The ML Testing Pyramid

Three layers, not one

Ordinary software has unit/integration/e2e tests. ML adds two axes — data and model behaviour — so a robust ML CI runs three families of test:

1. Code tests — the feature transforms, the training loop, the serving handler. Plain pytest.
2. Data tests — schema, ranges, null rates, category sets, and distribution checks on every incoming batch (Great Expectations, pandera, TFDV). A pipeline that trains on garbage ships garbage.
3. Behavioural tests — assert properties of the model, not just aggregate accuracy:
   • Invariance: changing an irrelevant feature shouldn't flip the prediction.
   • Directional: raising income should not increase default probability.
   • Minimum-functionality: obvious cases must be correct (a 30-year, fully-paid customer is not high-risk).

These behavioural checks (popularised by the CheckList paper) catch failures that aggregate accuracy hides.

Aggregate accuracy is a car's average speed — it hides that it stalls on every hill. Behavioural tests are the hill-start, the emergency-brake and the reverse-park checks: specific scenarios the model must never fail, however good the average looks.