Make it easier to deploy a mirror for uv-python command

[MeitarR]

The problem

Currently, in order to use the feature of uv to fetch Python versions (uv python install), in internal and air-gapped networks,
one needs to

1. download all the needed binaries (using https://github.com/astral-sh/uv/blob/main/scripts/create-python-mirror.py )
2. then get the resulting folder (14~ GB for a single uv version) to the target network - depending on the organization it may be a complicated stage, that is very hard (or even impossible) to automate
3. and serve it to everyone else in the target network.

a process that may take hours.

The big issue is that the paths of the CPython binaries are hardcoded in the binary and change almost every release of uv. Thus, the one who maintains the mirror must go through that process frequently, or else the feature of uv python install will break when updating the uv version locally.

(see https://github.com/astral-sh/uv/commits/main/crates/uv-python/download-metadata.json for the frequency)

Suggested solution

instead of hardcoded URLs, fetch the json externally.

I think that one way of solving this issue is, instead of hardcoding the list of URLs in the binary, you can fetch the list (json) https://github.com/astral-sh/uv/blob/main/crates/uv-python/download-metadata.json from external URL (probably from the repo), letting you update the available python binaries versions without the need to update uv, and let the users change the URL the binary uses to fetch the JSON (https://github.com/astral-sh/uv/blob/main/crates/uv-python/download-metadata.json), so it will work for every version of uv, and if you update your list, we can download and update it in the target network later without consequences.

[MeitarR]

@charliermarsh is this PR welcome? (And do you agree with the suggested solution?)

[fcwys]

I also hope to use this method. Currently, when configuring an image using python install mirror, due to the python-build-standalone version being updated to 20250317, the image site automatically synchronizes to the latest version. However, when installing Python using UV, the download still fails due to the 20250212 version.

UV Version：uv 0.6.5 (bcbcd0a 2025-03-06)

[Crypto-Spartan]

I have the same issue as @fcwys. Internal mirror is slow to get the updated python-build-standalone binaries and uv fails to download the latest tag. I work in an ephemeral environment in a corporate network, so i have an install script that reinstalls everything everyday. When uv breaks like this, I straight up can't use it at all.

I would love it if I could tell uv to use an older tag when downloading the python-build-standalone binaries, as sometimes we have the binaries from the previous week, but since uv is so picky about which tag it's downloading, it won't work.

[MeitarR]

After #10939 merged, this is the flow to setup and use the mirror in your network

Setup

on the internet-accessible-computer

1. Clone this repo
2. Download the Python binaries by running the create-python-mirror script with the wanted parameters
3. Transfer the mirror folder and the matching crates/uv-python/download-metadata.json file to your network

on your network

1. Put the content of the mirror folder on some server
2. Filter the download-metadata.json file only to have the versions you downloaded (I recommend doing it with jq command)
3. In the json, replace references to https://github.com/astral-sh/python-build-standalone/releases/download with the URL of the server you put the mirror in step 1
4. (suggestion) Also put the final download-metadata.json in the same server

Usage for users

1. fetch the modified download-metadata.json and save it in a known location (I recommend /etc/uv/download-metadata.json)
2. Set the environment variable UV_PYTHON_DOWNLOADS_JSON_URL to the location you saved the json in step 1 (recommand to add it to your .bashrc/.zshrc )
3. Use uv python as you would before

[Gankra]

I'm going to tenatively mark this issue as resolved since you can do the above process, but we'll keep an eye out for whether this can/should be more ergonomic/automatic.