First we should take a step back. Is there any machine learning done a quantum computer that cannot be efficiently simulated on a classical computer? The answer currently (2020) is no. In this respect quantum machine learning (which has many variants) is at the fundamental research phase. None of this is at a stage where it is at all considered something that should be mentioned in the same sentence with "production".
The last paragraph is probably a downer, but if you're a researcher it's an upper, because there are lots of open questions. There are a lot of really interesting ideas about how a quantum computer might be useful for machine learning. These differ on whether their data is quantum or classical, and how you put together different classical and quantum building blocks. The question is whether these will be useful for near term quantum computers (which have O(100) qubits and can computer for O(100) clock cycles before noise overwhelms them), and also whether they are interested for error correct quantum computers (think 5-15 years out). This is an area of active research. There is also a big caveat to what I said, and that is that while the quantum computations one might do can be classically simulated, in some cases, where your data is quantum, just being able do some small quantum computation on that quantum data can allow you to do things that you cannot do at all. These tasks are often "quantum sensor" problems, and are considered by many to be the first place practical quantum machine learning will be used.
So what is TensorFlow Quantum? It's a platform for doing research on quantum machine learning algorithms. It can be hooked up to a real quantum device, but where it excels right now is in performing classical simulations of quantum machine learning ideas. In particular it makes it very easy to do quantum machine learning with models that have quantum programs (quantum circuits) and classical machine learning (what people sometimes call hybrid algorithms) mixed together. It is integrated with TensorFlow and also with a framework for writing quantum programs Cirq. It is also backed by a highly efficient classical simulator of quantum programs, qsim. One advantage of TFQ is that it can leverage TensorFlow's ability to scale up training across a large distributed cluster. This, combined with the fast simulator, should allow researchers to quickly explore the space of possible research ideas.
Note that there are quite a few other quantum machine learning frameworks out there. These are all primarily aimed at helping researchers. Prominent ones are PennyLane and Microsoft's Quantum Machine Learning library in Q#. IBM also has nice tutorials on using their quantum programming framework for quantum machine learning.