dr-environment 0.1.0

DataRobot Execution Environment Builder

Homepage · PyPI · Documentation · Support

Build Codespace-ready execution environment Docker contexts for DataRobot App Framework recipes. Uses a customer-buildable Wolfi base with Python 3.11 and pre-warmed shared uv, npm, and Go dependency caches.

Note Early development. The generated Dockerfile and the CLI flags may change between releases. Released under DataRobot's open-source program; it does not carry an official support SLA. See SECURITY.md.

How this relates to DataRobot custom environments

The built-in environments in DataRobot Registry are the supported default, and the stock [DataRobot] Python 3.11 GenAI Agents environment is the right choice for most projects. This tool covers the air-gapped case: it produces an environment with every dependency cache baked in, so the image runs with no network access. Its output is uploaded through the same custom execution environment flow as any other image.

Prerequisites

• Docker
• uv
• The dr CLI, which hosts this plugin:

curl https://cli.datarobot.com/install | sh
# or: brew install datarobot-oss/taps/dr-cli

Lockfile validation runs each ecosystem's own tool, so a recipe may also need Node.js for a package.json, Go for a go.mod, or Task for an environment hook.

Install

uv tool install dr-environment

Make an execution environment

1. Generate the Docker context from your App Framework recipe:

dr environment recipe --recipe-path datarobot-agent-application

This writes docker_context/ in the current working directory (and docker_context.tar.gz unless you pass --no-tarball).

2. Change into the generated context:

cd docker_context

3. Build the image for linux/amd64 (required for DataRobot notebook kernels):

docker build --platform linux/amd64 -t exec-env .

On Apple Silicon, always pass --platform linux/amd64. An arm64 image fails at runtime with exec format error on start_server.sh.

4. Export the image to a tarball:

docker image save exec-env -o image.tar

5. Upload image.tar as a custom execution environment in DataRobot:

Usage reference

dr environment recipe --recipe-path /path/to/datarobot-agent-application

Defaults:

• Output: docker_context/ in the current working directory (override with --target)
• Archive: docker_context.tar.gz in the current working directory
• Versions: .datarobot/cli/versions.yaml

Options

dr environment recipe --recipe-path . --no-tarball

What it produces

docker_context/
├── Dockerfile                 # assembled from dockerfile.d/*
├── dockerfile.d/              # one fragment per build stage
├── build-deps/                # build-time manifests
├── kernel/                    # kernel deps, entrypoints, Jupyter assets
│   ├── requirements.txt       # kernel-only deps
│   ├── setup-caches.sh        # offline cache paths for login shells
│   ├── start_server_*.sh      # Codespaces and deployed-model entrypoints
│   └── agent/, extensions/    # kernel assets
└── components/<name>/         # per-component manifests

Lockfile policy

Build fails if a component has:

• pyproject.toml without uv.lock (fix: uv lock)
• package.json without package-lock.json (fix: npm install)
• go.mod without go.sum (fix: go mod tidy)
• Stale lockfiles (uv lock --check, npm ci --dry-run, or go mod verify failure)

Runtime cache behavior

The Dockerfile sets tool-native env vars:

• UV_CACHE_DIR=/opt/cache/uv
• NPM_CONFIG_CACHE=/opt/cache/npm
• NPM_CONFIG_PREFER_OFFLINE=true
• GOMODCACHE / GOCACHE

Per-component cache stages run uv sync --frozen --no-install-project against each uv.lock so runtime uv sync --offline can resolve lockfile wheels (including platform-specific builds like litellm).

The Dockerfile always sets strict offline env vars (UV_OFFLINE=1, NPM_CONFIG_OFFLINE=true, GOPROXY=off) in addition to shared cache paths. NOTEBOOKS_AIR_GAP=1 in setup-caches.sh applies the same overrides for login shells.

This is deliberate and always-on, not a toggle for the platform to opt out of. Without UV_OFFLINE=1, a cache miss makes uv fall through to whatever UV_INDEX_URL/mirror is configured in the environment — in an air-gapped install that host can be unreachable, or have no access to a certain library. So instead of a fast, clear "not found in the cache" error, the install hangs until the request times out or fails. Keeping this image strictly offline trades "install anything, sometimes slowly or not at all" for "install only what's baked, fail fast and clearly otherwise." A platform that needs to install extra, non-baked packages from its own reachable mirror should do so outside this image (e.g. in a derived image or a separate step), not by unsetting these vars here.

Component hooks

If a component Taskfile defines an environment task, dr environment recipe runs it with:

Development

uv pip install -e ".[dev]"
pytest
dr-environment --dr-plugin-manifest

See PLUGIN_TESTING.md and CONTRIBUTING.md.

License

Apache 2.0. See LICENSE and NOTICE.

That grant covers this repository's own code. It does not extend to the components the generated image fetches at build time, which keep their own terms, including Wolfi apk packages under GPL-2.0 and GPL-3.0, and the DataRobot Python SDK under the DataRobot Tool and Utility Agreement.