feat: Frontend and Runtime Request Cancellation Metrics (#7493)

Signed-off-by: Jacky <18255193+kthui@users.noreply.github.com>

lib/llm/src/grpc/service/openai.rs

@@ -21,7 +21,7 @@ use super::kserve::inference;
 // [gluo NOTE] These are common utilities that should be shared between frontends
 use crate::http::service::{
     disconnect::{ConnectionHandle, create_connection_monitor},
-    metrics::{Endpoint, InflightGuard, process_response_and_observe_metrics},
+    metrics::{CancellationLabels, Endpoint, InflightGuard, process_response_and_observe_metrics},
 };
 use dynamo_async_openai::types::{CompletionFinishReason, CreateCompletionRequest, Prompt};
 
@@ -54,14 +54,22 @@ pub async fn completion_response_stream(
     // create the context for the request
     // [WIP] from request id.
     let request_id = get_or_create_request_id(request.inner.user.as_deref());
+    let streaming = request.inner.stream.unwrap_or(false);
+    let cancellation_labels = CancellationLabels {
+        model: request.inner.model.clone(),
+        endpoint: "grpc_completions".to_string(),
+        request_type: if streaming { "stream" } else { "unary" }.to_string(),
+    };
     let request = Context::with_id(request, request_id.clone());
     let context = request.context();
 
     // create the connection handles
-    let (mut connection_handle, stream_handle) =
-        create_connection_monitor(context.clone(), Some(state.metrics_clone())).await;
-
-    let streaming = request.inner.stream.unwrap_or(false);
+    let (mut connection_handle, stream_handle) = create_connection_monitor(
+        context.clone(),
+        Some(state.metrics_clone()),
+        cancellation_labels,
+    )
+    .await;
     // update the request to always stream
     let request = request.map(|mut req| {
         req.inner.stream = Some(true);

lib/llm/src/grpc/service/tensor.rs

@@ -20,7 +20,7 @@ use validator::Validate;
 use crate::http::service::metrics::InflightGuard;
 use crate::http::service::{
     disconnect::{ConnectionHandle, create_connection_monitor},
-    metrics::{Endpoint, process_response_and_observe_metrics},
+    metrics::{CancellationLabels, Endpoint, process_response_and_observe_metrics},
 };
 
 use crate::protocols::tensor;
@@ -60,13 +60,22 @@ pub async fn tensor_response_stream(
 ) -> Result<impl Stream<Item = Annotated<NvCreateTensorResponse>>, Status> {
     // create the context for the request
     let request_id = get_or_create_request_id(request.id.as_deref());
+    let cancellation_labels = CancellationLabels {
+        model: request.model.clone(),
+        endpoint: Endpoint::Tensor.to_string(),
+        request_type: if streaming { "stream" } else { "unary" }.to_string(),
+    };
     let request = Context::with_id(request, request_id.clone());
     let context = request.context();
 
     // [gluo TODO] revisit metrics to properly expose it
     // create the connection handles
-    let (mut connection_handle, stream_handle) =
-        create_connection_monitor(context.clone(), Some(state.metrics_clone())).await;
+    let (mut connection_handle, stream_handle) = create_connection_monitor(
+        context.clone(),
+        Some(state.metrics_clone()),
+        cancellation_labels,
+    )
+    .await;
 
     // todo - make the protocols be optional for model name
     // todo - when optional, if none, apply a default

lib/llm/src/http/service/anthropic.rs

@@ -30,7 +30,7 @@ use tracing::Instrument;
 use super::{
     RouteDoc,
     disconnect::{ConnectionHandle, create_connection_monitor, monitor_for_disconnects},
-    metrics::{Endpoint, process_response_and_observe_metrics},
+    metrics::{CancellationLabels, Endpoint, process_response_and_observe_metrics},
     service_v2,
 };
 use crate::preprocessor::OpenAIPreprocessor;
@@ -125,12 +125,22 @@ async fn handler_anthropic_messages(
 
     // Create request context
     let request_id = get_or_create_request_id(None, &headers);
+    let streaming = request.stream;
+    let cancellation_labels = CancellationLabels {
+        model: request.model.clone(),
+        endpoint: Endpoint::AnthropicMessages.to_string(),
+        request_type: if streaming { "stream" } else { "unary" }.to_string(),
+    };
     let request = Context::with_id(request, request_id);
     let context = request.context();
 
     // Create connection handles
-    let (mut connection_handle, stream_handle) =
-        create_connection_monitor(context.clone(), Some(state.metrics_clone())).await;
+    let (mut connection_handle, stream_handle) = create_connection_monitor(
+        context.clone(),
+        Some(state.metrics_clone()),
+        cancellation_labels,
+    )
+    .await;
 
     let response =
         tokio::spawn(anthropic_messages(state, template, request, stream_handle).in_current_span())

lib/llm/src/http/service/disconnect.rs

@@ -33,7 +33,7 @@ use dynamo_runtime::engine::AsyncEngineContext;
 use futures::{Stream, StreamExt};
 use std::sync::Arc;
 
