Openlineage

OpenLineage Data Lineage

When to Use

Activate this skill when the task involves:

• Implementing data lineage tracking across Airflow, Spark, dbt, or Trino pipelines
• Setting up the OpenLineage spec emitters and a Marquez or OpenMetadata backend
• Designing column-level lineage for impact analysis
• Writing custom OpenLineage clients or enriching events with facets
• Debugging missing or broken lineage edges in the lineage graph
• Integrating lineage events with data catalogs (DataHub, Atlan, OpenMetadata)

Core Model

┌─────────────────────────────────────────────────────────────┐
│  OpenLineage Model                                          │
│                                                             │
│   Dataset ──────── Job ──────── Dataset                    │
│  (input)          (Run)         (output)                    │
│                                                             │
│   Each entity carries Facets — atomic metadata blocks:     │
│   • Schema facet: column names + types                      │
│   • Column-level lineage: input col → output col mapping   │
│   • Data quality assertions                                 │
│   • SQL query text                                          │
│   • Source code location                                    │
└─────────────────────────────────────────────────────────────┘

Events flow:  Pipeline Tool → OpenLineage Client → HTTP Transport → Backend
              (Airflow, Spark, dbt)                                 (Marquez / DataHub)

Lineage is collected passively — tools emit events without modifying pipeline logic.

RunEvent Specification

The canonical event JSON for an OpenLineage RunEvent:

{
  "eventType": "COMPLETE",
  "eventTime": "2024-03-15T10:30:00.000Z",
  "producer": "https://github.com/OpenLineage/OpenLineage/tree/1.0.0/integration/spark",
  "schemaURL": "https://openlineage.io/spec/1-0-5/OpenLineage.json#/definitions/RunEvent",

  "run": {
    "runId": "d46e465b-d358-4d32-83d4-df660ff614dd",
    "facets": {
      "nominalTime": {
        "_producer": "...",
        "_schemaURL": "...",
        "nominalStartTime": "2024-03-15T10:00:00Z",
        "nominalEndTime":   "2024-03-15T10:30:00Z"
      },
      "parent": {
        "_producer": "...",
        "_schemaURL": "...",
        "run":  {"runId": "e9c85741-93ab-4b27-9c8a-f3f4a3c0e001"},
        "job":  {"namespace": "airflow", "name": "etl_pipeline.run_spark_job"}
      }
    }
  },

  "job": {
    "namespace": "spark://spark-master:7077",
    "name": "silver.transform_orders",
    "facets": {
      "sql": {
        "_producer": "...",
        "_schemaURL": "...",
        "query": "INSERT INTO silver.orders SELECT id, customer_id, total FROM bronze.orders WHERE status = 'valid'"
      },
      "sourceCodeLocation": {
        "_producer": "...",
        "_schemaURL": "...",
        "type": "git",
        "url": "https://github.com/org/repo",
        "repoUrl": "https://github.com/org/repo",
        "path": "jobs/transform_orders.py",
        "version": "abc123"
      }
    }
  },

  "inputs": [
    {
      "namespace": "spark://spark-master:7077",
      "name": "bronze.orders",
      "facets": {
        "schema": {
          "_producer": "...",
          "_schemaURL": "...",
          "fields": [
            {"name": "id",          "type": "BIGINT"},
            {"name": "customer_id", "type": "BIGINT"},
            {"name": "total",       "type": "DECIMAL(10,2)"},
            {"name": "status",      "type": "VARCHAR"}
          ]
        },
        "dataSource": {
          "_producer": "...",
          "_schemaURL": "...",
          "name": "spark_iceberg",
          "uri":  "iceberg://lakehouse/bronze"
        }
      }
    }
  ],

  "outputs": [
    {
      "namespace": "spark://spark-master:7077",
      "name": "silver.orders",
      "facets": {
        "schema": {
          "_producer": "...",
          "_schemaURL": "...",
          "fields": [
            {"name": "id",          "type": "BIGINT"},
            {"name": "customer_id", "type": "BIGINT"},
            {"name": "total",       "type": "DECIMAL(10,2)"}
          ]
        },
        "columnLineage": {
          "_producer": "...",
          "_schemaURL": "...",
          "fields": {
            "id": {
              "inputFields": [
                {"namespace": "spark://spark-master:7077", "name": "bronze.orders", "field": "id"}
              ]
            },
            "customer_id": {
              "inputFields": [
                {"namespace": "spark://spark-master:7077", "name": "bronze.orders", "field": "customer_id"}
              ]
            },
            "total": {
              "inputFields": [
                {"namespace": "spark://spark-master:7077", "name": "bronze.orders", "field": "total"}
              ]
            }
          }
        },
        "outputStatistics": {
          "_producer": "...",
          "_schemaURL": "...",
          "rowCount": 150000,
          "size": 45000000
        }
      }
    }
  ]
}

Run States

Every job must emit at least START + (COMPLETE | FAIL | ABORT).

