Disable per-RPC timeout by default to prevent large upload retry loops

The GrpcStore rpc_timeout_s defaulted to 120 seconds, which is too
short for multi-GB uploads. This caused DeadlineExceeded errors that
triggered retries, restarting the upload and compounding the problem.

Dead connections are already detected by HTTP/2 keepalive (30s ping,
20s timeout) and TCP keepalive (30s) on each endpoint, so the per-RPC
total timeout is unnecessary for that purpose.

Setting rpc_timeout_s=0 now correctly disables the timeout instead of
silently falling through to the 120s default.

Fixes #2185

Author: MarcusSorealheis

nativelink-config/src/stores.rs

@@ -1126,10 +1126,16 @@ pub struct GrpcSpec {
     pub connections_per_endpoint: usize,
 
     /// Maximum time (seconds) allowed for a single RPC request (e.g. a
-    /// ByteStream.Write call) before it is cancelled. This prevents
-    /// individual RPCs from hanging forever on dead connections.
+    /// ByteStream.Write call) before it is cancelled.
     ///
-    /// Default: 120 (seconds)
+    /// A value of 0 (the default) disables the per-RPC timeout. Dead
+    /// connections are still detected by the HTTP/2 and TCP keepalive
+    /// mechanisms configured on each endpoint.
+    ///
+    /// For large uploads (multi-GB), either leave this at 0 or set it
+    /// large enough to accommodate the full transfer time.
+    ///
+    /// Default: 0 (disabled)
     #[serde(default, deserialize_with = "convert_duration_with_shellexpand")]
     pub rpc_timeout_s: u64,
 }

nativelink-store/src/grpc_store.rs

@@ -90,11 +90,7 @@ impl GrpcStore {
             endpoints.push(endpoint);
         }
 
-        let rpc_timeout = if spec.rpc_timeout_s > 0 {
-            Duration::from_secs(spec.rpc_timeout_s)
-        } else {
-            Duration::from_secs(120)
-        };
+        let rpc_timeout = Duration::from_secs(spec.rpc_timeout_s);
 
         Ok(Arc::new(Self {
             instance_name: spec.instance_name.clone(),