feat(metrics): monitor s3 sdk retry

grafana/risingwave-dev-dashboard.dashboard.py

@@ -555,6 +555,21 @@ def section_object_storage(outer_panels):
                         )
                     ],
                 ),
+                panels.timeseries_ops(
+                    "Operation Retry Rate",
+                    "",
+                    [
+                        panels.target(
+                            f"sum(irate({metric('aws_sdk_retry_counts')}[$__rate_interval])) by (instance, job, type)",
+                            "{{type}} - {{job}} @ {{instance}}",
+                        ),
+
+                        panels.target(
+                            f"sum(irate({metric('s3_read_request_retry_count')}[$__rate_interval])) by (instance, job, type)",
+                            "{{type}} - {{job}} @ {{instance}}",
+                        )
+                    ],
+                ),
                 panels.timeseries_dollar(
                     "Estimated S3 Cost (Realtime)",
                     "There are two types of operations: 1. GET, SELECT, and DELETE, they cost 0.0004 USD per 1000 "

src/cmd/Cargo.toml

@@ -21,6 +21,7 @@ normal = ["workspace-hack", "workspace-config", "task_stats_alloc"]
 [dependencies]
 anyhow = "1"
 clap = { version = "4", features = ["derive"] }
+prometheus = { version = "0.13" }
 risingwave_common = { path = "../common" }
 risingwave_compactor = { path = "../storage/compactor" }
 risingwave_compute = { path = "../compute" }

src/cmd/src/lib.rs

@@ -31,37 +31,39 @@ macro_rules! main {
         #[cfg_attr(coverage, no_coverage)]
         fn main() {
             let opts = clap::Parser::parse();
-            $crate::$component(opts);
+            let registry = prometheus::Registry::new();
+            $crate::$component(opts, registry);
         }
     };
 }
 
 // Entry point functions.
 
-pub fn compute(opts: ComputeNodeOpts) {
+pub fn compute(opts: ComputeNodeOpts, registry: prometheus::Registry) {
     risingwave_rt::init_risingwave_logger(
         risingwave_rt::LoggerSettings::new().enable_tokio_console(false),
+        registry.clone(),
     );
-    risingwave_rt::main_okk(risingwave_compute::start(opts));
+    risingwave_rt::main_okk(risingwave_compute::start(opts, registry));
 }
 
-pub fn meta(opts: MetaNodeOpts) {
-    risingwave_rt::init_risingwave_logger(risingwave_rt::LoggerSettings::new());
+pub fn meta(opts: MetaNodeOpts, registry: prometheus::Registry) {
+    risingwave_rt::init_risingwave_logger(risingwave_rt::LoggerSettings::new(), registry);
     risingwave_rt::main_okk(risingwave_meta::start(opts));
 }
 
-pub fn frontend(opts: FrontendOpts) {
-    risingwave_rt::init_risingwave_logger(risingwave_rt::LoggerSettings::new());
+pub fn frontend(opts: FrontendOpts, registry: prometheus::Registry) {
+    risingwave_rt::init_risingwave_logger(risingwave_rt::LoggerSettings::new(), registry);
     risingwave_rt::main_okk(risingwave_frontend::start(opts));
 }
 
-pub fn compactor(opts: CompactorOpts) {
-    risingwave_rt::init_risingwave_logger(risingwave_rt::LoggerSettings::new());
+pub fn compactor(opts: CompactorOpts, registry: prometheus::Registry) {
+    risingwave_rt::init_risingwave_logger(risingwave_rt::LoggerSettings::new(), registry);
     risingwave_rt::main_okk(risingwave_compactor::start(opts));
 }
 
-pub fn ctl(opts: CtlOpts) {
-    risingwave_rt::init_risingwave_logger(risingwave_rt::LoggerSettings::new());
+pub fn ctl(opts: CtlOpts, registry: prometheus::Registry) {
+    risingwave_rt::init_risingwave_logger(risingwave_rt::LoggerSettings::new(), registry);
 