-use crate::http::service::metrics::{ErrorType, InflightGuard, Metrics};
+use crate::http::service::metrics::{CancellationLabels, ErrorType, InflightGuard, Metrics};
 
 #[derive(Clone, Copy)]
 pub enum ConnectionStatus {
@@ -100,6 +100,7 @@ impl Drop for ConnectionHandle {
 pub async fn create_connection_monitor(
     engine_context: Arc<dyn AsyncEngineContext>,
     metrics: Option<Arc<Metrics>>,
+    cancellation_labels: CancellationLabels,
 ) -> (ConnectionHandle, ConnectionHandle) {
     // these oneshot channels monitor possible disconnects from the client in two different scopes:
     // - the local task (connection_handle)
@@ -113,6 +114,7 @@ pub async fn create_connection_monitor(
         connection_rx,
         stream_rx,
         metrics,
+        cancellation_labels,
     ));
 
     // Two handles, the first is armed, the second is disarmed
@@ -128,6 +130,7 @@ async fn connection_monitor(
     connection_rx: tokio::sync::oneshot::Receiver<ConnectionStatus>,
     stream_rx: tokio::sync::oneshot::Receiver<ConnectionStatus>,
     metrics: Option<Arc<Metrics>>,
+    cancellation_labels: CancellationLabels,
 ) {
     match connection_rx.await {
         Err(_) | Ok(ConnectionStatus::ClosedUnexpectedly) => {
@@ -135,6 +138,7 @@ async fn connection_monitor(
             tracing::trace!("Connection closed unexpectedly; issuing cancellation");
             if let Some(metrics) = &metrics {
                 metrics.inc_client_disconnect();
+                metrics.inc_cancellation(&cancellation_labels);
             }
             engine_context.kill();
         }
@@ -149,6 +153,7 @@ async fn connection_monitor(
             tracing::trace!("Stream closed unexpectedly; issuing cancellation");
             if let Some(metrics) = &metrics {
                 metrics.inc_client_disconnect();
+                metrics.inc_cancellation(&cancellation_labels);
             }
             engine_context.kill();
         }

lib/llm/src/http/service/metrics.rs

@@ -256,6 +256,7 @@ pub struct Metrics {
     model_kv_cache_block_size: IntGaugeVec,
     model_migration_limit: IntGaugeVec,
     model_migration_total: IntCounterVec,
+    model_cancellation_total: IntCounterVec,
 }
 
 // Inflight tracks requests from HTTP handler start until complete response is finished.
@@ -322,6 +323,13 @@ pub enum RequestType {
     Stream,
 }
 
+/// Labels for cancellation metrics
+pub struct CancellationLabels {
+    pub model: String,
+    pub endpoint: String,
+    pub request_type: String,
+}
+
 /// Status
 #[derive(PartialEq)]
 pub enum Status {
@@ -654,6 +662,15 @@ impl Metrics {
         )
         .unwrap();
 
+        let model_cancellation_total = IntCounterVec::new(
+            Opts::new(
+                frontend_metric_name(frontend_service::MODEL_CANCELLATION_TOTAL),
+                "Total number of request cancellations",
+            ),
+            &["model", "endpoint", "request_type"],
+        )
+        .unwrap();
+
         Metrics {
             request_counter,
             inflight_gauge,
@@ -674,6 +691,7 @@ impl Metrics {
             model_kv_cache_block_size,
             model_migration_limit,
             model_migration_total,
+            model_cancellation_total,
         }
     }
 
@@ -778,6 +796,7 @@ impl Metrics {
         registry.register(Box::new(self.model_kv_cache_block_size.clone()))?;
         registry.register(Box::new(self.model_migration_limit.clone()))?;
         registry.register(Box::new(self.model_migration_total.clone()))?;
+        registry.register(Box::new(self.model_cancellation_total.clone()))?;
 