Facet Reference

Job Facets

Run Facets

Dataset Facets

Marquez Backend

Marquez is the OpenLineage reference implementation — stores and visualizes lineage events.

Docker Compose

version: "3.8"
services:
  marquez-db:
    image: postgres:14
    environment:
      POSTGRES_DB: marquez
      POSTGRES_USER: marquez
      POSTGRES_PASSWORD: marquez
    volumes:
      - marquez-db:/var/lib/postgresql/data

  marquez:
    image: marquezproject/marquez:0.47.0
    environment:
      MARQUEZ_PORT: 5000
      MARQUEZ_ADMIN_PORT: 5001
      MARQUEZ_DB_HOST: marquez-db
      MARQUEZ_DB_PORT: 5432
      MARQUEZ_DB_NAME: marquez
      MARQUEZ_DB_USER: marquez
      MARQUEZ_DB_PASSWORD: marquez
    ports:
      - "5000:5000"   # API
      - "5001:5001"   # Admin
    depends_on:
      - marquez-db

  marquez-web:
    image: marquezproject/marquez-web:0.47.0
    environment:
      MARQUEZ_HOST: marquez
      MARQUEZ_PORT: 5000
    ports:
      - "3000:3000"   # UI
    depends_on:
      - marquez

volumes:
  marquez-db:

Marquez API

# List namespaces
curl http://localhost:5000/api/v1/namespaces | jq

# List jobs in namespace
curl "http://localhost:5000/api/v1/namespaces/airflow/jobs" | jq

# Get dataset with lineage
curl "http://localhost:5000/api/v1/namespaces/spark://master:7077/datasets/silver.orders" | jq

# Get lineage graph (upstream + downstream, depth=2)
curl "http://localhost:5000/api/v1/lineage?nodeId=dataset:spark://master:7077:silver.orders&depth=2" | jq

# Search datasets
curl "http://localhost:5000/api/v1/search?q=orders&type=DATASET" | jq

Airflow Integration

Installation

pip install apache-airflow-providers-openlineage

Configuration (airflow.cfg or environment variables)

[openlineage]
transport = {"type": "http", "url": "http://marquez:5000", "endpoint": "api/v1/lineage"}
namespace = airflow
disabled = false
disabled_for_operators = airflow.operators.empty.EmptyOperator

Or via environment variables:

export OPENLINEAGE_URL=http://marquez:5000
export OPENLINEAGE_NAMESPACE=airflow
export AIRFLOW__OPENLINEAGE__TRANSPORT='{"type": "http", "url": "http://marquez:5000", "endpoint": "api/v1/lineage"}'

What Airflow Captures Automatically

Custom Dataset Lineage in PythonOperator

from openlineage.client.run import Dataset
from openlineage.client.facet import SchemaDatasetFacet, SchemaField
from airflow.providers.openlineage.extractors.base import OperatorLineage

class MyPandasOperator(BaseOperator):
    def get_openlineage_facets_on_complete(self, ti):
        return OperatorLineage(
            inputs=[
                Dataset(
                    namespace="postgres://prod-db:5432",
                    name="public.customers",
                    facets={
                        "schema": SchemaDatasetFacet(
                            fields=[
                                SchemaField("id", "BIGINT"),
                                SchemaField("email", "VARCHAR"),
                            ]
                        )
                    },
                )
            ],
            outputs=[
                Dataset(
                    namespace="s3://data-lake",
                    name="silver/customers",
                )
            ],
        )

Spark Integration

Maven / pip

<!-- pom.xml or build.sbt for JVM Spark -->
<dependency>
  <groupId>io.openlineage</groupId>
  <artifactId>openlineage-spark_2.12</artifactId>
  <version>1.15.0</version>
</dependency>

