data-pipeline-builder

Overview

Designs and implements reliable ETL or ELT data pipelines. Covers the decision between ETL vs ELT, pipeline stages (extract, validate, transform, load, notify), idempotency patterns, dead-letter queues, orchestration with Apache Airflow or dbt, error handling and retries, and monitoring/ alerting for data freshness and quality.

When to Use This Skill

• Ingesting data from APIs, databases, files, or streams into a warehouse or another system.
• Building or refactoring data movement processes.
• The user mentions "data pipeline", "ETL", "ELT", "ingest", "Airflow DAG", or "dbt model".

Prerequisites

• Source and destination systems identified.
• Orchestrator (Airflow, Dagster, Prefect, or dbt + cron) or simple cron + scripts.
• Access to credentials for sources and destinations.
• Understanding of data volume and freshness requirements.

Steps

1. ETL vs ELT decision:
   • ETL: transform before load (smaller data, complex transforms, regulatory reasons).
   • ELT: load raw, transform in warehouse (modern, scalable, dbt-friendly).
2. Pipeline stages (template):
   • Extract: pull from source (API pagination, DB query, file download).
   • Validate: schema, nulls, duplicates, business rules (great expectations or custom).
   • Transform: clean, join, aggregate, enrich.
   • Load: idempotent upsert or append.
   • Notify / quality checks: freshness, row counts, anomaly detection.
3. Idempotency:
   • Use natural keys + upsert (MERGE, INSERT ... ON CONFLICT).
   • Watermark / high-water mark for incremental loads.
   • Deduplication keys.
4. Error handling & retries:
   • Dead-letter queue or "quarantine" table for bad rows.
   • Retry with exponential backoff + jitter (for transient API/DB errors).
   • Alert on persistent failures.
5. Orchestration:
   • Airflow DAG structure (tasks, dependencies, sensors, SLAs).
   • dbt: models, tests, snapshots, macros for incremental logic.
6. Monitoring:
   • Data freshness (last successful run time).
   • Volume anomalies.
   • Schema drift detection.
7. Output:
   • Pipeline architecture diagram (Mermaid or text).
   • Sample code for extract/validate/load (Python or SQL).
   • Airflow DAG example or dbt project structure.
   • Dead-letter and retry patterns.
   • Alerting rules.

Examples

A complete incremental ELT pipeline from a REST API (e.g., Stripe or Shopify) into BigQuery or Postgres using Python + dbt, with idempotent load, data quality tests, and an Airflow DAG wrapper is included.

Edge Cases & Error Handling

• Late-arriving data: Use watermark + reprocessing window.
• Schema changes in source: Schema evolution strategy (add columns as nullable, alert on breaking changes).
• Large backfills: Separate backfill jobs with different resource allocation.

Verification

1. Run the pipeline end-to-end on a small dataset.
2. Simulate a transient failure — retries succeed.
3. Simulate bad data — it goes to dead-letter, pipeline continues, alert fires.
4. Check that re-running the pipeline with the same source data produces identical results (idempotent).
5. Data freshness dashboard shows the expected last run time.
6. Success: Data arrives reliably, errors are isolated, and the pipeline can be re-run safely.

References

• Apache Airflow
• dbt
• Great Expectations
• The Data Engineering Cookbook