OK. If I understand correctly, your answer is the following: yarn.resources.* configuration variables are used by YARN localizer to make API and infrastructure classpath available, together with the application's own classpath, which is also determined by the YARN localizer. The question here is: how do we let the container JVM know the API/infrastructure classpaths when launching the container processes? If the API and infrastructure classpaths (i.e. installation path determined by the localizer) are customizable, then we would need to tell the container JVM those API/infra classpaths via some configuration variables as well, right? Hence, those configuration variable names need to be understood by the Samza application's code (which is run within the container) as well. If not, what's the mechanism that we will use to let the container JVM process to know where the YARN localizer has put API/infra classpaths?
Thanks! -Yi On Wed, Mar 11, 2020 at 8:09 PM Cameron Lee <cameronlee...@gmail.com> wrote: > The configuration variables are only used by the YARN localizer. The Samza > application will look for the framework resources in certain places in the > application's working directory when it needs to access them. My aim is to > do something similar to how "yarn.package.path" works. In other execution > environments, it is my understanding that "yarn.package.path" would get > replaced by a different environment-specific configuration key/value. > I agree that we should not use "yarn.resources.*" if the configurations are > not YARN-specific. Do you think that these resource localization configs > are generalizable to arbitrary environments? If so, does that mean > "yarn.package.path" is also generalizable? For example, what if some > execution environment does not use URLs to specify resource locations > (although maybe this isn't a reasonable concern to worry about?)? > > Thanks, > Cameron > > On Wed, Mar 11, 2020 at 4:43 PM Yi Pan <nickpa...@gmail.com> wrote: > > > Hi, Cameron, > > > > Thanks for the quick responses! Appreciate it. > > > > I am still having a concern on a): are those configuration variables used > > by YARN localizer or by Samza applications? If those are used only by the > > YARN localizer, I agree that we should keep those as yarn specific. > > Otherwise, I think that would still be better to name those as > > cluster.based.resources.*. The reason being: Samza applications are > > supposed to be able to run on different execution environments. Ideally, > > when we are deploying the same Samza application on YARN vs Mesos or > > managed K8s clusters, we should only need to change the configure values, > > not the configuration variable names and values. Does it make sense? > > Otherwise, we can schedule a conf call to clarify that. > > > > Thanks! > > > > -Yi > > > > On Tue, Mar 10, 2020 at 3:25 PM Cameron Lee <cameronlee...@gmail.com> > > wrote: > > > > > a) The "yarn.resources.*" configs are for localizing the necessary > > > resources into the working directory for the process. I felt that the > > > specific configuration format to specify these resources might be > > > YARN-specific (e.g. YARN has type and visibility configs for each of > its > > > resources), so a generic format might not apply. In a non-YARN case, > the > > > localization configs would need to be specified according to the > > technology > > > being used. > > > b) It is correct that the Avro version will need to be compatible with > > the > > > version that is used by the infrastructure, if infrastructure needs to > > use > > > Avro and pass the Avro object to the application. This is the case with > > any > > > serde technology that needs to be used. For the job coordinator, it is > > not > > > much of a concern anyways, since it is not doing serde of Avro > messages. > > > This may be more of a concern for general split deployment, which will > > > impact the processing containers, and will be a separate SEP. > > > c) It should work to leave infrastructure serdes in the infrastructure > > > classpath. The infrastructure serdes just see generic types (which are > > > java.lang.Object at runtime) for the messages, and they don't do > anything > > > with the concrete types, so in the infrastructure classes, the messages > > get > > > passed around as Object, but their concrete classes can still be loaded > > > from the application. As with (b), this is more of a concern for > general > > > split deployment, since the job coordinator doesn't do message serde. I > > > have run some tests regarding this classloading pattern, but we will do > > > further verification for general split deployment. > > > d) Yes, you are correct. Good catch. It should be "described above at > > > Application classloader". > > > > > > Thanks for all of your questions. I will clarify some details in the > doc > > > regarding your questions. > > > > > > Cameron > > > > > > On Mon, Mar 9, 2020 at 12:07 PM Yi Pan <nickpa...