Pip install Postgres – no Docker/Brew/apt
github.com
github.com
As that comment mentions:
> 5 years ago [...] Python packaging was a big mess.
> Today [...] I think Python's packaging story has never been stronger.
Poetry, & later Uv, are essential & invaluable milestones in that revolutionary story; Pip is terrible software & should be left in the past.
The comparison to other programs is missing the point of why pip is terrible. For that matter, last I checked Poetry relied on either vendored pip or something very much like it to do the actual package installation.
If you like workflows dictated by all-in-one tools, go for it. Don't force it on me. Some of us still see great value in the Unix philosophy.
The real problem is that pip doesn't even do a good job of its "one thing" (the actual package installation). It barely understands caching, has insecure and confusing default behaviours stemming from nigh-unfixable legacy issues, doesn't know how to inspect the installation environment properly (therefore relying on ugly hacks to install cross-environment at all), vendors a bunch of stuff it barely uses and often doesn't really need, and is overall horribly architected.
You've summarised many of pip's internally problems well - there's also the lack of any attempt at proper manifest management. This is partially an ecosystem problem (pre-PEP518), but the tools inform the specs as much as the specs inform the tools.
I think there's also a reasonable argument to be made that uv only does one thing, & that the big problems in this space before was that python required using 3 tools to awkwardly collaborate to get one thing done. That one thing is dependency management: pip to pull them, venv to load them, pyenv to pull & load one of them (the interpreter).
The other thing I don't like is having tens of megs of opaque compiled binary (that doesn't provide a useful library) when properly designed Python would perform more than acceptably for common use cases (tons of this is I/O-bound; a proper cache lets you just recursively hardlink a directory tree instead of unpacking an archive file; the bytecode compilation is implemented in C and could be parallelized with `multiprocess` as long as you don't require binary-level reproducibility for Python bytecode; etc.)