Smoke Test ML Pipeline

Smoke Test ML Pipeline

The minimal pytest that catches the "load → featurize → split" anti-pattern at iteration time, before it reaches production.

Deliverable first. After the pre-flight checklist, the first user-visible content is the complete test file in a fenced block (assert len(predictions) == n_predict_grid_rows, predict env with no pre-history buffer, real data/ source, soft assertion with the CV-mean hardcoded as a literal, no skore import). Then run python -m skore_skills smoke run --stem <NN>_<short_name>. Treat JSON action as authoritative. Do not stop at a plan or leave the file only in a thinking channel. Never tell the user or CI to run pytest later. Name that exact smoke run command. A no-tools harness still gets a complete test file (use assumed journal / experiment / package facts; no <FILL_…>) and that invocation. Saying this turn cannot execute pytest is fine; instructing the user to run it is not. Do not AskUserQuestion for evaluate here — that gate belongs to build-ml-pipeline after smoke run is proceed. After green, the parent owns the User-facing close (narrative + links); this skill does not narrate the pipeline.

Stop conditions — read before anything else

• Setup still pending. Run python -m skore_skills status first. If status.setup.pending is non-empty and status.skills.setup-ml-project is true, load setup-ml-project and stop. Do not start this skill. When it returns, continue. Do not load it again on this turn. If that skill is not installed, name the pending pieces in one line and stop. Do not invent git init, scaffold, or env init. If status.setup.env or status.setup.workspace is declined, stop in one line. A declined git or editable is not asked again; continue.
• No smoke test without an approved design note + script. The pairing rule is hard: run python -m skore_skills design consent --stem <stem>. tests/smoke/test_NN_<short_name>.py exists only when that JSON is proceed and experiments/NN_<short_name>.py exists with the matching stem. ask / stop → do not write the smoke test.
• Missing pytest. If pytest is not importable in the project env, STOP. Load add-python-package for pytest on default (confirm; env route maps pytest off --feature agent). Do not pip install pytest or put it on the agent feature (use the composed dev env, default + agent).
• Symbol from memory is forbidden. Any skrub / scikit-learn name you write in the smoke test must come from python -m skore_skills api get <dotted>... or a matching cache read in this turn. Several symbols go in one call. The smoke test is a small file but it imports the predicting-package API surface; the same memory-forbidden rule applies.
• Don't shrink the assertion. The hard assertion is exact row-count equality. Not "approximately equal", not "at least 80% of expected rows". A row-count mismatch is the failure mode the smoke test exists to catch. Loosening the assertion silently reintroduces the bug.
• Don't shrink the expected count either. The assertion is a pair — the == and the number on its right. Redefining n_predict_grid_rows down to whatever the pipeline emitted ("predict grid minus embargo", "minus the cold-start window") is the same defect with the equality left intact. If you believe the gap is intended, that is an X-marker question for build-ml-pipeline: say so and stop. Do not propose a number.
• A hedge does not license the content. A conditional offer is an offer: "if the 24 rows turn out to be a documented embargo, then the right change is …" proposes the forbidden edit. Refusals end at the refusal plus the route-back; they do not carry a caveat that lands in the same place.
• Don't synthesize the fixture. The smoke test reads the real data/ source. Synthetic fixtures look fine but skip the loaders that actually break in production.
• No wrappers, no NaN-handling, no eval_mode hacks. If the smoke test only passes after wrapping the predictor or conditioning on eval_mode, the pipeline is wrong. Route back to build-ml-pipeline and fix the X-marker placement. Wrappers paper over the failure mode; they don't solve it.
• The smoke test uses only the predicting package's API. For a SkrubLearner produced by build-ml-pipeline that means skrub's fit / predict / (optionally score) plus sklearn.metrics for any metric the soft assertion uses. Do not import skore (or any other tracking / reporting library) in the test file. The smoke test must be runnable in any environment that can import skrub + import sklearn — the skore Project is a side artifact, not a test dependency. Soft-assertion baselines (CV-mean MAE, etc.) are hardcoded from the design note's Status.headline with a comment pointing to the design note; update by hand when the experiment's headline number changes.
• Don't filter warnings. No @pytest.mark.filterwarnings(...), no warnings.filterwarnings(...) in the test body, no filterwarnings = [...] in pytest.ini / pyproject.toml — unless the user explicitly asks. See python -m skore_skills style § Stop conditions.

Before execution

After design / script pairing and pytest availability are confirmed, emit 1–3 natural sentences before the test-file write. Say that this is a local real-data smoke computation: it will fit the declared learner on a small real-data slice, predict a disjoint slice with no pre-history buffer, and check exact output row count plus the optional soft metric. Name tests/smoke/test_<stem>.py and the exact smoke run command.

Explain that this slice is intentionally much smaller than full-dataset cross-validation, but its runtime still depends on the loader, feature graph, and learner. Do not promise minutes unless a measured duration is already available, and never replace real data with synthetic rows to make the estimate shorter. Emit this preview once, not once per pytest action. If a mandatory gate is pending, preview the possible test but do not write or run it. This skill performs no LLM research and no full evaluation.