         Ok(())
     }
@@ -861,6 +880,20 @@ impl Metrics {
             .get()
     }
 
+    /// Increment the cancellation counter
+    pub fn inc_cancellation(&self, labels: &CancellationLabels) {
+        self.model_cancellation_total
+            .with_label_values(&[&labels.model, &labels.endpoint, &labels.request_type])
+            .inc();
+    }
+
+    /// Get the current cancellation count
+    pub fn get_cancellation_count(&self, labels: &CancellationLabels) -> u64 {
+        self.model_cancellation_total
+            .with_label_values(&[&labels.model, &labels.endpoint, &labels.request_type])
+            .get()
+    }
+
     /// Create a new [`InflightGuard`] for the given model and annotate if its a streaming request,
     /// and the kind of endpoint that was hit
     ///

lib/llm/src/http/service/openai.rs

@@ -36,7 +36,8 @@ use super::{
     disconnect::{ConnectionHandle, create_connection_monitor, monitor_for_disconnects},
     error::HttpError,
     metrics::{
-        Endpoint, ErrorType, EventConverter, process_response_and_observe_metrics,
+        CancellationLabels, Endpoint, ErrorType, EventConverter,
+        process_response_and_observe_metrics,
         process_response_using_event_converter_and_observe_metrics,
     },
     service_v2,
@@ -341,12 +342,22 @@ async fn handler_completions(
 
     // create the context for the request
     let request_id = get_or_create_request_id(request.inner.user.as_deref(), &headers);
+    let streaming = request.inner.stream.unwrap_or(false);
+    let cancellation_labels = CancellationLabels {
+        model: request.inner.model.clone(),
+        endpoint: Endpoint::Completions.to_string(),
+        request_type: if streaming { "stream" } else { "unary" }.to_string(),
+    };
     let request = Context::with_id(request, request_id);
     let context = request.context();
 
     // create the connection handles
-    let (mut connection_handle, stream_handle) =
-        create_connection_monitor(context.clone(), Some(state.metrics_clone())).await;
+    let (mut connection_handle, stream_handle) = create_connection_monitor(
+        context.clone(),
+        Some(state.metrics_clone()),
+        cancellation_labels,
+    )
+    .await;
 
     // possibly long running task
     // if this returns a streaming response, the stream handle will be armed and captured by the response stream
@@ -789,12 +800,22 @@ async fn handler_chat_completions(
 
     // create the context for the request
     let request_id = get_or_create_request_id(request.inner.user.as_deref(), &headers);
+    let streaming = request.inner.stream.unwrap_or(false);
+    let cancellation_labels = CancellationLabels {
+        model: request.inner.model.clone(),
+        endpoint: Endpoint::ChatCompletions.to_string(),
+        request_type: if streaming { "stream" } else { "unary" }.to_string(),
+    };
     let request = Context::with_id(request, request_id);
     let context = request.context();
 
     // create the connection handles
-    let (mut connection_handle, stream_handle) =
-        create_connection_monitor(context.clone(), Some(state.metrics_clone())).await;
+    let (mut connection_handle, stream_handle) = create_connection_monitor(
+        context.clone(),
+        Some(state.metrics_clone()),
+        cancellation_labels,
+    )
+    .await;
 
     let response =
         tokio::spawn(chat_completions(state, template, request, stream_handle).in_current_span())
@@ -1388,12 +1409,22 @@ async fn handler_responses(
 
     // create the context for the request
     let request_id = get_or_create_request_id(None, &headers);
+    let streaming = request.inner.stream.unwrap_or(false);
+    let cancellation_labels = CancellationLabels {
+        model: request.inner.model.clone().unwrap_or_default(),
+        endpoint: Endpoint::Responses.to_string(),
+        request_type: if streaming { "stream" } else { "unary" }.to_string(),
+    };
     let request = Context::with_id(request, request_id);
     let context = request.context();
 
     // create the connection handles
-    let (mut connection_handle, stream_handle) =
-        create_connection_monitor(context.clone(), Some(state.metrics_clone())).await;
+    let (mut connection_handle, stream_handle) = create_connection_monitor(
+        context.clone(),
+        Some(state.metrics_clone()),
+        cancellation_labels,
+    )
+    .await;
 
     let response =
         tokio::spawn(responses(state, template, request, stream_handle).in_current_span())
@@ -2054,8 +2085,16 @@ async fn video_stream(
     // video_stream returns the streaming body directly (graceful handler exit).
     // The stream_handle is armed below and lives inside the monitored stream so that
     // a client disconnect (body drop) signals the engine context to cancel.
-    let (mut connection_handle, mut stream_handle) =
-        create_connection_monitor(ctx.clone(), Some(state.metrics_clone())).await;
+    let (mut connection_handle, mut stream_handle) = create_connection_monitor(
+        ctx.clone(),
+        Some(state.metrics_clone()),
+        CancellationLabels {
+            model: model.clone(),
+            endpoint: Endpoint::Videos.to_string(),
+            request_type: "stream".to_string(),
+        },
+    )
+    .await;
     connection_handle.disarm();
 
     let mut http_queue_guard = Some(http_queue_guard);

lib/runtime/src/metrics/prometheus_names.rs

@@ -232,6 +232,9 @@ pub mod frontend_service {
     /// Total number of request migrations due to worker unavailability
     pub const MODEL_MIGRATION_TOTAL: &str = "model_migration_total";
 
