Mailing-List: contact general-help@hadoop.apache.org; run by ezmlm
Precedence: bulk
Reply-To: general@hadoop.apache.org
Received-SPF: neutral (nike.apache.org: local policy)
Date: Mon, 4 Mar 2013 22:15:05 -0800
From: Konstantin Boudnik <cos@apache.org>
To: general@hadoop.apache.org
Cc: dev@bigtop.apache.org
Subject: Re: [DISCUSS] stabilizing Hadoop releases wrt. downstream
Message-ID: <20130305061505.GF32383@tpx>
Mail-Followup-To: general@hadoop.apache.org, dev@bigtop.apache.org
References: 
 <CA+ULb+vU+i3HdA0ZmsG+N08+V7NiQtCdH678GwC3oicwfPU=Zg@mail.gmail.com>
 <DC1D0069-28C9-42A8-86C4-855EF818D8E1@hortonworks.com>
MIME-Version: 1.0
Content-Type: multipart/signed; micalg=pgp-sha256;
	protocol="application/pgp-signature"; boundary="uQr8t48UFsdbeI+V"
Content-Disposition: inline
In-Reply-To: <DC1D0069-28C9-42A8-86C4-855EF818D8E1@hortonworks.com>
User-Agent: Mutt/1.5.21 (2010-09-15)

--uQr8t48UFsdbeI+V
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: quoted-printable

Arun,

first of all, I don't think anyone is trying to put a blame on someone
else. E.g. I had similar experience with Oozie being broken because of
certain released changes in the upstream.

I am sure that most people in BigTop community - especially those who
share the committer-ship privilege in BigTop and other upstream
projects, including Hadoop, - would be happy to help with the
stabilization of the Hadoop base. The issue that a downstream
integration project is likely to have is - for once - the absence of
regularly published development artifacts. In the light of "it didn't
happen if there's no picture" here's a couple of examples:

  - 2.0.2-SNAPSHOT weren't published at all; only release 2.0.2-alpha artif=
acts were
  - 2.0.3-SNAPSHOT weren't published until Feb 29, 2013 (it happened just o=
nce)

So, technically speaking, unless an integration project is willing to
build and maintain its own artifacts, it is impossible to do any
preventive validation.

Which brings me to my next question: how do you guys address
"Integration is high on the list of *every* release". Again, please
don't get me wrong - I am not looking to lay a blame on or corner
anyone - I am really curious and would appreciate the input.


Vinod:

> As you yourself noted later, the pain is part of the 'alpha' status
> of the release. We are targeting +one of the immediate future
> releases to be a beta and so these troubles are really only the
> short +term.

I don't really want to get into the discussion about of what
constitutes the alpha and how it has delayed the adoption of Hadoop2
line. However, I want to point out that it is especially important for
"alpha" platform to work nicely with downstream consumers of the said
platform. For quite obvious reasons, I believe.

> I think there is a fundamental problem with the interaction of
> Bigtop with the downstream projects, if nothing else, with

BigTop is as downstream as it can get, because BigTop essentially
consumes all other component releases in order to produce a viable
stack. Technicalities aside...

> Hadoop. We never formalized on the process, will BigTop step in
> after an RC is up for vote or before? As I see it, it's happening

Bigtop essentially can give any component, including Hadoop, and
better yet - the set of components - certain guaratees about
compatibility and dependencies being included. Case in point is
missing commons libraries missed in 1.0.1 release that essentially
prevented HBase from working properly.

> after the vote is up, so no wonder we are in this state. Shall we
> have a pre-notice to Bigtop so that it can step in before?

The above is in contradiction with earlier statement of "Integration
is high on the list of *every* release". If BigTop isn't used for
integration testing, then how said integration testing is performed?
Is it some sort of test-patch process as Luke referred earlier?  And
why it leaves the room for the integration issues being uncaught?
Again, I am genuinely interested to know.

> these short term pains. I'd rather like us swim through these now
> instead of support broken APIs and features in our beta, having seen
> this very thing happen with 1.*.

I think you're mixing the point of integration with downstream and
being in an alpha phase of the development. The former isn't about
supporting "broken APIs" - it is about being consistent and avoid
breaking the downstream applicaitons without letting said applications
to accomodate the platform changes first.

Changes in the API, after all, can be relatively easy traced by
integration validation - this is the whole point of integration
testing. And BigTop does the job better then anything around, simply
because there's nothing else around to do it.

If you stay in shape-shifting "alpha" that doesn't integrate well for
a very long time, you risk to lose downstream customers' interest,
because they might get tired of waiting until a next stable API will
be ready for them.

> Let's fix the way the release related communication is happening
> across our projects so that we can all work together and make 2.X a
> success.

This is a very good point indeed! Let's start a separate discussion
thread on how we can improve the release model for coming Hadoop
releases, where we - as the community - can provide better guarantees
of the inter-component compatibility (sorry for an overused word).

