Managing Data Lineage with OpenLineage

When to use

• Setting up automated lineage across Airflow, dbt, and Spark.
• Answering "what breaks if I change this table?" (impact analysis).
• Scoping an incident's blast radius or a backfill's downstream footprint.
• Do NOT use for in-project dbt lineage only (dbt docs cover that; use this for cross-tool, cross-system lineage).

Core model

OpenLineage describes each pipeline run as Job (the transformation), Run (one execution), and Datasets (inputs/outputs), optionally with column-level mappings. Emitters send these events to a backend (e.g. Marquez) that builds the graph.

Workflow

- [ ] Add the OpenLineage integration to each producer (Airflow/dbt/Spark)
- [ ] Point emitters at a collector (Marquez or vendor endpoint)
- [ ] Verify jobs/datasets/edges appear with column-level detail where supported
- [ ] Use the graph for impact analysis before changes
- [ ] Use it during incidents to scope downstream impact and backfills

1. Instrument producers. Airflow (OpenLineage provider), dbt (dbt-ol/native), and Spark (OpenLineage listener) emit events automatically — prefer these over hand-rolled lineage.
2. Centralize. Send events to Marquez or a lineage platform so the whole graph is queryable in one place.
3. Validate coverage — confirm datasets, edges, and column-level lineage render for critical flows.
4. Impact analysis — before changing/deprecating a table, traverse downstream to find affected models, dashboards, and owners.
5. Incident scoping — from a bad source dataset, list every downstream consumer to size the fix and the backfill.

Patterns

Airflow emission — enable the OpenLineage provider and set the transport endpoint via env/config; DAG runs then emit job/dataset events without code changes.

Use lineage to scope a backfill — the set of downstream datasets reachable from the corrected table is exactly the set that must be reprocessed (pairs with designing-backfills-and-replays).

Column-level lineage — when a single column is wrong, column mappings pinpoint which downstream columns depend on it, narrowing the investigation.

Common pitfalls

• Manual lineage docs — drift immediately; use automated emitters.
• Partial instrumentation — gaps make the graph misleading; instrument every producer in a critical path.
• Lineage without ownership — knowing what breaks is only useful if you can find who to tell; attach owners to datasets.
• Ignoring column-level lineage where available — table-level alone forces broad, slow investigations.
• Treating lineage as a diagram, not a tool — wire it into change review and incident response, not just a wiki page.