| It was definitely cultural. The engineering hierarchy was: 1. C++ engineers thought the only real language was C++ 2. Java engineers thought the only real languages were C++ or Java 3. Python engineers either thought the only real languages were Python, C++ or Java or some of them thought only Python At that time (I don't know about now), Google had a designation of a frontend softare engineer ("FE SWE") and you'd see interview feedback where a particular interviewer would be neutral on a candidate and soft reject them by explicitly stating they were maybe good enough to be an FE SWE, even though the official stance was FE SWEs had to pass the regular SWE standard plus some extra. Basically, anything JS/CSS/HTML related was very much looked down upon by many. Blaze went through some growing pains. At a time it was a full Python interpreter and then an interpreter of a subset of Python and ultimately not really Python at all. There was a time when you had to do things with genrules, which was an awful user expeience and it broke caching, two reasons why they got rid of it, ultimately. But Blaze made sense because you had to be able to compile things outside of Java, Python and (later) Go where Java and Go in particular had better build systems for purely Java and Go (respectively) code bases. It got better once there were tools for auto-generating Blaze config (ie the java_library build units). Where Blaze was horrible was actually with protobufs. auto-generated code wasn't stored in the repo (unlike Facebook). There were protobuf versions (although, even by 2010, most things were protobuf version 2) but there were also API versions. And they weren't compatible. So Java had API version 1 (mutable) and 2 (immutable) and if you needed to use some team's protobuf but they'd never updated to API v2, you'd either have to make everything v1 or do some horrible hacks or create your own build units for v2. But I digress. Python essesntially got supplanted by Go outside of DS/ML. The code review tool was originally written in Python (ie Mondrian) before being rewritten in (IIRC) GWT (ie Critique). For a very long time Critique lacked features Mondrian had. Personally, I was always sympathetic to avoiding large Python code bases just for the lack of strict typing. You ended up having to write unit tests for spelling mistakes. I can't speak to Tensorflow vs PyTorch. I suspect this institutional disdain for Python was probably a factor in GvR leaving to go join Dropbox. |