`uv add`/`uv sync`/... do not respect the active virtualenv, should they?

[akx]

trafficstars

Related to #1495, #5229, but not exactly the same IMO.

Like #1495, I prefer to have my virtualenvs located centrally instead of in project directories. I have a Fish shell function called v <NAME> that activates a virtualenv ~/envs/<NAME> if available and cds to the related project directory ~/build/<NAME> if available.

I wanted to try out uv add and uv sync and was initially somewhat confused when they didn't seem to do anything, only to realize that they currently always work on the ~/.venv virtualenv, and don't even stop to consider a currently active virtualenv to work on.

There's apparently a workaround to symlink .venv to the centrally-located venv, but that also seems a bit messy too.

I think it should be at least an (envvar or config-file) option to have the "project venv" tools (run, add, remove, sync, etc.) consider the currently active virtualenv, if any.

Pretty minimal Docker container repro:

~ $ docker run -it python:3.12 bash
root@7ab1d7a6054d:/# pip install uv
Successfully installed uv-0.3.3
root@7ab1d7a6054d:/# mkdir ~/envs
root@7ab1d7a6054d:/# uv venv ~/envs/myenv
Using Python 3.12.4 interpreter at: usr/local/bin/python3
Creating virtualenv at: root/envs/myenv
root@7ab1d7a6054d:/# source ~/envs/myenv/bin/activate
(myenv) root@7ab1d7a6054d:/# echo $VIRTUAL_ENV
/root/envs/myenv
(myenv) root@7ab1d7a6054d:/# mkdir -p ~/projects/myproject
(myenv) root@7ab1d7a6054d:/# cd ~/projects/myproject
(myenv) root@7ab1d7a6054d:~/projects/myproject# uv init
Initialized project `myproject`
(myenv) root@7ab1d7a6054d:~/projects/myproject# uv add ruff
Using Python 3.12.4 interpreter at: /usr/local/bin/python3
Creating virtualenv at: .venv
Resolved 2 packages in 275ms
   Built myproject @ file:///root/projects/myproject
Prepared 2 packages in 983ms
Installed 2 packages in 1ms
 + myproject==0.1.0 (from file:///root/projects/myproject)
 + ruff==0.6.2
(myenv) root@7ab1d7a6054d:~/projects/myproject#
(myenv) root@7ab1d7a6054d:~/projects/myproject# uv pip list
(myenv) root@7ab1d7a6054d:~/projects/myproject#

[MithicSpirit]

I'm a bit of a configuration junkie, so my immediate opinion is "yes—under a configuration option". If a configuration option is deemed too complex for this, I think it'd still be nice to make it the default behavior to support the "external venv" workflow, though a message should be printed if the active venv seems to be outside the current project. If the current project has a venv, but the active venv is different, I'm not sure which should be used by uv; maybe it should just result in an error.

[zanieb]

At the very least, I think we should warn if there's an active virtual environment that isn't the target.

We're very hesitant to sync to active virtual environments. I'll be considering this as we hear more use cases.

[ramarnat]

This is confusing to me, If I do:

uv venv /opt/venv && \
uv sync

Shouldnt the uv sync respect that I created the venv in the first step, it does not seem to do so. What is the use case for uv venv, just uv pip?

[zanieb]

Yeah, uv venv isn't really for use with the project interface. It's for manual management of virtual environments, which is generally the purview of uv pip.

[ramarnat]

here's our use case that is tripping me up. We have a docker image that gets built and we use an activated venv in there. During development/testing, this docker image is also used and we mount the local development directory ./app:/app

I dont want the dev/test to use the local ./app/.venv because that was built for osx, and it should use the activated venv.

For now I have worked around it by setting up the build to use a link from app/.venv to /opt/venv and it seems to work.

[zanieb]

Yep! We agree the Docker volume mount is a big problem and we are working on designing a recommended solution.

Does adding .venv to .dockerignore not work?

[ramarnat]

.dockerignore does not impact a mount. So for dev we are doing -v ./app:/app when running the docker image created with the activated venv.

[zanieb]

Yep, just struggled with that myself.

Have you tried docker compose watch? Is that an option for you?

[zanieb]

Also, have you tried --volume .:/app --volume /app/.venv? This appears to be working for me.

[ramarnat]

ah, watch is something is new, and did not know about. I'll try both options.

[zanieb]

Thanks! Let me know if you have any problems — I'm writing some examples up with those.

