Should uv.find_uv_bin() be able to find /usr/bin/uv?

[musicinmybrain]

I have almost finished preparing a uv package for Fedora. As written, the uv.find_uv_bin() Python function would not be able to find the executable /usr/bin/uv that this package would provide:

• First it checks

  uv/python/uv/__init__.py

  Line 13 in 2288ff7

  path = os.path.join(sysconfig.get_path("scripts"), uv_exe)

  which evaluates to /usr/local/bin/uv.
• Then it checks

  uv/python/uv/__init__.py

  Line 26 in 2288ff7

  path = os.path.join(sysconfig.get_path("scripts", scheme=user_scheme), uv_exe)

  where user_scheme is "posix_user", which evaluates to '/home/myusername/.local/bin/uv'.
• Then it checks

  uv/python/uv/__init__.py

  Lines 31 to 32 in 2288ff7

  	pkg_root = os.path.dirname(os.path.dirname(__file__))
  	target_path = os.path.join(pkg_root, "bin", uv_exe)

  which evaluates to something like /usr/lib/python3.13/bin/uv or /usr/lib64/python3.13/bin/uv. (I currently get the latter, but since the uv Python library is actually arch-independent when it doesn’t bundle the uv executable, I should probably ensure it’s the former.)

---

I should be able to fix this by something like:

diff --git a/python/uv/__init__.py b/python/uv/__init__.py
index 781eee4f..f2603785 100644
--- a/python/uv/__init__.py
+++ b/python/uv/__init__.py
@@ -10,6 +10,12 @@ def find_uv_bin() -> str:
 
     uv_exe = "uv" + sysconfig.get_config_var("EXE")
 
+    bindir_path = sysconfig.get_config_var("BINDIR")
+    if bindir_path is not None:
+        path = os.path.join(bindir_path, uv_exe)
+        if os.path.isfile(path):
+            return path
+
     path = os.path.join(sysconfig.get_path("scripts"), uv_exe)
     if os.path.isfile(path):
         return path

or, probably better,

diff --git a/python/uv/__init__.py b/python/uv/__init__.py
index 781eee4f..80953b2e 100644
--- a/python/uv/__init__.py
+++ b/python/uv/__init__.py
@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import os
+import shutil
 import sys
 import sysconfig
 
@@ -10,6 +11,10 @@ def find_uv_bin() -> str:
 
     uv_exe = "uv" + sysconfig.get_config_var("EXE")
 
+    path = shutil.which(uv_exe)
+    if path is not None:
+        return path
+
     path = os.path.join(sysconfig.get_path("scripts"), uv_exe)
     if os.path.isfile(path):
         return path

Is this a change that you think belongs in uv upstream, or is this something I should carry as a downstream patch?

[charliermarsh]

Is that binary path exposed anywhere on sysconfig or distutils? I think the goal of this method is to find the Python that's installed in the current environment, so which feels a little too permissive.

[charliermarsh]

As a separate question: when you install via the package manager, are you installing the Python sources or just the uv binary? (Sort of asking: does this even need to work?)

[musicinmybrain]

From the uv source RPM, I plan to build a binary RPM (what people actually install) also called uv, which will install /usr/bin/uv, plus ancillary files that should ship with the executable: tab-completion support, basic documentation like README.md, license files, etc..

Then I plan to have another binary RPM, built from the same source RPM, called python3-uv. This provides the importable Python uv package, i.e. https://github.com/astral-sh/uv/tree/main/python/uv, along with the .dist-info metadata, and which has an exact-version dependency on the uv binary RPM built from the same source RPM. It can therefore rely on “referencing” the system-wide /usr/bin/uv. For RPM package users, this should work as if they had installed the package from PyPI. We do need this to work, because system packages like hatch will be depending on uv via a Python dependency. This system python3-uv should always find the system /usr/bin/uv rather than something else in the user’s path, which, come to think of it, is an argument against shutil.which and in favor of sysconfig.get_config_var("BINDIR") for our use case. (For similar reasons, we always adjust #!/usr/bin/env python shebangs in RPM-installed Python scripts to #!/usr/bin/python3 to ensure they run with the system Python interpreter.)

If you think that finding /usr/bin/uv is a need that is specific to this kind of distribution packaging, then it’s easy enough to make the adaptation downstream only. However (unless I’m missing something), it does seem a little strange to support finding uv in /usr/local/bin but not in /usr/bin.

[musicinmybrain]

However (unless I’m missing something), it does seem a little strange to support finding uv in /usr/local/bin but not in /usr/bin.

Looking again at the actual code, I suppose that while

uv/python/uv/__init__.py

Line 13 in 2288ff7

path = os.path.join(sysconfig.get_path("scripts"), uv_exe)

evaluates to /usr/local/bin/uv in a system-wide installation like this, that’s not the situation you are trying to cover with that case. It’s really intended for /path/to/virtualenv/bin/uv. And furthermore, in the virtualenv case, you presumably don’t want to find /usr/bin/uv (or/usr/local/bin/uv) first.

So maybe this really is a situation where a downstream-only patch is the right approach to ensure the system Python uv library finds the system uv executable. However, if you think there’s anything to be improved here in the general case, I’m happy to offer a PR.

[charliermarsh]

Yeah, I think it's right for this to be a downstream-only patch. Sorry if it causes you any inconvenience but that's where I'm ending up too. (Thanks for walking me through it.)

[musicinmybrain]

Yeah, I think it's right for this to be a downstream-only patch. Sorry if it causes you any inconvenience but that's where I'm ending up too. (Thanks for walking me through it.)

Thanks for talking through it with me. You helped me clarify my reasoning, and I’m now much more confident that the first proposed downstream patch from #4451 (comment) is a good approach.