     // Note: Use a simple current thread runtime for ctl.
     // When there's a heavy workload, multiple thread runtime seems to respond slowly. May need

src/cmd_all/Cargo.toml

@@ -23,6 +23,7 @@ anyhow = "1"
 clap = { version = "4", features = ["cargo", "derive"] }
 console = "0.15"
 const-str = "0.5"
+prometheus = { version = "0.13" }
 risingwave_cmd = { path = "../cmd" }
 risingwave_common = { path = "../common" }
 risingwave_compactor = { path = "../storage/compactor" }

src/cmd_all/src/bin/risingwave.rs

@@ -61,14 +61,14 @@ impl Component {
         fn parse_opts<T: FromArgMatches>(matches: &ArgMatches) -> T {
             T::from_arg_matches(matches).map_err(|e| e.exit()).unwrap()
         }
-
+        let registry = prometheus::Registry::new();
         match self {
-            Self::Compute => compute(parse_opts(matches)),
-            Self::Meta => meta(parse_opts(matches)),
-            Self::Frontend => frontend(parse_opts(matches)),
-            Self::Compactor => compactor(parse_opts(matches)),
-            Self::Ctl => ctl(parse_opts(matches)),
-            Self::Playground => playground(parse_opts(matches)),
+            Self::Compute => compute(parse_opts(matches), registry),
+            Self::Meta => meta(parse_opts(matches), registry),
+            Self::Frontend => frontend(parse_opts(matches), registry),
+            Self::Compactor => compactor(parse_opts(matches), registry),
+            Self::Ctl => ctl(parse_opts(matches), registry),
+            Self::Playground => playground(parse_opts(matches), registry),
         }
     }
 
@@ -141,11 +141,11 @@ fn main() -> Result<()> {
     Ok(())
 }
 
-fn playground(opts: PlaygroundOpts) {
+fn playground(opts: PlaygroundOpts, registry: prometheus::Registry) {
     let settings = risingwave_rt::LoggerSettings::new()
         .enable_tokio_console(false)
         .with_target("risingwave_storage", Level::WARN);
-    risingwave_rt::init_risingwave_logger(settings);
+    risingwave_rt::init_risingwave_logger(settings, registry);
     risingwave_rt::main_okk(risingwave_cmd_all::playground(opts)).unwrap();
 }
 

src/cmd_all/src/playground.rs

@@ -215,8 +215,9 @@ pub async fn playground(opts: PlaygroundOpts) -> Result<()> {
                 opts.insert(0, "compute-node".into());
                 tracing::info!("starting compute-node thread with cli args: {:?}", opts);
                 let opts = risingwave_compute::ComputeNodeOpts::parse_from(opts);
-                let _compute_handle =
-                    tokio::spawn(async move { risingwave_compute::start(opts).await });
+                let _compute_handle = tokio::spawn(async move {
+                    risingwave_compute::start(opts, prometheus::Registry::new()).await
+                });
             }
             RisingWaveService::Frontend(mut opts) => {
                 opts.insert(0, "frontend-node".into());

src/common/Cargo.toml

@@ -89,6 +89,7 @@ tokio = { version = "0.2", package = "madsim-tokio", features = [
 toml = "0.7"
 tonic = { version = "0.2", package = "madsim-tonic" }
 tracing = "0.1"
+tracing-subscriber = "0.3.16"
 twox-hash = "1"
 url = "2"
 uuid = "1.2.2"

src/common/src/metrics.rs

@@ -13,8 +13,12 @@
 // limitations under the License.
 
 use hytra::TrAdder;
-use prometheus::core::{Atomic, GenericGauge};
-
+use prometheus::core::{Atomic, AtomicU64, GenericCounter, GenericGauge};
+use prometheus::{register_int_counter_with_registry, Registry};
+use tracing::Subscriber;
+use tracing_subscriber::layer::Context;
+use tracing_subscriber::registry::LookupSpan;
+use tracing_subscriber::Layer;
 pub struct TrAdderAtomic(TrAdder<i64>);
 
 impl Atomic for TrAdderAtomic {
@@ -44,3 +48,49 @@ impl Atomic for TrAdderAtomic {
 }
 
