diff --git a/ucm/observability.py b/ucm/observability.py
index 6bc36099e..b7409dfed 100644
--- a/ucm/observability.py
+++ b/ucm/observability.py
@@ -66,12 +66,16 @@ def __init__(self, model_name, worker_id, config_path):
         Load metrics config from YAML file (config_path),
         register metrics using prometheus_client, and start a thread to get updated metrics.
         """
+        # Always initialize instance attributes so that shutdown() and
+        # __del__() work even when we skip metric registration below.
+        self.config = self._load_config(config_path)
+        self.log_interval = self.config.get("log_interval", 10)
+        self.is_running = False
+        self.thread = None
+
         if _metric_mappings:
             logger.warning("Metrics are already registered, skipping re-registration.")
             return
-        # Load metrics config
-        self.config = self._load_config(config_path)
-        self.log_interval = self.config.get("log_interval", 10)
 
         # Set up histogram max length
         histogram_max_length = self.config.get("histogram_max_length", 10000)
@@ -97,9 +101,10 @@ def __init__(self, model_name, worker_id, config_path):
         # Initialize metrics based on config
         self._init_metrics_from_config()
 
-        # Start thread to update metrics
+        # Start daemon thread to update metrics so it won't block
+        # process exit if shutdown() is not explicitly called.
         self.is_running = True
-        self.thread = threading.Thread(target=self.update_stats_loop)
+        self.thread = threading.Thread(target=self.update_stats_loop, daemon=True)
         self.thread.start()
 
     def _register_metrics_by_type(self, metric_type):
@@ -190,7 +195,8 @@ def update_stats_loop(self):
 
     def shutdown(self):
         self.is_running = False
-        self.thread.join()
+        if self.thread is not None:
+            self.thread.join()
 
     def __del__(self):
         try: