// the find
dabeaz/dataklasses
A different spin on dataclasses.
A proof-of-concept by David Beazley that generates dataclass-like dunder methods but caches the bytecode rather than re-running exec() per class, making imports 15-20x faster. It's less than 100 lines and intentionally ships zero features beyond __init__, __repr__, __eq__, and __match_args__. This is a teaching tool and minimalist alternative, not a stdlib replacement.
The bytecode caching insight is genuinely clever — Python functions with structurally identical bytecode share the same code object regardless of argument names, so you only exec() once per class shape. The performance claim is plausible and benchmarked in perf.py. At under 100 lines, it's trivially auditable and safe to vendor. The README is refreshingly honest about what it is.
No installation path, no package on PyPI — the intended distribution model is copy-paste, which means you own maintenance forever. No support for defaults, default_factory, field metadata, frozen classes, post_init, or inheritance — anything beyond the basics requires you to write it yourself. The last commit was January 2022, and 'you maintain it' is the literal documented support policy. Works fine as a sketch of an idea; genuinely rough as a production dependency.