I share your concern but I have saved so much time with uv already that I figure ill ride it till the VC enshitification kills the host.
Hopefully at the point the community is centralized enough to move in one direction.
Hopefully at the point the community is centralized enough to move in one direction.
But the increased resource usage is real. Now around 10% of our builds get OOM killed because the build container isn't provisioned big enough to handle uv's excessive memory usage. I've considered reducing the number of available threads to try throttle the non-deterministic allocation behavior, but that would presumably make it slower too, so instead we just click the re-run job button. Even with that manual intervention 10% of the time, it is so much faster than pip it's worth it.
(I work on uv)
We run on-prem k8s and do the pip install stage in a 2CPU/4GB Gitlab runner, which feels like it should be sufficient for the uv:python3.12-bookworm image. We have about 100 deps that aside from numpy/pandas/pyarrow are pretty lightweight. No GPU stuff. I tried 2CPU/8GB runners but it still OOMed occasionally so didn't seem worth using up those resources for the normal case. I don't know enough about the uv internals to understand why it's so expensive, but it feels counter-intuitive because the whole venv is "only" around 500MB.
Did you try reducing the concurrency limit?