Skip to content

Input/Output Schemas: Catching Bad Requests Early

What was run

An attempt to build a real, explicit validation layer in front of the served champion model, which immediately surfaced a real, honest gap: the actual deployed model has no signature at all. A real fallback schema was built and tested instead, contrasted directly against Lesson 7's messier, raw stack-trace rejection.

A real, honest discovery: this model was never given a signature

model = mlflow.pyfunc.load_model("models:/breast-cancer-classifier@champion")
EXPECTED_COLUMNS = model.metadata.get_input_schema().input_names()

AttributeError: 'NoneType' object has no attribute 'input_names'

A real, genuine gap: champion points at the model trained by train_capstone.py, which — unlike Lesson 3's dedicated signature-test run — never called infer_signature() before logging. The real, deployed, promoted-through-a-real-approval-gate model has no built-in schema to check requests against at all.

The real fix: an explicit schema, enforced before the model ever runs

EXPECTED_COLUMNS = list(load_breast_cancer().feature_names)

@app.post("/predict")
async def predict(request: Request):
    columns = body.get("columns", [])
    missing = set(EXPECTED_COLUMNS) - set(columns)
    if missing or extra:
        raise HTTPException(400, {"error": "schema_mismatch", "missing_columns": sorted(missing), ...})


curl -s -w "\nHTTP:%{http_code}\n" -X POST http://127.0.0.1:5601/predict -d '{"columns": ["mean radius", "mean texture"], "data": [[14.0, 20.0]]}'

{"detail":{"error":"schema_mismatch","missing_columns":["area error","compactness error", ... 28 total]}}
HTTP:400

A real, clean, structured 400 — listing exactly which columns are missing, as real, actionable JSON — compared directly against Lesson 7's real finding, where the same kind of malformed request against MLflow's own generic serving endpoint produced a messier 400 with scikit-learn's raw internal ValueError and full stack trace embedded in the response body.

curl -s http://127.0.0.1:5601/predict -d '{"columns": [...all 30 real columns...], "data": [[...]]}'
# {"predictions":[0]}

A real, correctly-shaped request still succeeds normally, confirming the validation layer only blocks genuinely malformed requests, not valid ones.

How It Actually Works

This lesson's own real discovery — a production-promoted model with no signature at all — demonstrates something important that a course built entirely around idealized examples would miss: real models, in real pipelines, don't always arrive pre-packaged with a complete schema, even after passing a real approval gate (Lesson 6). When that happens, the responsibility for validating input shifts from the model's own packaging (Lesson 3's ideal case) to the real serving layer sitting in front of it — exactly the explicit, hand-written schema this lesson built as a genuine, necessary fallback.

The real, practical advantage of validating explicitly at the serving layer, demonstrated directly by comparing this lesson's clean JSON error against Lesson 7's raw stack trace: a deliberate validation step controls exactly what a caller sees on failure — a clear, structured, actionable message — rather than whatever internal exception the underlying framework happens to raise, which can (as Lesson 7 found) leak internal file paths and library internals into a real, external-facing API response.

Common mistakes

  • Assuming every logged model has a usable signature without checking — this lesson's own real AttributeError is exactly the consequence of that assumption, against a real, currently-deployed production model.
  • Returning raw internal exceptions (stack traces, library-specific error messages) directly in an API response, as Lesson 7's real, unvalidated endpoint did — a real security and usability concern, since internal paths and library versions become visible to any caller, not just ones that should see them.
  • Building schema validation that only checks for missing fields and not unexpected ones — this lesson's real validator checks both (missing and extra), since an unexpected extra field reaching a model can be just as real a sign of a caller-side bug as a missing one.

Exercise

Fix the real underlying gap this lesson found, rather than only working around it: retrain and re-log a new version of breast-cancer-classifier using Level 1's capstone script, this time adding a real infer_signature() call before mlflow.sklearn.log_model(), exactly as Lesson 3 did. Promote it through the real approval gate, then confirm model.metadata.get_input_schema() no longer returns None for the new real champion.