 pub type TrAdderGauge = GenericGauge<TrAdderAtomic>;
+
+pub struct CustomLayer {
+    pub aws_sdk_retry_counts: GenericCounter<AtomicU64>,
+    pub aws_http_timeout_retry_counts: GenericCounter<AtomicU64>,
+}
+
+impl<S> Layer<S> for CustomLayer
+where
+    S: Subscriber + for<'a> LookupSpan<'a>,
+{
+    fn on_event(&self, event: &tracing::Event<'_>, _ctx: Context<'_, S>) {
+        if event.metadata().target() == "aws_smithy_client::retry"
+            && event.metadata().level() == &tracing::Level::DEBUG
+        {
+            self.aws_sdk_retry_counts.inc();
+        }
+
+        if event.metadata().target() == "http_timeout_retry"
+            && event.metadata().level() == &tracing::Level::DEBUG
+        {
+            self.aws_http_timeout_retry_counts.inc();
+        }
+    }
+}
+
+impl CustomLayer {
+    pub fn new(registry: Registry) -> Self {
+        let aws_sdk_retry_counts = register_int_counter_with_registry!(
+            "aws_sdk_retry_counts",
+            "Total number of aws sdk retry happens",
+            registry
+        )
+        .unwrap();
+
+        let aws_http_timeout_retry_counts = register_int_counter_with_registry!(
+            "aws_http_timeout_retry_counts",
+            "Total number of s3 http timeout retry happens",
+            registry
+        )
+        .unwrap();
+        Self {
+            aws_sdk_retry_counts,
+            aws_http_timeout_retry_counts,
+        }
+    }
+}

src/compute/src/lib.rs

@@ -178,7 +178,10 @@ use std::pin::Pin;
 use crate::server::compute_node_serve;
 
 /// Start compute node
-pub fn start(opts: ComputeNodeOpts) -> Pin<Box<dyn Future<Output = ()> + Send>> {
+pub fn start(
+    opts: ComputeNodeOpts,
+    registry: prometheus::Registry,
+) -> Pin<Box<dyn Future<Output = ()> + Send>> {
     // WARNING: don't change the function signature. Making it `async fn` will cause
     // slow compile in release mode.
     Box::pin(async move {
@@ -200,7 +203,7 @@ pub fn start(opts: ComputeNodeOpts) -> Pin<Box<dyn Future<Output = ()> + Send>>
         tracing::info!("advertise addr is {}", advertise_addr);
 
         let (join_handle_vec, _shutdown_send) =
-            compute_node_serve(listen_addr, advertise_addr, opts).await;
+            compute_node_serve(listen_addr, advertise_addr, opts, registry).await;
 
         for join_handle in join_handle_vec {
             join_handle.await.unwrap();

src/compute/src/server.rs

@@ -81,6 +81,7 @@ pub async fn compute_node_serve(
     listen_addr: SocketAddr,
     advertise_addr: HostAddr,
     opts: ComputeNodeOpts,
+    registry: prometheus::Registry,
 ) -> (Vec<JoinHandle<()>>, Sender<()>) {
     // Load the configuration.
     let config = load_config(&opts.config_path, Some(opts.override_config.clone()));
@@ -161,7 +162,7 @@ pub async fn compute_node_serve(
 
     let mut sub_tasks: Vec<(JoinHandle<()>, Sender<()>)> = vec![];
     // Initialize the metrics subsystem.
-    let registry = prometheus::Registry::new();
+
     monitor_process(&registry).unwrap();
     let source_metrics = Arc::new(SourceMetrics::new(registry.clone()));
     let hummock_metrics = Arc::new(HummockMetrics::new(registry.clone()));

src/object_store/Cargo.toml

@@ -33,6 +33,7 @@ tokio = { version = "0.2", package = "madsim-tokio", features = [
 ] }
 tokio-retry = "0.3"
 tracing = "0.1"
+tracing-subscriber = { version = "0.3", features = ["env-filter"] }
 # This crate is excluded from hakari (see hakari.toml) after hdfs is introduced...
 #
 # [target.'cfg(not(madsim))'.dependencies]

src/object_store/src/object/object_metrics.rs

@@ -25,6 +25,7 @@ pub struct ObjectStoreMetrics {
     pub operation_latency: HistogramVec,
     pub operation_size: HistogramVec,
     pub failure_count: GenericCounterVec<AtomicU64>,
+    pub request_retry_count: GenericCounterVec<AtomicU64>,
 }
 
 impl ObjectStoreMetrics {
@@ -81,12 +82,21 @@ impl ObjectStoreMetrics {
         )
         .unwrap();
 
+        let request_retry_count = register_int_counter_vec_with_registry!(
+            "s3_read_request_retry_count",
+            "The number of retry times of object store request",
+            &["type"],
+            registry
+        )
+        .unwrap();
+
         Self {
             write_bytes,
             read_bytes,
             operation_latency,
             operation_size,
             failure_count,
+            request_retry_count,
         }
     }