[RFC][DISCUSS] TVM Apache Incubation

[tqchen]

Dear community, happy new year! As the TVM community grows larger, we have heard voices on the possibility of joining a software foundation. As per our community tradition, I open this thread for discussing the possible futures of the community. In particular, we would like to discuss the possibility to propose TVM to Apache Incubator.

Please share your thoughts, concerns, ideas.

Why

The TVM currently runs like an Apache project already, with merit-based governance model and a strong emphasis of community. So it is natural to do such a transition. Markus, our potential Apache mentor kindly gave a talk during TVM conference. You can find his slides here

Why Not

Apache does not magically make things happen. I have seen good and bad things happen in Apache communities. Sometimes it is simply due to the conflict of interest happened in bad ways, and not necessarily has a direct connection to incubation itself. As a diverse community, we believe we have already experienced ways to work together in a positive and collaborative way, so I hope such risk is minimized.

Another potential risk could be changing ways of development. Historically, Apache projects favor a single way of development:

• Use gitbox(which get mirrored to github)
• Use JIRA instead of github issues.
• Mail-list centric culture
  • use dev@ for development
  • many projects uses user@ mail-list for general support discussion.

These restrictions are being lifted recently and Apache encourages each project to follow their way of development(e.g. use github issues). One of the main philosophy behind these choices is to make sure that the development is not dominated by a single party. For example, face to face or one to one hangout call makes it hard for other parties to participate. So we need to ask developers to summarize their discussions and provide RFC to the community via public archive(e.g. github issues). By making all the development and discussion history public, we enable community members who maybe not have time to participate face to face meetings to join the development smoothly as well. Our community adopts the same philosophy(develop in a publically archivable place), with slight variants:

• We use github primary for development.
• We use github issues for design discussions and RFCs
• We use a discourse forum https://discuss.tvm.ai for general support discussions

My understanding is that the Apache is flexible enough as long as we follow the general principle. Likely we will need to subscribe github issues to dev@, but can keep the discourse forum and use github for development

How it could affect us

We shall continue to grow the community and make TVM awesome for everyone. Because the TVM already follows the Apache-style governance model, the transition should be smooth. There will be a disruptive period when we transition the codebase

The following things will happen, based on my previous experience with Apache projects,
if we decide to incubate

• We will send out a proposal to Apache incubator to see if they would like to sponsor the project
• We will migrate the project repo to apache/incubator-tvm
• Apache generally encourages each community to use their ways of development, as long as we follow the apache-way, we can keep our current style of development
  • Likely we can keep the current way of development, with GitHub and RFCs
  • A dev@ mail-list will be created, which mirrors GitHub issues, and RFCs can also be discussed there. dev@ mail-list is important in Apache culture, many projects do design discussion there. So far, we use github issues like the similar way people use dev@, so mirroring GH issues to dev@ would likely give us the most smooth transition.

The current committers will be invited as to Apache committers. The current PMC already follows the Apache style of voting in committers, and we only need to migrate the PMC mail-list. While reviewer is not an official role in Apache projects, we can still propose to maintaining our reviewer list and use this mechanism to help us find potential committers.

Share your thoughts

As per community tradition, all major proposals will be discussed as RFC by members of the community. We will open this thread for a week for discussions on whether should we join the process. I have invited our potential mentors(@markusweimer, @sscdotopen, @bgchun) to join the discussion as well.

I will send out the incubating proposal for public RFC as well if we agree that that is a good idea. cc @dmlc/tvm-comitter

Please share your thoughts, examples:

• Let us do apache incubation
• I do not care
• No-op, the current way is OK, let us just keep the project as it is
• I have a question on ...

[were]

+1
(The essence of human beings is a repeater)

[masahi]

+1

[icemelon]

+1

[Laurawly]

+1

[tqchen]

Thanks for everyone who responded so far, I have updated the thread to also discuss the cons, and removed the +1, -1 so it does not look like a vote.

If you have some specific questions, or thoughts, please feel free to share them.

[srkreddy1238]

+1

[ZihengJiang]

The transition sounds good to me. One thing is, will Apache require us to transition the website/CI hosting?

[tqchen]

re. CI/website, as far as I understand, the main website(https://tvm.ai) will need to be migrated to https://tvm.incubator.apache.org or https://tvm.apache.org . Likely we can keep our own CIs due to the special requirement to setup the pipelines and discuss forum.

Regardless of where things go, the general principle is that the project will be owned by ASF and resources are owned and shared by the community

[Ravenwater]

what are the success statistics for Apache projects that have University-paid versus Company-paid principals? I am worried that a project with this level of complexity needs a continuously engaged engineering team of significant cost and skill level to keep the core healthy.

[tqchen]

I think whether a contributor comes from university or company does not matter. It is dangerous to impose bias toward either one. Note that a contributors' background does not necessarily imply their skill level. I have personally seen proliferate open source contributors from both backgrounds.

Our merit-based principle means that whoever made sustained high-quality contributions will get added to as a committer and gain more weight in the decision.

The quality of the codebase really boils down to whether can we have a diverse and healthy community that upholds a high standard, and works together closely. I think we are working hard toward that :)

[nhynes]

a project with this level of complexity needs a continuously engaged engineering team

I think that's already what we have! Even without incubation and a community-based dev team, TVM is already a significant way towards not only fast ML on the usual devices, but also heretofore unexplored ones with VTA. Indeed, TVM's development closely parallels that of MXNet, which continues to thrive, and actually drops some of the procedural/tooling baggage that can hinder contribution.

Furthermore, although there are certainly success stories of paid core devs (e.g., Neovim), keeping the project maintained exclusively by the community aligns its goals with both the ASF and those same people who use it.

Of course, one unique aspect of TVM is that its stakeholders come from both industry and academia. Very often, too, contributors represent both interests (e.g., those at industry research labs). (And, personally, I'd be more reassured by an army of capable grad students spending their 25 hour days hacking on TVM :). Overall, I think that there's no worry that the TVM codebase would grow anything but more cohesive.

[sscdotopen]

Sebastian here from the ASF. I would like to note that many successful Apache projects such as Flink or Spark managed to attract a mix of (paid) academic and industry contributors. A goal of the incubation process in Apache is to broaden the set of stakeholders in a project so that it does not exclusively rely on the benevolence of a single contributing entity. Best, Sebastian

…

On 09.01.19 08:41, Theodore Omtzigt wrote: what are the success statistics for Apache projects that have University-paid versus Company-paid principals? I am worried that a project with this level of complexity needs a continuously engaged engineering team of significant cost and skill level to keep the core healthy. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#2401 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAZAa9DaxfrItE_ueWA_OnI8foapS7KXks5vBfGLgaJpZM4Z2hE8>.

[tmoreau89]

I'm all for this proposed transition. And regarding academic-based vs. industry-based engineering team, I think we have the right mix to keep the project healthy and growing!

[xqdan]

Let us do apache incubation