How I Migrated 5 Python Projects to UV in One Hour and Cut CI Time by 5×

Why pip feels slow

pip is a Python program, so each invocation incurs a ~300 ms interpreter start‑up cost; this overhead dominates tiny operations like version checks. pip also runs under the GIL, making download, unpack, and install phases strictly serial. Finally, pip copies every package into each virtual environment, duplicating data across projects.

UV’s three architectural accelerators

UV is a single 50 MB Rust binary that eliminates interpreter start‑up, uses Tokio async runtime with reqwest for true multi‑core parallelism, and stores a single copy of each package in ~/.cache/uv linked into environments via hard links, turning installation into a millisecond‑scale file‑system pointer operation.

PubGrub – UV’s dependency‑resolution brain

UV replaces pip’s naive back‑tracking search with the PubGrub algorithm (implemented in pubgrub‑rs). PubGrub learns conflict‑driven clauses (CDCL): when a version conflict is discovered, it records a rule that forbids the offending combination and never retries it. This reduces the search space dramatically; for example, resolving Apache Airflow’s tree takes 0.21 s with UV versus 13.89 s with pip‑tools (≈65× faster).

UV also adds a Forking Resolver to handle platform‑specific markers, splitting the resolution graph into separate branches for different platforms and writing both into a universal uv.lock file.

Step‑by‑step migration

# macOS/Linux
curl -LsSf https://astral.sh/uv/install.sh | sh
# or via pip (once)
pip install uv
uv --version  # e.g. uv 0.11.14

Initialize a project and import existing requirements:

# In project root
uv init
uv add -r requirements.txt  # reads requirements.txt, writes pyproject.toml and uv.lock
uv sync  # creates virtual env, installs all deps, caches globally

Running the project becomes a single command:

uv run python main.py

Real‑world benchmark

Installing the Django full stack drops from 45 s (cold pip cache) to 3 s (cold UV cache) and from 12 s (warm pip cache) to 0.5 s (warm UV cache). A FastAPI project with 57 dependencies goes from 45 s (pip) to 8 s (UV cold) and 0.8 s (UV warm). CI time for a five‑project suite fell from 9 min to 4.5 min.

CI configuration swap

Replace the standard actions/setup-python step with astral-sh/setup-uv@v6 and change pip install -r requirements.txt to uv sync. The Action automatically caches ~/.cache/uv, keeping CI sync under 5 s.

Cross‑platform lockfile

uv.lock

records platform markers and resolution‑markers, allowing the same lockfile to be used on macOS and Linux without manual marker edits. This mirrors Cargo’s Cargo.lock approach and eliminates the “pip freeze” snapshot problem.

Strategic relevance

In March 2026 OpenAI acquired Astral (the UV maintainer) to embed a fast, deterministic Python environment manager into Codex agents. Millisecond‑scale uv add + uv run enables AI agents to install and execute code without waiting for pip’s 45‑second installs, marking a shift from code‑generation quality to tooling control in the AI‑coding race.

Takeaway

UV is not a fleeting hype tool; it consolidates six separate Python utilities into one binary, delivers 10‑100× speed gains, provides a universal lockfile, and is now backed by OpenAI’s AI‑agent strategy. Learning UV equips developers to focus on higher‑level Python work rather than environment friction.