Ask HN: Is tensorflow ready?

I've been trying to get up and running with TensorFlow for image classification and while I have some working code, it pales in comparison to the non-working code I have accumulated over the past few weeks. I have an older retrain.py example script working for building models, but trying to actually serve it has been a whole other ball of wax.

It seems every tutorial (outside the 'hello world' level), or resource, is broken or otherwise inaccessible. A number of sites I needed to signup for (kaggle, nvidia) had entirely broken processes for participation. And the TF documentation has a way of saying a lot without saying much at all.

Trying to use tensorflow_model_server has been a tailchase of signatures, request formats and arbitrary incompatible saved model formats has gotten me nowhere. I tried to trace back the evolution of some example code with git blame and it was just PR's with empty descriptions. (i would expect better from google?)

Oracle released graphpipe to streamline model serving - which seems like a great solution on the surface, but its been impossible to get running, also poorly documented, and in the end a waste of time for me.

Usually, for most technology, I can push through the pain and get something going. But Everything TF and related seems novel, poorly documented, and hacky. I've been at it for weeks and still haven't made the kind of progress i'd expect to have made by now.

I've never had THIS much trouble with any technology and it's starting to stress me out. I know I'm ignorant of the deeper ML theory, but holy crap how does anyone use this? Is Tensorflow just this hard for everyone? Because it presents itself as a framework for developers-like-me™ to get into ML, but it has been nothing like that for me.