uv pip compile perf regression in 0.2.6

[blueraft]

There's a significant performance regression in uv 0.2.6 for projects that rely on setuptools_scm.

MRE

1. Initialize a blank git repo.
2. Copy the following pyproject.toml.

[build-system]
requires = ["setuptools<69.3.0", "setuptools-scm"]
build-backend = "setuptools.build_meta"

[project]
name = 'foobar'
dynamic = ["version"]
requires-python = ">=3.9"

dependencies = ["flask"]

[tool.setuptools_scm]

Benchmarks from an Ubuntu machine (I am able to reproduce this on MacOS too):

1. uv 0.2.6

❯ uv --version
uv 0.2.6
❯ hyperfine 'uv pip compile pyproject.toml'
Benchmark 1: uv pip compile pyproject.toml
  Time (mean ± σ):     604.1 ms ±   5.2 ms    [User: 515.1 ms, System: 96.1 ms]
  Range (min … max):   596.4 ms … 611.8 ms    10 runs

2. uv 0.2.5

❯ uv --version
uv 0.2.5
❯ hyperfine 'uv pip compile pyproject.toml'
Benchmark 1: uv pip compile pyproject.toml
  Time (mean ± σ):      58.2 ms ±   1.3 ms    [User: 41.6 ms, System: 23.7 ms]
  Range (min … max):    55.1 ms …  61.0 ms    50 runs

[konstin]

The cause of this regression is we unified some code paths and now require a static version field to read a pyproject.toml statically, instead we're now building the package to get the metadata including the version.

[charliermarsh]

We can fix this though it's sort of a pain.

[blueraft]

Thank you!