Merge pull request #1053 from Concordium/fix/load-shed

Add load-shedding to the GRPC V2 service.
Concordium · Oct 13, 2023 · f0f3eca · f0f3eca
2 parents 91a7f8a + d472d01
commit f0f3eca
Show file tree

Hide file tree

Showing 4 changed files with 50 additions and 20 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -2,6 +2,13 @@
 
 ## Unreleased changes
 
+## 6.1.7
+
+ - Add load-shedding to the V2 GRPC API. In particular, if at the time of the
+  request the node is already handling more than
+  `CONCORDIUM_NODE_GRPC2_MAX_CONCURRENT_REQUESTS` requests then the incoming
+  request will be immediately rejected.
+
 ## 6.1.6
 
 - Fix a regression in the start up time. When upgrading from an earlier version, the first start-up

diff --git a/concordium-node/Cargo.lock b/concordium-node/Cargo.lock
diff --git a/concordium-node/Cargo.toml b/concordium-node/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "concordium_node"
-version = "6.1.6" # must be kept in sync with 'is_compatible_version' in 'src/configuration.rs'
+version = "6.1.7" # must be kept in sync with 'is_compatible_version' in 'src/configuration.rs'
 description = "Concordium Node"
 authors = ["Concordium <[email protected]>"]
 exclude = [".gitignore", ".gitlab-ci.yml", "test/**/*","**/**/.gitignore","**/**/.gitlab-ci.yml"]
@@ -78,7 +78,7 @@ tempfile = { version = "3.1" }
 tonic = { version = "0.9", features = ["tls"] }
 tonic-reflection = "0.9"
 tower-http = { version = "0.4", features = ["trace", "metrics"] }
-tower = "0.4"
+tower = {version = "0.4", features = ["load-shed"]}
 tonic-web = "0.9"
 prost = "0.11"
 tokio = { version = "1.20", features = ["macros", "rt-multi-thread", "signal", "io-util", "time"] }

diff --git a/concordium-node/src/grpc2.rs b/concordium-node/src/grpc2.rs
@@ -1064,12 +1064,18 @@ pub mod server {
                     .http2_keepalive_timeout(Some(std::time::Duration::from_secs(
                         config.keepalive_timeout,
                     )))
-                    .layer(log_layer)
-                    .layer(tower::limit::ConcurrencyLimitLayer::new(config.max_concurrent_requests))
-                    // Note: the in-flight request layer applies after the limit layer. This is what we want so that the
-                    // metric reflects the actual number of in-flight requests.
+                    // Note: the in-flight request layer applies first here. Since we are using a load-shed
+                    // layer just below this corresponds very directly to the number of requests being actually handled.
+                    // The technical reason for this is that we cannot really stack the in flight requests layer
+                    // below the stats layer since we want to transform some `Err` responses in the stats layer
+                    // to Ok responses with a meaningful gRPC status code,
+                    // but since the in flight request layer adds a guard to count in-flight requests this would
+                    // mean we'd have to construct such a guard in the response, which is not possible.
                     .layer(in_flight_request_layer)
-                    .layer(stats_layer);
+                    .layer(stats_layer)
+                    .layer(tower::load_shed::LoadShedLayer::new())
+                    .layer(tower::limit::ConcurrencyLimitLayer::new(config.max_concurrent_requests))
+                    .layer(log_layer);
                 if let Some(identity) = identity {
                     builder = builder
                         .tls_config(ServerTlsConfig::new().identity(identity))
@@ -2560,19 +2566,21 @@ fn get_grpc_code_label(code: tonic::Code) -> &'static str {
 /// Actual middleware implementation updating the stats.
 /// The middleware is called once for each gRPC request, even for the streaming
 /// gRPC methods.
-impl<S> tower::Service<hyper::Request<hyper::Body>> for StatsMiddleware<S>
+impl<S, Body: hyper::body::HttpBody + Unpin + Send + Default>
+    tower::Service<hyper::Request<hyper::Body>> for StatsMiddleware<S>
 where
     S: tower::Service<
             hyper::Request<hyper::Body>,
-            Response = hyper::Response<tonic::body::BoxBody>,
+            Response = hyper::Response<Body>,
+            Error = tower::BoxError,
         > + Clone
         + Send
         + 'static,
     S::Future: Send + 'static,
 {
-    type Error = S::Error;
+    type Error = tower::BoxError;
     type Future = futures::future::BoxFuture<'static, Result<Self::Response, Self::Error>>;
-    type Response = S::Response;
+    type Response = hyper::Response<Body>;
 
     fn poll_ready(
         &mut self,
@@ -2599,15 +2607,30 @@ where
                 // Time taken for the inner service to send back a response, meaning for
                 // streaming gRPC methods this is the duration for it to first return a stream.
                 let duration = request_received.elapsed().as_secs_f64();
-                if result.is_err() {
-                    grpc_request_duration
-                        .with_label_values(&[
-                            endpoint_name.as_str(),
-                            get_grpc_code_label(tonic::Code::Internal),
-                        ])
-                        .observe(duration);
+                match result {
+                    Err(e) => {
+                        // If the load shed service terminated the request this will be signalled as
+                        // an Err(Overloaded). So record resource exhaustion
+                        // in the metrics.
+                        let (code, response) = if e.is::<tower::load_shed::error::Overloaded>() {
+                            // return a response with empty body of the correct type. `to_http`
+                            // constructs a response with a `BoxBody` but
+                            // here we need a more general one to make the service generic enough.
+                            let new_response =
+                                tonic::Status::resource_exhausted("Too many concurrent requests.")
+                                    .to_http()
+                                    .map(|_| Default::default());
+                            (tonic::Code::ResourceExhausted, Ok(new_response))
+                        } else {
+                            (tonic::Code::Internal, Err(e))
+                        };
+                        grpc_request_duration
+                            .with_label_values(&[endpoint_name.as_str(), get_grpc_code_label(code)])
+                            .observe(duration);
+                        return response;
+                    }
+                    Ok(result) => (result, duration),
                 }
-                (result?, duration)
             };
 
             // Check if the gRPC status header is part of the HTTP headers, if not check for