Tool uv monorepo

UV Monorepo Development with Workspaces

Complete guide for building and managing Python monorepos using uv's workspace functionality.

Overview

uv's workspace feature enables true monorepo architecture for Python projects, solving the historical challenge of managing multiple packages in a single repository. Inspired by Rust's Cargo, workspaces provide:

• Unified dependency resolution: Single uv.lock file for entire repository
• Version consistency: Eliminates version drift across packages
• Automatic editable installs: Code changes propagate instantly
• Fast resolution: Rust-powered solver 10-100x faster than pip
• Standard compliance: Built on PEP 621, PEP 735

When to use uv workspaces:

• ✅ Microservices sharing common libraries
• ✅ Multi-package applications (CLI + core + services)
• ✅ Internal library ecosystems
• ✅ Projects requiring strict dependency consistency

Key resources:

• Official docs: https://docs.astral.sh/uv/concepts/projects/workspaces/
• Projects guide: https://docs.astral.sh/uv/

---

Core Concepts

1. Unified Lock File

The most powerful feature: a single uv.lock at repository root that:

• Resolves all workspace members' dependencies into one conflict-free graph
• Eliminates "version drift" where independent projects use different versions
• Forces all components to use identical versions of shared dependencies (e.g., pydantic, fastapi)

When you run uv lock, the system evaluates the entire workspace and creates a mathematically consistent dependency solution.

2. Python Version Constraints

Workspaces enforce a single requires-python for the entire repository, calculated as the intersection of all members' requirements:

Package	requires-python	Workspace Result
root	>=3.10	-
service-a	>=3.11	>=3.11
service-b	>=3.12	>=3.12 (strictest wins)

This ensures all members can coexist in the shared virtual environment (.venv).

3. Workspace vs Path Dependencies

Two approaches for managing related packages:

Feature	Workspace	Path Dependencies
Lock file	Single uv.lock for all	Separate per project
Python version	Unified (intersection)	Independent per project
Virtual env	Single shared .venv	Individual .venv per project
Consistency	Enforced	Flexible
Best for	Tightly coupled services	Highly independent projects

---

Directory Structure

Basic Layout

my-monorepo/
├── pyproject.toml       # Workspace root config
├── uv.lock              # Unified lock file
├── .venv/               # Shared virtual environment
└── packages/
    ├── core/
    │   ├── pyproject.toml
    │   └── src/core/
    ├── api/
    │   ├── pyproject.toml
    │   └── src/api/
    └── cli/
        ├── pyproject.toml
        └── src/cli/

Root Configuration

pyproject.toml (workspace root):

[project]
name = "my-monorepo-workspace"  # Must be unique from members!
version = "0.1.0"
requires-python = ">=3.10"

[tool.uv.workspace]
members = ["packages/*"]        # Glob patterns
# exclude = ["packages/legacy/*"]  # Optional exclusions

[tool.uv]
package = false                 # Virtual root (not installable)

[dependency-groups]
dev = [
    "pytest>=7.4",
    "ruff>=0.1",
    "mypy>=1.0",
]

Critical: Name Collision

The workspace root name must be unique from all member names. If both root and a member use my-app, uv sync fails with:

Error: Duplicate workspace member: my-app

Use descriptive names: my-app-workspace for root, my-app for actual package.

---

Package Dependencies

Declaring Internal Dependencies

To make api depend on core, use two-step declaration:

packages/api/pyproject.toml:

[project]
name = "api"
dependencies = [
    "core",  # Standard PEP 621 declaration
]

[tool.uv.sources]
core = { workspace = true }  # uv-specific: resolve from workspace

Why two declarations?

• [project.dependencies]: Standard metadata, readable by all tools
• [tool.uv.sources]: Routing table for uv-specific resolution

If [tool.uv.sources] is missing, uv provides helpful error:

Error: Package 'core' is a workspace member but missing sources entry

Automatic Editable Install

With workspace = true, uv automatically installs members in editable mode. Code changes in core are instantly reflected in api without reinstallation.

