hadoop-yarn-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Allen Wittenauer ...@altiscale.com>
Subject Re: Looking to a Hadoop 3 release
Date Fri, 06 Mar 2015 05:21:56 GMT

Is there going to be a general upgrade of dependencies?  I'm thinking of jetty & jackson
in particular.

On Mar 5, 2015, at 5:24 PM, Andrew Wang <andrew.wang@cloudera.com> wrote:

> I've taken the liberty of adding a Hadoop 3 section to the Roadmap wiki
> page. In addition to the two things I've been pushing, I also looked
> through Allen's list (thanks Allen for making this) and picked out the
> shell script rewrite and the removal of HFTP as big changes. This would be
> the place to propose features for inclusion in 3.x, I'd particularly
> appreciate help on the YARN/MR side.
> 
> Based on what I'm hearing, let me modulate my proposal to the following:
> 
> - We avoid cutting branch-3, and release off of trunk. The trunk-only
> changes don't look that scary, so I think this is fine. This does mean we
> need to be more rigorous before merging branches to trunk. I think
> Vinod/Giri's work on getting test-patch.sh runs on non-trunk branches would
> be very helpful in this regard.
> - We do not include anything to break wire compatibility unless (as Jason
> says) it's an unbelievably awesome feature.
> - No harm in rolling alphas from trunk, as it doesn't lock us to anything
> compatibility wise. Downstreams like releases.
> 
> I'll take Steve's advice about not locking GA to a given date, but I also
> share his belief that we can alpha/beta/GA faster than it took for Hadoop
> 2. Let's roll some intermediate releases, work on the roadmap items, and
> see how we're feeling in a few months.
> 
> Best,
> Andrew
> 
> On Thu, Mar 5, 2015 at 3:21 PM, Siddharth Seth <sseth@apache.org> wrote:
> 
>> I think it'll be useful to have a discussion about what else people would
>> like to see in Hadoop 3.x - especially if the change is potentially
>> incompatible. Also, what we expect the release schedule to be for major
>> releases and what triggers them - JVM version, major features, the need for
>> incompatible changes ? Assuming major versions will not be released every 6
>> months/1 year (adoption time, fairly disruptive for downstream projects,
>> and users) -  considering additional features/incompatible changes for 3.x
>> would be useful.
>> 
>> Some features that come to mind immediately would be
>> 1) enhancements to the RPC mechanics - specifically support for AsynRPC /
>> two way communication. There's a lot of places where we re-use heartbeats
>> to send more information than what would be done if the PRC layer supported
>> these features. Some of this can be done in a compatible manner to the
>> existing RPC sub-system. Others like 2 way communication probably cannot.
>> After this, having HDFS/YARN actually make use of these changes. The other
>> consideration is adoption of an alternate system ike gRpc which would be
>> incompatible.
>> 2) Simplification of configs - potentially separating client side configs
>> and those used by daemons. This is another source of perpetual confusion
>> for users.
>> 
>> Thanks
>> - Sid
>> 
>> 
>> On Thu, Mar 5, 2015 at 2:46 PM, Steve Loughran <stevel@hortonworks.com>
>> wrote:
>> 
>>> Sorry, outlook dequoted Alejandros's comments.
>>> 
>>> Let me try again with his comments in italic and proofreading of mine
>>> 
>>> On 05/03/2015 13:59, "Steve Loughran" <stevel@hortonworks.com<mailto:
>>> stevel@hortonworks.com>> wrote:
>>> 
>>> 
>>> 
>>> On 05/03/2015 13:05, "Alejandro Abdelnur" <tucu00@gmail.com<mailto:
>>> tucu00@gmail.com><mailto:tucu00@gmail.com>> wrote:
>>> 
>>> IMO, if part of the community wants to take on the responsibility and
>> work
>>> that takes to do a new major release, we should not discourage them from
