Why can schema evolution break a scientific analysis?

Short answer

A column can retain its name while changing type, units, meaning or missing-value rules. Analyses may still run but yield incorrect scientific conclusions.

Why

Consider a concentration column that changes from mg/L to micrograms/mL. The numeric values are equivalent only after unit-aware conversion, whereas a different change such as ng/mL would require a scaling factor. A pipeline expecting sample_id as text may also strip leading zeros after upstream inference changes it to an integer.

Protect consumers with explicit typed contracts including field meaning, units, identifiers, controlled vocabularies, null semantics and permitted ranges. Test backward compatibility before promotion, record schema versions and migrate old files deliberately rather than coercing silently. Columnar formats preserve physical types but cannot enforce all biological meaning on their own.

Keep test fixtures for historical schema versions and define what should fail fast.

Common mistakes

Accepting an unchanged column name as evidence that its measurement meaning is unchanged

The interviewer might ask next

What checks would detect a supplier silently switching units on a lab concentration field?