Add retry_with_exponential_backoff method; Use in tonic logs client

Author: AaronRM

opentelemetry-otlp/src/exporter/http/mod.rs

@@ -388,7 +388,7 @@ impl OtlpHttpClient {
         logs: LogBatch<'_>,
     ) -> Result<(Vec<u8>, &'static str, Option<&'static str>), String> {
         use opentelemetry_proto::tonic::collector::logs::v1::ExportLogsServiceRequest;
-        let resource_logs = group_logs_by_resource_and_scope(logs, &self.resource);
+        let resource_logs = group_logs_by_resource_and_scope(&logs, &self.resource);
         let req = ExportLogsServiceRequest { resource_logs };
 
         let (body, content_type) = match self.protocol {

opentelemetry-otlp/src/exporter/tonic/logs.rs

@@ -1,3 +1,4 @@
+use std::sync::Arc;
 use core::fmt;
 use opentelemetry::otel_debug;
 use opentelemetry_proto::tonic::collector::logs::v1::{
@@ -13,6 +14,8 @@ use opentelemetry_proto::transform::logs::tonic::group_logs_by_resource_and_scop
 
 use super::BoxInterceptor;
 
+use crate::retry::{retry_with_exponential_backoff, RetryPolicy};
+
 pub(crate) struct TonicLogsClient {
     inner: Mutex<Option<ClientInner>>,
     #[allow(dead_code)]
@@ -58,41 +61,61 @@ impl TonicLogsClient {
 
 impl LogExporter for TonicLogsClient {
     async fn export(&self, batch: LogBatch<'_>) -> OTelSdkResult {
-        let (mut client, metadata, extensions) = match self.inner.lock().await.as_mut() {
-            Some(inner) => {
-                let (m, e, _) = inner
-                    .interceptor
-                    .call(Request::new(()))
-                    .map_err(|e| OTelSdkError::InternalFailure(format!("error: {e:?}")))?
-                    .into_parts();
-                (inner.client.clone(), m, e)
-            }
-            None => return Err(OTelSdkError::AlreadyShutdown),
+        let policy = RetryPolicy {
+            max_retries: 3,
+            initial_delay_ms: 100,
+            max_delay_ms: 1600,
+            jitter_ms: 100,
         };
 
-        let resource_logs = group_logs_by_resource_and_scope(batch, &self.resource);
+        let batch = Arc::new(batch);
 
-        otel_debug!(name: "TonicLogsClient.ExportStarted");
+        retry_with_exponential_backoff(policy, "TonicLogsClient.Export", {
+            let batch = Arc::clone(&batch);
+            let inner = &self.inner;
+            let resource = &self.resource;
+            move || {
+                let batch = Arc::clone(&batch);
+                Box::pin(async move {
+                    let (mut client, metadata, extensions) = match inner.lock().await.as_mut() {
+                        Some(inner) => {
+                            let (m, e, _) = inner
+                                .interceptor
+                                .call(Request::new(()))
+                                .map_err(|e| OTelSdkError::InternalFailure(format!("error: {e:?}")))?
+                                .into_parts();
+                            (inner.client.clone(), m, e)
+                        }
+                        None => return Err(OTelSdkError::AlreadyShutdown),
+                    };
 
-        let result = client
-            .export(Request::from_parts(
-                metadata,
-                extensions,
-                ExportLogsServiceRequest { resource_logs },
-            ))
-            .await;
+                    let resource_logs = group_logs_by_resource_and_scope(&*batch, resource);
 
-        match result {
-            Ok(_) => {
-                otel_debug!(name: "TonicLogsClient.ExportSucceeded");
-                Ok(())
-            }
-            Err(e) => {
-                let error = format!("export error: {e:?}");
-                otel_debug!(name: "TonicLogsClient.ExportFailed", error = &error);
-                Err(OTelSdkError::InternalFailure(error))
+                    otel_debug!(name: "TonicsLogsClient.ExportStarted");
+
+                    let result = client
+                        .export(Request::from_parts(
+                            metadata,
+                            extensions,
+                            ExportLogsServiceRequest { resource_logs },
+                        ))
+                        .await;
+
+                    match result {
+                        Ok(_) => {
+                            otel_debug!(name: "TonicLogsClient.ExportSucceeded");
+                            Ok(())
+                        }
+                        Err(e) => {
+                            let error = format!("export error: {e:?}");
+                            otel_debug!(name: "TonicLogsClient.ExportFailed", error = &error);
+                            Err(OTelSdkError::InternalFailure(error))
+                        }
+                    }
+                })
             }
-        }
+        })
+        .await
     }
 
