Use compute_metric_availability in AxClient.fit_model (#5111)

Summary:
Pull Request resolved: #5111

The previous data check in `fit_model` only verified that the DataFrame
was non-empty (`lookup_data().df.empty`), which would pass even when data
existed for only a subset of required metrics. This could allow model
fitting to proceed with incomplete data, leading to downstream errors.

Replace the manual check with `compute_metric_availability()` from
`ax.core.utils`, which inspects per-trial metric coverage against the
optimization config's required metrics. `fit_model` now raises
`DataRequiredError` unless at least one completed trial has data for
**all** required metrics (`MetricAvailability.COMPLETE`).

Reviewed By: saitcakmak

Differential Revision: D98208718

fbshipit-source-id: 9c5b898a92e8656516e19af5a12174528e6ce9e4

Author: sdaulton

ax/service/ax_client.py

@@ -40,6 +40,7 @@
     TParameterization,
     TParamValue,
 )
+from ax.core.utils import compute_metric_availability, MetricAvailability
 from ax.early_stopping.strategies import BaseEarlyStoppingStrategy
 from ax.early_stopping.utils import estimate_early_stopping_savings
 from ax.exceptions.constants import CHOLESKY_ERROR_ANNOTATION
@@ -1407,9 +1408,14 @@ def fit_model(self) -> None:
             raise DataRequiredError(
                 "At least one trial must be completed with data to fit a model."
             )
-        if self.experiment.lookup_data(trial_indices=completed_trial_indices).df.empty:
+        availability = compute_metric_availability(
+            experiment=self.experiment,
+            trial_indices=completed_trial_indices,
+        )
+        if not any(v == MetricAvailability.COMPLETE for v in availability.values()):
             raise DataRequiredError(
-                "At least one completed trial must have data attached to fit a model."
+                "At least one completed trial must have data for all required "
+                "metrics to fit a model."
             )
         self.generation_strategy.fit(experiment=self.experiment)
 

ax/service/tests/test_ax_client.py

@@ -2581,6 +2581,28 @@ def test_get_model_predictions_no_next_trial_no_completed_trial(self) -> None:
         ):
             ax_client.get_model_predictions()
 
+    def test_fit_model_partial_metric_data(self) -> None:
+        """Test that fit_model raises when completed trials only have data for
+        a subset of required metrics."""
+        ax_client = _set_up_client_for_get_model_predictions_no_next_trial()
+        # Attach a trial and complete it with data for only one of the two
+        # required metrics (test_metric1 is the objective, test_metric2 is the
+        # constraint). We bypass complete_trial() because it marks the trial as
+        # failed when required metrics are missing. Instead, we attach data and
+        # mark completed directly, simulating the case where the data check at
+        # completion time is skipped (e.g., data is attached asynchronously).
+        trial: TParameterization = {"x1": 0.1, "x2": 0.1}
+        _parameters, trial_index = ax_client.attach_trial(trial)
+        ax_trial = ax_client.get_trial(trial_index)
+        ax_trial.update_trial_data(raw_data={"test_metric1": (1.0, 0.0)})
+        ax_trial.mark_completed()
+
+        with self.assertRaisesRegex(
+            DataRequiredError,
+            "At least one completed trial must have data for all required metrics",
+        ):
+            ax_client.fit_model()
+
     def test_get_model_predictions_no_next_trial_filtered(self) -> None:
         ax_client = _set_up_client_for_get_model_predictions_no_next_trial()
         _attach_completed_trials(ax_client)