One thing I’m hoping is an outcome of the surveys and ensuing discussions: a sober look at which parts of the infrastructure are effectively on a self sustaining path and which are reliant on large capital infusions to ensure their continued viability. For the latter, we need to seriously ask the question of whether such infusions will be made, and if not, make practical decisions accordingly.
My rationale for concern on this is simple but someone recently pointed out to me that it may not be apparent to everyone: the present state came from a large and mission-oriented investment by a couple of big players sponsoring large amounts of engineering work (most notably a federation of interests from Google but also others).
In large part, that investment no longer exists. Even those with a historic interest and goodwill cannot afford to maintain everything that was built. Even in the cases where some of us have large discretion to invest in the commons, it is not mission-oriented but tactical. And it must have a positive cost benefit that accrues to us.
To speak plainly, there are several parts of the MLIR infrastructure that I believe are well into self-sustaining territory and represent a fantastic value – both as a user and as a way for community minded investors like ourselves to give back. I’d like to make sure those are set up for success and self sustenance so they continue to be that way well into the future and well after the historic folks who hold the knowledge have passed the torch.
But there are others, like the tensor compiler domain which were never finished and have a lot of other issues that are derived straight from their origin story as a tensorflow compiler. As someone who in prior lives literally employed or sponsored many of the engineers who created that work, I am in a somewhat unique position to know how much it cost – and further to have an estimate of both the additional capital needed to make it fulfill its original mission and the corresponding maintenance costs. I remain one of the biggest employers of people attached to that work, and what sobers me is that without some reset of scope or mission, I can’t afford it. Further, because it is in the category of “much assembly required”, I can’t even join up with other like minded people and expect to get a positive payout because any investment upstream has a very small marginal benefit to the products we build on this part of the technology (and brings a lot of costs on its own). That isn’t to say we don’t do that but that it isn’t enough. Example, example.
I think there are still ways out of this but it requires us to use more of the tools in our toolbox than technical arguments or pretending that we are near the point of sustainability. We need to drastically manage the scope and make sure that the books balance for the investors. That has an obvious tie in to governance, but in my mind it also ties back to restructuring: I’ve been leading software projects for decades and I don’t know how to effect a change like that while embedded into a project like MLIR. Further, as a person who wants good things for LLVM generally, it really feels wrong to me to run effectively a sandbox project as part of a mature whole. There are many reasons for this that are perhaps only obvious to those of us who have managed large groups of people, but they are almost to the same level of natural law as most of the hard technical discussions and issues.
A wise person in the LLVM community once told me that if you want to build a compiler, the first thing you should do is get an MBA. That is because that over the timeframe and ruinous costs involved, it isn’t sufficient to expect that a purely technical vision will get there. You’ve got to have a plan for the investment and factor that in strongly. That’s where we are with some of this, in my opinion.
I’m not sure if this helps, but I got some feedback recently that some of these angles may not be obvious to a lot of people in the community, so I decided to try to share more of what I’ve been able to see over the last five years of this project.