fix replicas query

deployment/vllm/autoscaler.yml

@@ -18,6 +18,7 @@ spec:
       containers:
         - name: autoscaler
           image: ghcr.io/wmgeolab/k8s-autoscaler:latest # Update this to your image
+          imagePullPolicy: Always
           securityContext:
             runAsGroup: 50036
             runAsUser: 237827
@@ -38,6 +39,10 @@ spec:
           env:
             - name: KUBECONFIG
               value: /app/.kube/config
+            - name: UVICORN_LOG_LEVEL
+              value: "debug"
+            - name: LOG_LEVEL
+              value: "debug"
             - name: VLLM_SERVICE_HOST
               value: "vllm-svc"
             - name: VLLM_SERVICE_PORT

k8s-autoscaler/k8s_autoscaler/api/routes.py

@@ -39,13 +39,13 @@ async def health_check(
 ):
     """Health check endpoint."""
     phase = await kube.get_pod_phase()
-    current_replicas, desired_replicas = await kube.get_replicas()
+    current_replicas = await kube.get_replicas()
     return {
         "status": "healthy",
         "vllm_status": phase,
         "vllm_running": phase == PodPhase.RUNNING,
         "current_replicas": current_replicas,
-        "desired_replicas": desired_replicas,
+        # "desired_replicas": desired_replicas,
         "last_activity": time.strftime(
             "%Y-%m-%d %H:%M:%S", time.localtime(state.last_activity)
         ),

k8s-autoscaler/k8s_autoscaler/kubernetes.py

@@ -27,6 +27,8 @@ async def execute(self, cmd: str) -> tuple[bool, str]:
             output = stdout.decode().strip() if success else stderr.decode().strip()
             if not success:
                 logger.error(f"kubectl command failed: {output}")
+
+            logger.debug(f"Command output: `{output}`")
             return success, output
         except Exception as e:
             logger.error(f"kubectl command failed: {e}")
@@ -60,27 +62,31 @@ async def scale_deployment(self, replicas: int) -> bool:
             logger.error(f"Invalid replica count: {replicas}")
             return False
 
+        # success, output = await self.execute(
+        #     f"scale deployment {self.settings.vllm_deployment} --replicas={replicas}"
+        # )
+
         success, output = await self.execute(
-            f"scale deployment {self.settings.vllm_deployment} --replicas={replicas}"
+            f"patch deployment {self.settings.vllm_deployment} -p '{{\"spec\":{{\"replicas\":{replicas}}}}}'"
         )
+
         if success:
             logger.info(f"Successfully scaled deployment to {replicas} replicas")
         return success
 
-    async def get_replicas(self) -> tuple[int, int]:
-        """Get current and desired replica counts."""
+    async def get_replicas(self) -> int:
+        """Get desired replica count."""
         cmd = (
             f"get deployment {self.settings.vllm_deployment} "
-            "-o jsonpath='{.status.replicas} {.spec.replicas}'"
+            "-o jsonpath='{.spec.replicas}'"
         )
         success, output = await self.execute(cmd)
         if success and output:
             try:
-                current, desired = map(int, output.split())
-                return current, desired
+                return int(output)
             except ValueError:
-                logger.error(f"Failed to parse replica counts: {output}")
-        return -1, -1
+                logger.error(f"Failed to parse replica count: {output}")
+        return -1
 
     async def deployment_exists(self) -> bool:
         """Check if the vLLM deployment exists."""

k8s-autoscaler/k8s_autoscaler/vllm.py

@@ -56,7 +56,7 @@ async def monitor_inactivity(self):
                 if not await self.kube.deployment_exists():
                     break
 
-                current_replicas, _ = await self.kube.get_replicas()
+                current_replicas = await self.kube.get_replicas()
                 if (
                     time.time() - self.state.last_activity
                     > self.settings.inactivity_timeout