Data Quality Validation

Ocupación

Categorías

Ingeniería de Datos

3. Dataform Quality Tests

Use Dataform assertions to catch data quality issues:

• assert_ prefix files run as data quality checks
• Returns rows that FAIL the assertion (empty result = passing)
• Integrate into CI/CD pipeline before promoting data to production

4. Schema Validation Patterns

• Validate schema on ingestion using Dataflow side outputs
• Reject malformed records to a dead-letter GCS bucket or Pub/Sub topic
• Use BigQuery table schema with REQUIRED mode for mandatory fields
• For JSON: validate against JSON Schema before writing

5. Data Freshness Monitoring

• Cloud Monitoring custom metrics for pipeline lag
• BigQuery scheduled queries to check MAX(created_at) vs current time
• Dataplex data quality tasks for ongoing freshness checks
• Alert via Cloud Monitoring → Pub/Sub → Cloud Functions for remediation

Key Concepts

• Data Catalog — metadata management, tagging, business glossary, lineage
• Dataplex — unified data governance layer over GCS + BigQuery data lakes
• Dataform — version-controlled SQL transformations with built-in testing
• INFORMATION_SCHEMA.TABLE_STORAGE — monitor table sizes and freshness in BigQuery

Checklist

• NULL checks on required fields?
• Range/format validation on critical columns?
• Duplicate detection on primary key?
• Dead-letter path for rejected records?
• Freshness monitoring with alerting?
• Schema registered in Data Catalog?

Output Format

• 🔴 Critical — no validation on ingested data, no dead-letter for bad records
• 🟡 Warning — no freshness monitoring, no schema enforcement on ingestion
• 🟢 Suggestion — Dataform assertions for ongoing quality, Dataplex for governance

Exam Tips

• Dataform = SQL-based ELT + testing in BigQuery (like dbt for GCP)
• Dataplex = governance + quality across the data lake (GCS + BigQuery)
• Dead-letter pattern = invalid records → separate GCS path for human review
• Data Catalog = metadata, not data quality (but integrates with Cloud DQ)
• Schema validation should happen at ingestion (Dataflow) not after writing to BigQuery