Add configurable max_task_poll_retries to WorkerConfig

client/src/lib.rs

@@ -232,7 +232,9 @@ impl Default for RetryConfig {
 }
 
 impl RetryConfig {
-    pub(crate) const fn task_poll_retry_policy() -> Self {
+    /// The default retry policy for task poll requests (workflow, activity, nexus).
+    /// Retries indefinitely with exponential backoff (200ms initial, 2x multiplier, 10s max).
+    pub const fn task_poll_retry_policy() -> Self {
         Self {
             initial_interval: Duration::from_millis(200),
             randomization_factor: 0.2,

client/src/retry.rs

@@ -25,6 +25,7 @@ const LONG_POLL_FATAL_GRACE: Duration = Duration::from_secs(60);
 pub struct RetryClient<SG> {
     client: SG,
     retry_config: Arc<RetryConfig>,
+    task_poll_retry_override: Option<Arc<RetryConfig>>,
 }
 
 impl<SG> RetryClient<SG> {
@@ -33,8 +34,17 @@ impl<SG> RetryClient<SG> {
         Self {
             client,
             retry_config: Arc::new(retry_config),
+            task_poll_retry_override: None,
         }
     }
+
+    /// Override the retry policy used for task poll requests (workflow, activity, nexus).
+    /// By default, task polls use an unlimited-retry policy. Use this to set a finite
+    /// `max_retries` so the worker exits after repeated poll failures.
+    pub fn with_task_poll_retry_config(mut self, config: RetryConfig) -> Self {
+        self.task_poll_retry_override = Some(Arc::new(config));
+        self
+    }
 }
 
 impl<SG> RetryClient<SG> {
@@ -71,7 +81,10 @@ impl<SG> RetryClient<SG> {
             retry_short_circuit = ext.get::<NoRetryOnMatching>().cloned();
         }
         let retry_cfg = if call_type == CallType::TaskLongPoll {
-            RetryConfig::task_poll_retry_policy()
+            self.task_poll_retry_override
+                .as_deref()
+                .cloned()
+                .unwrap_or_else(RetryConfig::task_poll_retry_policy)
         } else {
             (*self.retry_config).clone()
         };

core-api/src/worker.rs

@@ -122,6 +122,14 @@ pub struct WorkerConfig {
     #[builder(default = "Duration::from_secs(5)")]
     pub local_timeout_buffer_for_activities: Duration,
 
+    /// Maximum number of retries for task poll requests (workflow, activity, nexus) before the
+    /// error is propagated. 0 means unlimited retries (default, preserving existing behavior).
+    /// Setting this to a nonzero value causes the worker to exit with an error after that many
+    /// consecutive poll failures, allowing the orchestrator (e.g. Kubernetes) to restart the pod
+    /// with fresh connections.
+    #[builder(default = "0")]
+    pub max_task_poll_retries: usize,
+
     /// Any error types listed here will cause any workflow being processed by this worker to fail,
     /// rather than simply failing the workflow task.
     #[builder(default)]

core/src/lib.rs

@@ -134,7 +134,14 @@ pub(crate) fn init_worker_client(
     if let Some(ref id_override) = config.client_identity_override {
         client.options_mut().identity.clone_from(id_override);
     }
-    RetryClient::new(client, RetryConfig::default())
+    let retry_client = RetryClient::new(client, RetryConfig::default());
+    if config.max_task_poll_retries > 0 {
+        let mut poll_cfg = RetryConfig::task_poll_retry_policy();
+        poll_cfg.max_retries = config.max_task_poll_retries;
+        retry_client.with_task_poll_retry_config(poll_cfg)
+    } else {
+        retry_client
+    }
 }
 
 /// Creates a unique sticky queue name for a worker, iff the config allows for 1 or more cached