[inoa-jboliveira]

Hi, I was going to add a similar issue here. I think more than respecting current active virtualenv, they should also respect the VIRTUAL_ENV environment variable.

I would prefer if we could select the virtualenv folder in uv.toml or pyproject.toml

For historical reasons, on some projects the name of the env folder is relevant (and .venv is also an arbitrary name)

[ramarnat]

Also, have you tried --volume .:/app --volume /app/.venv? This appears to be working for me.

that one did not work for me for one of the cases I am trying, because then I cant do uv sync to add the dev dependencies, without it trying to add everything again, and not use what's already there in /opt/venv. If I didnt have to add the development deps, then I think that would work.

the watch with this stanza seems to be working. Have to go back and change the flow that calls docker directly to use docker compose to see if this will round out the existing cases to use uv instead of pip

    develop:
      watch:
        - action: sync
          path: ./app
          target: /app
          ignore:
            - .venv/  

[zanieb]

that one did not work for me for one of the cases I am trying, because then I cant do uv sync to add the dev dependencies, without it trying to add everything again, and not use what's already there in /opt/venv. If I didnt have to add the development deps, then I think that would work.

Can you share a concrete example? I'm not quite following and am happy to debug.

[ramarnat]

assuming mydockerimage was created in two steps, first a build stage with /opt/venv as the activated venv, with /app/.venv symlinked to /opt/env and uses uv sync --locked -no-dev to build out the venv. Then in the second stage, looks like this:

FROM python:3.12.3-slim AS runtime

ARG APP_VERSION=default
# Make sure we use the virtualenv:
ENV PATH="/opt/venv/bin:$PATH"
ENV APP_VERSION=${APP_VERSION}
RUN echo "Building image on version: ${APP_VERSION}"

# needed if used make test-docker etc.
ENV VIRTUAL_ENV="/opt/venv"
COPY --from=build /bin/uv /bin/uv

WORKDIR /app
COPY --from=build /opt/venv /opt/venv
COPY . /app

EXPOSE 8000

CMD ["python", "-m", "app.main"]

Then if I try the docker run:

docker run  -v .:/app -v /app/.venv -w /app -it mydockerimage bash -c "uv sync --locked"

which if I am understanding your suggestion correctly, it will override /app/.venv with an empty directory, because /app/.venv does not exist locally, It fails with:

warning: Ignoring existing virtual environment linked to non-existent Python interpreter: .venv/bin/python3
Using Python 3.12.3 interpreter at: /usr/local/bin/python3
error: failed to remove directory `/app/.venv`
  Caused by: Resource busy (os error 16)

[zanieb]

Sorry I think I'll need to see a complete minimal example. Perhaps you could derive a small multi-stage image as you describe from this (work in progress) example: https://github.com/astral-sh/uv-docker-example

[ramarnat]

ok, let me adapt that example in a fork, and push it in a few

[ramarnat]

ok, I see the bit that I was missing. I need to copy the /app folder from the build over to the runtime stage, and that will bring the /app/.venv volume over for it to be mounted. Let me play around with this a bit more.

[JanMurmann]

This would be very helpful for us, too.

Our use case is that we run tox in our pipeline and the first step is to verify and install the requirements for the unit tests. So basically we would like to do uv sync --extra test --locked with the current tox virtualenv being the target of the sync.

[akx]

We're very hesitant to sync to active virtual environments.

I kind of understand (though not exactly fully).

I think I would be pretty happy if the sync/add/... commands accepted e.g. --allow-active-environment (or more broadly, I guess, --use-environment=default/active..?) which could be set via envvar (UV_ALLOW_ACTIVE_ENVIRONMENT=1, UV_USE_ENVIRONMENT=active?) or other system-wide config, so I could set that in my usual dev environment.

(I suppose --use-environment=active should also imply requiring an active virtualenv, and the tools wouldn't then create the .venv default venv.)

[FBen3]

We're very hesitant to sync to active virtual environments.

Can you explain the reasoning behind this? I thought most users activate a virtual environment to ensure they're not modifying anything globally.

similar to @akx I keep my virtual environments outside of the repo to save space. Then have an alias to activate whichever environment I need or want to test. It would be nice to have an --active-env flag that can be used with sync/add/... commands, or set some config, to allow for this.

[DanCardin]

It seems kind of wild to not respect VIRTUAL_ENV, for any command honestly. I'm not sure I can imagine a drawback. in the happy path, no one is setting it, or otherwise activating the virtualenv. but in every other case, it's the path of least surprise to me.

