diff --git a/sgl-model-gateway/src/core/worker_registry.rs b/sgl-model-gateway/src/core/worker_registry.rs
index be79597719f5..f78663427c59 100644
--- a/sgl-model-gateway/src/core/worker_registry.rs
+++ b/sgl-model-gateway/src/core/worker_registry.rs
@@ -8,6 +8,7 @@ use dashmap::DashMap;
 use uuid::Uuid;
 
 use crate::core::{ConnectionMode, RuntimeType, Worker, WorkerType};
+use crate::observability::metrics::RouterMetrics;
 
 /// Unique identifier for a worker
 #[derive(Debug, Clone, Hash, Eq, PartialEq)]
@@ -138,8 +139,8 @@ impl WorkerRegistry {
                 conn_workers.retain(|id| id != worker_id);
             }
 
-            // TODO we may even remove it from Prometheus exports
             worker.set_healthy(false);
+            RouterMetrics::remove_worker_metrics(worker.url());
 
             Some(worker)
         } else {
diff --git a/sgl-model-gateway/src/observability/metrics.rs b/sgl-model-gateway/src/observability/metrics.rs
index 2fa4fcb6641b..c8d8ba11b1a2 100644
--- a/sgl-model-gateway/src/observability/metrics.rs
+++ b/sgl-model-gateway/src/observability/metrics.rs
@@ -536,6 +536,15 @@ impl RouterMetrics {
         .increment(1);
     }
 
+    // TODO delete the metrics (instead of setting them to zero)
+    pub fn remove_worker_metrics(worker_url: &str) {
+        gauge!("sgl_router_cb_state","worker" => worker_url.to_string()).set(0.0);
+        gauge!("sgl_router_worker_health","worker" => worker_url.to_string()).set(0.0);
+        gauge!("sgl_router_worker_load","worker" => worker_url.to_string()).set(0.0);
+        gauge!("sgl_router_running_requests","worker" => worker_url.to_string()).set(0.0);
+        gauge!("sgl_router_tree_size","worker" => worker_url.to_string()).set(0.0);
+    }
+
     pub fn set_job_queue_depth(depth: usize) {
         gauge!("sgl_router_job_queue_depth").set(depth as f64);
     }