🔄 Airflow DQ Pipeline

Production-grade data pipeline using Write-Audit-Publish pattern — the same approach used at Netflix, Airbnb, and Spotify.

🎯 The Problem

Most pipelines write directly to production. If bad data slips through, stakeholders see it before you can fix it.

✅ The Solution

Never write to production directly. Instead:

Sensor → Staging → DQ Checks → Production → Cleanup

📊 Pipeline Architecture

Task	What It Does	Why It Matters
wait_for_polygon_tickers	Waits for upstream data	Don't run on empty tables
fetch_to_staging	Loads to staging table	Isolate unvalidated data
run_dq_checks	4 quality validations	Catch issues early
exchange_to_production	Atomic partition swap	Zero-downtime publish
cleanup_staging	Drops temp table	No orphan data

🛡️ Data Quality Checks

✓ Row count > 0           # Table not empty
✓ No NULL close prices    # Required field present  
✓ All prices > 0          # Business logic validation
✓ All volumes >= 0        # No negative volumes

If ANY check fails → pipeline stops → bad data never reaches production.

🔧 Tech Stack

• Orchestration: Apache Airflow
• Data Lake: PyIceberg + AWS Glue Catalog
• Storage: S3 (Iceberg format)
• Pattern: Write-Audit-Publish (idempotent, backfillable)

🧠 Key Concepts Demonstrated

• ✅ Idempotency (same input = same output)
• ✅ Partition-scoped overwrites (preserve history)
• ✅ Sensor-based dependencies
• ✅ Staging table isolation
• ✅ Automated data quality gates

📚 Credit

Pattern learned from Zach Wilson — ex-Netflix, ex-Airbnb Data Engineer — via DataExpert.io bootcamp.

🔗 Related Projects

• airflow-soda-integration - Soda CLI + Airflow for DQ checks
• LuBot.ai - AI analytics platform with 17 nightly DQ workers

---

⭐ If this helped you, star the repo!