@gmail.com> wrote: > > > > > > > Hi, Cameron, > > > > > > > > Sorry to chime in late. Overall, looks great! I do have a few > > > > suggestions/questions before I can cast my vote here: > > > > a) for the configuration variable names, why are we limiting > ourselves > > to > > > > yarn.resource.*? We have changed some of the configuration variables > > from > > > > yarn specific to non-yarn specific. I would love to keep that > > consistent > > > > (i.e. gradually moving all our yarn-specific configuration variables > to > > > > non-yarn-specifc names) > > > > b) for the avro case as referred to in the delegation case in the > > > > Infrastructure classloader, if we delegate the object deserialization > > > class > > > > to the application classloader, would it be possible that the > > application > > > > provides an non-compatible version of avro class than the ones used > > > within > > > > the "infrastructure plugins" and hence causing runtime exception in > the > > > > infrastructure plugin? Or is the solution being: do not directly use > > > serde > > > > classes in the infrastructure code? > > > > c) following the description of infrastructure classloader flow, > where > > > > should we expect the serde classes? In the application classpath, I > > > guess? > > > > So, does that mean that we should exclude serde classes (including > > > > SerializableSerde and JsonSerdeV2) in the Samza infrastructure > package, > > > and > > > > tell the users to package them in application package? > > > > d) I am a bit confused about the description on "multiple" > application > > > > classloaders on the job coordinator: one is for the describe flow and > > the > > > > other is in the "Application" classloader, instead of "API" > > classloader, > > > > right? > > > > > > > > Best, > > > > > > > > -Yi > > > > > > > > > > > > On Wed, Mar 4, 2020 at 11:32 AM Ke Wu <ke.wu...@gmail.com> wrote: > > > > > > > > > +1. > > > > > > > > > > Thanks for driving this effort. > > > > > > > > > > Best, > > > > > Ke > > > > > > > > > > > On Mar 3, 2020, at 6:28 PM, Jagadish Venkatraman < > > > > jagadish1...@gmail.com> > > > > > wrote: > > > > > > > > > > > > +1 binding. > > > > > > > > > > > > Thanks Cameron. I look forward to this feature taking our "Stream > > > > > > Processing as a service" offering to the next level. > > > > > > > > > > > > Cheers > > > > > > > > > > > > On Tuesday, March 3, 2020, Prateek Maheshwari < > prate...@utexas.edu > > > > > > > > wrote: > > > > > > > > > > > >> +1 (binding) from me. Thanks for contributing this feature. > > Looking > > > > > forward > > > > > >> to having dependency isolation and to the ability to upgrade the > > > > > framework > > > > > >> independently from an application. > > > > > >> > > > > > >> Thanks, > > > > > >> Prateek > > > > > >> > > > > > >> On Fri, Feb 28, 2020 at 10:48 AM Cameron Lee < > > > cameronlee...@gmail.com > > > > > > > > > > >> wrote: > > > > > >> > > > > > >>> Hi all, > > > > > >>> > > > > > >>> This is a call for a vote on SEP-24: Cluster-based Job > > Coordinator > > > > > >>> Dependency Isolation. Thanks to everyone who reviewed the > > proposal > > > > and > > > > > >>> provided feedback. > > > > > >>> > > > > > >>> I have addressed comments on the SEP, and I am not aware of any > > > > further > > > > > >>> major questions or objections, so I am starting this vote. > > > > > >>> > > > > > >>> SEP link: > > > > > >>> > > > > > >>> https://cwiki.apache.org/confluence/display/SAMZA/SEP- > > > > > >> 24%3A+Cluster-based+Job+Coordinator+Dependency+Isolation > > > > > >>> > > > > > >>> Discuss thread: > > > > > >>> > > > > > >>> > > https://mail-archives.apache.org/mod_mbox/samza-dev/202001.mbox/% > > > > > >> 3cCAMja7KeGcRZ3H95Rxk5XE= > > > 60zxm6jxjkjuwwxmgmadpfbyk...@mail.gmail.com > > > > %3e > > > > > >>> There was also some discussion through comments on the SEP page > > > (see > > > > > >>> Resolved Comments). > > > > > >>> > > > > > >>> Please vote: > > > > > >>> [ ] +1 approve > > > > > >>> [ ] +0 no opinion > > > > > >>> [ ] -1 disapprove (and reason why) > > > > > >>> > > > > > >>> Thank you, > > > > > >>> Cameron > > > > > >>> > > > > > >> > > > > > > > > > > > > > > > > > > -- > > > > > > Jagadish > > > > > > > > > > > > > > > > > > > >
