SIENTIAPDE-1645: Expose detailed model training information via a new Prometheus gauge. This gauge, sientia_training_info, records metadata such as dataset sizes, feature count, and evaluation metrics (MSE, MAE, R2) along with the training run's timestamp.

model_manager/activities/training.py

@@ -262,6 +262,18 @@ class Training(SientiaMonitoring):
                     train_result.r2_val
                 )
 
+            mm_metrics.SIENTIA_TRAINING_INFO.labels(
+                pod_id=labels['pod_id'],
+                model_name=train_params.model_name,
+                model_type=train_params.model_type,
+                dataset_train_rows=str(len(train_result.train_data)),
+                dataset_val_rows=str(len(train_result.val_data)),
+                feature_count=str(len(train_params.variable_columns)),
+                mse=str(train_result.mse_val) if train_result.mse_val is not None else '',
+                mae=str(train_result.mae_val) if train_result.mae_val is not None else '',
+                r2=str(train_result.r2_val) if train_result.r2_val is not None else '',
+            ).set(time.time() * 1000)
+
             self.info(f'Starting MLflow run for {train_params.model_type}', metadata)
             with self.mlflow_repository.start_run(
                 model_name=train_params.model_name,

model_manager/metrics.py

@@ -27,6 +27,22 @@ APP_UP = Gauge(
 
 _TRAINING_LABELS = ['pod_id', 'model_name', 'model_type']
 
+SIENTIA_TRAINING_INFO = Gauge(
+    'sientia_training_info',
+    'Metadata and execution timestamp (ms) of the last successful model training run',
+    [
+        'pod_id',
+        'model_name',
+        'model_type',
+        'dataset_train_rows',
+        'dataset_val_rows',
+        'feature_count',
+        'mse',
+        'mae',
+        'r2',
+    ],
+)
+
 SIENTIA_TRAINING_MODEL_TRAINED_TOTAL = Counter(
     'sientia_training_model_trained_total',
     'Number of successfully completed model training runs',

tests/test_metrics.py

@@ -329,3 +329,51 @@ def test_sientia_training_feature_count_is_gauge():
     assert isinstance(SIENTIA_TRAINING_FEATURE_COUNT, Gauge)
     assert SIENTIA_TRAINING_FEATURE_COUNT._name == 'sientia_training_feature_count'
     _assert_training_labels(SIENTIA_TRAINING_FEATURE_COUNT)
+
+
+def test_sientia_training_info_is_gauge():
+    from prometheus_client import Gauge
+
+    from model_manager.metrics import SIENTIA_TRAINING_INFO
+
+    assert isinstance(SIENTIA_TRAINING_INFO, Gauge)
+    assert SIENTIA_TRAINING_INFO._name == 'sientia_training_info'
+
+    expected_labels = {
+        'pod_id', 'model_name', 'model_type',
+        'dataset_train_rows', 'dataset_val_rows', 'feature_count',
+        'mse', 'mae', 'r2',
+    }
+    assert expected_labels == set(SIENTIA_TRAINING_INFO._labelnames)
+
+
+def test_sientia_training_info_set_value():
+    import time
+
+    from model_manager.metrics import SIENTIA_TRAINING_INFO
+
+    ts = time.time() * 1000
+    SIENTIA_TRAINING_INFO.labels(
+        pod_id='test-pod',
+        model_name='my_model',
+        model_type='linear',
+        dataset_train_rows='1000',
+        dataset_val_rows='200',
+        feature_count='5',
+        mse='0.01',
+        mae='0.08',
+        r2='0.95',
+    ).set(ts)
+
+    value = SIENTIA_TRAINING_INFO.labels(
+        pod_id='test-pod',
+        model_name='my_model',
+        model_type='linear',
+        dataset_train_rows='1000',
+        dataset_val_rows='200',
+        feature_count='5',
+        mse='0.01',
+        mae='0.08',
+        r2='0.95',
+    )._value._value
+    assert abs(value - ts) < 2000