+    /// Total number of request cancellations
+    pub const MODEL_CANCELLATION_TOTAL: &str = "model_cancellation_total";
+
     /// Active decode blocks (KV cache blocks) per worker
     /// Gauge metric tracking current KV cache block utilization for each worker
     pub const WORKER_ACTIVE_DECODE_BLOCKS: &str = "worker_active_decode_blocks";
@@ -346,6 +349,9 @@ pub mod work_handler {
     /// Total number of errors in work handler processing
     pub const ERRORS_TOTAL: &str = "errors_total";
 
+    /// Total number of requests cancelled by work handler (client stop/kill or disconnect)
+    pub const CANCELLATION_TOTAL: &str = "cancellation_total";
+
     /// Network transit: frontend send to backend receive (wall-clock, cross-process)
     pub const NETWORK_TRANSIT_SECONDS: &str = "network_transit_seconds";
 

lib/runtime/src/pipeline/network/ingress/push_handler.rs

@@ -24,6 +24,7 @@ pub struct WorkHandlerMetrics {
     pub request_bytes: IntCounter,
     pub response_bytes: IntCounter,
     pub error_counter: IntCounterVec,
+    pub cancellation_total: IntCounter,
 }
 
 impl WorkHandlerMetrics {
@@ -34,6 +35,7 @@ impl WorkHandlerMetrics {
         request_bytes: IntCounter,
         response_bytes: IntCounter,
         error_counter: IntCounterVec,
+        cancellation_total: IntCounter,
     ) -> Self {
         Self {
             request_counter,
@@ -42,6 +44,7 @@ impl WorkHandlerMetrics {
             request_bytes,
             response_bytes,
             error_counter,
+            cancellation_total,
         }
     }
 
@@ -90,6 +93,12 @@ impl WorkHandlerMetrics {
             metrics_labels,
         )?;
 
+        let cancellation_total = metrics.create_intcounter(
+            work_handler::CANCELLATION_TOTAL,
+            "Total number of requests cancelled by work handler",
+            metrics_labels,
+        )?;
+
         Ok(Self::new(
             request_counter,
             request_duration,
@@ -97,6 +106,7 @@ impl WorkHandlerMetrics {
             request_bytes,
             response_bytes,
             error_counter,
+            cancellation_total,
         ))
     }
 }
@@ -218,6 +228,7 @@ where
         let mut publisher = tcp::client::TcpClient::create_response_stream(
             request.context(),
             control_msg.connection_info,
+            self.metrics().map(|m| m.cancellation_total.clone()),
         )
         .await
         .map_err(|e| {

lib/runtime/src/pipeline/network/tcp.rs

@@ -124,8 +124,11 @@ mod tests {
         let context_rank1 = Context::with_id((), context_rank0.id().to_string());
 
         // connect to the server socket
-        let mut send_stream =
-            client::TcpClient::create_response_stream(context_rank1.context(), connection_info)
+        let mut send_stream = client::TcpClient::create_response_stream(
+            context_rank1.context(),
+            connection_info,
+            None,
+        )
         .await
         .unwrap();
         println!("Client connected");

lib/runtime/src/pipeline/network/tcp/client.rs

@@ -12,6 +12,8 @@ use tokio::{
 };
 use tokio_util::codec::{FramedRead, FramedWrite};
 
+use prometheus::IntCounter;
+
 use super::{CallHomeHandshake, ControlMessage, TcpStreamConnectionInfo};
 use crate::engine::AsyncEngineContext;
 use crate::pipeline::network::{
@@ -63,6 +65,7 @@ impl TcpClient {
     pub async fn create_response_stream(
         context: Arc<dyn AsyncEngineContext>,
         info: ConnectionInfo,
+        cancellation_counter: Option<IntCounter>,
     ) -> Result<StreamSender> {
         let info =
             TcpStreamConnectionInfo::try_from(info).context("tcp-stream-connection-info-error")?;
@@ -97,7 +100,12 @@ impl TcpClient {
         // captured by the monitor task
         let (alive_tx, alive_rx) = tokio::sync::oneshot::channel::<()>();
 
-        let reader_task = tokio::spawn(handle_reader(framed_reader, context.clone(), alive_tx));
+        let reader_task = tokio::spawn(handle_reader(
+            framed_reader,
+            context.clone(),
+            alive_tx,
+            cancellation_counter,
+        ));
 
