Data Quality Dimensions in Production
What you'll learn: The five core dimensions that define data quality in production ML systems and how they impact model reliability.
Why Data Quality Matters Beyond Drift
You've learned to detect input drift and concept drift—but even when distributions stay stable, poor data quality can silently degrade your model's performance. Production data quality goes beyond "is this data different?" to ask "is this data good?"
The Five Core Dimensions
1. Completeness
Are all expected features present? Missing values might indicate upstream pipeline failures. A profile that's 95% complete in training but only 70% complete in production signals a critical issue.
2. Accuracy
Does the data reflect reality correctly? This is hard to verify without ground truth, but sanity checks matter: negative ages, future timestamps, or sensor readings outside physical limits all indicate accuracy problems.
3. Consistency
Do related fields agree with each other? If age is 25 but birth_year is 1960, something's wrong. Consistency also spans time—the same user ID should have stable demographic attributes across requests.
4. Timeliness
Is the data fresh enough for your use case? A real-time fraud detector fed yesterday's transaction data is useless. Timeliness issues often appear as timestamp drift or staleness alerts.
5. Validity
Does data conform to expected formats and constraints? Valid data follows schema rules: enums match expected values, strings match regex patterns, numerical ranges stay within bounds.
The Production Perspective
Think of these dimensions as a health checklist. Your model might handle minor violations during training, but production compounds these issues at scale. A 1% validity error across millions of predictions becomes thousands of bad decisions.
Key Takeaway: Monitor all five data quality dimensions—completeness, accuracy, consistency, timeliness, and validity—as they directly impact model reliability even when drift detection shows stable distributions.