I think core developers should do whatever they want. Nobody is entitled to anything from a developer of an OSS project, unless there is a contract and/or money involved.
Nobody seriously uses TensorFlow without Keras, except in situations when they have to make custom modifications precisely to support broken things that the core devs won’t fix.
Contrast this with Keras and it’s night and day. Keras actually has support guidelines and a GitHub project and, even if lacking capacity, actually cares about its users instead of being willfully hostile to users like TensorFlow is. The original creator of Keras, Francois Chollet, can even be found very often in GitHub issue threads assisting totally novice newcomers to Keras and explaining tips and tricks with custom layers, etc.
I compare it to the 'old days' of software: it used to be common to have a phone number to call. Then it became email only and people complained there should be a support phone number. Then it was $0.99 apps and maybe you got a support forum. Now in the era of freemium, you're lucky if you get an online community support forum... and people complain when nobody monitors it. When you move to free, the support budget goes to $0 which leaves developers to shoulder support in addition to everything else. Unsurprisingly, many don't want to do this. If you're willing to pay, sure you can still get anything you want (up to and including on-site support for enterprise stuff if you have deep enough pockets) but as long as the default support model is clear at least one can make an informed decision. Support is one of those things that the $300+ (phone support), $30+ (email support), $10- (forum?), $0 (good luck) pays/paid for after all...
In the case of TensorFlow in particular, it allows forward thinking people to realize it’s not a safe project to bet the farm on at all, not even anywhere close, even without factoring in Google’s tendency to sunset projects regardless of their relative success or community needs.
This is why nobody’s using it. Keras gives you the option of other backends, and there’s a large shift to PyTorch as well. Even being new and not as featureful, you simply just cannot stake the long term health of a project on the constant rug pull that is TensorFlow.
I only care about what core devs “should” do from the perspective of “what keeps a project healthy.”
If someone wants to optimize that choice for something else, like their personal enjoyment of hacking on a new feature, I think it would have unfortunate consequences, but other than stating my feeling about that, I couldn’t care less and it seems irrelevant to the discussion.
That’s why I characterize it as an ego problem. As far as the whole rest of the project and user base is concerned, what some particular person wants just doesn’t matter.
Yes, they’re free to make their choices that way. Then it just takes that particular case out of the discussion we’re having here about providing user support.