Fiddler AI schema mismatch between training and production data

Fiddler compares the production data schema against the baseline (training) schema, and it flags three things: feature names that do not match, data types that changed, and features present in training but missing from production data. The fix is usually in the serving pipeline rather than in Fiddler: make sure the feature engineering code that runs in production emits exactly the same feature names and types as the training pipeline, and check for silently dropped columns, for example a join that started returning nulls or a renamed upstream field. If you use the Databricks Feature Store integration, define drift monitoring off the feature store definitions so both sides share one schema source.

Context: Fiddler is flagging schema mismatches on my production model even though training and serving use the same pipeline. What exactly does Fiddler compare, and how do I fix it?

Source: Fiddler is flagging schema mismatches on my production model even though training and serving use the same pipeline. What exactly does Fiddler compare, and how do I fix it?

Verified resolution: Fiddler compares the production data schema against the baseline (training) schema, and it flags three things: feature names that do not match, data types that changed, and features present in training but missing from production data. The fix is usually in the serving pipeline rather than in Fiddler: make sure the feature engineering code that runs in production emits exactly the same feature names and types as the training pipeline, and check for silently dropped columns, for example a join that started returning nulls or a renamed upstream field. If you use the Databricks Feature Store integration, define drift monitoring off the feature store definitions so both sides share one schema source.

Source URL: https://docs.fiddler.ai/integrations/ml-platforms-and-tools/ml-platforms#schema-mismatches