         // transport specific handshake message
         let handshake = CallHomeHandshake {
@@ -213,9 +221,11 @@ async fn handle_reader(
     framed_reader: FramedRead<tokio::io::ReadHalf<tokio::net::TcpStream>, TwoPartCodec>,
     context: Arc<dyn AsyncEngineContext>,
     alive_tx: tokio::sync::oneshot::Sender<()>,
+    cancellation_counter: Option<IntCounter>,
 ) -> FramedRead<tokio::io::ReadHalf<tokio::net::TcpStream>, TwoPartCodec> {
     let mut framed_reader = framed_reader;
     let mut alive_tx = alive_tx;
+    let mut cancellation_counted = false;
     loop {
         tokio::select! {
             msg = framed_reader.next() => {
@@ -233,9 +243,17 @@ async fn handle_reader(
 
                                 match msg {
                                     ControlMessage::Stop => {
+                                        if let Some(counter) = &cancellation_counter && !cancellation_counted {
+                                            counter.inc();
+                                            cancellation_counted = true;
+                                        }
                                         context.stop();
                                     }
                                     ControlMessage::Kill => {
+                                        if let Some(counter) = &cancellation_counter && !cancellation_counted {
+                                            counter.inc();
+                                            cancellation_counted = true;
+                                        }
                                         context.kill();
                                     }
                                     ControlMessage::Sentinel => {
@@ -256,6 +274,11 @@ async fn handle_reader(
                     }
                     None => {
                         tracing::debug!("tcp stream closed by server");
+                        // If no Stop/Kill was received, this is a cancellation where frontend
+                        // dropped the connection
+                        if let Some(counter) = &cancellation_counter && !cancellation_counted {
+                            counter.inc();
+                        }
                         break;
                     }
                 }
@@ -768,10 +791,9 @@ mod tests {
 
         // Spawn the reader task
         let controller_clone = controller.clone();
-        let reader_handle =
-            tokio::spawn(
-                async move { handle_reader(framed_reader, controller_clone, alive_tx).await },
-            );
+        let reader_handle = tokio::spawn(async move {
+            handle_reader(framed_reader, controller_clone, alive_tx, None).await
+        });
 
         // Send Stop control message from server
         framed_server
@@ -805,10 +827,9 @@ mod tests {
 
         // Spawn the reader task
         let controller_clone = controller.clone();
-        let reader_handle =
-            tokio::spawn(
-                async move { handle_reader(framed_reader, controller_clone, alive_tx).await },
-            );
+        let reader_handle = tokio::spawn(async move {
+            handle_reader(framed_reader, controller_clone, alive_tx, None).await
+        });
 
         // Send Kill control message from server
         framed_server
@@ -842,7 +863,9 @@ mod tests {
 
         // Spawn the reader task
         let reader_handle =
-            tokio::spawn(async move { handle_reader(framed_reader, controller, alive_tx).await });
+            tokio::spawn(
+                async move { handle_reader(framed_reader, controller, alive_tx, None).await },
+            );
 
         // Drop the alive_rx to close the channel (simulating writer finishing)
         drop(alive_rx);
@@ -869,7 +892,9 @@ mod tests {
 
         // Spawn the reader task
         let reader_handle =
-            tokio::spawn(async move { handle_reader(framed_reader, controller, alive_tx).await });
+            tokio::spawn(
+                async move { handle_reader(framed_reader, controller, alive_tx, None).await },
+            );
 
         // Close the framed server to signal EOF to the client
         framed_server.close().await.unwrap();
@@ -896,10 +921,9 @@ mod tests {
 
         // Spawn the reader task
         let controller_clone = controller.clone();
-        let reader_handle =
-            tokio::spawn(
-                async move { handle_reader(framed_reader, controller_clone, alive_tx).await },
-            );
+        let reader_handle = tokio::spawn(async move {
+            handle_reader(framed_reader, controller_clone, alive_tx, None).await
+        });
 
         // Send multiple Stop messages (first one will stop, subsequent ones are no-ops)
         framed_server
@@ -937,10 +961,9 @@ mod tests {
 
         // Spawn the reader task
         let controller_clone = controller.clone();
-        let reader_handle =
-            tokio::spawn(
-                async move { handle_reader(framed_reader, controller_clone, alive_tx).await },
-            );
+        let reader_handle = tokio::spawn(async move {
+            handle_reader(framed_reader, controller_clone, alive_tx, None).await
+        });
 
         // Send Stop first, then Kill
         framed_server

tests/fault_tolerance/cancellation/test_sglang.py

@@ -20,6 +20,8 @@ from tests.fault_tolerance.cancellation.utils import (
     poll_for_pattern,
     read_streaming_responses,
     send_cancellable_request,
+    verify_frontend_cancellation_metrics,
+    verify_runtime_cancellation_metrics,
 )
 from tests.utils.constants import FAULT_TOLERANCE_MODEL_NAME
 from tests.utils.managed_process import ManagedProcess
@@ -242,7 +244,7 @@ def test_request_cancellation_sglang_aggregated(
                 ),
             ]
 
-            for request_type, description in test_scenarios:
+            for idx, (request_type, description) in enumerate(test_scenarios):
                 logger.info(f"Testing {description.lower()}...")
 
