Do I need a commercial data observability platform to start?

No. You can cover most risk with dbt tests, Great Expectations or Soda for deeper checks, Prometheus for metrics, and your existing orchestrator for gating. Add a vendor later if you need ML-based anomaly detection at scale.

How do I pick thresholds without false positives?

Baseline for two weeks, then set thresholds using trailing averages and percent bands. Page on SLOs (e.g., freshness >30 min for 10 min) and send everything else as warnings.

Who should get paged for data quality failures?

The data product owner. Tag models with owners in code and route alerts via Alertmanager/PagerDuty based on those tags. Shared channels cause alert fatigue and no accountability.

What metrics convince executives this is working?

MTTR for data incidents, SLO compliance rate, number of prevented downstream builds, and ‘data downtime minutes’ on executive metrics. Tie each incident to a dollar impact where possible.

How do we prevent schema drift from breaking models?

Stop using SELECT *. Add dbt contracts (`constraints/enforced`), enable GE/Soda schema checks, and run CI on PRs for sources and models. For sources like Kafka, use a schema registry and enforce compatibility.

Data-engineering · Oct 17, 2025 · 9 minute read

The 7 a.m. Dashboard That Lied — And the Data Quality Guardrails That Shut It Up

If a single schema change can nuke your CFO’s dashboard, you don’t have analytics—you have a roulette wheel. Here’s the data quality monitoring that actually prevents downstream failures, with code you can ship this week.

Back to all posts

The 7 a.m. Dashboard That Lied — And the Data Quality Guardrails That Shut It Up

Key takeaways

Implementation checklist