hadoop-yarn-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Wangda Tan <wheele...@gmail.com>
Subject Re: [VOTE] Merge YARN-1197 container resize into trunk
Date Wed, 23 Sep 2015 03:14:01 GMT
Hi Karthik,

Let me elaborate more to make you feel better of this change, don't be
scared by the size of the patch :)

Common RM/Scheduler part:
- AbstractYarnScheduler new logic only.
- AppSchedulingInfo new logic only.
- RMContainer / RMNode state machine, new logic only
- SchedulerApplicationAttempt / Allocation, refactoring to existing
reservation logic so increase request reservation can reuse it, and
refactored to simply updating container token / pull container part so
increase/decrease/new-allocation can reuse same code.

FairScheduler:
- Small change since we updated how to pull container updated token. I
believe it will be a straightforward change for you if you take a closer
look at it.

CapacityScheduler:
- Most changes are separate logic or small refactorings, most complex
allocation logic stays within IncreaseContainerAllocator.java.

Please let me know where you want to get more details of implementations.

I strongly suggest you to take a glance at the diff, we have already worked
on the merge for the past one week, and we've paid a lot of extra time to
keep YARN-1197 sync with trunk in the past several months. After this merge
finished, a couple of weeks needed to finish end-to-end test and some other
extra tests, it won't affect our upcoming branch-2 release.

I would not prefer to merge to trunk only, all people working on RM side
will be affected, we're very carefully avoid such divergence of RM in
trunk/branch-2. Since nobody wants to create two different patches for
every RM changes. And also, after this finished, other efforts can happen
in parallel such as YARN-4091.

Let me know if you have any other questions/concerns.

Thanks,
Wangda

On Tue, Sep 22, 2015 at 7:26 PM, Karthik Kambatla <kasha@cloudera.com>
wrote:

> I am sorry, but merging a potentially disruptive change to branch-2 without
> end-to-end tests seems too disruptive to me.
>
> I do agree with you on the potential inconvenience of having to post
> different patches for trunk and branch-2, but I would rather have that
> inconvenience than the risk of merging something that hasn't been
> thoroughly tested.
>
> On Tue, Sep 22, 2015 at 6:18 PM, Wangda Tan <wheeleast@gmail.com> wrote:
>
> > Hi Karthik,
> >
> > Thanks for comments! However, I think only merge to trunk may not work,
> > this patch involves thousands lines of code changes in scheduler side,
> only
> > putting that to trunk could lead to trunk/branch-2 totally incompatible
> for
> > resource manager. I think most of the code changes are new to scheduler
> > instead of modifying existed logic, they're not very tricky to me. And
> when
> > 2.8 will be released is not planned yet, at least we have a couple of
> > months to make sure this feature becomes usable and not cause existing
> > behavior regressions.
> >
> > Sounds good to you?
> >
> > Wangda
> >
> >
> > On Tue, Sep 22, 2015 at 3:48 PM, Karthik Kambatla <kasha@cloudera.com>
> > wrote:
> >
> > > +1 on merging to trunk. It would be nice to have some amount of testing
> > > done before the merge, but I understand how merging to trunk would
> likely
> > > speed up the testing efforts.
> > >
> > > Let us not merge into branch-2 until after we have done a fair bit of
> > > testing, and are comfortable including it in a release. While the code
> > > mostly appears to not mess with existing scheduling logic, I am
> concerned
> > > about regressions to existing scheduling behavior.
> > >
> > >
> > >
> > > On Tue, Sep 22, 2015 at 1:28 PM, Karthik Kambatla <kasha@cloudera.com>
> > > wrote:
> > >
> > > > By the way, for the purposes of merge vote, I believe a committer's
> > vote
> > > > is binding. So, Wangda and Zhihai's votes should be binding. :)
> > > >
> > > > On Tue, Sep 22, 2015 at 11:38 AM, Zhihai Xu <zxu@cloudera.com>
> wrote:
> > > >
> > > >> +1 (non-binding)
> > > >>
> > > >> thanks
> > > >> Zhihai Xu
> > > >>
> > > >> On Tue, Sep 22, 2015 at 12:10 AM, Xuan Gong <xgong@hortonworks.com>
> > > >> wrote:
> > > >>
> > > >> > +1 Binding
> > > >> >
> > > >> > Thanks
> > > >> >
> > > >> > Xuan Gong
> > > >> >
> > > >> > > On Sep 22, 2015, at 12:03 AM, Junping Du <jdu@hortonworks.com>
> > > wrote:
> > > >> > >
> > > >> > > +1. (Binding).
> > > >> > >
> > > >> > > Thanks,
> > > >> > >
> > > >> > > Junping
> > > >> > > ________________________________________
> > > >> > > From: Wangda Tan <wheeleast@gmail.com>
> > > >> > > Sent: Thursday, September 17, 2015 3:19 AM
> > > >> > > To: yarn-dev@hadoop.apache.org
> > > >> > > Subject: Re: [VOTE] Merge YARN-1197 container resize into
trunk
> > > >> > >
> > > >> > > +1 (non-binding),
> > > >> > >
> > > >> > > Thanks Jian starting this thread. This can minimize effort
of
> > works
> > > >> > across branches.
> > > >> > >
> > > >> > > To clarify, this feature is end-to-end code  completed,
we have
> > API,
> > > >> > rm/nm implementations patches committed, but we haven't tested
it
> > > >> > end-to-end. Filed YARN-4175 to create an example program to test
> it
> > > >> > end-to-end.
> > > >> > >
> > > >> > > Regards,
> > > >> > > Wangda
> > > >> > >
> > > >> > >> On Sep 16, 2015, at 6:30 PM, Jian He <jhe@hortonworks.com>
> > wrote:
> > > >> > >>
> > > >> > >> Hi All,
> > > >> > >>
> > > >> > >> Thanks Meng Ding and Wangda Tan for all the hard work
!
> > > >> > >>
> > > >> > >> I would like to call a vote to merge YARN-1197 container
resize
> > > into
> > > >> > trunk.
> > > >> > >>
> > > >> > >> Key idea:
> > > >> > >> This feature adds the ability for AM to change container
> resource
> > > >> size
> > > >> > at runtime.
> > > >> > >>
> > > >> > >> Details:
> > > >> > >> - This feature is tracked at
> > > >> > https://issues.apache.org/jira/browse/YARN-1197
> > > >> > >> - It’s currently developed at a separate branch:
> > > >> > https://github.com/apache/hadoop/commits/YARN-1197
> > > >> > >> - A uber patch(https://issues.apache.org/jira/browse/YARN-4157
> )
> > > >> > generated from YARN-1197 to run against trunk  shows all unit
> tests
> > > have
> > > >> > passed.
> > > >> > >> - This feature now can work end-to-end.
> > > >> > >> -  All the unresolved jiras under YARN-1197 will be
the next
> > step.
> > > >> > >>
> > > >> > >> Thanks,
> > > >> > >> Wangda Tan & Meng Ding & Jian He
> > > >> > >
> > > >> > >
> > > >> >
> > > >> >
> > > >>
> > > >
> > > >
> > >
> >
>

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