                 # Send the request (non-blocking)
@@ -291,6 +293,17 @@ def test_request_cancellation_sglang_aggregated(
 
                 logger.info(f"{description} detected successfully")
 
+                # Verify cancellation metrics after each scenario
+                verify_frontend_cancellation_metrics(
+                    frontend_port=frontend.frontend_port,
+                    request_type=request_type,
+                    expected_count=1,
+                )
+                verify_runtime_cancellation_metrics(
+                    worker_system_port=worker.system_port,
+                    expected_count=idx + 1,
+                )
+
 
 @pytest.mark.timeout(300)  # 3x average
 @pytest.mark.gpu_2
@@ -396,3 +409,19 @@ def test_request_cancellation_sglang_decode_cancel(
                 logger.info(
                     "Chat completion stream cancellation in decode phase detected successfully"
                 )
+
+                # Verify cancellation metrics
+                verify_frontend_cancellation_metrics(
+                    frontend_port=frontend.frontend_port,
+                    request_type="chat_completion_stream",
+                    expected_count=1,
+                )
+                verify_runtime_cancellation_metrics(
+                    worker_system_port=decode_worker.system_port,
+                    expected_count=1,
+                )
+                verify_runtime_cancellation_metrics(
+                    worker_system_port=prefill_worker.system_port,
+                    expected_count=0,
+                    component="prefill",
+                )

tests/fault_tolerance/cancellation/test_trtllm.py

@@ -22,6 +22,8 @@ from tests.fault_tolerance.cancellation.utils import (
     poll_for_pattern,
     read_streaming_responses,
     send_cancellable_request,
+    verify_frontend_cancellation_metrics,
+    verify_runtime_cancellation_metrics,
 )
 from tests.utils.constants import FAULT_TOLERANCE_MODEL_NAME
 from tests.utils.managed_process import ManagedProcess
@@ -208,7 +210,7 @@ def test_request_cancellation_trtllm_aggregated(
                 ),
             ]
 
-            for request_type, description in test_scenarios:
+            for idx, (request_type, description) in enumerate(test_scenarios):
                 logger.info(f"Testing {description.lower()}...")
 
                 # Send the request (non-blocking)
@@ -248,6 +250,18 @@ def test_request_cancellation_trtllm_aggregated(
 
                 logger.info(f"{description} detected successfully")
 
+                # Verify cancellation metrics after each scenario
+                verify_frontend_cancellation_metrics(
+                    frontend_port=frontend.frontend_port,
+                    request_type=request_type,
+                    expected_count=1,
+                )
+                verify_runtime_cancellation_metrics(
+                    worker_system_port=worker.system_port,
+                    expected_count=idx + 1,
+                    component="tensorrt_llm",
+                )
+
 
 @pytest.mark.timeout(195)  # 3x average
 def test_request_cancellation_trtllm_decode_cancel(
@@ -332,6 +346,23 @@ def test_request_cancellation_trtllm_decode_cancel(
                     "Chat completion stream cancellation in decode phase detected successfully"
                 )
 
+                # Verify cancellation metrics
+                verify_frontend_cancellation_metrics(
+                    frontend_port=frontend.frontend_port,
+                    request_type="chat_completion_stream",
+                    expected_count=1,
+                )
+                verify_runtime_cancellation_metrics(
+                    worker_system_port=decode_worker.system_port,
+                    expected_count=1,
+                    component="tensorrt_llm",
+                )
+                verify_runtime_cancellation_metrics(
+                    worker_system_port=prefill_worker.system_port,
+                    expected_count=0,
+                    component="prefill",
+                )
+
 
 @pytest.mark.timeout(195)  # 3x average
 def test_request_cancellation_trtllm_prefill_cancel(
@@ -424,6 +455,23 @@ def test_request_cancellation_trtllm_prefill_cancel(
                     "Completion request cancellation during prefill phase detected successfully"
                 )
 
+                # Verify cancellation metrics
+                verify_frontend_cancellation_metrics(
+                    frontend_port=frontend.frontend_port,
+                    request_type="completion",
+                    expected_count=1,
+                )
+                verify_runtime_cancellation_metrics(
+                    worker_system_port=decode_worker.system_port,
+                    expected_count=0,
+                    component="tensorrt_llm",
+                )
+                verify_runtime_cancellation_metrics(
+                    worker_system_port=prefill_worker.system_port,
+                    expected_count=1,
+                    component="prefill",
+                )
+
 