---

Development Dependencies

PEP 735 Dependency Groups

For dev tools (pytest, ruff, mypy), use dependency groups in root:

[dependency-groups]
dev = [
    "pytest>=7.4",
    "pytest-cov>=4.1",
    "ruff>=0.1",
    "mypy>=1.0",
]

Benefits:

• Not included in build artifacts (wheels, sdists)
• Shared across all packages: uv sync --group dev
• Prevents version inconsistencies (e.g., different linters per service)

Install:

# Install all dependencies + dev group
uv sync --group dev

# Install only production dependencies
uv sync

---

Testing with Pytest

Import File Mismatch Problem

In monorepos with multiple tests/ directories, pytest defaults cause errors:

import file mismatch:
  imported module 'test_helpers' has this __file__: .../packages/cli/tests/test_helpers.py
  which is not the same as: .../packages/core/tests/test_helpers.py

Cause: Pytest's prepend mode inserts test directories into sys.path, causing name collisions.

Solution: Importlib Mode

Root pyproject.toml:

[tool.pytest.ini_options]
addopts = ["--import-mode=importlib"]

This uses Python's importlib to import tests directly without modifying sys.path, eliminating name collisions.

Important: Do NOT add __init__.py to test directories when using importlib mode. This causes silent test skipping (tests are cached under one path but collected from another).

---

Docker Optimization

Problem: Monolithic Dependencies

A single uv.lock contains dependencies for ALL services. Naive Docker builds include unnecessary dependencies, inflating image size and attack surface.

Solution: Multi-Stage Build with Export

Step 1: Extract Service-Specific Dependencies

FROM ghcr.io/astral-sh/uv:python3.12-alpine AS builder

WORKDIR /app
COPY uv.lock pyproject.toml ./
COPY packages/ ./packages/

# Extract only dependencies for 'api' service
RUN uv export --frozen --directory packages/api -o requirements.txt && \
    # Remove workspace member references (e.g., -e ./packages/core)
    sed -i '/^-e/d' requirements.txt

Step 2: Install External Dependencies (Cached Layer)

RUN --mount=type=cache,target=/root/.cache/uv \
    uv pip sync requirements.txt --no-cache --compile-bytecode

Step 3: Copy Internal Libraries and Application Code

# Copy shared libraries first (changes less frequently)
COPY packages/core/ ./packages/core/

# Copy application code last (changes most frequently)
COPY packages/api/ ./packages/api/

# Install as editable
RUN uv pip install -e ./packages/api/

Benefits:

• External dependencies cached in Docker layer
• Application code changes don't invalidate dependency layer
• --compile-bytecode pre-generates .pyc files for faster startup

Layer Caching Strategy

Stage	Command	Cache Behavior
Base	Copy uv binary	Rarely changes
Dependencies	uv export + uv pip sync	Only invalidated when uv.lock changes
Shared libs	Copy packages/core/	Invalidated when core changes
Application	Copy packages/api/	Invalidated on every app code change

---

CI/CD Best Practices

Global Cache Strategy (GitHub Actions)

Problem: Per-PR caches waste storage and slow down CI.

Solution: Single cache from main branch, read-only for PRs.

Workflow 1: Main Branch (Cache Write)

name: Build Cache
on:
  push:
    branches: [main]
  schedule:
    - cron: '0 0 * * 0'  # Weekly refresh

jobs:
  cache:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: astral-sh/setup-uv@v5
        with:
          enable-cache: false  # Manual cache control

      - name: Sync dependencies
        run: uv sync --all-groups

      - name: Save cache
        uses: actions/cache/save@v4
        with:
          path: |
            ~/.cache/uv
            .venv
          key: uv-${{ runner.os }}-${{ hashFiles('uv.lock') }}

