Databricks event reviewer

Directory examples/python/features/databricks_events

This incident-review node receives Databricks notifications through an emitter and calls the separate Databricks integration node. Responsibilities are split as follows:

Capabilities

Capability	Kind	Signature	Module	Summary
`review_databricks_event`	traversal	`(event: Dict[str, Any] = {}, event_id: str = '', event_type: str = '', payload: Dict[str, Any] = {}, databricks: Dict[str, Any] = {}, trigger: TriggerContext \| None = None) -> dict`	`examples/python/features/databricks_events/agent.py`	Review one Databricks notification (server-normalized fields, or a raw payload in ``event``): fetch the run and a metric summary, decide a verdict.
`summarize_metric_change`	traversal	`(metric_name: str, time_window: str) -> dict`	`examples/python/features/databricks_events/agent.py`	Direct-call example: plan an investigation through the Databricks node; model review only with a configured endpoint.

Commands

Taken from this example's own README and module docstrings.

python -m atfgraph.worker --module examples.python.features.databricks_events.agent:app
python -m atfgraph.worker --module integrations.databricks.node.agent:app &

Files

examples/python/features/databricks_events/README.md
examples/python/features/databricks_events/__init__.py
examples/python/features/databricks_events/agent.py
examples/python/features/databricks_events/agentnode.yaml

README

Rendered from examples/python/features/databricks_events/README.md; the source file is not modified.

This incident-review node receives Databricks notifications through an emitter and calls the separate Databricks integration node. Responsibilities are split as follows:

The graph server verifies the webhook and admits work.
The databricks node owns Jobs, SQL and AI Functions calls.
incident_reviewer composes those capabilities and returns a rule-based verdict: incident, watch, no_action or needs_review.

The optional summary calls databricks.ai_query only when DATABRICKS_AI_ENDPOINT is configured. Otherwise it reports review_skipped, without inventing a model answer.

Run (fixture mode, no workspace)

Start a daemon with DATABRICKS_WEBHOOK_SECRET in its environment, then run the fixture nodes from the repository root:

DATABRICKS_FIXTURE_DIR=integrations/databricks/fixtures DATABRICKS_WAREHOUSE_ID=wh-fixture \
  python -m atfgraph.worker --module integrations.databricks.node.agent:app &
python -m atfgraph.worker --module examples.python.features.databricks_events.agent:app

Find the reviewer's /hooks/{emitter_id} URL in its log and send a job-notification payload:

curl -X POST "$URL" -H "Authorization: Bearer $DATABRICKS_WEBHOOK_SECRET" \
  -H 'content-type: application/json' \
  --data @integrations/databricks/fixtures/jobs_on_failure.json

For this fixture, inspect verdict: incident, failed task load, the state message and metric rows. The path should show two calls to the Databricks node.

Emitter source

The example defaults DATABRICKS_TRIGGER_SOURCE to generic_bearer, requiring the sender's matching Bearer secret. The reviewer normalizes its raw body. A configured databricks source uses Basic verification, source normalization, run/event deduplication and workspace/job filters. Verify that the source is enabled in your server before selecting it; the default bridge does not demonstrate that provider-specific route.

tests/test_databricks_integration.py uses a real daemon with fixture Databricks services. It does not verify live workspace notifications.