airflow-floe 0.1.5

Airflow connector for Floe CLI (manifest-first orchestration)

Floe + Airflow (MVP example)

This folder contains:

• a reusable Python connector package (src/airflow_floe)
• example DAGs (dags/) that consume the package

For local setup of both Dagster and Airflow with isolated virtual environments, see:

• orchestrators/LOCAL_DEV.md

Contents

• INTEGRATION_SPEC.md: contract between Airflow and Floe CLI
• schemas/: JSON Schemas for XCom payloads and manifest contract
• src/airflow_floe/manifest.py: manifest loader (floe.manifest.v1)
• src/airflow_floe/runtime.py: shared runtime helpers (manifest context, run event parsing, summary loading)
• src/airflow_floe/hooks.py: reusable manifest hook (FloeManifestHook)
• src/airflow_floe/operators.py: reusable run hook/operator (FloeRunHook, FloeRunOperator)
• example/config.yml: small Floe config for demo
• dags/floe_example_operator_dag.py: example DAG using FloeRunOperator

Configuration

The connector is driven entirely by environment variables. No code changes are needed to switch between single-manifest, multi-manifest, or fallback mode.

Env var	What it does
FLOE_MANIFEST	Primary: path to a floe.manifest.v1 JSON — enables asset registration at parse time
FLOE_MANIFEST_DIR	Multi-manifest mode: directory of *.manifest.json files; one DAG per file
FLOE_CONFIG	Fallback run config used when manifest loading fails — DAG still executes floe run, but no assets are registered
FLOE_CMD	Override the floe binary path (default: floe on $PATH)

Manifest loading failure: If FLOE_MANIFEST points to a path that cannot be loaded (file not found, invalid JSON, schema error), build_dag_manifest_context_or_empty catches the exception and returns an empty asset context — the DAG still runs floe run using FLOE_CONFIG as the config path, but no assets are registered or materialized.

Note: Setting only FLOE_CONFIG without a valid FLOE_MANIFEST does not guarantee "no assets". When FLOE_MANIFEST is unset, the example DAG defaults to the bundled example manifest (example/manifest.airflow.json), which registers its own assets. To run with no assets, either point FLOE_MANIFEST to a path that does not exist, or remove the default fallback from your DAG code.

Multi-manifest mode: Each JSON file in FLOE_MANIFEST_DIR becomes a separate DAG. The DAG ID is derived from the manifest file name — for example orders.manifest.json → DAG floe_orders. Entity tasks within the DAG are generated from the manifest's entities[] list.

Quick usage

1. Make sure floe is available in PATH (or set FLOE_CMD).

2. (Optional but recommended) install the connector package:

   python -m pip install -e orchestrators/airflow-floe
   

3. Point Airflow DAGs folder to orchestrators/airflow-floe/dags.

4. Set env vars:

   export FLOE_CMD="floe"
   export FLOE_MANIFEST="/absolute/path/to/orchestrators/airflow-floe/example/manifest.airflow.json"
   # optional multi-manifest mode (1 manifest => 1 DAG):
   # export FLOE_MANIFEST_DIR="/absolute/path/to/orchestrators/airflow-floe/example/manifests"
   # optional fallback (no assets registered):
   # export FLOE_CONFIG="/absolute/path/to/orchestrators/airflow-floe/example/config.yml"
   

5. Generate manifest from Floe config:

   floe manifest generate \
     -c orchestrators/airflow-floe/example/config.yml \
     --output orchestrators/airflow-floe/example/manifest.airflow.json
   

   For multi-manifest mode (one DAG per manifest), generate domain manifests:

   floe manifest generate \
     -c orchestrators/airflow-floe/example/config.hr.yml \
     --output orchestrators/airflow-floe/example/manifests/hr.manifest.json
   
   floe manifest generate \
     -c orchestrators/airflow-floe/example/config.sales.yml \
     --output orchestrators/airflow-floe/example/manifests/sales.manifest.json
   

6. Trigger DAG floe_example_operator. In multi-manifest mode (FLOE_MANIFEST_DIR), DAGs are generated as floe_<manifest_name>.

Notes

• The DAG uses Floe CLI contracts directly and expects:
  • run log schema: floe.log.v1
  • terminal event: run_finished
• Assets are created at parse time from FLOE_MANIFEST and materialized when run tasks finish.
• floe_example_operator also publishes asset events when manifest assets are available.
• The returned task payload shape follows floe.airflow.run.v1.
• Local run summaries emitted as local://... are resolved and loaded by the connector runtime helpers.
• Floe NDJSON stdout/stderr are streamed into task logs (Audit/Task Log view) and still parsed for run_finished.