Workflow 2: PR Branches (Cache Read-Only)

name: Test
on: pull_request

jobs:
  test:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - name: Restore cache
        uses: actions/cache/restore@v4
        with:
          path: |
            ~/.cache/uv
            .venv
          key: uv-${{ runner.os }}-${{ hashFiles('uv.lock') }}
          restore-keys: |
            uv-${{ runner.os }}-

      - uses: astral-sh/setup-uv@v5
      - run: uv sync  # Uses cache, downloads only diff
      - run: uv run pytest

Benefits:

• PRs never pollute cache
• Always clean state from main
• Minimal storage usage

---

Common Commands

Workspace Management

# Initialize new workspace
uv init --package my-monorepo
cd my-monorepo

# Add workspace member
uv init --lib packages/core
uv init --lib packages/api

# Configure workspace in root pyproject.toml
# Add: [tool.uv.workspace] members = ["packages/*"]

# Sync all dependencies
uv sync

# Sync with dev tools
uv sync --group dev

# Run command in workspace context
uv run python -c "import core; import api"

# Run from specific package directory
cd packages/api
uv run uvicorn main:app

Building & Publishing

# Build specific package
uv build --package api

# Build all packages
uv build --all-packages

# Check version
uv version --package core

# Publish to PyPI
uv publish --package core

---

Migration from Poly-repo

Pre-Migration: Version Alignment

Critical: Before creating workspace, align all dependency versions across services.

1. Update all services to latest compatible versions:

# In each service
uv sync --upgrade
uv run pytest  # Verify no breakage

1. Resolve version conflicts:

If Service A uses Django 4.2 and Service B uses Django 5.1, choose one version and update all services.

1. Create workspace structure:

mkdir -p packages
mv service-a packages/
mv service-b packages/

1. Configure root workspace:

Create root pyproject.toml with workspace configuration.

1. Convert path dependencies:

Replace relative path dependencies with workspace = true.

Expected outcome: uv lock will surface any hidden version conflicts that were masked by isolated environments.

---

Troubleshooting

Error: Duplicate workspace member

Cause: Root and member have same name.

Solution: Rename root to project-workspace.

Error: Missing sources entry

Cause: Package in [project.dependencies] but not in [tool.uv.sources].

Solution: Add { workspace = true } to sources.

Pytest import file mismatch

Cause: Default prepend mode + duplicate test file names.

Solution: Add --import-mode=importlib to pytest config.

CUDA/GPU dependencies in monorepo

Solution: Use platform-specific members or path dependencies for GPU-specific code.

---

Best Practices

1. Name virtual roots descriptively: my-app-workspace not my-app
2. Use dependency groups: Keep dev tools in [dependency-groups]
3. Importlib mode always: Set --import-mode=importlib in pytest config
4. Docker multi-stage: Use uv export to extract service-specific deps
5. Cache from main: GitHub Actions cache strategy from main branch only
6. Align before migration: Update all versions before creating workspace

---

Summary

uv workspaces provide:

Core Benefits:

• 10-100x faster dependency resolution (Rust-powered)
• Zero version drift through unified lock file
• Instant code propagation with automatic editable installs
• Standard compliance (PEP 621, PEP 735)
• Simplified CI/CD through single environment

Key Features:

• Single uv.lock for entire repository
• Unified Python version constraint (intersection)
• Shared virtual environment (.venv)
• Automatic workspace member discovery
• Built-in Docker optimization patterns

Resources:

• Official docs: https://docs.astral.sh/uv/concepts/projects/workspaces/
• Practical guide: https://dev.to/aws/3-things-i-wish-i-knew-before-setting-up-a-uv-workspace-30j6
• Advanced CI/CD: https://www.reddit.com/r/Python/comments/1iy4h5k/cracking_the_python_monorepo_build_pipelines_with/

uv workspaces eliminate Python's historical monorepo pain points, bringing Cargo-like simplicity and performance to multi-package projects.