I frequently use quick swapping of virutalenvs under different names to fast test across multiple versions of python. I personally do so with my own tool (currently, would love to stop using it) prp -n p39; pytest, prp -n p310; pytest, where my venvs live somewhere else entirely....Granted with uv, uv sync --python 3.9 takes all of 1s to swap, it does unnecessarily kill and recreate the venv each time, and output 2 pages of text.

Somewhat tangentially related, I submitted https://github.com/astral-sh/uv/issues/1495 a while back basically to facilitate this workflow. Where ideally i wouldn't even necessarily be activating any of the venvs, if i could be doing something like uv -p 3.9 sync, uv -p 3.9 run pytest(then swap to -p 3.10); and if under the hood it was just reading/writing to different venv paths external to the project folder (ideally based on pwd and the python version, since ./.venv clobbers itself, and any other local path starts creating clutter). Then in that world, I might no longer have a need or reason to activate a venv in the first place.

[charliermarsh]

... if under the hood it was just reading/writing to different venv paths external to the project folder (ideally based on pwd and the python version, since ./.venv clobbers itself, and any other local path starts creating clutter). Then in that world, I might no longer have a need or reason to activate a venv in the first place.

Just a brief drive-by comment: this is something we've considered but cut from scope for last week's release :)

[blast-hardcheese]

👋 Big +1 for the principle of least surprise when it comes to not just using the active VIRTUAL_ENV.

I'm sympathetic towards not wanting to try and heal a broken virtualenv (if indeed this is the motivation), and there are at least a few places in the documentation where I've seen "unexpected behavior" if uv ends up executing a different python binary than the one living in the virtualenv for one reason or another.

As a workaround, I've discovered that .venv being a symlink does not seem to hinder operation, so prp -n p39 can be replaced with ln (for example, or in my case attempting to adapt .venv to .pythonlibs which is the standard venv location on Replit)

Considering venvs are already roughly ephemeral, even if uv completely wrecks an existing virtualenv I think this'll strengthen the ecosystem in the long run due to increased expectation of interoperability.

(New user here, I'd like to take the opportunity to express my gratitude both to the team as well as all contributors for the fantastic tool 🙂)

[jamesharris-garmin]

I'm really interested in how this effects the tox/nox workflow since, prior to this release I have been using nox for most of my projects/script management.

Do we have any suggestions on how to handle similar workflows with the project/workspace interface?

[JonathanRayner]

I would prefer it if e.g. uv sync would respect VIRTUAL_ENV.

It was the most surprising thing I found when starting to use uv - I commonly have active virtual environments that I want to be obeyed by uv commands. I think the fact that the environment is active is good indication that the user intends to use it and if they don't, that's a user error.

It also is a surprising difference from uv pip install, which does respect VIRTUAL_ENV. I think it would be great if uv sync had similar/consistent behavior.

[inoa-jboliveira]

I don't know if anyone got my suggestion above, but it would be great if there was a config in pyproject.toml to refer to the virtual env path besides the VIRTUAL_ENV variable (of course an active env would take priority). This makes it possible to run uv sync/add as intended without activating the env manually.

[tool.uv]
virtual_env = '.my_venv'

or

[tool.uv]
virtual_env = '/usr/local/my_global_env'

I'm very eager to use this new add/sync interface! Thank you guys for such cool project

[DanCardin]

If that was going to be a feature, uv-specific (i.e. not project-specific, i.e. not pyproject.toml) config would make a lot more sense to me. The virtualenv usage/location is definitely in the realm of developer-specific concern imo, not project-specific concern.

[inoa-jboliveira]

Maybe you are thinking about libraries, but it is common for a project meant to run directly to have several strong opinions on where things are located.

According to the latest 0.4.0 release notes, it seems to be a goal of uv to cater to theses projects.

Direct quotes from the maintainer:

Most users are not developing libraries that need to be packaged and published to PyPI. Instead, they're building applications using web frameworks, or running collections of Python scripts in the project's root directory.

This release adds first-class support for Python projects that are not designed as Python packages (e.g., web applications, data science projects, etc.).

[DanCardin]

i mean, it's ultimately up to them, but it seems weird to me for the project (versus myself or the tool) to have control over where my venv goes, regardless of the usecase. it totally makes sense for uv to have strong opinions or for the user. but i just really dont see a reason for the project itself to care (particularly with a tool like uv, where you can write your project tooling to use uv run or whatever and expect it to do the right thing).