fix(router): make HTTP pool idle timeout configurable

docs/advanced_features/sgl_model_gateway.md

@@ -593,6 +593,14 @@ Response:
 
 ## Reliability and Flow Control
 
+### HTTP Client Pool
+
+Configure the idle timeout for pooled upstream HTTP connections:
+
+| Parameter | Default | Description |
+|-----------|---------|-------------|
+| `--pool-idle-timeout-secs` | 50 | Idle timeout in seconds for pooled upstream HTTP connections. Can also be set with `SMG_POOL_IDLE_TIMEOUT_SECS`. |
+
 ### Retries
 
 Configure exponential backoff retries:

docs_new/docs/advanced_features/sgl_model_gateway.mdx

@@ -944,6 +944,32 @@ Response:
 ***
 ## Reliability and Flow Control
 
+### HTTP Client Pool
+
+Configure the idle timeout for pooled upstream HTTP connections:
+
+<table style={{width: "100%", borderCollapse: "collapse", tableLayout: "fixed"}}>
+  <colgroup>
+    <col style={{width: "34%"}} />
+    <col style={{width: "33%"}} />
+    <col style={{width: "33%"}} />
+  </colgroup>
+  <thead>
+    <tr style={{borderBottom: "2px solid #d55816"}}>
+      <th style={{textAlign: "left", padding: "10px 12px", fontWeight: 700, whiteSpace: "nowrap", backgroundColor: "rgba(255,255,255,0.02)"}}>Parameter</th>
+      <th style={{textAlign: "left", padding: "10px 12px", fontWeight: 700, whiteSpace: "nowrap", backgroundColor: "rgba(255,255,255,0.05)"}}>Default</th>
+      <th style={{textAlign: "left", padding: "10px 12px", fontWeight: 700, whiteSpace: "nowrap", backgroundColor: "rgba(255,255,255,0.02)"}}>Description</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>`--pool-idle-timeout-secs`</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>50</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>Idle timeout in seconds for pooled upstream HTTP connections. Can also be set with `SMG_POOL_IDLE_TIMEOUT_SECS`.</td>
+    </tr>
+  </tbody>
+</table>
+
 ### Retries
 
 Configure exponential backoff retries:

sgl-model-gateway/README.md

@@ -726,6 +726,7 @@ Router flags map to these values:
 - `--redis-retention-days` (env: `REDIS_RETENTION_DAYS`). Set to `-1` for persistent storage (default: 30 days).
 
 ## Reliability & Flow Control
+- **HTTP Client Pool**: Upstream HTTP connection pool idle timeout defaults to 50 seconds. Configure via `--pool-idle-timeout-secs` or `SMG_POOL_IDLE_TIMEOUT_SECS`.
 - **Retries**: Default max retries = 5 with exponential backoff (`--retry-max-retries`, `--retry-initial-backoff-ms`, `--retry-max-backoff-ms`, `--retry-backoff-multiplier`, `--retry-jitter-factor`). Retries trigger on 408/429/500/502/503/504.
 - **Circuit Breakers**: Per worker thresholds (`--cb-failure-threshold`, `--cb-success-threshold`, `--cb-timeout-duration-secs`, `--cb-window-duration-secs`). Disable via `--disable-circuit-breaker`.
 - **Rate Limiting**: Token bucket driven by `--max-concurrent-requests`. Set `--rate-limit-tokens-per-second` to override refill rate. Configure request queue via `--queue-size` and `--queue-timeout-secs`; queued requests observe FIFO order and respect cancellation.

sgl-model-gateway/src/app_context.rs

@@ -329,7 +329,7 @@ impl AppContextBuilder {
         let has_tls_config = config.client_identity.is_some() || !config.ca_certificates.is_empty();
 
         let mut client_builder = Client::builder()
-            .pool_idle_timeout(Some(Duration::from_secs(50)))
+            .pool_idle_timeout(Some(Duration::from_secs(config.pool_idle_timeout_secs)))
             .pool_max_idle_per_host(500)
             .timeout(Duration::from_secs(timeout_secs))
             .connect_timeout(Duration::from_secs(10))

sgl-model-gateway/src/config/builder.rs

@@ -187,6 +187,11 @@ impl RouterConfigBuilder {
         self
     }
 
+    pub fn pool_idle_timeout_secs(mut self, timeout: u64) -> Self {
+        self.config.pool_idle_timeout_secs = timeout;
+        self
+    }
+
     // ==================== Rate Limiting ====================
 
     pub fn max_concurrent_requests(mut self, max: i32) -> Self {

sgl-model-gateway/src/config/types.rs

@@ -7,6 +7,8 @@ use serde::{Deserialize, Serialize};
 use super::ConfigResult;
 use crate::core::ConnectionMode;
 