>>> doing that.
>>> 
>>> Having multiple major branches active is a standard practice.
>>> 
>>> Looking @ 2.x, the major work (HDFS HA, YARN) meant that it did take a
>>> long time to get out, and during that time 0.21, 0.22, got released and
>>> ignored; 0.23 picked up and used in production.
>>> 
>>> The 2.04-alpha release was more of a troublespot as it got picked up
>>> widely enough to be used in products, and changes were made between that
>>> alpha & 2.2 itself which raised compatibility issues.
>>> 
>>> For 3.x I'd propose
>>> 
>>> 
>>>  1.  Have less longevity of 3.x alpha/beta artifacts
>>>  2.  Make clear there are no guarantees of compatibility from alpha/beta
>>> releases to shipping. Best effort, but not to the extent that it gets in
>>> the way. More succinctly: we will care more about seamless migration from
>>> 2.2+ to 3.x than from a 3.0-alpha to 3.3 production.
>>>  3.  Anybody who ships code based on 3.x alpha/beta to recognise and
>>> accept policy (2). Hadoop's "instability guarantee" for the 3.x
>> alpha/beta
>>> phase
>>> 
>>> As well as backwards compatibility, we need to think about Forwards
>>> compatibility, with the goal being:
>>> 
>>> Any app written/shipped with the 3.x release binaries (JAR and native)
>>> will work in and against a 3.y Hadoop cluster, for all x, y in Natural
>>> where y>=x  and is-release(x) and is-release(y)
>>> 
>>> That's important, as it means all server-side changes in 3.x which are
>>> expected to to mandate client-side updates: protocols, HDFS erasure
>>> decoding, security features, must be considered complete and stable
>> before
>>> we can say is-release(x). In an ideal world, we'll even get the semantics
>>> right with tests to show this.
>>> 
>>> Fixing classpath hell downstream is certainly one feature I am +1 on.
>> But:
>>> it's only one of the features, and given there's not any design doc on
>> that
>>> JIRA, way too immature to set a release schedule on. An alpha schedule
>> with
>>> no-guarantees and a regular alpha roll, could be viable, as new features
>> go
>>> in and can then be used to experimentally try this stuff in branches of
>>> Hbase (well volunteered, Stack!), etc. Of course instability guarantees
>>> will be transitive downstream.
>>> 
>>> 
>>> This time around we are not replacing the guts as we did from Hadoop 1 to
>>> Hadoop 2, but superficial surgery to address issues were not considered
>> (or
>>> was too much to take on top of the guts transplant).
>>> 
>>> For the split brain concern, we did a great of job maintaining Hadoop 1
>> and
>>> Hadoop 2 until Hadoop 1 faded away.
>>> 
>>> And a significant argument about 2.0.4-alpha to 2.2 protobuf/HDFS
>>> compatibility.
>>> 
>>> 
>>> Based on that experience I would say that the coexistence of Hadoop 2 and
>>> Hadoop 3 will be much less demanding/traumatic.
>>> 
>>> The re-layout of all the source trees was a major change there, assuming
>>> there's no refactoring or switch of build tools then picking things back
>>> will be tractable
>>> 
>>> 
>>> Also, to facilitate the coexistence we should limit Java language
>> features
>>> to Java 7 (even if the runtime is Java 8), once Java 7 is not used
>> anymore
>>> we can remove this limitation.
>>> 
>>> +1; setting javac.version will fix this
>>> 
>>> What is nice about having java 8 as the base JVM is that it means you can
>>> be confident that all Hadoop 3 servers will be JDK8+, so downstream apps
>>> and libs can use all Java 8 features they want to.
>>> 
>>> There's one policy change to consider there which is possibly, just
>>> possibly, we could allow new modules in hadoop-tools to adopt Java 8
>>> languages early, provided everyone recognised that "backport to branch-2"
>>> isn't going to happen.
>>> 
>>> -Steve
>>> 
>>> 
>> 


Mime
View raw message