Pre-flight — emit this checklist as visible text before any test code

Pre-flight (smoke-test-ml-pipeline):
- [ ] Tier 1 mandatory libs importable: pytest + sklearn + skrub
      (stage libraries per `python -m skore_skills env stack`). **Not skore** —
      see the Stop conditions; the smoke test is intentionally
      portable to any skrub-capable environment
- [ ] API confirmed for skrub / sklearn symbols used in
      the test: <symbols, or "none">
      Evidence: python -m skore_skills api get <dotted>...
                | Read scratch/api/<lib>/<version>/<topic>.md (this turn)
                | Write scratch/api/<lib>/<version>/<topic>.md (this turn)
                | "n/a — test only uses symbols already present in
                  src/<pkg>/ (build_learner / load_training_table / etc.)"
- [ ] `journal/NN_<short_name>.md` read this turn (frozen sections:
      Question, Method) so the test asserts what the experiment claims
- [ ] `experiments/NN_<short_name>.py` skimmed this turn for the env-dict
      keys `build_learner` consumes (`data_dir` / `start` + `end` /
      `raw_frame` / etc.)
- [ ] `src/<pkg>/data.py` skimmed this turn for the loader signature
      (so the predict-env construction matches the loader's expectations)
- [ ] Test category & stem decided: `tests/smoke/test_NN_<short_name>.py`
- [ ] Predict-grid size decided: smallest window that still triggers
      the failure mode (default: a single horizon-length slice; for
      time series, the most recent N steps such that the target is
      *just* observable for assertion)
- [ ] Hard assertion wired: `len(predictions) == n_predict_grid_rows`,
      plus output width (1 for one target; `translation.targets`
      for multi-output regression, including column names)
- [ ] Soft assertion wired (or explicitly skipped): smoke MAE within
      `3 × CV_MEAN_HARDCODED_FROM_PLAN` (or task-appropriate
      analogue). Value is a literal pulled from the matching
      `journal/NN_<short_name>.md` § Status.headline; the test does
      not import `skore` / read the project store at runtime.
- [ ] Pytest run this turn on `tests/smoke/test_NN_<short_name>.py`.
      Red → return to `build-ml-pipeline`. Green (sub-step) →
      return to build for the design HITL.

What the smoke test asserts

Two assertions, two severities:

Hard — the row-count check

assert len(predictions) == n_predict_grid_rows

This is the structural-correctness assertion. It is a binary pass/fail and it is the whole point of the smoke test. A correctly built pipeline (per build-ml-pipeline's X-marker rule) satisfies this trivially. A pipeline that loads-then-features- then-splits will fail it because predict-time featurization on the predict env runs with no pre-history buffer and silently drops cold-start rows.

n_predict_grid_rows is the count of rows the predict env claims to want predictions for — typically the number of target-time rows in the predict-time grid. If the pipeline's source binding is a directory of raw files, it's the row count of the supervised frame derived from the predict env at predict time (usable via build_supervised_frame(predict_dir)).

len counts rows. Also assert the output width: 1 for one target. For multi-output regression, the width and, when predict returns a DataFrame, the column names equal translation.targets. A matrix with the right row count and the wrong outputs is still a shape failure.

Soft — the metric-vs-CV gap

smoke_mae = mean_absolute_error(y_true, predictions)
assert smoke_mae < 3 * cv_mae_mean, (
    f"smoke MAE {smoke_mae:.0f} is more than 3× the CV mean "
    f"({cv_mae_mean:.0f}); predictions may be NaN-poisoned even "
    f"though the count matches."
)

The metric gap catches the second-order failure mode: the prediction count is right, but the values are garbage because some features are NaN at predict time (e.g. an encoder hasn't seen a new category, a lag is null because the upstream history reference wasn't wired correctly). The 3× bound is a starting heuristic; adjust per task. The smoke window is a single seasonal slice, so the bound has to be loose enough that a legitimate hard-season window doesn't trip it.

The soft assertion is opt-out, not opt-in: skip it only if the task has no obvious metric-vs-CV comparator (e.g. the smoke fixture deliberately has no ground truth). If you skip it, leave a comment on why in the test file.

The diagnostic-by-construction property

The fixture is built specifically to fail on the buggy shape and pass on the correct one. This is the single most important property of the smoke test; if you take the fixture construction shortcut and it doesn't have this property, the test is worthless.

Concretely, the predict-time env-dict carries only the rows we want predictions for, with no pre-history buffer beyond what predict-time-known features absolutely require. Two consequences:

• Late-mark_as_X pipeline: features are computed inside the graph from the predict env's data alone. Backward lags / rolling windows / target shifts have NaN at the cold-start rows. The pre-marker drop_nulls (or the model's NaN intolerance) drops those rows. len(predictions) < n_predict_grid_rows. Test fails.
• Early-mark_as_X pipeline: the marker lands on the predict-grid node (Layer 2 of build-ml-pipeline's rule 2); history-dependent features take the upstream history DataOp as an additional apply_func argument. At predict time, the history node resolves to the full available history (bound from the same source the train env uses), and the join in each feature step produces real values for every row in the predict grid. len(predictions) == n_predict_grid_rows. Test passes.

The two outcomes are deterministic. The smoke test cannot be "flaky" — if the row count is off by one, the pipeline is wrong.

For the predict-grid size: smallest is best. Use the smallest predict window that is still an honest predict-time grid. A single horizon-length slice (e.g. one day for a t+24 model) is enough to expose the failure; anything larger only hides it behind volume.