+pub const DEFAULT_POOL_IDLE_TIMEOUT_SECS: u64 = 50;
+
 /// Main router configuration
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct RouterConfig {
@@ -28,6 +30,8 @@ pub struct RouterConfig {
     pub log_dir: Option<String>,
     pub log_level: Option<String>,
     pub request_id_headers: Option<Vec<String>>,
+    #[serde(default = "default_pool_idle_timeout_secs")]
+    pub pool_idle_timeout_secs: u64,
     /// Set to -1 to disable rate limiting
     pub max_concurrent_requests: i32,
     pub queue_size: usize,
@@ -119,6 +123,10 @@ fn default_l1_max_memory() -> usize {
     50 * 1024 * 1024 // 50MB
 }
 
+fn default_pool_idle_timeout_secs() -> u64 {
+    DEFAULT_POOL_IDLE_TIMEOUT_SECS
+}
+
 impl TokenizerCacheConfig {
     /// Returns Some(self) if any caching is enabled, None otherwise.
     /// Use this when passing cache config to tokenizer registration workflow.
@@ -492,6 +500,7 @@ impl Default for RouterConfig {
             log_dir: None,
             log_level: None,
             request_id_headers: None,
+            pool_idle_timeout_secs: default_pool_idle_timeout_secs(),
             max_concurrent_requests: -1,
             queue_size: 100,
             queue_timeout_secs: 60,
@@ -613,6 +622,10 @@ mod tests {
         assert!(config.trace_config.is_none());
         assert!(config.log_dir.is_none());
         assert!(config.log_level.is_none());
+        assert_eq!(
+            config.pool_idle_timeout_secs,
+            DEFAULT_POOL_IDLE_TIMEOUT_SECS
+        );
     }
 
     #[test]
@@ -662,6 +675,22 @@ mod tests {
         assert!(deserialized.trace_config.is_none());
     }
 
+    #[test]
+    fn test_router_config_pool_idle_timeout_deserialization_default() {
+        let config = RouterConfig::default();
+        let mut json = serde_json::to_value(&config).unwrap();
+        json.as_object_mut()
+            .unwrap()
+            .remove("pool_idle_timeout_secs");
+
+        let deserialized: RouterConfig = serde_json::from_value(json).unwrap();
+
+        assert_eq!(
+            deserialized.pool_idle_timeout_secs,
+            default_pool_idle_timeout_secs()
+        );
+    }
+
     #[test]
     fn test_routing_mode_is_pd_mode() {
         let regular = RoutingMode::Regular {

sgl-model-gateway/src/main.rs

@@ -8,7 +8,7 @@ use smg::{
         CircuitBreakerConfig, ConfigError, ConfigResult, DiscoveryConfig, HealthCheckConfig,
         HistoryBackend, ManualAssignmentMode, MetricsConfig, OracleConfig, PolicyConfig,
         PostgresConfig, RedisConfig, RetryConfig, RouterConfig, RoutingMode, TokenizerCacheConfig,
-        TraceConfig,
+        TraceConfig, DEFAULT_POOL_IDLE_TIMEOUT_SECS,
     },
     core::ConnectionMode,
     observability::{
@@ -298,6 +298,16 @@ struct CliArgs {
     #[arg(long, num_args = 0.., help_heading = "Request Handling")]
     cors_allowed_origins: Vec<String>,
 
+    // ==================== HTTP Client Pool ====================
+    /// Idle timeout in seconds for pooled upstream HTTP connections
+    #[arg(
+        long,
+        env = "SMG_POOL_IDLE_TIMEOUT_SECS",
+        default_value_t = DEFAULT_POOL_IDLE_TIMEOUT_SECS,
+        help_heading = "HTTP Client Pool"
+    )]
+    pool_idle_timeout_secs: u64,
+
     // ==================== Rate Limiting ====================
     /// Maximum concurrent requests (-1 to disable)
     #[arg(long, default_value_t = -1, help_heading = "Rate Limiting")]
@@ -972,6 +982,7 @@ impl CliArgs {
             .request_timeout_secs(self.request_timeout_secs)
             .worker_startup_timeout_secs(self.worker_startup_timeout_secs)
             .worker_startup_check_interval_secs(self.worker_startup_check_interval)
+            .pool_idle_timeout_secs(self.pool_idle_timeout_secs)
             .max_concurrent_requests(self.max_concurrent_requests)
             .queue_size(self.queue_size)
             .queue_timeout_secs(self.queue_timeout_secs)