# Download JAR for spark-submit
wget https://repo1.maven.org/maven2/io/openlineage/openlineage-spark_2.12/1.15.0/openlineage-spark_2.12-1.15.0.jar

spark-submit Configuration

spark-submit \
  --jars /opt/spark/jars/openlineage-spark_2.12-1.15.0.jar \
  --conf spark.extraListeners=io.openlineage.spark.agent.OpenLineageSparkListener \
  --conf spark.openlineage.transport.type=http \
  --conf spark.openlineage.transport.url=http://marquez:5000 \
  --conf spark.openlineage.transport.endpoint=/api/v1/lineage \
  --conf spark.openlineage.namespace=spark://spark-master:7077 \
  --conf spark.openlineage.parentJobNamespace=airflow \
  --conf spark.openlineage.parentJobName=etl_dag.run_spark_transform \
  --conf spark.openlineage.parentRunId=d46e465b-d358-4d32-83d4-df660ff614dd \
  my_etl_job.py

PySpark Programmatic Configuration

from pyspark.sql import SparkSession

spark = SparkSession.builder \
    .appName("silver_transform") \
    .config("spark.extraListeners",
            "io.openlineage.spark.agent.OpenLineageSparkListener") \
    .config("spark.openlineage.transport.type", "http") \
    .config("spark.openlineage.transport.url", "http://marquez:5000") \
    .config("spark.openlineage.transport.endpoint", "/api/v1/lineage") \
    .config("spark.openlineage.namespace", "spark://spark-master:7077") \
    .getOrCreate()

# All subsequent spark.read / spark.write / df.createOrReplaceTempView
# and SQL queries are automatically tracked
df = spark.read.table("bronze.orders")
result = df.filter("status = 'valid'").select("id", "customer_id", "total")
result.writeTo("silver.orders").append()
# → Emits RunEvent with inputs=[bronze.orders], outputs=[silver.orders]

What Spark Captures

dbt Integration

pip install openlineage-dbt

profiles.yml — no changes needed; OL is configured separately

# Run dbt with OpenLineage emission
export OPENLINEAGE_URL=http://marquez:5000
export OPENLINEAGE_NAMESPACE=dbt_prod
dbt run --target prod

# Or via dbt-openlineage package flags
dbt-ol run --target prod

dbt_project.yml metadata captured

# These model-level configs appear in lineage facets
models:
  my_project:
    staging:
      +meta:
        owner: "data-team"
        tags: ["daily", "customers"]

What dbt Captures

Custom Python Emitter

from openlineage.client import OpenLineageClient
from openlineage.client.run import (
    RunEvent, RunState, Run, Job,
    Dataset, InputDataset, OutputDataset,
)
from openlineage.client.facet import (
    SchemaDatasetFacet, SchemaField,
    SqlJobFacet, NominalTimeRunFacet,
    ColumnLineageDatasetFacet, ColumnLineageDatasetFacetFieldsAdditional,
    ColumnLineageDatasetFacetFieldsAdditionalInputFields,
    OutputStatisticsOutputDatasetFacet,
)
from openlineage.client.transport.http import HttpTransport, HttpConfig
import uuid
from datetime import datetime, timezone

client = OpenLineageClient(
    transport=HttpTransport(
        HttpConfig(url="http://marquez:5000", endpoint="api/v1/lineage")
    )
)

run_id = str(uuid.uuid4())
now = datetime.now(timezone.utc).isoformat()

# Emit START
client.emit(RunEvent(
    eventType=RunState.START,
    eventTime=now,
    run=Run(runId=run_id),
    job=Job(namespace="my-pipeline", name="daily_revenue"),
    producer="https://github.com/org/pipeline",
    inputs=[InputDataset(namespace="postgres://db:5432", name="public.orders")],
    outputs=[OutputDataset(namespace="s3://warehouse", name="gold/revenue")],
))

