I believe the correct answer is "it depends" :-) If the frameworks are running for long times, then it shouldn't matter (much) if you call reviveOffers or not, because after the reject filter expires, it will be just like calling reviveOffers. I guess that the master will be able to consolidate the offers and re-offer it to a single framework if the reject filter expired, AND the resources from both offers are currently not used by any framework.
On Tue, Apr 7, 2015 at 3:13 PM, David Greenberg <dsg123456...@gmail.com> wrote: > Additionally, it was my understanding that it's possible (when multiple > frameworks are connected) for mesos to offer one offer to one framework, > and the other offer to the other framework. In this case, they'd never get > consolidated and merged together. We've observed this in our cluster as it > runs for a long time; however, I don't think our frameworks call > reviveOffers. Would reviveOffers change the situation? > > On Tue, Apr 7, 2015 at 1:50 AM Itamar Ostricher <ita...@yowza3d.com> > wrote: > >> Got it. Thanks! >> >> On Mon, Apr 6, 2015 at 9:02 PM, Vinod Kone <vinodk...@apache.org> wrote: >> >>> To clarify David's answer, you should only get <16, 8> offer until the >>> filter on <0, 8> is active. Once the filter expires (or you call >>> reviveOffers), Mesos will consolidate those resources and send a <16, 16> >>> offer. You don't need to restart the master for the aggregation. >>> >>> On Mon, Apr 6, 2015 at 10:46 AM, Itamar Ostricher <ita...@yowza3d.com> >>> wrote: >>> >>>> Thanks David! >>>> >>>> I'd like to make sure I understand you correctly. >>>> I will get both <16,8> & <0,8> offers, or just the <16,8> offer? >>>> (because I previously rejected the <0,8> offer, and did not call >>>> reviveOffers) >>>> >>>> On Mon, Apr 6, 2015 at 3:06 PM, David Greenberg <dsg123456...@gmail.com >>>> > wrote: >>>> >>>>> You'll actually see that as 2 offers: <16,8> and <0,8>. Your framework >>>>> may need to consolidate those offers if it requires more than 8gb of >>>>> memory >>>>> for the next task; you can do that by calling launchTasks with both >>>>> offers. >>>>> Besides combining offers in the framework, if you fail over the master, >>>>> the >>>>> offers would also be consolidated. >>>>> On Mon, Apr 6, 2015 at 7:29 AM Itamar Ostricher <ita...@yowza3d.com> >>>>> wrote: >>>>> >>>>>> Say my scheduler received a resource offers from slave S with <16cpu, >>>>>> 16GiB mem>, and called launchTasks on this offer with utilization of >>>>>> <16cpu, 8GiB mem>. >>>>>> From what I see (with mesos 0.21), the left over <0cpu, 8GiB mem> is >>>>>> considered rejected, and will be re-offered after the rejection filter, >>>>>> or >>>>>> after I call reviveOffers. >>>>>> >>>>>> What I'm not sure about, is what happens when the task that was >>>>>> launched completes, the rejection filter is still applicable, and I don't >>>>>> call reviveOffers. >>>>>> Once the <16cpu, 8GiB mem> is available, will I get these resources >>>>>> as an offer, or will the master combine it with the other 8GiB mem and >>>>>> offer me <16cpu, 16GiB> again? >>>>>> >>>>>> Thanks, >>>>>> - Itamar. >>>>>> >>>>> >>>> >>> >>