 @pytest.mark.xfail(reason="Test fails only on CI", strict=False)
 @pytest.mark.timeout(195)  # 3x average
@@ -523,3 +571,20 @@ def test_request_cancellation_trtllm_kv_transfer_cancel(
                 logger.info(
                     "Workers are functional after cancellation during KV transfer"
                 )
+
+                # Verify cancellation metrics
+                verify_frontend_cancellation_metrics(
+                    frontend_port=frontend.frontend_port,
+                    request_type="completion",
+                    expected_count=1,
+                )
+                verify_runtime_cancellation_metrics(
+                    worker_system_port=decode_worker.system_port,
+                    expected_count=1,
+                    component="tensorrt_llm",
+                )
+                verify_runtime_cancellation_metrics(
+                    worker_system_port=prefill_worker.system_port,
+                    expected_count=0,
+                    component="prefill",
+                )

tests/fault_tolerance/cancellation/test_vllm.py

@@ -21,6 +21,8 @@ from tests.fault_tolerance.cancellation.utils import (
     poll_for_pattern,
     read_streaming_responses,
     send_cancellable_request,
+    verify_frontend_cancellation_metrics,
+    verify_runtime_cancellation_metrics,
 )
 from tests.utils.constants import FAULT_TOLERANCE_MODEL_NAME
 from tests.utils.managed_process import ManagedProcess
@@ -242,7 +244,7 @@ def test_request_cancellation_vllm_aggregated(
                 ),
             ]
 
-            for request_type, description in test_scenarios:
+            for idx, (request_type, description) in enumerate(test_scenarios):
                 logger.info(f"Testing {description.lower()}...")
 
                 # Send the request (non-blocking)
@@ -282,6 +284,17 @@ def test_request_cancellation_vllm_aggregated(
 
                 logger.info(f"{description} detected successfully")
 
+                # Verify cancellation metrics after each scenario
+                verify_frontend_cancellation_metrics(
+                    frontend_port=frontend.frontend_port,
+                    request_type=request_type,
+                    expected_count=1,
+                )
+                verify_runtime_cancellation_metrics(
+                    worker_system_port=worker.system_port,
+                    expected_count=idx + 1,
+                )
+
 
 @pytest.mark.timeout(150)  # 3x average
 @pytest.mark.nightly
@@ -365,6 +378,22 @@ def test_request_cancellation_vllm_decode_cancel(
                     "Chat completion stream cancellation in decode phase detected successfully"
                 )
 
+                # Verify cancellation metrics
+                verify_frontend_cancellation_metrics(
+                    frontend_port=frontend.frontend_port,
+                    request_type="chat_completion_stream",
+                    expected_count=1,
+                )
+                verify_runtime_cancellation_metrics(
+                    worker_system_port=decode_worker.system_port,
+                    expected_count=1,
+                )
+                verify_runtime_cancellation_metrics(
+                    worker_system_port=prefill_worker.system_port,
+                    expected_count=0,
+                    component="prefill",
+                )
+
 
 @pytest.mark.timeout(150)  # 3x average
 @pytest.mark.nightly
@@ -458,3 +487,19 @@ def test_request_cancellation_vllm_prefill_cancel(
                 logger.info(
                     "Completion request cancellation during prefill phase detected successfully"
                 )
+
+                # Verify cancellation metrics
+                verify_frontend_cancellation_metrics(
+                    frontend_port=frontend.frontend_port,
+                    request_type="completion",
+                    expected_count=1,
+                )
+                verify_runtime_cancellation_metrics(
+                    worker_system_port=decode_worker.system_port,
+                    expected_count=0,
+                )
+                verify_runtime_cancellation_metrics(
+                    worker_system_port=prefill_worker.system_port,
+                    expected_count=1,
+                    component="prefill",
+                )

tests/fault_tolerance/cancellation/utils.py

@@ -298,6 +298,170 @@ def read_streaming_responses(
     )
 