Cos

On Fri, Mar 01, 2013 at 10:58AM, Arun C Murthy wrote:
> I feel this is being blown out of proportion.
>=20
> Integration is high on the list of *every* release. In future, if anyone =
or
> bigtop wants to help, running integration tests on a hadoop RC and provid=
ing
> feedback would be very welcome. I'm pretty sure I will stop an RC if it
> means it breaks and Oozie or HBase or Pig or Hive and re-spin it. For e.g.
> see recent efforts to do a 2.0.4-alpha.
>=20
> With hadoop-2.0.3-alpha we discovered 3 *bugs* - making it sound like we
> intentionally disregard integation issues is very harsh.
>=20
> Please also see other thread where we discussed stabilizing APIS, protoco=
ls
> etc. for the next 'beta' release.
>=20
> Arun
>=20
> On Feb 26, 2013, at 5:43 PM, Roman Shaposhnik wrote:
>=20
> > Hi!
> >=20
> > for the past couple of releases of Hadoop 2.X code line the issue
> > of integration between Hadoop and its downstream projects has
> > become quite a thorny issue. The poster child here is Oozie, where
> > every release of Hadoop 2.X seems to be breaking the compatibility
> > in various unpredictable ways. At times other components (such
> > as HBase for example) also seem to be affected.
> >=20
> > Now, to be extremely clear -- I'm NOT talking about the *latest* version
> > of Oozie working with the *latest* version of Hadoop, instead
> > my observations come from running previous *stable*  releases
> > of Bigtop on top of Hadoop 2.X RCs.
> >=20
> > As many of you know Apache Bigtop aims at providing a single
> > platform for integration of Hadoop and Hadoop ecosystem projects.
> > As such we're uniquely positioned to track compatibility between
> > different Hadoop releases with regards to the downstream components
> > (things like Oozie, Pig, Hive, Mahout, etc.). Every single single RC
> > we've been pretty diligent at trying to provide integration-level feedb=
ack
> > on the quality of the upcoming release,  but it seems that our efforts
> > don't quite suffice in Hadoop 2.X stabilizing.
> >=20
> > Of course, one could argue that while Hadoop 2.X code line was
> > designated 'alpha' expecting much in the way of perfect integration
> > and compatibility was NOT what the Hadoop community was
> > focusing on. I can appreciate that view, but what I'm interested in
> > is the future of Hadoop 2.X not its past. Hence, here's my question
> > to all of you as a Hadoop community at large:
> >=20
> > Do you guys think that the project have reached a point where integrati=
on
> > and compatibility issues should be prioritized really high on the list
> > of things that make or break each future release?
> >=20
> > The good news, is that Bigtop's charter is in big part *exactly* about
> > providing you with this kind of feedback. We can easily tell you when
> > Hadoop behavior, with regard to downstream components, changes
> > between a previous stable release and the new RC (or even branch/trunk).
> > What we can NOT do is submit patches for all the issues. We are simply
> > too small a project and we need your help with that.
> >=20
> > I truly believe that we owe it to the downstream projects, and in the
> > second half of this email I will try to convince you of that.
> >=20
> > We all know that integration projects are impossible to pull off
> > unless there's a general consensus between all of the projects involved
> > that they indeed need to work with each other. You can NOT force
> > that notion, but you can always try to influence. This relationship
> > goes both ways.
> >=20
> > Consider a question in front of the downstream communities
> > of  whether or not to adopt Hadoop 2.X as the basis. To answer
> > that question each downstream project has to be reasonably
> > sure that their concerns will NOT fall on deaf ears and that
> > Hadoop developers are, essentially, 'ready' for them to pick
> > up Hadoop 2.X. I would argue that so far the Hadoop community
> > had gone out of its way to signal that 2.X codeline is NOT
> > ready for the downstream.
> >=20
> > I would argue that moving forward this is a really unfortunate
> > situation that may end up undermining the long term success
> > of Hadoop 2.X if we don't start addressing the problem. Think
> > about it -- 90% of unit tests that run downstream on Apache
> > infrastructure are still exercising Hadoop 1.X underneath.
> > In fact, if you were to forcefully make, lets say, HBase's
> > unit tests run on top of Hadoop 2.X quite a few of them
> > are going to fail. Hadoop community is, in effect, cutting
> > itself off from the biggest source of feedback -- its downstream
> > users. This in turn:
> >=20
> >   * leaves Hadoop project in a perpetual state of broken
> >     windows syndrome.
> >=20
> >   * leaves Apache Hadoop 2.X releases in a state considerably
> >     inferior to the releases *including* Apache Hadoop done by the
> >     vendors. The users have no choice but to alight themselves
> >     with vendor offerings if they wish to utilize latest Hadoop functio=
nality.
> >     The artifact that is know as Apache Hadoop 2.X stopped being
> >     a viable choice thus fracturing the user community and reducing
> >     the benefits of a commonly deployed codebase.
> >=20
> >    * leaves downstream projects of Hadoop  in a jaded state where
> >      they legitimately get very discouraged and frustrated and eventual=
ly
> >      give up thinking that -- well, we work with one release of Hadoop
> >      (the stable one Hadoop 1.X) and we shall wait for the Hadoop
> >      community to get their act together.
> >=20
> > In my view (shared by quite a few members of the Apache Bigtop) we
> > can definitely do better than this if we all agree that the proposed
> > first 'beta' release of Hadoop 2.0.4 is the right time for it to happen.
> >=20
> > It is about time Hadoop 2.X community wins back all those end users
> > and downstream projects that got left behind during the alpha
> > stabilization phase.
> >=20
> > Thanks,
> > Roman.
>=20
> --
> Arun C. Murthy
> Hortonworks Inc.
> http://hortonworks.com/
>=20
>=20

--uQr8t48UFsdbeI+V
Content-Type: application/pgp-signature; name="signature.asc"
Content-Description: Digital signature

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.11 (GNU/Linux)

iF4EAREIAAYFAlE1jWkACgkQenyFlstYjhL0zAD8DbH4t4mOs0RzfgdQ5Xunc6qv
Tb6kgX+1gzC8BO3GkzUA/j22W4Ug8U3vbR/iDs9grNeKsrKA9zes2ekj2gEWY29G
=FOBz
-----END PGP SIGNATURE-----

--uQr8t48UFsdbeI+V--