     fn shutdown_with_timeout(&self, _timeout: time::Duration) -> OTelSdkResult {

opentelemetry-otlp/src/lib.rs

@@ -365,6 +365,7 @@ mod metric;
 #[cfg(feature = "trace")]
 #[cfg(any(feature = "http-proto", feature = "http-json", feature = "grpc-tonic"))]
 mod span;
+mod retry;
 
 pub use crate::exporter::Compression;
 pub use crate::exporter::ExportConfig;

opentelemetry-otlp/src/retry.rs

@@ -0,0 +1,47 @@
+use std::future::Future;
+use std::time::{Duration, SystemTime};
+use opentelemetry::otel_warn;
+use tokio::time::sleep;
+
+pub(crate) struct RetryPolicy {
+    pub max_retries: usize,
+    pub initial_delay_ms: u64,
+    pub max_delay_ms: u64,
+    pub jitter_ms: u64,
+}
+
+fn generate_jitter(max_jitter: u64) -> u64 {
+    let now = SystemTime::now();
+    let nanos = now.duration_since(SystemTime::UNIX_EPOCH).unwrap().subsec_nanos();
+    nanos as u64 % (max_jitter + 1)
+}
+
+pub(crate) async fn retry_with_exponential_backoff<F, Fut, T, E>(
+    policy: RetryPolicy,
+    operation_name: &str,
+    mut operation: F,
+) -> Result<T, E>
+where
+    F: FnMut() -> Fut,
+    E: std::fmt::Debug,
+    Fut: Future<Output = Result<T, E>>,
+{
+    let mut attempt = 0;
+    let mut delay = policy.initial_delay_ms;
+
+    loop {
+        match operation().await {
+            Ok(result) => return Ok(result),
+            Err(err) if attempt < policy.max_retries => {
+                attempt += 1;
+                // Log the error and retry after a delay with jitter
+                otel_warn!(name: "OtlpRetry", message = format!("Retrying operation {:?} due to error: {:?}", operation_name, err));
+                let jitter = generate_jitter(policy.jitter_ms);
+                let delay_with_jitter = std::cmp::min(delay + jitter, policy.max_delay_ms);
+                sleep(Duration::from_millis(delay_with_jitter)).await;
+                delay = std::cmp::min(delay * 2, policy.max_delay_ms);
+            }
+            Err(err) => return Err(err),
+        }
+    }
+}

opentelemetry-proto/src/transform/logs.rs

@@ -165,8 +165,8 @@ pub mod tonic {
         }
     }
 
-    pub fn group_logs_by_resource_and_scope(
-        logs: LogBatch<'_>,
+    pub fn group_logs_by_resource_and_scope<'a>(
+        logs: &'a LogBatch<'a>,
         resource: &ResourceAttributesWithSchema,
     ) -> Vec<ResourceLogs> {
         // Group logs by target or instrumentation name
@@ -271,7 +271,7 @@ mod tests {
         let resource: ResourceAttributesWithSchema = (&resource).into(); // Convert Resource to ResourceAttributesWithSchema
 
         let grouped_logs =
-            crate::transform::logs::tonic::group_logs_by_resource_and_scope(log_batch, &resource);
+            crate::transform::logs::tonic::group_logs_by_resource_and_scope(&log_batch, &resource);
 
         assert_eq!(grouped_logs.len(), 1);
         let resource_logs = &grouped_logs[0];
@@ -291,7 +291,7 @@ mod tests {
         let log_batch = LogBatch::new(&logs);
         let resource: ResourceAttributesWithSchema = (&resource).into(); // Convert Resource to ResourceAttributesWithSchema
         let grouped_logs =
-            crate::transform::logs::tonic::group_logs_by_resource_and_scope(log_batch, &resource);
+            crate::transform::logs::tonic::group_logs_by_resource_and_scope(&log_batch, &resource);
 
         assert_eq!(grouped_logs.len(), 1);
         let resource_logs = &grouped_logs[0];