Why Your tf.function Isn’t Working the Way You Think (and How to Fix It)

This guide breaks down the limitations of TensorFlow’s tf.function, especially when dealing with Python side effects, global variable mutations, iterators, and recursion. Through detailed examples, it shows why some code executes unpredictably, how state leaks can crash your workflow, and why recursive functions aren’t supported. It also explains best practices and workarounds—like using TensorFlow-native objects, tf.init_scope, and tf.data APIs—to ensure consistent, performant, and portable graph execution.