, or `Conflict` 
> maybe as well. We'll pick one eventually.
> This approach is clearly not ideal, since we would prefer to rescind as 
> little offers as possible.
> The challenges of implementing the ideal solution in the current state is 
> described in the document above.
> TODO(mpark): Add more comments and test cases.
> Diffs
> -----
>   src/master/http.cpp 3a1598fad4db03e5f62fd4a6bd26b2bedeee4070 
>   src/master/master.hpp 827d0d599912b2936beb9615610f627f6c9a2d43 
>   src/master/master.cpp 5b5e3c37d4433c8524db267866aebc0a35a181f1 
>   src/master/validation.hpp 469d6f56c3de28a34177124aae81ce24cb4ad160 
>   src/master/validation.cpp 9d128aa1b349b018b8e4a1916434d848761ca051 
> Diff: https://reviews.apache.org/r/35702/diff/
> Testing
> -------
> `make check`
> Thanks,
> Michael Park

Reply via email to