Thank you Renan. Supporting a single executor per scheduler will suffice for our current needs. Btw, could the JIRA be updated since the change has been integrated?
-Bharath On Thu, May 12, 2016 at 10:26 PM, Renan DelValle <[email protected]> wrote: > > Hi Bharath, > > On Thu, May 12, 2016 at 10:42 AM, Bharath Ravi Kumar <[email protected]> > wrote: > >> 1) Thanks for pointing to the changes. I'll await their stabilization / >> GA. >> >> 2) I realize I stated it imprecisely. I meant the executor itself could >> run external to the task (e.g. the executor runs in a mesos containerizer >> but is "external" to the docker container which it launches). However, I >> realize it's not running as a regular host process without controls. I see >> this model followed in the experimental compose executor: >> https://github.com/mesos/docker-compose-executor/ >> The jira to support external executors in aurora appears inactive though: >> https://issues.apache.org/jira/browse/AURORA-1376 >> I also recall reading about the work to let the executor run externally, >> which would be very useful for such scenarios. >> > Support for custom executors has been completed. The flag must be enabled > and a config file must be provided: > -custom_executor_config [file must exist, file must be readable] > Path to custom executor settings configuration file. > > (org.apache.aurora.scheduler.configuration.executor.ExecutorModule.custom_executor_config) > > A config file looks like this: > https://gist.github.com/rdelval/e5e41a7a9d7da243b440ef37564a0af9 > > Should be noted that you will no longer be able to get the web ui from the > thermos observer when using a custom executor that is not thermos or based > on thermos. > > That having been said., what hasn't been implemented, is running multiple > executors at on a single instance of aurora-scheduler. Unfortunately my > bandwidth got clobbered around the time I began work on this and it wasn't > a simple feature to implement. I still hope to finish work on this if there > is still interest. > > -Renan > >> >> > Thanks, >> Bharath >> >> >> On Thu, May 12, 2016 at 4:32 AM, Maxim Khutornenko <[email protected]> >> wrote: >> >>> 1) The refactoring of resource management should make it easier to model >>>> custom resources. I can see why the resource model is bounded & strongly >>>> typed, and not free form (i.e. config based). Is the refactoring planned >>>> for commit/merge any time soon? >>> >>> >>> It's currently in progress. If you are curious about the direction, >>> below is the list of the relevant changes so far (most recent first): >>> >>> https://reviews.apache.org/r/47261/ >>> >>> https://reviews.apache.org/r/47050/ >>> >>> https://reviews.apache.org/r/46997/ >>> >>> https://reviews.apache.org/r/46948/ >>> >>> https://reviews.apache.org/r/46810 >>> >>> https://reviews.apache.org/r/46716/ >>> >>> https://reviews.apache.org/r/46459/ >>> >>> https://reviews.apache.org/r/46064/ >>> https://reviews.apache.org/r/46057/ >>> >>> 2) For supporting custom (operator privileged) pre/post task launch >>>> steps on a host, I'm considering extending/replacing thermos (based on my >>>> initial understanding). I'm assuming that since the executor runs external >>>> to the container that is used to launch tasks, it'd have the privilege & >>>> visibility to carry out such operator-specified administrative steps. >>>> Thoughts? >>> >>> >>> As mentioned above, the executor has to be packed into the user volume >>> at this point. Joshua Cohen is currently exploring the possibility of >>> running executor externally, so perhaps he could give you more details >>> when/if he has some results. >>> >>> On Wed, May 11, 2016 at 3:46 PM, Mauricio Garavaglia < >>> [email protected]> wrote: >>> >>>> >>>> >>>> On Tue, May 10, 2016 at 10:47 PM, Bharath Ravi Kumar < >>>> [email protected]> wrote: >>>> >>>>> Hi Maxim, >>>>> >>>>> Thanks for responding. >>>>> >>>>> 1) The refactoring of resource management should make it easier to >>>>> model custom resources. I can see why the resource model is bounded & >>>>> strongly typed, and not free form (i.e. config based). Is the refactoring >>>>> planned for commit/merge any time soon? >>>>> >>>>> 2) For supporting custom (operator privileged) pre/post task launch >>>>> steps on a host, I'm considering extending/replacing thermos (based on my >>>>> initial understanding). I'm assuming that since the executor runs external >>>>> to the container that is used to launch tasks, it'd have the privilege & >>>>> visibility to carry out such operator-specified administrative steps. >>>>> Thoughts? >>>>> >>>> >>>> Not really, the executors runs internally in the container. >>>> >>>> >>>>> >>>>> Thanks, >>>>> Bharath >>>>> >>>>> On Mon, May 9, 2016 at 10:23 PM, Maxim Khutornenko <[email protected]> >>>>> wrote: >>>>> >>>>>> These steps need to run external to the docker container being >>>>>>> launched. Based on the documentation, I couldn't find a way to achieve >>>>>>> these in Apache Aurora. Can someone let me know if this is indeed >>>>>>> possible? >>>>>>> >>>>>> >>>>>> >>>>>> There is no support for this sort of functionality in Aurora, at >>>>>> least not yet (ideas/proposals are welcome!). You can accomplish >>>>>> *some* of that by configuring your custom logic to run within your >>>>>> task image though. >>>>>> >>>>>> I also wanted to know if arbitrary resources modeled in Mesos can be >>>>>>> expressed in an Aurora job definition, assuming that resource isolation >>>>>>> is >>>>>>> handled externally. >>>>>> >>>>>> >>>>>> We are currently working on generalizing resource management in >>>>>> Aurora to simplify adding new resource types. That said, having a fully >>>>>> dynamic (e.g. config file driven) support for adding arbitrary resource >>>>>> types is not our goal at the moment. You can find more details here >>>>>> <https://docs.google.com/document/d/1J9SIswRMpVKQpnlvJAMAJtKfPP7ZARFknuyXl-2aZ-M> >>>>>> . >>>>>> >>>>>> On Fri, May 6, 2016 at 6:51 AM, Bharath Ravi Kumar < >>>>>> [email protected]> wrote: >>>>>> >>>>>>> Hi, >>>>>>> >>>>>>> I need the ability to run a few pre-lauch steps (e.g. create a >>>>>>> custom work directory) and post launch steps (e.g. docker inspect) on >>>>>>> the >>>>>>> physical host when launching a docker container. These steps need to run >>>>>>> external to the docker container being launched. Based on the >>>>>>> documentation, I couldn't find a way to achieve these in Apache Aurora. >>>>>>> Can >>>>>>> someone let me know if this is indeed possible? >>>>>>> >>>>>>> I also wanted to know if arbitrary resources modeled in Mesos can be >>>>>>> expressed in an Aurora job definition, assuming that resource isolation >>>>>>> is >>>>>>> handled externally. >>>>>>> >>>>>>> Thanks, >>>>>>> Bharath >>>>>>> >>>>>> >>>>>> >>>>> >>>> >>> >> >
