data-pipeline-quality

Testing Pyramid for Data

Run tests in this order. Cheapest and fastest first:

Schema tests are cheap. Run them on every pipeline execution. Business rule tests are mid-tier — run them on staging and production. Integration tests are expensive — run them on a schedule (daily or pre-release).

dbt Test Patterns

Generic tests for reusable checks. Apply across models:

models:
  - name: fct_encounters
    columns:
      - name: encounter_id
        tests: [not_null, unique]
      - name: encounter_type
        tests:
          - accepted_values:
              values: ['inpatient', 'outpatient', 'emergency', 'observation']
      - name: patient_id
        tests:
          - relationships:
              to: ref('dim_patient')
              field: patient_id

Custom generic test for row count tolerance:

{% test row_count_within_tolerance(model, min_count, max_count) %}
select count(*) as row_count
from {{ model }}
having count(*) < {{ min_count }} or count(*) > {{ max_count }}
{% endtest %}

Singular tests for business logic specific to one model. Use singular tests when the logic doesn't generalize.

Data Contracts

A data contract is a product spec for your data. It defines what consumers can depend on.

contract:
  name: fct_encounters
  version: 2
  owner: data-platform-team
  sla:
    freshness: "< 4 hours from source update"
    completeness: ">= 99.5% of expected rows"
    accuracy: ">= 99.9% match to source of record"
  schema:
    encounter_id: {type: bigint, nullable: false, unique: true}
    patient_id: {type: bigint, nullable: false}
    encounter_date: {type: date, nullable: false}

Producer responsibilities: Meet the SLA, notify consumers before breaking changes, version the schema.

Consumer expectations: Query only contracted fields, respect the grain, report quality issues.

Circuit Breakers

Wire quality gates into pipeline stages. When a check fails, block the pipeline and alert.

• Schema violation: Block immediately. Bad schema corrupts everything downstream.
• Row count outside tolerance: Block and alert. Investigate before proceeding.
• Freshness SLA breach: Alert the on-call. Don't block unless downstream consumers can't tolerate stale data.
• Business rule failure: Block if the failure rate exceeds threshold (e.g., >1% of rows). Alert if below.

Cross-reference data-quality-assessment for the 4-stage quality model (detect → assess → respond → prevent).

Never auto-heal data quality issues in production. Alert, block, investigate. Auto-fixes mask root causes and erode trust faster than stale data.

Monitoring

Track quality over time, not just point-in-time pass/fail:

• Row count trends: Compare expected and observed counts within the same scope. Abrupt or gradual changes suggest checks; their shape does not establish source or pipeline root cause. Test competing explanations.
• Freshness: Separate source-event time, arrival time and processing time against the consumer's requirement. A successful run can process stale or incomplete inputs. Track the relevant lag distribution and coverage.
• Anomaly detection: Choose statistical or fixed thresholds for the risk, seasonality and available history. Confirm expected ranges with the owner; a statistical alert does not replace an acceptance rule.

Cross-References

For quality scoring methodology (the 5-dimension rubric and maturity model), see data-quality-assessment. This skill covers how to automate those checks in your pipeline.