diff --git a/src/health/mod.rs b/src/health/mod.rs
new file mode 100644
index 0000000..fcb722b
--- /dev/null
+++ b/src/health/mod.rs
@@ -0,0 +1 @@
+pub mod tcp;
diff --git a/src/health/tcp.rs b/src/health/tcp.rs
new file mode 100644
index 0000000..9993d7b
--- /dev/null
+++ b/src/health/tcp.rs
@@ -0,0 +1,41 @@
+use std::{sync::atomic::Ordering, time::Duration};
+
+use crate::state::{app::SharedAppState, backend::BackendState};
+use anyhow::Result;
+use tokio::{net::TcpStream, time::sleep};
+use tracing::info;
+
+// This will evolve later into:
+// - retries
+// - thresholds
+// - latency tracking
+// - richer health states
+pub async fn check_backend_status(backend: &BackendState) -> Result<()> {
+    let backend_address = { format!("{}:{}", backend.config.host, backend.config.port) };
+
+    match TcpStream::connect(&backend_address).await {
+        Ok(_) => {
+            backend.healthy.store(true, Ordering::Relaxed);
+            info!("backend {} healthy", backend.config.id);
+        }
+        Err(_) => {
+            backend.healthy.store(false, Ordering::Relaxed);
+            info!("backend {} unreachable", backend.config.id);
+        }
+    }
+
+    Ok(())
+}
+
+pub async fn start_health_checker(state: SharedAppState) {
+    loop {
+        let state = state.read().await;
+        for upstream in &state.upstreams {
+            for backend in &upstream.backends {
+                let _ = check_backend_status(backend).await;
+            }
+        }
+        drop(state);
+        sleep(Duration::from_secs(5)).await;
+    }
+}
diff --git a/src/lib.rs b/src/lib.rs
index 28f5d37..89058d0 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -1,5 +1,6 @@
 pub mod algorithms;
 pub mod common;
 pub mod config;
+pub mod health;
 pub mod proxy;
 pub mod state;
diff --git a/src/main.rs b/src/main.rs
index 4508d63..7c3c5bb 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -4,6 +4,7 @@
 use anyhow::{Result, bail};
 use laminar::{
     config::{loader::load_config, validator::validate_config},
+    health::tcp::start_health_checker,
     proxy::tcp::start_tcp_proxy,
     state::app::{AppState, SharedAppState},
 };
@@ -39,6 +40,13 @@ async fn main() -> Result<()> {
     // for upstream in &state.upstreams {
     //     info!("upstream '{}' initialized with {} backends", upstream.id, upstream.backends.len());
     // }
+
+    let health_state = shared_state.clone();
+
+    tokio::spawn(async move {
+        start_health_checker(health_state).await;
+    });
+
     let listener_address = format!("{listener_host}:{listener_port}");
 
     start_tcp_proxy(&listener_address, shared_state).await?;
diff --git a/src/proxy/tcp.rs b/src/proxy/tcp.rs
index 39f32cf..1999dec 100644
--- a/src/proxy/tcp.rs
+++ b/src/proxy/tcp.rs
@@ -30,7 +30,13 @@ async fn handle_connection(mut stream: TcpStream, state: SharedAppState) -> anyh
     let backend_address = {
         let state = state.read().await;
         let upstream = &state.upstreams[0];
-        let backend = upstream.next_backend();
+        let backend = match upstream.next_backend() {
+            Some(backend) => backend,
+            None => {
+                error!("no healthy backend available");
+                return Ok(());
+            }
+        };
         format!("{}:{}", backend.config.host, backend.config.port)
     };
 
@@ -38,8 +44,6 @@ async fn handle_connection(mut stream: TcpStream, state: SharedAppState) -> anyh
 
     let mut backend_stream = TcpStream::connect(&backend_address).await?;
 
-    drop(state);
-
     copy_bidirectional(&mut stream, &mut backend_stream).await?;
 
     Ok(())
diff --git a/src/state/app.rs b/src/state/app.rs
index db63271..b71b9ee 100644
--- a/src/state/app.rs
+++ b/src/state/app.rs
@@ -12,9 +12,15 @@ pub struct UpstreamPool {
 
 impl UpstreamPool {
     // Very naive round robin.
-    pub fn next_backend(&self) -> &BackendState {
-        let index = self.current_index.fetch_add(1, Ordering::Relaxed);
-        &self.backends[index % self.backends.len()]
+    pub fn next_backend(&self) -> Option<&BackendState> {
+        for _ in 0..self.backends.len() {
+            let index = self.current_index.fetch_add(1, Ordering::Relaxed);
+            let backend = &self.backends[index % self.backends.len()];
+            if backend.healthy.load(Ordering::Relaxed) {
+                return Some(backend);
+            }
+        }
+        None
     }
 }
 // Central shared runtime state for the entire load balancer.
diff --git a/src/state/backend.rs b/src/state/backend.rs
index dd20e3f..1a8a128 100644
--- a/src/state/backend.rs
+++ b/src/state/backend.rs
@@ -1,3 +1,5 @@
+use std::sync::atomic::AtomicBool;
+
 // use std::sync::Arc;
 // use tokio::sync::RwLock;
 use crate::config::BackendServerConfig;
@@ -13,7 +15,7 @@ pub struct BackendState {
     // Temporary boolean health indicator.
     // This will later evolve into a richer health state machine:
     // Healthy -> Unhealthy -> Recovering
-    pub healthy: bool,
+    pub healthy: AtomicBool,
 
     // This becomes important for least-connections balancing.
     pub active_connections: usize,
@@ -24,6 +26,11 @@ pub struct BackendState {
 
 impl BackendState {
     pub fn new(config: BackendServerConfig) -> Self {
-        Self { config, healthy: true, active_connections: 0, failed_health_checks: 0 }
+        Self {
+            config,
+            healthy: AtomicBool::new(true),
+            active_connections: 0,
+            failed_health_checks: 0,
+        }
     }
 }