diff --git a/src/fetch.rs b/src/fetch.rs
index a8935f5..5679836 100644
--- a/src/fetch.rs
+++ b/src/fetch.rs
@@ -97,9 +97,28 @@ impl Client {
             node_url.unwrap_or(format!("{LOCAL}:{port}"))
         };
         log::info!("connecting to {base_url}");
-        let client = reqwest::Client::builder()
+        let mut builder = reqwest::Client::builder()
             .timeout(Duration::from_secs(args.request_timeout_seconds))
-            .connect_timeout(Duration::from_secs(args.request_timeout_seconds)) // Connection establishment timeout
+            .connect_timeout(Duration::from_secs(args.request_timeout_seconds)); // Connection establishment timeout
+        if args.node_disable_conn_pool {
+            if use_esplora {
+                // The flag is node-only; applying it to Esplora would force a fresh
+                // TCP + TLS handshake per request — a real regression — so ignore it
+                // here, but make the ignored setting loud rather than silent.
+                log::warn!(
+                    "**********************************************************************\n\
+                     * --node-disable-conn-pool (NODE_DISABLE_CONN_POOL) is set together   \n\
+                     * with --use-esplora. This flag only affects the local node           \n\
+                     * connection and is being IGNORED for the Esplora backend; keep-alive \n\
+                     * pooling stays ENABLED.                                              \n\
+                     **********************************************************************"
+                );
+            } else {
+                // No keep-alive reuse: each request opens a fresh connection.
+                builder = builder.pool_max_idle_per_host(0);
+            }
+        }
+        let client = builder
             .build()
             .with_context(|| "Failed to create HTTP client with timeout")?;
 
@@ -688,6 +707,110 @@ mod test {
         assert_eq!(result[&6], 2.0);
     }
 
+    /// Spawn a minimal HTTP/1.1 server on loopback that counts the TCP
+    /// connections it accepts and answers every request with the same canned
+    /// blockhash, keeping the connection alive for any follow-up requests.
+    /// Returns the bound address and the accept counter.
+    async fn spawn_counting_node() -> (
+        std::net::SocketAddr,
+        std::sync::Arc<std::sync::atomic::AtomicUsize>,
+    ) {
+        use std::sync::atomic::{AtomicUsize, Ordering};
+        use std::sync::Arc;
+        use tokio::io::{AsyncReadExt, AsyncWriteExt};
+        use tokio::net::TcpListener;
+
+        let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+        let addr = listener.local_addr().unwrap();
+        let connections = Arc::new(AtomicUsize::new(0));
+        let counter = connections.clone();
+
+        tokio::spawn(async move {
+            // A valid 64-hex blockhash so `block_hash` parses the body successfully.
+            const HASH_HEX: &str =
+                "000000000019d6689c085ae165831e934ff763ae46a2a6c172b3f1b60a8ce26f";
+            let response = format!(
+                "HTTP/1.1 200 OK\r\nContent-Length: {}\r\n\r\n{HASH_HEX}",
+                HASH_HEX.len()
+            );
+            loop {
+                let (mut socket, _) = match listener.accept().await {
+                    Ok(pair) => pair,
+                    Err(_) => break,
+                };
+                counter.fetch_add(1, Ordering::SeqCst);
+                let response = response.clone();
+                tokio::spawn(async move {
+                    let mut buf = [0u8; 1024];
+                    // Serve each request on this connection until the client closes it,
+                    // so a reused keep-alive connection is counted only once.
+                    'conn: loop {
+                        let mut req = Vec::new();
+                        loop {
+                            match socket.read(&mut buf).await {
+                                Ok(0) | Err(_) => break 'conn, // client closed the connection
+                                Ok(n) => {
+                                    req.extend_from_slice(&buf[..n]);
+                                    if req.windows(4).any(|w| w == b"\r\n\r\n") {
+                                        break; // full request headers received
+                                    }
+                                }
+                            }
+                        }
+                        if socket.write_all(response.as_bytes()).await.is_err() {
+                            break;
+                        }
+                    }
+                });
+            }
+        });
+
+        (addr, connections)
+    }
+
+    fn node_client(addr: std::net::SocketAddr, disable_conn_pool: bool) -> Client {
+        let mut args = Arguments::default();
+        args.network = Network::Bitcoin;
+        args.use_esplora = false;
+        args.node_url = Some(format!("http://{addr}"));
+        args.rpc_user_password = Some("user:pass".to_string()); // satisfies is_valid()
+        args.request_timeout_seconds = 30; // Default::default() leaves this 0, which is_valid() rejects
+        args.node_disable_conn_pool = disable_conn_pool;
+        Client::new(&args).unwrap()
+    }
+
+    /// Default (pooled) behavior: sequential requests reuse a single keep-alive connection.
+    #[tokio::test]
+    async fn test_conn_pool_enabled_reuses_connection() {
+        use std::sync::atomic::Ordering;
+
+        let (addr, connections) = spawn_counting_node().await;
+        let client = node_client(addr, false);
+
+        for _ in 0..3 {
+            client.block_hash(0).await.unwrap().unwrap();
+        }
+
+        assert_eq!(connections.load(Ordering::SeqCst), 1);
+    }
+
+    /// With the pool disabled, every request establishes its own connection, so a
+    /// locality-aware L4 proxy re-evaluates the upstream on each request instead of
+    /// staying pinned to whichever backend the first connection landed on.
+    #[tokio::test]
+    async fn test_conn_pool_disabled_opens_fresh_connection_each_request() {
+        use std::sync::atomic::Ordering;
+
+        let (addr, connections) = spawn_counting_node().await;
+        let client = node_client(addr, true);
+
+        for _ in 0..3 {
+            client.block_hash(0).await.unwrap().unwrap();
+        }
+
+        assert_eq!(connections.load(Ordering::SeqCst), 3);
+    }
+
     #[tokio::test]
     #[ignore = "connects to prod server"]
     async fn test_client_esplora() {
diff --git a/src/server/mod.rs b/src/server/mod.rs
index 82dfbbb..17dfc50 100644
--- a/src/server/mod.rs
+++ b/src/server/mod.rs
@@ -140,6 +140,11 @@ pub struct Arguments {
     #[arg(env, long, default_value = "30")]
     pub request_timeout_seconds: u64,
 
+    /// Disable HTTP keep-alive connection pooling to the node, forcing a fresh connection per request.
+    /// Node-only: ignored (with a warning) when --use-esplora is set.
+    #[arg(env, long)]
+    pub node_disable_conn_pool: bool,
+
     /// Timeout in seconds for reading incoming HTTP request headers (protects against slowloris attacks)
     #[arg(env, long, default_value = "10")]
     pub header_read_timeout_seconds: u64,
@@ -187,6 +192,7 @@ impl std::fmt::Debug for Arguments {
             .field("enable_db_statistics", &self.enable_db_statistics)
             .field("cache_control_seconds", &self.cache_control_seconds)
             .field("request_timeout_seconds", &self.request_timeout_seconds)
+            .field("node_disable_conn_pool", &self.node_disable_conn_pool)
             .field(
                 "header_read_timeout_seconds",
                 &self.header_read_timeout_seconds,