You can set acceptable error rates for your data quality rules by defining the maximum percentage of records that can fail a given rule before action is required — calibrated to the business impact of errors, not to a generic "best practice" percentage.
The most common data quality monitoring mistake is setting every rule to a 0% error rate threshold. When everything is always in violation, alerts become noise. Teams stop paying attention. The monitoring system becomes the reason no one notices when quality actually degrades.
The Two Factors That Determine a Good Threshold
Factor 1: Business impact of errors in this field. A 1% error rate on customer email addresses has a different business impact than a 1% error rate on an internal notes field. Higher-impact fields deserve tighter thresholds.
Sohovi tracks quality trends across runs and alerts you when a metric — null rate, duplicate count, score — moves outside its normal range.
Factor 2: Historical baseline for this field. If your email validity rate has consistently been 97.5% for the past year, a threshold of 95% gives meaningful headroom. A threshold of 99% would generate constant false alarms for normal variation.
A Practical Threshold Framework
| Tier | Field examples | Recommended threshold | |---|---|---| | Mission-critical | Customer email, order ID, financial amounts | 98–99% | | Important business | Phone number, company name, address | 90–96% | | Useful but non-critical | Job title, industry, demographic data | 75–85% | | Optional enrichment | Any field that's often blank by design | 50–70% |
Setting Thresholds Based on Historical Baseline
- Calculate the average failure rate for the field over the past 3–6 months
- Add a buffer of 5–10 percentage points to create your alert threshold
- This catches genuine degradation while ignoring normal variation
Frequently Asked Questions
Q: What is a data quality error rate threshold? A threshold is the maximum percentage of records that can fail a specific validation rule before an alert is triggered. It distinguishes acceptable background noise from a real quality problem worth investigating.
Q: Why not set all error rate thresholds to 0%? A 0% threshold generates constant alerts for normal data variation. When everything is always in violation, the monitoring system produces noise rather than signal. Teams stop paying attention.
Q: How do I know if a threshold is set correctly? A well-calibrated threshold should trigger alerts rarely enough that each alert is taken seriously, but frequently enough to catch genuine problems. If you're getting alerts more than weekly for routine variation, your threshold is too tight.
Q: Should all fields have the same error rate threshold? No. The right threshold varies by field importance, usage frequency, and historical baseline. A customer email field needs a tighter threshold than an optional enrichment field.
Q: What's the difference between an alert threshold and a hard-stop threshold? An alert threshold triggers a notification. A hard-stop threshold halts a process until quality is restored. Hard-stop thresholds are for fields where continued processing with poor quality would cause unacceptable damage.
Q: How often should I review and adjust my thresholds? Review quarterly for the first year of monitoring, then annually once you have a stable baseline. Also review after any significant change to data sources or how the data is used.
Q: What should I do when a field consistently runs just below threshold? Investigate whether the threshold is wrong or there's a real quality problem. If historical data shows the field has never reached the threshold, it's set too high. If it used to be above threshold and recently declined, that's a genuine quality issue.
Q: Is a percentage threshold always the right way to measure data quality? Not always. For some fields, absolute count thresholds are more meaningful — "more than 50 invalid email addresses per import" may be more actionable than "more than 2%" if import sizes vary significantly.
Q: How do I set thresholds for new datasets with no historical baseline? Use the tier-based defaults as a starting point. Monitor for 60–90 days before adjusting. Document that thresholds are provisional until you have a stable baseline.
Q: Can data quality thresholds change over time as the business grows? Yes. A 2% error rate acceptable when you had 5,000 customers may represent 10,000 individual errors when you have 500,000 — a very different business impact. Revisit thresholds as your data volume changes.
Setting the right thresholds converts data quality monitoring from a nuisance into a useful early warning system. Start with tier-based defaults, calibrate to your historical baseline, and review quarterly.
