From adbf9ef1339d0b4d96bb206d3477be9090cd0509 Mon Sep 17 00:00:00 2001
From: zhuofeng <zhuofeng@yunshan.net.cn>
Date: Mon, 30 Mar 2026 16:37:34 +0800
Subject: [PATCH] fix: Agent is unaware of forced Server restarts, causing
 retransmission timeout latency of ~15 minutes.

---
 agent/src/sender/uniform_sender.rs | 36 ++++++++++++++++++++++++++++++
 1 file changed, 36 insertions(+)

diff --git a/agent/src/sender/uniform_sender.rs b/agent/src/sender/uniform_sender.rs
index 909d712292a..252f3916d9c 100644
--- a/agent/src/sender/uniform_sender.rs
+++ b/agent/src/sender/uniform_sender.rs
@@ -27,6 +27,8 @@ use std::sync::{
 use std::thread::{self, JoinHandle};
 use std::time::{Duration, Instant, SystemTime};
 
+use socket2::TcpKeepalive;
+
 use arc_swap::access::Access;
 use lazy_static::lazy_static;
 use log::{debug, error, info, warn};
@@ -433,6 +435,10 @@ impl<T: Sendable> UniformSender<T> {
     const TCP_WRITE_TIMEOUT: u64 = 3; // s
     const QUEUE_READ_TIMEOUT: u64 = 3; // s
     const DEFAULT_RECONNECT_INTERVAL: u8 = 10; // s
+    const TCP_KEEPALIVE_TIME: u64 = 30; // s, idle time before first keepalive probe
+    const TCP_KEEPALIVE_INTERVAL: u64 = 10; // s, interval between keepalive probes
+    const TCP_KEEPALIVE_RETRIES: u32 = 3; // number of keepalive probes before declaring dead
+    const TCP_USER_TIMEOUT: u64 = 100; // s, max time for unacked data before TCP gives up
 
     pub fn new(
         id: usize,
@@ -596,6 +602,36 @@ impl<T: Sendable> UniformSender<T> {
                     conn.tcp_stream.take();
                     return;
                 }
+                #[cfg(target_os = "linux")]
+                {
+                    // Wrap the raw fd temporarily to configure socket options; forget the
+                    // wrapper so it does not close the fd owned by TcpStream.
+                    let sock = socket2::SockRef::from(&*tcp_stream);
+                    // SO_KEEPALIVE: detect idle dead connections (server went away silently).
+                    // After TCP_KEEPALIVE_TIME seconds of idle the kernel sends probes every
+                    // TCP_KEEPALIVE_INTERVAL seconds; TCP_KEEPALIVE_RETRIES failed probes
+                    // close the connection (~60 s total with these defaults).
+                    let keepalive = TcpKeepalive::new()
+                        .with_time(Duration::from_secs(Self::TCP_KEEPALIVE_TIME))
+                        .with_interval(Duration::from_secs(Self::TCP_KEEPALIVE_INTERVAL))
+                        .with_retries(Self::TCP_KEEPALIVE_RETRIES);
+                    if let Err(e) = sock.set_tcp_keepalive(&keepalive) {
+                        debug!("{} sender tcp stream set keepalive failed {}", self.name, e);
+                    }
+                    // TCP_USER_TIMEOUT: limit how long unacknowledged data may remain in
+                    // flight before the kernel gives up.  This caps the TCP retransmission
+                    // storm that would otherwise last ~15 min (tcp_retries2 = 15) when the
+                    // remote end vanishes without sending RST/FIN.  100 s corresponds
+                    // roughly to tcp_retries2 = 5.
+                    if let Err(e) =
+                        sock.set_tcp_user_timeout(Some(Duration::from_secs(Self::TCP_USER_TIMEOUT)))
+                    {
+                        debug!(
+                            "{} sender tcp stream set user timeout failed {}",
+                            self.name, e
+                        );
+                    }
+                }
                 info!(
                     "{} sender tcp connection to {}:{} succeed.",
                     self.name, conn.dest_ip, conn.dest_port