Recent Optimizations in Python's Reference Counting
rushter.comThis seems like a pretty big deal that should have gotten more attention in the 3.14 release. It's fairly buried in the "What's New" document, but apparently the original issue is at https://github.com/python/cpython/issues/130704 and corresponding PR https://github.com/python/cpython/pull/130708 .
Has improving CPython performance become a huge focus in just the past five years or so, or is that just a perception issue on my end?
A developer wrote a paper in 2020 about how to make Python substantially (5x) faster.
Microsoft then funded a 'Fast CPython' team that included Guido, to realize that goal. They disbanded the team in June.
https://devblogs.microsoft.com/python/python-311-faster-cpyt...
Because of how much of a cornerstone python generally has been in AI circles, performance improvements have gotten a lot more attention in the past few years.
Is python performance actually relevant to AI use cases? It's my understanding that all of the actual number crunching is done with native code.
Sure, but that doesn't mean there's no ergonomic benefits to the surrounding code being faster. Plus there's just more eyes and attention on the ecosystem.
Guido gave up the idea of keeping CPython interpreter simple. It's now a complex beast with JIT and a lot of small optimizations.