weird behaviour of uv sync and the uv cache

[kforner]

Summary

When I preinstall a package using uv pip install, the cache seems to be populated but uv sync does not use it.
Here's a Dockerfile:

FROM ubuntu:24.04
COPY --from=ghcr.io/astral-sh/uv:0.9.8 /uv /uvx /bin/
USER ubuntu
RUN uv python install 3.14
WORKDIR /home/ubuntu

RUN uv venv && uv pip install pandas && rm -rf .venv
COPY --chown=ubuntu:ubuntu py_dev/project_only/pyproject.toml .

RUN ls  -ls ~/.cache/uv/wheels-v5/pypi/

RUN rm  -rf .venv && uv sync --link-mode copy -v 2>&1 | grep pandas | grep  Identified
RUN rm  -rf .venv && uv sync --link-mode copy -v 2>&1 | grep pandas 

Here's the pyproject.toml:

[project]
name = "dummy"
version = "0.1.0"
requires-python = ">=3.14"
dependencies = [ 
    "pandas"
]

and here's the interesting part of the docker build output:

#16 [stage-0 10/12] RUN ls  -ls ~/.cache/uv/wheels-v5/pypi/
#16 0.157 total 24
#16 0.157 4 drwxr-xr-x 2 ubuntu ubuntu 4096 Nov 20 10:28 numpy
#16 0.157 4 drwxr-xr-x 2 ubuntu ubuntu 4096 Nov 20 10:28 pandas
#16 0.157 4 drwxr-xr-x 2 ubuntu ubuntu 4096 Nov 20 10:28 python-dateutil
#16 0.157 4 drwxr-xr-x 2 ubuntu ubuntu 4096 Nov 20 10:28 pytz
#16 0.157 4 drwxr-xr-x 2 ubuntu ubuntu 4096 Nov 20 10:28 six
#16 0.157 4 drwxr-xr-x 2 ubuntu ubuntu 4096 Nov 20 10:28 tzdata
#16 DONE 0.2s

#17 [stage-0 11/12] RUN rm  -rf .venv && uv sync --link-mode copy -v 2>&1 | grep pandas | grep  Identified
#17 18.57 DEBUG Identified uncached distribution: pandas==2.3.3
#17 DONE 18.8s

#18 [stage-0 12/12] RUN rm  -rf .venv && uv sync --link-mode copy -v 2>&1 | grep pandas
#18 0.416 DEBUG Registry requirement already cached: pandas==2.3.3
#18 0.416  + pandas==2.3.3
#18 DONE 0.6s

• the first uv sync execution states that DEBUG Identified uncached distribution: pandas==2.3.3
• the second run: DEBUG Registry requirement already cached: pandas==2.3.3
  Since pandas was already installed using uv pip install, and the cache populated, I can't understand what's happening.

Platform

docker on linux

Version

uv 0.9.8

Python version

No response

[kforner]

Note that substituting uv pip install pandas by uv add pandas fixes that behaviour.

[zanieb]

How do you know they're resolving to the same version of pandas? What version does uv pip install pandas install?

[konstin]

I can reproduce this with any package, e.g. sniffio. When using torch it become notable that we don't share the cache between uv pip install and uv sync, but we do share it inside the interface. That shouldn't happen, cause they are the same file and should go through the same preparation and install code.

[konstin]

MRE (works with sniffio too, but pandas or torch shows that it's not just a logging error):

uv cache clean && uv venv -c && uv pip install nvidia-nccl-cu12 && rm -rf .venv # Fresh
uv venv -c && uv pip install nvidia-nccl-cu12 && rm -rf .venv # Cached
rm -rf .venv && rm -f pyproject.toml && uv init && uv add nvidia-nccl-cu12 # Fresh, but it should be cached
rm -rf .venv && rm -f pyproject.toml && uv init && uv add nvidia-nccl-cu12 # Cached

[charliermarsh]

I think this is because in uv pip, we don't compute hashes by default, but uv sync needs them to validate -- so we end up re-downloading.

[charliermarsh]

In other words, I think this works if you add --verify-hashes to uv pip install.

[kforner]

Thanks, I confirm that --verify-hashes fixes it.
Is there a reason it is not enabled by default?