Why I Replaced Poetry with uv for almost everything

I used to be that person who would smugly paste a poetry install command into every Dockerfile, CI, etc. I even wrote a blog post last year praising Poetry—it still felt like the obvious default—until one night, after skimming a few posts about uv and waiting on slow builds after every tiny tweak, I installed it out of curiosity. The build that had been taking two minutes finished in fourteen seconds. 🤯

That single data point made me pause. Over the next few weeks I migrated a side project, then a few repository, and eventually every new workflow on my laptop and I started to apply. Today uv is the first tool I install on a fresh dev box—right after git and zsh.

⚙️ What exactly is uv?

Consider uv your all-in-one solution for Python dependency and environment management:

Why does it feel faster? The team rewrote the entire dependency solver in Rust and removed a lot of legacy edge‑cases that only matter to 0.1 % of packages. I rarely dive into the implementation details of python dependency tools, but seeing uv sync finish before the coffee machine warms up still feels like cheating.

TL;DR — uv takes the everyday pain points of Python tooling and solves them with a single binary that happens to be lightning‑fast.

🚀 Five‑minute starter guide

I’ve tested these steps on macOS, Ubuntu, and a Windows. They’re all you need to get productive:

  # 1 — install
  curl -LsSf https://astral.sh/uv/install.sh | sh   # macOS / Linux
  # iwr https://astral.sh/uv/install.ps1 | iex      # Windows PowerShell

  # 2 — create a fresh project
  uv init ml-exp && cd ml-exp

  # 3 — add your libraries
  uv add scikit-learn xgboost optuna matplotlib pandas

  # 4 — lock & sync for reproducibility
  uv lock && uv sync

  # 5 — run code inside the managed venv
  uv run python train.py

🧪 Where uv earned its keep in my MLOps/DevOps life

1. CI pipelines — Switching from Poetry’s dependency install to uv sync can reduce each test job by ~6 minutes, adding up to roughly one full engineer-day of CI time saved over 300 runs a week.

2. Dockerized experiments — Swapping poetry install for uv sync inside Dockerfiles shaved minutes off image builds—suddenly, iterating on model containers felt as fast as editing notebooks.

3. Colleague laptops — New hires clone the repo, run uv sync, and are ready to code before they’ve even signed in to their email.

🐍 Multi‑python made boring

Even with lightweight projects, hopping between toolchains is friction-free now:

  # one–time interpreter install
  uv python install 3.10 3.11 3.12

  # slim envs per stack
  uv venv pipeline-prod   --python 3.10   # MLflow + Kedro
  uv venv rag-lab         --python 3.11   # LangChain + ChromaDB
  uv venv orchestrator    --python 3.12   # Prefect + FastAPI

Need to lock a repo? uv python pin 3.11 drops a .python-version file so your IDE, pre-commit, and CI all agree, no more surprise interpreter conflicts between projects.

💡 Pro tips & common issues

⚠️ Minor snags worth knowing:

• SSH-based Git dependencies — Resolving a private GitHub package over SSH failed silently until I manually exported:

    export GIT_SSH_COMMAND='ssh -i ~/.ssh/id_ed25519'
  

  After that, uv handled it fine, but the initial error message gave no hint it was SSH-related.

• No editable installs (-e) yet — uv doesn’t support pip install -e ., which is a blocker if you rely on live reload during local development. You’ll need to work around it with manual symlinks or fall back to pip for now.

• Slight CLI surprises — If you run uv pip install ... outside a uv-managed venv, it installs into your global interpreter, unlike pip, which often warns or errors. It’s easy to forget you’re not in a managed shell, so double-check where you’re installing.

⚔️ Head‑to‑head speed numbers

uv = quickest result in each benchmark; bold values mark where uv beats every other tool by a wide margin.

What each scenario means

• Warm Installation – Re-installing with wheels already cached locally.
• Cold Installation – First-time install on a clean cache.
• Warm Resolution – Updating/locking deps after a previous successful run.
• Cold Resolution – Resolving a lock file on a brand-new machine with no cached metadata.

Where to find the raw benchmark

Check out their GitHub benchmarks here for more details.

✅ When to reach for uv, and when not to

uv is a no‑brainer for:

• Heterogeneous data‑science repos where setup time kills flow.
• CI pipelines that pay per minute.
• Teams who secretly hate reading internal Confluence pages about pyenv hacks.

Hold off if:

• You package a public library and rely on Poetry’s build command (for now; UV build is on their roadmap).
• You need editable installs (pip install -e) with custom PEP 660 backends. support is experimental.

🔗 Further reading & watching

• The concise official docs are surprisingly good.
• Astral’s launch post explains the design philosophy.

Final thoughts

Every few years the Python ecosystem gets a ‘why didn’t we have this earlier?’ moment—black, then ruff, and now uv. Swapping it in took a lazy Saturday afternoon and eliminated half a dozen lines from my Makefile. If your workflow still juggles pip, venv, pyenv, and a dash of pipx, give uv an hour. Chances are you’ll save many more.