 
+def _parse_frontend_cancellation_metric(
+    metrics_text: str, model_name: str, endpoint: str, request_type: str
+) -> int:
+    """
+    Parse the frontend cancellation metric from Prometheus metrics text.
+
+    Args:
+        metrics_text: Raw Prometheus metrics text
+        model_name: The model name label value
+        endpoint: The endpoint label value (e.g. "completions", "chat_completions")
+        request_type: The request_type label value ("stream" or "unary")
+
+    Returns:
+        The metric count, or 0 if not found
+    """
+    for line in metrics_text.splitlines():
+        if not line.startswith("dynamo_frontend_model_cancellation_total{"):
+            continue
+        if (
+            f'endpoint="{endpoint}"' in line
+            and f'model="{model_name}"' in line
+            and f'request_type="{request_type}"' in line
+        ):
+            parts = line.rsplit(None, 1)
+            if len(parts) == 2:
+                try:
+                    return int(float(parts[1]))
+                except ValueError:
+                    pass
+
+    return 0
+
+
+def _parse_runtime_cancellation_metric(
+    metrics_text: str,
+    namespace: str = "dynamo",
+    component: str = "backend",
+    endpoint: str = "generate",
+) -> int:
+    """
+    Parse the runtime cancellation metric from Prometheus metrics text.
+
+    The metric is dynamo_component_cancellation_total with auto-injected
+    labels (dynamo_namespace, dynamo_component, dynamo_endpoint).
+
+    Args:
+        metrics_text: Raw Prometheus metrics text
+        namespace: Expected dynamo_namespace label value
+        component: Expected dynamo_component label value
+        endpoint: Expected dynamo_endpoint label value
+
+    Returns:
+        The metric count, or 0 if not found
+    """
+    for line in metrics_text.splitlines():
+        if not line.startswith("dynamo_component_cancellation_total{"):
+            continue
+        if (
+            f'dynamo_namespace="{namespace}"' in line
+            and f'dynamo_component="{component}"' in line
+            and f'dynamo_endpoint="{endpoint}"' in line
+        ):
+            parts = line.rsplit(None, 1)
+            if len(parts) == 2:
+                try:
+                    return int(float(parts[1]))
+                except ValueError:
+                    pass
+
+    return 0
+
+
+def _resolve_cancellation_labels(request_type: str) -> tuple[str, str]:
+    """
+    Map a test request type to frontend metric labels.
+
+    Args:
+        request_type: One of "completion", "chat_completion", "chat_completion_stream"
+
+    Returns:
+        (endpoint, request_type_label) tuple
+    """
+    mapping = {
+        "completion": ("completions", "unary"),
+        "chat_completion": ("chat_completions", "unary"),
+        "chat_completion_stream": ("chat_completions", "stream"),
+    }
+    if request_type not in mapping:
+        pytest.fail(f"Unknown request type: {request_type}")
+    return mapping[request_type]
+
+
+def verify_frontend_cancellation_metrics(
+    frontend_port: int,
+    request_type: str,
+    expected_count: int = 0,
+) -> None:
+    """
+    Verify frontend cancellation metrics.
+
+    Args:
+        frontend_port: Port where the frontend /metrics is served
+        request_type: The test request type ("completion", "chat_completion", "chat_completion_stream")
+        expected_count: Expected cancellation count for this request type
+    """
+    endpoint, req_type_label = _resolve_cancellation_labels(request_type)
+
+    frontend_metrics_url = f"http://localhost:{frontend_port}/metrics"
+    try:
+        response = requests.get(frontend_metrics_url, timeout=5)
+        response.raise_for_status()
+    except requests.RequestException as e:
+        pytest.fail(
+            f"Failed to fetch frontend metrics from {frontend_metrics_url}: {e}"
+        )
+
+    frontend_text = response.text
+    count = _parse_frontend_cancellation_metric(
+        frontend_text, FAULT_TOLERANCE_MODEL_NAME, endpoint, req_type_label
+    )
+
+    logger.info(
+        f"Frontend cancellation metrics - endpoint={endpoint}, "
+        f"request_type={req_type_label}: {count}"
+    )
+
+    assert count == expected_count, (
+        f"Frontend: expected {expected_count} cancellations "
+        f"for endpoint={endpoint}, request_type={req_type_label}, "
+        f"but got {count}"
+    )
+
+
+def verify_runtime_cancellation_metrics(
+    worker_system_port: int,
+    expected_count: int = 0,
+    component: str = "backend",
+) -> None:
+    """
+    Verify runtime (worker) cancellation metrics.
+
+    Args:
+        worker_system_port: Port where the worker /metrics is served
+        expected_count: Expected cumulative cancellation count
+        component: The dynamo_component label value (e.g. "backend", "prefill")
+    """
+    worker_metrics_url = f"http://localhost:{worker_system_port}/metrics"
+    try:
+        response = requests.get(worker_metrics_url, timeout=5)
+        response.raise_for_status()
+    except requests.RequestException as e:
+        pytest.fail(f"Failed to fetch worker metrics from {worker_metrics_url}: {e}")
+
+    worker_text = response.text
+    count = _parse_runtime_cancellation_metric(worker_text, component=component)
+
+    logger.info(f"Runtime cancellation metrics (component={component}): {count}")
+
+    assert count == expected_count, (
+        f"Runtime (component={component}): expected {expected_count} cancellations, "
+        f"but got {count}"
+    )
+
+
 def read_log_content(log_path: str | None) -> str:
     """Read log content from a file"""
     if log_path is None: