diff --git a/CHANGELOG.md b/CHANGELOG.md
index 3fd1328..220aa9f 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -9,7 +9,14 @@ All notable changes to this project will be documented in this file.
 - Handle Redis connection errors (e.g. broken pipe during master failover) gracefully instead of panicking.
   Previously, `get_queued_query_count` would `.unwrap()` on the Redis result, causing a panic that poisoned
   the metrics `RwLock`, cascading into further panics and leaving pods unresponsive ([#111]).
-
+- Recover from a black-holed Redis connection instead of hanging until the liveness probe restarts the pod.
+  When the node running the Redis master is drained, the TCP connection silently stops delivering packets
+  without ever being closed, so the only error trino-lb sees is a response timeout. The redis crate's
+  `ConnectionManager` does not reconnect on timeouts (only on dropped-connection errors), so it kept reusing
+  the dead connection forever. trino-lb now enables TCP keepalive / `TCP_USER_TIMEOUT` on the Redis socket and
+  runs a background health check that rebuilds the connection itself when a periodic ping fails ([#109]).
+
+[#109]: https://github.com/stackabletech/trino-lb/issues/109
 [#111]: https://github.com/stackabletech/trino-lb/pull/111
 
 ## [0.6.0] - 2026-02-17
diff --git a/trino-lb-persistence/Cargo.toml b/trino-lb-persistence/Cargo.toml
index 1c8bd61..287a444 100644
--- a/trino-lb-persistence/Cargo.toml
+++ b/trino-lb-persistence/Cargo.toml
@@ -25,3 +25,6 @@ tokio.workspace = true
 tracing.workspace = true
 trait-variant.workspace = true
 url.workspace = true
+
+[dev-dependencies]
+tokio = { workspace = true, features = ["macros", "rt"] }
diff --git a/trino-lb-persistence/src/redis/mod.rs b/trino-lb-persistence/src/redis/mod.rs
index cdf36a2..c9cd1c9 100644
--- a/trino-lb-persistence/src/redis/mod.rs
+++ b/trino-lb-persistence/src/redis/mod.rs
@@ -1,18 +1,23 @@
 use std::{
     fmt::Debug,
     num::TryFromIntError,
+    sync::{Arc, PoisonError, RwLock},
     time::{Duration, SystemTime, SystemTimeError, UNIX_EPOCH},
 };
 
-use futures::{TryFutureExt, future::try_join_all};
+use futures::{
+    TryFutureExt,
+    future::{BoxFuture, try_join_all},
+};
 use redis::{
-    AsyncCommands, Client, RedisError, Script,
+    AsyncCommands, Client, IntoConnectionInfo, RedisError, Script,
     aio::{ConnectionManager, ConnectionManagerConfig, MultiplexedConnection},
     cluster::{ClusterClientBuilder, ClusterConfig},
     cluster_async::ClusterConnection,
+    io::tcp::{TcpSettings, socket2::TcpKeepalive},
 };
 use snafu::{OptionExt, ResultExt, Snafu};
-use tracing::{Instrument, debug, debug_span, info, instrument};
+use tracing::{Instrument, debug, debug_span, info, instrument, warn};
 use trino_lb_core::{
     TrinoClusterName, TrinoLbQueryId, TrinoQueryId,
     config::RedisConfig,
@@ -26,6 +31,40 @@ use crate::Persistence;
 const REDIS_CONNECTION_TIMEOUT: Duration = Duration::from_secs(10);
 const REDIS_RESPONSE_TIMEOUT: Duration = Duration::from_secs(10);
 
