Imported from curiositech/some_claude_skills (.claude/skills/data-pipeline-engineer/SKILL.md). Install upstream with npx skills add curiositech/some_claude_skills --skill data-pipeline-engineer. Copyright stays with the author.

Data Pipeline Engineer

Expert data engineer specializing in ETL/ELT pipelines, streaming architectures, data warehousing, and modern data stack implementation.

Quick Start

1. Identify sources - data formats, volumes, freshness requirements
2. Choose architecture - Medallion (Bronze/Silver/Gold), Lambda, or Kappa
3. Design layers - staging → intermediate → marts (dbt pattern)
4. Add quality gates - Great Expectations or dbt tests at each layer
5. Orchestrate - Airflow DAGs with sensors and retries
6. Monitor - lineage, freshness, anomaly detection

Core Capabilities

Architecture Patterns

Medallion Architecture (Recommended)

BRONZE (Raw)     → Exact source copy, schema-on-read, partitioned by ingestion
      ↓ Cleaning, Deduplication
SILVER (Cleansed) → Validated, standardized, business logic applied
      ↓ Aggregation, Enrichment
GOLD (Business)   → Dimensional models, aggregates, ready for BI/ML

Lambda vs Kappa

• Lambda: Batch + Stream layers → merged serving layer (complex but complete)
• Kappa: Stream-only with replay → simpler but requires robust streaming

Reference Examples

Full implementation examples in ./references/:

Anti-Patterns (10 Critical Mistakes)

1. Full Table Refreshes

Symptom: Truncate and rebuild entire tables every run Fix: Use incremental models with is_incremental(), partition by date

2. Tight Coupling to Source Schemas

Symptom: Pipeline breaks when upstream adds/removes columns Fix: Explicit source contracts, select only needed columns in staging

3. Monolithic DAGs

Symptom: One 200-task DAG running 8 hours Fix: Domain-specific DAGs, ExternalTaskSensor for dependencies

4. No Data Quality Gates

Symptom: Bad data reaches production before detection Fix: Great Expectations or dbt tests at each layer, block on failures

5. Processing Before Archiving

Symptom: Raw data transformed without preserving original Fix: Always land raw in Bronze first, make transformations reproducible

6. Hardcoded Dates in Queries

Symptom: Manual updates needed for date filters Fix: Use Airflow templating (e.g., ds variable) or dynamic date functions

7. Missing Watermarks in Streaming

Symptom: Unbounded state growth, OOM in long-running jobs Fix: Add withWatermark() to handle late-arriving data

8. No Retry/Backoff Strategy

Symptom: Transient failures cause DAG failures Fix: retries=3, retry_exponential_backoff=True, max_retry_delay

9. Undocumented Data Lineage

Symptom: No one knows where data comes from or who uses it Fix: dbt docs, data catalog integration, column-level lineage

10. Testing Only in Production

Symptom: Bugs discovered by stakeholders, not engineers Fix: dbt --target dev, sample datasets, CI/CD for models

Quality Checklist

Pipeline Design:

• Incremental processing where possible
• Idempotent transformations (re-runnable safely)
• Partitioning strategy defined and documented
• Backfill procedures documented

Data Quality:

• Tests at Bronze layer (schema, nulls, ranges)
• Tests at Silver layer (business rules, referential integrity)
• Tests at Gold layer (aggregation checks, trend monitoring)
• Anomaly detection for volumes and distributions

Orchestration:

• Retry and alerting configured
• SLAs defined and monitored
• Cross-DAG dependencies use sensors
• max_active_runs prevents parallel conflicts

Operations:

• Data lineage documented
• Runbooks for common failures
• Monitoring dashboards for pipeline health
• On-call procedures defined

Validation Script

Run ./scripts/validate-pipeline.sh to check:

• dbt project structure and conventions
• Airflow DAG best practices
• Spark job configurations
• Data quality setup

External Resources

• dbt Best Practices
• Airflow Best Practices
• Great Expectations Docs
• Delta Lake Guide
• Kafka Streams