Home | About | Sematext search-lucene.com search-hadoop.com
 Search Hadoop and all its subprojects:

Switch to Threaded View
Hadoop, mail # general - [DISCUSS] Spin out MR, HDFS and YARN as their own TLPs and disband Hadoop umbrella project


Copy link to this message
-
Re: [DISCUSS] Spin out MR, HDFS and YARN as their own TLPs and disband Hadoop umbrella project
Todd Lipcon 2012-08-31, 17:06
On Fri, Aug 31, 2012 at 9:58 AM, Robert Evans <[EMAIL PROTECTED]> wrote:
> The problem there is that YARN depends on Common, and MapReduce depends on
> YARN, so we would either have a circular dependency or we would have to
> split off MapRedcue too.

I haven't been in the MR codebase much of late, so I'll defer to your
judgment here: would it be feasible to have an abstraction layer for
"cluster manager" separated out into a pile of interfaces? Then we
could leave MR inside Hadoop, and Yarn would have an "MR->Yarn
binding" module. I'm not sure where the line would be drawn, but one
possibility would be to separate out the MR _task_ code from the MR
scheduling code (AM, Job Submission, etc)

Again would be a large project, but as Eli said, it would help make MR
more "relocatable" onto other cluster schedulers like Mesos (or even a
traditional grid scheduler). Another possible boon there would be
something I've discussed with Arun a few times: it would be cool if we
could get the new MR task code (in particular the rewritten reduce,
but also some of the new exciting work that Tsuyoshi and Mariappan are
doing) running in the context of an MR1 cluster.

-Todd

>
> On 8/31/12 11:54 AM, "Eli Collins" <[EMAIL PROTECTED]> wrote:
>
>>How about a proposal to just spin YARN off as a TLP?  Rationale:
>>
>>1. YARN started as a separate project and has a more independent
>>community than Common/HDFS/MR (per below these communities do not
>>divide at sub-project boundaries) that appears to want to be even more
>>independent.
>>
>>2. YARN is technically much easier to separate from the rest of the
>>code base (than separating Common and HDFS for example). Separating it
>>out will also help accelerate other efforts like MR2 support for
>>Apache Mesos.
>>
>>3. It side steps a number of thorny issues (how to handle branch-1,
>>how to handle what Hadoop is wrt enforcing trademark, who to remove
>>people from the Hadoop PMC, etc) that haven't been addressed in any of
>>these proposals.
>>
>>4. It's a proof point - if you can't make the case for YARN then
>>there's no way we're going to make a case for splitting the other
>>projects (this thread).
>>
>>Ie this doesn't have to be an all-or-nothing proposition for all
>>sub-projects, since the communities don't fall on sub-project
>>boundaries.
>>
>>Thanks,
>>Eli
>>
>>On Tue, Aug 28, 2012 at 7:33 PM, Mattmann, Chris A (388J)
>><[EMAIL PROTECTED]> wrote:
>>> [decided to minimize traffic and to simply put this in one thread]
>>>
>>> Hi Guys,
>>>
>>> See the recent discussion on these threads:
>>>
>>> YARN as its own Hadoop "sub project": http://s.apache.org/WW1
>>> Maintain a single committer list for the Hadoop project:
>>>http://s.apache.org/Owx
>>>
>>> ...and just pay attention to the Hadoop project over the last 3-4
>>>years. It's operating
>>> as a single project, that's masking separate communities that
>>>themselves are really
>>> separate ASF projects.
>>>
>>> At the ASF, this has been a problem area called "umbrella" projects and
>>>over the years,
>>> all I've seen from them is wasted bandwidth, artificial barriers and
>>>the inventions of
>>> new ways to perform process mongering and to reduce the fun in
>>>developing software
>>> at this fantastic foundation.
>>>
>>> I've talked about umbrella projects enough. We've diverted conversation
>>>enough.
>>> Enough people have tried to act like there is some technical mumbo
>>>jumbo that is
>>> preventing the eventual act of higher power that I myself hope comes
>>>should these
>>> discussions prove unfruitful through normal means.
>>>
>>> *these. are. separate. projects.*
>>>
>>>*there.are.not.blocker.issues.from.spinning.out.these.projects.as.their.o
>>>wn.communities*
>>>
>>> In this email: http://s.apache.org/rSm
>>>
>>> And in the 2 subsequent follow ons in that thread, I've outlined a
>>>process that I'll copy
>>> through below for splitting these projects into their own TLPs:
>>>
>>> -----snip
>>> Process:
>>>
>>> 0. [DISCUSS] thread for <TLP name> in which you talk about #1 and #2

Todd Lipcon
Software Engineer, Cloudera