+// TCP keepalive / user-timeout settings for the Redis socket.
+//
+// The redis crate's [`ConnectionManager`] only reconnects when a command fails with a
+// dropped-connection error (broken pipe, connection reset, ...). When the Kubernetes node running
+// the Redis master is drained, the established TCP connection turns into a black hole: packets are
+// silently dropped and no FIN/RST ever arrives, so the only error trino-lb ever sees is a response
+// *timeout* - which the redis crate does not treat as a reason to reconnect. Without these socket
+// options the connection would stay wedged until the liveness probe restarts the pod (see issue
+// #109). Enabling TCP keepalive (and TCP_USER_TIMEOUT) lets the kernel tear the dead socket down on
+// its own, shrinking the detection window; the [`Reconnectable`] wrapper then rebuilds it.
+const REDIS_TCP_KEEPALIVE_TIME: Duration = Duration::from_secs(5);
+const REDIS_TCP_KEEPALIVE_INTERVAL: Duration = Duration::from_secs(2);
+const REDIS_TCP_KEEPALIVE_RETRIES: u32 = 3;
+const REDIS_TCP_USER_TIMEOUT: Duration = Duration::from_secs(12);
+
+/// How often the background health check pings Redis to detect a black-holed connection (see the
+/// note on [`REDIS_TCP_KEEPALIVE_TIME`]) and rebuild it via [`Reconnectable`]. Such a connection
+/// only ever produces response *timeouts*, which the redis crate does not treat as a reason to
+/// reconnect, so without this proactive check it would stay wedged until the liveness probe
+/// restarts the pod (see issue #109).
+const REDIS_HEALTH_CHECK_INTERVAL: Duration = Duration::from_secs(5);
+
+/// TCP settings applied to every Redis connection (and every reconnect), see the constants above.
+fn tcp_settings() -> TcpSettings {
+    TcpSettings::default()
+        .set_keepalive(
+            TcpKeepalive::new()
+                .with_time(REDIS_TCP_KEEPALIVE_TIME)
+                .with_interval(REDIS_TCP_KEEPALIVE_INTERVAL)
+                .with_retries(REDIS_TCP_KEEPALIVE_RETRIES),
+        )
+        .set_user_timeout(REDIS_TCP_USER_TIMEOUT)
+}
+
 const LAST_QUERY_COUNT_FETCHER_UPDATE_KEY: &str = "lastQueryCountFetcherUpdate";
 
 const BINCODE_CONFIG: bincode::config::Configuration = bincode::config::standard();
@@ -129,6 +168,73 @@ pub enum Error {
     },
 }
 
+/// Builds a fresh Redis connection. Used by [`Reconnectable`] to rebuild a connection that went bad.
+type ConnectionFactory<T> =
+    Arc<dyn Fn() -> BoxFuture<'static, Result<T, RedisError>> + Send + Sync>;
+
+/// A Redis connection that can be transparently rebuilt when it goes bad.
+///
+/// The redis crate's [`ConnectionManager`] only reconnects on dropped-connection errors, not on
+/// timeouts, so a black-holed connection (e.g. after a Redis node drain, see issue #109 and the
+/// note on [`REDIS_TCP_KEEPALIVE_TIME`]) would otherwise time out forever. This wrapper lets us
+/// rebuild the underlying connection ourselves; the background health check in
+/// [`RedisPersistence::spawn_health_check`] decides *when* to rebuild.
+struct Reconnectable<T> {
+    current: RwLock<Arc<T>>,
+    factory: ConnectionFactory<T>,
+    /// Guarantees only a single rebuild runs at a time.
+    reconnecting: tokio::sync::Mutex<()>,
+}
+
+impl<T> Reconnectable<T>
+where
+    T: Send + Sync + 'static,
+{
+    fn new(initial: T, factory: ConnectionFactory<T>) -> Self {
+        Self {
+            current: RwLock::new(Arc::new(initial)),
+            factory,
+            reconnecting: tokio::sync::Mutex::new(()),
+        }
+    }
+
+    /// Returns the connection currently in use.
+    fn current(&self) -> Arc<T> {
+        Arc::clone(&self.current.read().unwrap_or_else(PoisonError::into_inner))
+    }
+
+    /// Rebuilds the connection, replacing the one currently in use.
+    ///
+    /// `used` is the connection that was found to be bad. This is a no-op if another rebuild is
+    /// already in progress, or if the connection has already been replaced since `used` was
+    /// obtained (i.e. another rebuild already happened). This keeps overlapping failure reports
+    /// from rebuilding the connection over and over.
+    async fn reconnect(&self, used: &Arc<T>) {
+        let Ok(_guard) = self.reconnecting.try_lock() else {
+            // Another task is already rebuilding the connection.
+            return;
+        };
+
+        if !Arc::ptr_eq(&self.current(), used) {
+            // The connection was already replaced while we waited for the lock.
+            return;
+        }
+
+        match (self.factory)().await {
+            Ok(connection) => {
+                *self.current.write().unwrap_or_else(PoisonError::into_inner) = Arc::new(connection);
+                info!("Successfully rebuilt the Redis connection");
+            }
+            Err(error) => {
+                warn!(
+                    ?error,
+                    "Failed to rebuild the Redis connection, will retry on the next health check"
+                );
+            }
+        }
+    }
+}
+
 /// This Redis implementation works against Redis clusters. It uses a single connection that is shared between all
 /// operations for best performance. However, this makes atomic operations hard as their are some pitfalls regarding
 /// `WATCH` in combination with `MULTI` and `EXEC` documented
@@ -140,7 +246,7 @@ pub struct RedisPersistence<R>
 where
     R: AsyncCommands + Clone,
 {
-    connection: R,
+    connection: Arc<Reconnectable<R>>,
     compare_and_set_script: Script,
 
