# ML Engine

Job-based evaluation engine (Python 3.13). Polls the Arthur Platform for jobs, loads data through source connectors (S3, GCS, BigQuery, Snowflake, ODBC databases), and computes metrics. Flow: `job_agent.py` (polling) → `job_runner.py` (thread/process runners) → `job_executor.py` → `job_executors/*`.

Second mode: with `ML_ENGINE_DISCOVERY_CONFIG` set, `job_agent.main` runs standalone agent discovery instead — no Platform, no GenAI Engine. Flow: `standalone/discovery_config.py` (file) → `standalone/agent.py` (interval scheduler) → `standalone/scan.py` → the shared `run_source_scan` → `standalone/sinks/*` (Splunk HEC, webhook). See [docs/standalone-discovery.md](docs/standalone-discovery.md).

## Commands

```bash
# The GenAI Engine client is generated, not committed — required before uv sync
cd scripts && ./openapi_client_utils.sh generate python && ./openapi_client_utils.sh install python && cd ..
./scripts/install_db_dependencies.sh     # ODBC / Oracle drivers

uv sync --group dev --group linters
uv run python src/ml_engine/job_agent.py
uv run pytest tests/unit
uv run mypy src/ml_engine
./scripts/lint.sh
```

Needs `ARTHUR_API_HOST`, `ARTHUR_CLIENT_ID`, `ARTHUR_CLIENT_SECRET` to talk to the platform.

## Gotchas

- Regenerate the GenAI Engine client whenever the GenAI Engine API changes.
- Job executors should be stateless and idempotent where possible.
- Health check endpoints must respond quickly — slow responses cause container restarts.
- Discovery connectors are listed in `discovery.source_connectors()` and must declare `SENSITIVE_FIELDS`, or standalone mode refuses them; sink targets are listed in `standalone.sinks.Destination`, one module each.
- A connector with no Platform vendor goes in `discovery.standalone_connectors()` instead, so Platform executors never see it. `local` is one: it reads the endpoint collector's `inventory.json` from a mounted folder, for Deckard's quickstart, and depends on the collector still publishing that file under that name.