fix prophet detector when irregular data

Summary:
Fix the Prophet detector when data is irregular.
We resort to using all data if too much data is removed by outlier removal.

Reviewed By: yazovskymeta

Differential Revision: D67392638

fbshipit-source-id: b48abd49f07965df43020b2a8c235dea0d9eb266

kats/detectors/prophet_detector.py

@@ -473,12 +473,19 @@ def fit(
         data_df = timeseries_to_prophet_df(total_data)
 
         if self.remove_outliers:
-            data_df = self._remove_outliers(
+            updated_data_df = self._remove_outliers(
                 data_df,
                 self.outlier_threshold,
                 uncertainty_samples=self.outlier_removal_uncertainty_samples,
                 vectorize=self.vectorize,
             )
+            if len(updated_data_df) < 2:
+                logging.warning(
+                    f"Outlier removal left {len(updated_data_df)} out of"
+                    f" {len(data_df)} data points. Reverting to use all data."
+                )
+            else:
+                data_df = updated_data_df
         # seasonalities depends on current time series
         self.seasonalities_to_fit = seasonalities_processing(
             data_df[PROPHET_TIME_COLUMN], self.seasonalities

kats/tests/detectors/test_prophet_detector.py

@@ -526,8 +526,8 @@ def test_irregular_data_intervals(self, data_multiplier: float) -> None:
         )
 
         model = ProphetDetectorModel(remove_outliers=True)
-        with self.assertRaises(ValueError):
-            model.fit(irregular_ts)
+        model.fit(irregular_ts)
+        # No ValueError raised
 
     def test_default_score_func(self) -> None:
         """Test that 'deviation_from_predicted_val' is used by default