There are well designed abstractions that are helpful. It's very hard to come up with them.
Mycrotoque aims at those who quickly jump to introduce new abstractions into their workflow before they understand what they are doing. It very often results in pain. It helps in the short term, because you are up and running faster but these shiny new frameworks are often immature and, like Potemkin villages, only look good from the correct angle, and have no meat. Once you try to use them for custom scenarios, you end up having to route around the abstraction layer because it wasn't designed so well.
To take an example outside of cloud management etc, in deep learning frameworks: there were competing libraries some years ago: caffe, theano, tensorflow etc. François Chollet decided to create an abstraction layer that could unify multiple low level frameworks. Called Keras. Did it work? No. Because the different implementations were different enough that they couldn't usefully be coerced under the same abstraction layer. I mean it worked for some time, to some extent. But very soon after, they dropped support for anything but Tensorflow. Creating good abstractions is very hard especially if the underlying layer is unstable and gets big conceptual changes. With kernels it's not the case. We've had them in similar Form for like half a century.