Data Pipeline Architecture Patterns

Pipeline Architectures

ETL vs ELT (CRITICAL)

• ETL: Extract → Transform (staging) → Load
  • Traditional, on-premise data warehouses
  • Pre-aggregation, complex transformations
• ELT: Extract → Load (raw) → Transform (in warehouse)
  • Cloud warehouses (Snowflake, BigQuery)
  • Leverage warehouse compute power

Lambda Architecture

• Batch layer: historical data processing
• Speed layer: real-time stream processing
• Serving layer: merge batch + real-time views
• Complexity: maintain two codebases

Kappa Architecture

• Stream-only processing
• Single codebase for batch + real-time
• Reprocessing via replay
• Simpler than Lambda

Medallion Architecture

• Bronze: Raw data (append-only)
• Silver: Cleaned, conformed data
• Gold: Business-level aggregations
• Databricks pattern

Orchestration Patterns

DAG-Based Orchestration

• Airflow, Prefect, Dagster
• Task dependencies as DAG
• Retries, backfills, scheduling

Event-Driven Orchestration

• Kafka, Pub/Sub triggers
• Real-time, low-latency
• Decoupled producers/consumers

Hybrid Orchestration

• Scheduled batch + event-driven streams
• Example: Airflow DAG triggered by Kafka event

Data Quality Frameworks

Data Contracts (CRITICAL)

• Define schema, freshness, volume expectations
• Producer-consumer agreement
• Break build on violation

Validation Frameworks

• Great Expectations: Python-based expectations
• dbt tests: SQL-based tests
• Soda: YAML-based checks

Data Lineage

• Track data origin and transformations
• Debug data quality issues
• Compliance and auditing

Idempotency Patterns

Idempotent Design (CRITICAL)

• Same input → same output (no side effects)
• Upserts instead of inserts
• Partition replacement instead of append

Deduplication

• Use unique keys
• Window-based deduplication
• Consumer group offset management

References

• Data Engineering Design Patterns