Repository navigation
Reuse code objects for similar dataclass definitions #100930
Description
Activity
I’ve been looking at dataclasses and this interests me. What would need to happen to take this over the finish-line?
For starters, an analysis of memory usage and cache hit ratio for common dataclass usage patterns.
Thanks! I’m assuming this would take some legwork, as pyperformance lacks a dataclass benchmark (if that issue is still relevant)?
I admit to not being super knowledgeable on the subject, but I like digging into topics. I’ll report back when I have more to contribute.
Closing in favor of #109870, which is the current direction
dataclassesseems to be heading in.From my understanding, this could be complementary with the combined exec rework.
Combined exec speeds up initial creation, this could further speed up creation of subsequent similar dataclasses.
Maybe once that has landed, we could rework this if there is interest.
Reacted by Brandt Bucher
A little over a year ago, @dabeaz came up with a cool way of speeding up
dataclasscreation by avoiding unnecessaryexeccalls. Essentially, his proof-of-conceptdataklassesmodule caches code objects for methods of "similarly-shaped" dataclasses, and patches them with the correct names:https://github.com/dabeaz/dataklasses
I have a working prototype of a similar idea for the stdlib
dataclassesmodule over in #92650. It basically doubles the speed ofdataclassdefinitions.CC @ericvsmith
Linked PRs
dataclasscode caching #92650