Session 29: Property drain: data accessors

Baseline: session 28 finished the modeling-verb drain (16 verbs). Session 29 promotes the user-facing data accessor properties off self._legacy onto a snapshot in self._fit_state. The public API surface no longer requires self._legacy to exist on read paths.

CHANGED — engine#

• CHANGED — packages/engine/pycaret/core/experiment.py — Experiment.fit now calls _snapshot_fit_state() after self._legacy.setup() returns. The snapshot captures references (not copies) to seven legacy attributes in self._fit_state, a dict-backed cache.
• CHANGED — Experiment.X / X_train / X_test / y / y_train / y_test / preprocess_pipeline now read from self._fit_state instead of dispatching to self._legacy.<attr> on every access. _require_fitted() is still called to maintain the NotFittedError contract.
• CHANGED — Defensive getattr(legacy, name, None) in _snapshot_fit_state() lets the same code path work across task types. Clustering / anomaly experiments don't have y_test; the snapshot stores None for missing slots.

ADDED — tests#

• ADDED — packages/engine/tests/test_session29_property_drain.py — 4 tests:
  • test_data_properties_do_not_call_legacy_after_fit — the property drain-lock. Wraps every drained self._legacy.<X> accessor with a raise-on-read sentinel post-fit; the 7 properties continue to return correct values, proving they no longer touch the legacy holder.
  • test_data_properties_clustering_y_is_none — clustering experiments don't have a target; y / y_train / y_test come back None.
  • test_data_properties_require_fit — every accessor raises NotFittedError on an unfit experiment.
  • test_fit_state_returns_equivalent_data_to_legacy — sanity check on shape + columns + identity for the singleton preprocess_pipeline.

INTERNAL#

• INTERNAL — References vs deep copies. _fit_state holds references to the legacy attribute values, not deep copies. Mutating exp.X_train propagates to the underlying frame, matching legacy semantics. This decision was made deliberately: copying would (a) double memory for large datasets and (b) break code that does exp.X_train.iloc[5:10] = …. The cost is that _fit_state is implicitly tied to the legacy lifetime; once setup() is itself drained (last step before deleting pycaret/internal/pycaret_experiment/), the references will hold the data the new fit path produces directly.
• INTERNAL — Why dict not dataclass. A dataclass would force a fixed schema; the snapshot-as-dict tolerates task-specific gaps (clustering's y_test=None) without needing typed Optional annotations everywhere. When the state-holder migration finishes (post-4.0.0), the dict can be promoted to a typed FitState dataclass; for now the dict shape gives flexibility.
• INTERNAL — The drain-lock test pattern, now applied to properties. Sessions 22-28 used the monkeypatch self._legacy.<verb> → raise pattern for verb drains. Session 29 generalises it to attribute reads: object.__setattr__(legacy, name, _BoomDescriptor()) shadows the legacy property with a raise-on-read object. The try/except AttributeError handles slot-bound legacy attributes that can't be shadowed (best-effort). The test passes because the drain genuinely doesn't touch those callsites — the sentinels are never read.

Session 29 delta summary#