# ... do actual work ...

# Emit COMPLETE with column lineage
client.emit(RunEvent(
    eventType=RunState.COMPLETE,
    eventTime=datetime.now(timezone.utc).isoformat(),
    run=Run(runId=run_id),
    job=Job(
        namespace="my-pipeline",
        name="daily_revenue",
        facets={"sql": SqlJobFacet(query="SELECT order_date, SUM(total) AS revenue FROM orders GROUP BY 1")},
    ),
    producer="https://github.com/org/pipeline",
    inputs=[
        InputDataset(
            namespace="postgres://db:5432",
            name="public.orders",
            facets={
                "schema": SchemaDatasetFacet(fields=[
                    SchemaField("order_date", "DATE"),
                    SchemaField("total", "NUMERIC"),
                ])
            },
        )
    ],
    outputs=[
        OutputDataset(
            namespace="s3://warehouse",
            name="gold/revenue",
            facets={
                "schema": SchemaDatasetFacet(fields=[
                    SchemaField("order_date", "DATE"),
                    SchemaField("revenue", "NUMERIC"),
                ]),
                "columnLineage": ColumnLineageDatasetFacet(
                    fields={
                        "order_date": ColumnLineageDatasetFacetFieldsAdditional(
                            inputFields=[ColumnLineageDatasetFacetFieldsAdditionalInputFields(
                                namespace="postgres://db:5432",
                                name="public.orders",
                                field="order_date",
                            )]
                        ),
                        "revenue": ColumnLineageDatasetFacetFieldsAdditional(
                            inputFields=[ColumnLineageDatasetFacetFieldsAdditionalInputFields(
                                namespace="postgres://db:5432",
                                name="public.orders",
                                field="total",
                                transformationType="AGGREGATE",
                                transformationDescription="SUM",
                            )]
                        ),
                    }
                ),
                "outputStatistics": OutputStatisticsOutputDatasetFacet(
                    rowCount=365,
                    size=4096,
                ),
            },
        )
    ],
))

Namespace Conventions

Consistent namespaces are critical — if Airflow and Spark use different namespace strings for the same dataset, lineage will show two disconnected nodes.

Dataset names follow <schema>.<table> or <database>.<schema>.<table> conventions — match exactly what the SQL engine uses.

Transport Types

# Composite transport example: HTTP + file
from openlineage.client.transport.composite import CompositeTransport, CompositeConfig

transport = CompositeTransport(CompositeConfig(transports=[
    {"type": "http", "url": "http://marquez:5000", "endpoint": "api/v1/lineage"},
    {"type": "file", "log_file_path": "/var/log/openlineage.jsonl"},
]))

Impact Analysis Pattern

Once lineage is in Marquez, use the API to answer "what breaks if I drop column X?":

import requests

def get_downstream_jobs(namespace: str, dataset_name: str, depth: int = 5) -> list[dict]:
    """Return all jobs that read from the given dataset (transitively)."""
    resp = requests.get(
        "http://marquez:5000/api/v1/lineage",
        params={
            "nodeId": f"dataset:{namespace}:{dataset_name}",
            "depth": depth,
        },
    )
    graph = resp.json()
    return [
        node for node in graph["graph"]
        if node["type"] == "JOB" and node["id"] != f"dataset:{namespace}:{dataset_name}"
    ]

# Example: find all downstream jobs of silver.orders
downstream = get_downstream_jobs(
    namespace="spark://spark-master:7077",
    dataset_name="silver.orders",
)
for job in downstream:
    print(job["id"], job["data"]["latestRun"]["state"])

Anti-Patterns

1. Different namespace strings for the same database — Airflow uses postgres://host:5432, Spark uses postgresql — graph shows two unconnected clusters. Standardize namespaces across all tools.

2. Emitting only COMPLETE events without START — many backends require START to create the run record; COMPLETE alone is silently dropped or creates orphan records.

3. Not linking Spark/dbt child runs to parent Airflow run — lineage is fragmented per system. Always pass parentJobNamespace, parentJobName, parentRunId to Spark and dbt.

4. Omitting nominalTime facet on scheduled jobs — makes time-partitioned lineage unusable. Use the Airflow logical date as nominalStartTime.

5. Using file transport in production — JSON files grow unboundedly and are not queryable. Use HTTP transport to Marquez.

6. Relying solely on job-level lineage — table-level lineage is insufficient for column impact analysis. Instrument SQL-based jobs with openlineage-sql to get column lineage.

7. Custom dataset names that diverge from SQL names — e.g., calling a dataset "orders_v2" while SQL uses silver.orders creates phantom nodes. Use the actual schema.table name.

8. Not handling FAIL events — a job that silently dies without emitting FAIL leaves runs in START state forever. Always wrap job logic in try/except and emit FAIL on error.

References to Consult When Needed

• OpenLineage spec: openlineage.io/spec
• Marquez project: marquezproject.ai
• Airflow provider: airflow.apache.org/docs/apache-airflow-providers-openlineage/
• Spark integration: openlineage.io/docs/integrations/spark
• dbt integration: openlineage.io/docs/integrations/dbt
• openlineage-sql (SQL parser): github.com/OpenLineage/OpenLineage/tree/main/integration/sql