Correction, outside of the ML groups and anything related to CI/infra. Especially if you remember that Starlark[0] is a dialect of Python.
And that’s without going into all the outside-of-g3 code (of which there is a metric ton, especially if you worked with any teams that deal with hardware or third-party/acquisition stuff).
Starlark is not Python. It feels like Python when you write it, but it's very different in a lot of ways that really matter.
A lot of people don't realize that the issues that python haters (myself included) have aren't generally about the look and feel of the language, but about how many sharp edges the language has for maintenance and scaling. Python 4 could fix all of these things if they ever did it.
I agree with you on this, but just saying, there is plenty of actual python (not starlark) at Google, all over the most crucial places (which a lot of people who haven’t touched those don’t even realize exist).
Personal example - the fleet of hardware prototypes (that I used to work on) used for hardware-in-the-loop testing (basically the hardware version of CI/CD) was pretty much reliant on python. Anything that was compact enough to be accomplished by a script and generic enough (open this serial connection, write to that memory address, etc.), the de-facto default choice was almost always python.
But I fully agree with you otherwise, in a sense that I haven’t seen much of a gigantic python codebase consisting of a bunch of interconnected python modules, like you would see with other popular languages at Google (e.g., C++/Java/JS).
And that’s without going into all the outside-of-g3 code (of which there is a metric ton, especially if you worked with any teams that deal with hardware or third-party/acquisition stuff).
0. https://github.com/bazelbuild/starlark