This is a great discussion; thanks everyone. >From my perspective, the functionality to execute pipelines on both Storm and JStorm is very welcome and a big step forward for Beam.
I'm not an expert on the Storm/JStorm differences, but one vs. two runners discussion sounds more like a packaging / experience question than a deep technical problem. I'm guessing that a fair amount of functionality can be shared between both runners -- it would be a shame not to share these parts. Even further, with a possible merger on the horizon, it would make sense to plan and design for an (eventual) unified solution. In the meanwhile, whether the functionality is released in one runner (and, say, controlled with a flag) or two runners (without a flag) sounds minor. But, it would be great if the code that can be shared -- be written in a way that is shareable. On Tue, Apr 11, 2017 at 10:35 PM, 刘键(Basti Liu) <basti...@alibaba-inc.com> wrote: > Hi Taylor, > > It is glad to see your opinion. > After the open source of Beam, there are a lot of interests in Beam from > our internal users in Alibaba and other companies in China, which promotes > us to provide the support of JStorm runner. But since the implementation of > Storm runner is out of date, and over the past year many new features or > different solution(especially for exactly once and state) were introduced > in JStorm, we have to start the separate development of JStorm runner. > Currently, we have finished a prototype(support most PTransforms, window > and trigger of Beam) as Pei mentioned in another email, and the full > testing is still on-going. Some users has built up their trial topology on > it in Alibaba. But for further improvement, we still need the help of > review from Beam community to ensure the correctness, and get notification > of any broken or un-compatible update of Beam evolves. That is the reason > why we decide to commit JStorm runner into Beam repository. > > For personal understanding, the JStorm runner is not a duplicated effort. > The major part of JStorm runner is probably reused in Storm. Some other > parts like exactly once and state needs a propagation. When Storm community > plan to restart the development of Storm runner, we'd like to help on this, > as a part of merging JStorm features planned before. At that time, we can > discuss whether merging JStorm feature or propagation is required. > Looking forward to the better collaboration between Beam, Storm and JStorm. > > Regards > Jian Liu(Basti) > > -----Original Message----- > From: P. Taylor Goetz [mailto:ptgo...@apache.org] > Sent: Tuesday, April 11, 2017 1:48 AM > To: dev@beam.apache.org; d...@storm.apache.org > Subject: Apache Storm/JStorm Runner(s) for Apache Beam > > Note: cross-posting to dev@beam and dev@storm > > I’ve seen at least two threads on the dev@ list discussing the JStorm > runner and my hope is we can expand on that discussion and cross-pollinate > with the Storm/JStorm/Beam communities as well. > > A while back I created a very preliminary proof of concept of getting a > Storm Beam runner working [1]. That was mainly an exercise for me to > familiarize myself with the Beam API and discover what it would take to > develop a Beam runner on top of Storm. That code is way out of date (I was > targeting Beam’s HEAD before the 0.2.0 release, and a lot of changes have > since taken place) and didn’t really work as Jian Liu pointed out. It was a > start, that perhaps could be further built upon, or parts harvested, etc. I > don’t have any particular attachment to that code and wouldn’t be upset if > it were completely discarded in favor of a better or more extensible > implementation. > > What I would like to see, and I think this is a great opportunity to do > so, is a closer collaboration between the Apache Storm and JStorm > communities. For those who aren’t familiar with those projects’ > relationship, I’ll start with a little history… > > JStorm began at Alibaba as a fork of Storm (pre-Apache?) with Storm’s > Clojure code reimplemented in Java. The rationale behind that move was that > Alibaba had a large number of Java developers but very few who were > proficient with Clojure. Moving to pure Java made sense as it would expand > the base of potential contributors. > > In late 2015 Alibaba donated the JStorm codebase to the Apache Storm > project, and the Apache Storm PMC committed to converting its Clojure code > to Java in order to incorporate the code donation. At the time there was > one catch — Apache Storm had implemented comprehensive security features > such as Kerberos authentication/authorization and multi-tenancy in its > Clojure code, which greatly complicated the move to Java and incorporation > of the JStorm code. JStorm did not have the same security features. A > number of JStorm developers have also become Storm PMC members. > > Fast forward to today. The Storm community has completed the bulk of the > move to Java and the next major release (presumably 2.0, which is currently > under discussion) will be largely Java-based. We are now in a much better > position to begin incorporating JStorm’s features, as well as implementing > new features necessary to support the Beam API (such as support for bounded > pipelines, among other features). > > Having separate Apache Storm and JStorm beam runner implementations > doesn’t feel appropriate in my personal opinion, especially since both > projects have expressed an ongoing commitment to bringing JStorm’s > additional features, and just as important, community, to Apache Storm. > > One final note, when the Storm community initially discussed developing a > Beam runner, the general consensus was do so within the Storm repository. > My current thinking is that such an effort should take place within the > Beam community, not only since that is the development pattern followed by > other runner implementations (Flink, Apex, etc.), but also because it would > serve to increase collaboration between Apache projects (always a good > thing!). > > I would love to hear opinions from others in the Storm/JStorm/Beam > communities. > > -Taylor= > >
