hadoop-yarn-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Subramaniam V K <subru...@gmail.com>
Subject Re: [DISCUSS] YARN-5079 : Native YARN framework layer for services and Apache Slider
Date Wed, 20 Jul 2016 00:27:34 GMT
+1 on the proposal as:

1) We definitely need first class support for long running services on YARN.
2) We need to surface the support we are building in an easy to consume
format to app developers as YARN APIs IMHO is too low-level for direct
end-developer consumption.
3) Slider is the obvious choice for (2) (echoing Vinod/Gour).
3) At the same time we need to make sure that other *frameworks *like
Apache REEF can also build directly on the native services support in YARN
as there might be custom scenarios (in our case ML services) that cannot
all be satisfied by Slider alone.

I would like to think of Slider integration as a production grade reference
implementation of services on YARN akin to Nexus/Surface lines of devices
for Android/Windows :).


On Tue, Jul 19, 2016 at 4:51 PM, jonmaron@gmail.com <jonmaron@gmail.com>
wrote:

> +1 - this approach definitely addresses an important need.
>
> The project has gone through at least a couple of development/repo
> approaches:
>
> 1)  git-flow with master/develop branches
> 2)  abandonment of the master branch and work/releases based on the develop
> branch.
>
> I have no strong objections to the approach above, but given the existence
> of the two branches above, and some developer familiarity with the
> 'develop' branch, perhaps we could:
>
> 1)  Synch the master branch to the develop branch (not sure which git
> mechanism would best achieve this)
> 2)  Leverage the develop branch to create the agent/app-package based
> Slider distribution
>
> In this way there'a a legacy branch available while develop continues to
> proceed on the existing and familiar development branch.
>
> Just a thought.  Like I said - no real strong feelings :)
>
> On Mon, Jul 18, 2016 at 3:23 PM, Gour Saha <gsaha@hortonworks.com> wrote:
>
> > +1 forwarding from the discussion in Slider DL
> >
> > Note: On the Slider side, we plan to create a branch corresponding to
> this
> > YARN branch. In this, we intend to remove all the pieces which will move
> > to the YARN branch (Slider core, AM, client). We will retain the Agent
> and
> > app-packages which will then depend on the new hadoop-slider module (in
> > addition to the existing hadoop modules that it already depends on). This
> > will create a single view of Slider codebase, exactly as it stands today,
> > fully supporting the current app-packages. Slider can even release its
> > future versions from this new branch, once the hadoop-sldier module
> > reaches a matured state. It will be an easier path for existing Slider
> > users/app-owners to move to the future stable state of Slider completely
> > off of Hadoop YARN codebase. All that would be expected is to migrate the
> > app-packages to the agent-less version. Of course the Slider community
> > will do the migration for the current app-packages in the Slider
> codebase.
> >
> > -Gour
> >
> > On 7/14/16, 7:36 PM, "Vinod Kumar Vavilapalli" <vinodkv@apache.org>
> wrote:
> >
> > >Hi, Hadoop YARN community!
> > >
> > >(Cross-posting across Hadoop and Slider communities)
> > >
> > >I opened a JIRA a little while ago to pursue a native YARN framework for
> > >services: https://issues.apache.org/jira/browse/YARN-5079.
> > >
> > >It is part of a bigger effort that a bunch of us YARN community members
> > >are interested in making progress on: YARN-4692 - [Umbrella] Simplified
> > >and first-class support for services in YARN.
> > >
> > >The idea is that with our current attention on making services
> > >first­-class, it's time to take a fresh look at how we can make Apache
> > >Hadoop YARN support services well out of the box. I’ve been looking at
> > >various possibilities - ranging from a custom new framework room scratch
> > >to using one of the existing projects - and stopped at Apache Slider
> > >(http://slider.incubator.apache.org) given its association with some of
> > >the YARN community members (Steve Loughran, Devaraj Das, Arun C Murthy,
> > >myself etc.).
> > >
> > >Slider client & AM already handles a great deal of the functionality
> that
> > >we need. I posit that assimilating the client, ApplicationMaster etc of
> > >an existing framework like Apache Slider can serve our purpose really
> > >well. My early informal discussions about this with few Hadoop and
> Slider
> > >community members yielded generally favourable feedback.
> > >
> > >The Apache Slider incubator community also discussed this and expressed
> > >generally positive interest in YARN taking up Slider’s key pieces, you
> > >can see that discussion here: https://s.apache.org/0hoh.
> > >
> > >So in summary, we are looking to the following
> > >
> > >- Code
> > >   ― ‘Graduate' key pieces (Slider client, AM) of Apache Slider into
> > >Apache Hadoop for providing a native services experiences in YARN
> > >   ― Leave for now some of the pieces behind in Apache Slider - (a)
> > >Slider agent as we won’t need it, (b) Slider packages that need more
> > >deliberation in terms of where they will live in the long term.
> > >   ― Create a branch in YARN, copy this code over into a new module(s),
> > >and work towards completing a functioning app running on YARN.
> > >
> > >- Communities & releases
> > >   ― Good thing is that many of Apache Slider community members are
> > >already seasoned folk in the Apache Hadoop ecosystem projects. For those
> > >committers & PMC in Slider that are not yet Hadoop committers / PMC,
> > >without complicating things much, a proposed path forward is active
> > >participation in the branch (as branch committers?) and eventually in
> > >mainline YARN and thus go through a natural progression to committership
> > >/ PMC. Given that most of the members are stalwarts in the Apache
> > >communities, this should be a cinch IMO.
> > >   ― The work on this new code can start, and depending on its state,
> and
> > >assuming that the experiment succeeds, can be merged into trunk and
> later
> > >picked up in the next nearest & feasible Apache Hadoop release.
> > >   ― While the work on forked-over-code goes on till we have a
> > >functioning app, the current Apache Slider project continues to live,
> > >with supporting releases etc. At some point in the future, when things
> > >become clearer, more decisions can be taken on the (parts left behind in
> > >the) incubating project’s future.
> > >
> > >Thoughts?
> > >
> > >Thanks
> > >+Vinod
> >
> >
>
>
> --
> Sent from some random computer....
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message