From 5b69dc4206cc4f0964868e4a637eedaa33bd5be9 Mon Sep 17 00:00:00 2001
From: Zilinghan <zl52@illinois.edu>
Date: Wed, 4 Mar 2026 12:01:50 -0600
Subject: [PATCH 1/2] fix mlflow logger

---
 src/driver/continuous_monitor.py | 26 +++++++++++++++++++++++---
 src/logger/mlflow_logger.py      | 23 ++++++++++++++++++++---
 2 files changed, 43 insertions(+), 6 deletions(-)

diff --git a/src/driver/continuous_monitor.py b/src/driver/continuous_monitor.py
index 299578c..7c2c15d 100644
--- a/src/driver/continuous_monitor.py
+++ b/src/driver/continuous_monitor.py
@@ -265,6 +265,12 @@ def _handle_drift(self, drift_signal: DriftSignal) -> None:
         self.logger.info(
             f"==== DRIFT DETECTED (Event #{self.drift_event_count})! ====", level=0
         )
+        # Log data timestamp range if the harness tracks it
+        timerange = getattr(self.modelHarness, "current_window_timerange", None)
+        if timerange is not None:
+            self.logger.info(
+                f"\tData time range: {timerange[0]} → {timerange[1]}", level=1
+            )
         self.logger.info(
             f"\tRegime: {drift_signal.regime.value if drift_signal.regime else 'N/A'}",
             level=1,
@@ -328,17 +334,31 @@ def _log_metrics(self, drift_signal: DriftSignal, metric_value: float) -> None:
         """
         flops_perf = self.flops_profiler.get_performance()
 
-        # Log all drift metrics including performance in a single call
+        # Log all drift metrics in a single call.
+        # detected=0 is sampled at 10% to reduce log volume; detected=1 is
+        # always included so no true drift events are dropped.
+        log_detected = drift_signal.drift_detected or (np.random.random() <= 0.1)
         self.logger.stage("drift")
+        # Include data timestamp range if available from the harness
+        timerange = getattr(self.modelHarness, "current_window_timerange", None)
+        ts_fields = {}
+        if timerange is not None:
+            ts_fields["data_time_start"] = timerange[0]
+            ts_fields["data_time_end"] = timerange[1]
         self.logger.log(
             {
-                "detected": drift_signal.drift_detected,
+                **(
+                    {"detected": int(drift_signal.drift_detected)}
+                    if log_detected
+                    else {}
+                ),
                 "score": drift_signal.drift_score,
                 "regime": (drift_signal.regime.value if drift_signal.regime else "N/A"),
                 "confidence": (
                     drift_signal.confidence if drift_signal.confidence else "N/A"
                 ),
                 f"metric_{self.metric_idx}": metric_value,
+                **ts_fields,
                 **{f"cperf_{k}": v for k, v in flops_perf.items()},
             },
-        )
+        )
\ No newline at end of file
diff --git a/src/logger/mlflow_logger.py b/src/logger/mlflow_logger.py
index f6b7278..4ffbbd9 100644
--- a/src/logger/mlflow_logger.py
+++ b/src/logger/mlflow_logger.py
@@ -174,7 +174,7 @@ def log(
 
         # Add step metrics
         prefixed_metrics["step"] = current_step
-        if self._current_stage:
+        if self._current_stage and increment:
             prefixed_metrics[f"{self._current_stage}.step"] = self._stage_steps[
                 self._current_stage
             ]
@@ -187,13 +187,30 @@ def log(
         if self.enabled and self.run:
             mlflow = self._get_mlflow()
             # Filter to only numeric values (MLflow only accepts numbers for metrics)
+            # Exclude bools: bool is a subclass of int in Python, but casting
+            # True/False to float(1.0/0.0) corrupts categorical semantics.
+            # Instead, booleans are converted to int (0/1) separately.
             numeric_metrics = {
                 k: float(v)
                 for k, v in prefixed_metrics.items()
-                if isinstance(v, (int, float)) and k != "step"
+                if isinstance(v, (int, float))
+                and not isinstance(v, bool)
+                and k != "step"
             }
+            # Convert booleans to int (0/1) so they remain discrete metrics
+            bool_metrics = {
+                k: int(v) for k, v in prefixed_metrics.items() if isinstance(v, bool)
+            }
+            numeric_metrics.update(bool_metrics)
             if numeric_metrics:
-                mlflow.log_metrics(numeric_metrics, step=current_step)
+                # WandB uses the global step as x-axis, however, MLflow uses the per-stage step so that
+                # sparse global-step values do not cause MLflow to downsample away the rare detected=1 events.
+                log_step = (
+                    self._stage_steps[self._current_stage]
+                    if self._current_stage
+                    else current_step
+                )
+                mlflow.log_metrics(numeric_metrics, step=log_step)
 
     def save(
         self,

From c50445c76401c7cc99d94ddd26ca52560e86d993 Mon Sep 17 00:00:00 2001
From: Ana Gainaru <ana.gainaru@gmail.com>
Date: Wed, 4 Mar 2026 18:14:40 -0500
Subject: [PATCH 2/2] Pass the CI

---
 src/driver/continuous_monitor.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/driver/continuous_monitor.py b/src/driver/continuous_monitor.py
index 7c2c15d..a015da3 100644
--- a/src/driver/continuous_monitor.py
+++ b/src/driver/continuous_monitor.py
@@ -361,4 +361,4 @@ def _log_metrics(self, drift_signal: DriftSignal, metric_value: float) -> None:
                 **ts_fields,
                 **{f"cperf_{k}": v for k, v in flops_perf.items()},
             },
-        )
\ No newline at end of file
+        )