Reason of cascaded kill in a group

[Zhitao Li <zh...@gmail.com>]

Hi,

We are considering adding a new reason to StatusUpdate::Reason
<https://github.com/apache/mesos/blob/master/include/mesos/mesos.proto#L2395>,
to reflect the case when a task in a task group is killed cascaded:

Currently, if a task fails in a task group, other active tasks in the same
group will see *TASK_KILLED* without any custom reason. We would like to
provide a custom reason like *REASON_TASK_GROUP_KILLED* to distinguish
whether the task is killed upon request of scheduler or upon a cascaded
failure.


Question to framework maintainer: does any framework depends the value of
this reason? If not, we probably can just change the reason without a
opt-in mechanism from framework (i.e, a new framework capability).

Please let me know if your framework as such a dependency.

Thanks!


-- 
Cheers,

Zhitao Li

Re: Reason of cascaded kill in a group

[Zhitao Li <zh...@gmail.com>]

Hi Benjamin,

Yes that's what I meant: adding a new reason for such cascaded kill.

On Tue, Apr 10, 2018 at 1:17 PM, Benjamin Mahler <bm...@apache.org> wrote:

> Are you saying that there was no reason previously, and there would be a
> reason after the change? If so, adding a reason where one did not exist is
> safe from a backwards compatibility perspective.
>
> On Mon, Apr 9, 2018 at 10:32 AM, Zhitao Li <zh...@gmail.com> wrote:
>
>> Hi,
>>
>> We are considering adding a new reason to StatusUpdate::Reason
>> <https://github.com/apache/mesos/blob/master/include/mesos/mesos.proto#L2395>,
>> to reflect the case when a task in a task group is killed cascaded:
>>
>> Currently, if a task fails in a task group, other active tasks in the
>> same group will see *TASK_KILLED* without any custom reason. We would
>> like to provide a custom reason like *REASON_TASK_GROUP_KILLED* to
>> distinguish whether the task is killed upon request of scheduler or upon a
>> cascaded failure.
>>
>>
>> Question to framework maintainer: does any framework depends the value
>> of this reason? If not, we probably can just change the reason without a
>> opt-in mechanism from framework (i.e, a new framework capability).
>>
>> Please let me know if your framework as such a dependency.
>>
>> Thanks!
>>
>>
>> --
>> Cheers,
>>
>> Zhitao Li
>>
>
>


-- 
Cheers,

Zhitao Li

Re: Reason of cascaded kill in a group

[Benjamin Mahler <bm...@apache.org>]

Are you saying that there was no reason previously, and there would be a
reason after the change? If so, adding a reason where one did not exist is
safe from a backwards compatibility perspective.

On Mon, Apr 9, 2018 at 10:32 AM, Zhitao Li <zh...@gmail.com> wrote:

> Hi,
>
> We are considering adding a new reason to StatusUpdate::Reason
> <https://github.com/apache/mesos/blob/master/include/mesos/mesos.proto#L2395>,
> to reflect the case when a task in a task group is killed cascaded:
>
> Currently, if a task fails in a task group, other active tasks in the same
> group will see *TASK_KILLED* without any custom reason. We would like to
> provide a custom reason like *REASON_TASK_GROUP_KILLED* to distinguish
> whether the task is killed upon request of scheduler or upon a cascaded
> failure.
>
>
> Question to framework maintainer: does any framework depends the value of
> this reason? If not, we probably can just change the reason without a
> opt-in mechanism from framework (i.e, a new framework capability).
>
> Please let me know if your framework as such a dependency.
>
> Thanks!
>
>
> --
> Cheers,
>
> Zhitao Li
>