Dev Environment Setup

Bring a checkout from "just cloned" to "can build the backend, the UI, and the ingestion framework" with one call.

When to Activate

• "set up my dev environment", "onboard me", "I just cloned this repo"
• "set up this worktree" (a Claude Code worktree does not inherit the main repo's venv)
• A build fails on a missing or wrong-version tool: antlr4: command not found, class file has wrong version, Unsupported engine … node, ModuleNotFoundError: metadata.generated
• "check my environment", "why doesn't make generate work"

The One Call

./scripts/dev_setup.sh          # or: make dev_setup

Everything is idempotent — re-running skips what is already correct. Useful flags (pass through the Makefile as make dev_setup ARGS="--slim -y"):

Process

Step 1 — Diagnose before changing anything

./scripts/dev_setup.sh --check

Read the warnings. They name the exact missing piece; do not install anything the check reports as already present.

Step 2 — Choose the depth

• Backend/UI work only, or a new worktree → ./scripts/dev_setup.sh --slim -y
• Connector or ingestion work → ./scripts/dev_setup.sh -y (full [all-dev-env])
• Ask the user before running the full install if they are on a metered/slow link — it downloads every connector's dependencies.

Step 3 — Enter the environment

The script writes .dev-env.local.sh (gitignored) at the repo root:

source .dev-env.local.sh

It exports JAVA_HOME, puts ~/.local/bin (the ANTLR CLI) and the right Node on PATH, and activates the venv. Tell the user to source it in each new shell, or add it to their shell rc.

Step 4 — Verify with real commands

Do not claim success without output. Minimum bar:

source .dev-env.local.sh
make prerequisites                              # all ✓
python -c "import metadata.generated.schema.entity.data.table"   # models generated
mvn -q -pl openmetadata-spec install -DskipTests                 # backend toolchain
cd openmetadata-ui/src/main/resources/ui && yarn tsc --noEmit --version  # UI deps

What the script actually does

1. Detects platform → brew / apt / dnf / yum / pacman / zypper.
2. Installs the build toolchain plus the headers the ingestion wheels compile against (libffi, openssl, sasl/gssapi, krb5, libpq, librdkafka, unixodbc, libxml2/xslt).
3. Ensures Java 21, Maven ≥ 3.6, Node 22, Yarn 1.x, Python ≥ 3.10, ANTLR 4.9.2, Docker.
4. Creates env/, then runs the CLAUDE.md bootstrap sequence: make install_dev_env → make generate → make yarn_install_cache → make install_test precommit_install → make prerequisites.
5. Writes .dev-env.local.sh and, when mise is active, a gitignored .mise.local.toml so mise's shell hook does not restore incompatible global Java/Node versions.

Troubleshooting

Maven fails with TypeTag :: UNKNOWN

This usually means Maven is running on a newer JDK even though Java 21 is installed. Run the setup again and source .dev-env.local.sh; the generated environment places $JAVA_HOME/bin first on PATH. When mise is installed, setup also writes .mise.local.toml, preventing mise's prompt hook from immediately restoring a newer global JDK.

make prerequisites fails on macOS with "declare -A is not supported" scripts/check_prerequisites.sh needs bash ≥ 4; macOS ships 3.2. brew install bash (the setup script does this and then runs the check through the brew bash).

java: command not found after installing on macOS Homebrew's openjdk@21 is keg-only — never symlinked onto PATH. Use export JAVA_HOME="$(brew --prefix openjdk@21)/libexec/openjdk.jdk/Contents/Home" and put $JAVA_HOME/bin first. .dev-env.local.sh already does this.

antlr4: command not found, or generated parsers are rejected at runtime The ANTLR CLI must be exactly 4.9.2 — it has to match the pinned antlr4-python3-runtime and the JS runtime. A distro antlr4 of any other version produces parsers those runtimes reject. Fix: make install_antlr_cli ANTLR_INSTALL_DIR="$HOME/.local/bin" (checksum-pinned, no sudo).

make generate fails or silently produces nothing It must run inside the venv (source env/bin/activate) and only from the repo root — it is a root-only target and does not exist under ingestion/. It wipes and rebuilds ingestion/src/metadata/generated, so a partial run leaves an unimportable tree; re-run it rather than hand-patching.

Node is not version 22 Some distro releases provide Node 18/20 while rolling installations may already have Node 24/26. The current UI tree includes i18next-parser@9.4.0, which accepts Node 18, 20, or 22 and rejects newer majors. The script uses mise when it is already installed, otherwise it falls back to a user-local nvm install of Node 22 — neither path needs root. Re-source .dev-env.local.sh afterwards.

Yarn resolves to 3.x/4.x (Berry) This repo is Yarn Classic (engines.yarn: ^1.22.0). npm i -g yarn@1.22.22. A Berry yarn will fail yarn install --frozen-lockfile.

spaCy/Thinc/Blis fails while installing build dependencies on Python 3.13/3.14 The project metadata has no Python ceiling, but the current native dev dependency versions do not publish wheels for those minors and their fallback source build fails. The setup script uses Python 3.11 by default and currently builds its venv only with Python 3.10–3.12. Re-run the setup; it will offer to rebuild an incompatible existing env/ and provision Python 3.11 through uv if needed.

cx_Oracle fails to build It is installed deliberately with --no-build-isolation (see ingestion/Makefile). Use the make targets rather than a bare pip install.

Docker daemon unreachable on Linux sudo systemctl enable --now docker and sudo usermod -aG docker $USER, then log out and back in — group membership is not picked up by the current shell.

Claude Code worktree has no venv Worktrees do not copy env/. Either run ./scripts/dev_setup.sh --slim inside the worktree, or symlink the main repo's: ln -s /path/to/main-repo/env env.

--skip-tools and still failing The user may not have sudo. Report exactly which packages are missing and the one-line install command for their manager; do not attempt privileged installs they did not approve.

Guardrails

• Never run the script with sudo. It escalates only for the specific package-manager calls that need it.
• Do not edit .dev-env.local.sh by hand — it is regenerated on every run.
• Do not pip install ingestion dependencies directly; use the make targets so the pins (setuptools<81, cx_Oracle build isolation) are honored.
• If a step fails, surface the real command output. Do not report "environment ready" from a run that produced warnings.