We use Julia for almost all numeric/data-oriented stuff now, and have moved a lot of the Python code to Julia. We still use Python for other parts of the codebase, like connecting to exchanges and doing the actual trading. At this point 99% of new code I write is Julia, and for the company as a whole it’s about 50-50. The performance differential has become even more ridiculous, an experiment that used to cost ~$125 to run now costs ~.33 cents, and most of those improvements wouldn’t really have been possible in Python.
Interop is still a bit painful, in particular because any Python code called through Julia is not thread-safe, and even forcing Python code to run on the main thread doesn’t solve all the issues.