     /// Sometimes we need to do stuff for all cluster groups, so we need to store them to iterate over them
@@ -158,17 +264,32 @@ impl RedisPersistence<ConnectionManager> {
             .set_connection_timeout(Some(REDIS_CONNECTION_TIMEOUT))
             .set_response_timeout(Some(REDIS_RESPONSE_TIMEOUT));
 
-        let client = Client::open(config.endpoint.as_str()).context(CreateClientSnafu)?;
-        let connection = client
-            .get_connection_manager_with_config(redis_config)
-            .await
-            .context(CreateClientSnafu)?;
-
-        Ok(Self {
+        // The TCP settings live on the `ConnectionInfo` (not on `ConnectionManagerConfig`, which
+        // does not expose them), and are re-applied on every reconnect the manager performs.
+        let connection_info = config
+            .endpoint
+            .as_str()
+            .into_connection_info()
+            .context(CreateClientSnafu)?
+            .set_tcp_settings(tcp_settings());
+        let client = Client::open(connection_info).context(CreateClientSnafu)?;
+
+        let factory: ConnectionFactory<ConnectionManager> = {
+            let client = client.clone();
+            Arc::new(move || {
+                let client = client.clone();
+                let redis_config = redis_config.clone();
+                Box::pin(async move { client.get_connection_manager_with_config(redis_config).await })
+            })
+        };
+
+        let connection = factory().await.context(CreateClientSnafu)?;
+
+        Ok(Self::new_with_connection(
             connection,
-            compare_and_set_script: compare_and_set_script(),
+            factory,
             cluster_groups,
-        })
+        ))
     }
 }
 
@@ -184,24 +305,32 @@ impl RedisPersistence<ClusterConnection<MultiplexedConnection>> {
             .set_response_timeout(REDIS_RESPONSE_TIMEOUT);
 
         let client = ClusterClientBuilder::new([config.endpoint.as_str()])
+            .tcp_settings(tcp_settings())
             .build()
             .context(CreateClientSnafu)?;
-        let connection = client
-            .get_async_connection_with_config(redis_config)
-            .await
-            .context(CreateClientSnafu)?;
 
-        Ok(Self {
+        let factory: ConnectionFactory<ClusterConnection<MultiplexedConnection>> = {
+            let client = client.clone();
+            Arc::new(move || {
+                let client = client.clone();
+                let redis_config = redis_config.clone();
+                Box::pin(async move { client.get_async_connection_with_config(redis_config).await })
+            })
+        };
+
+        let connection = factory().await.context(CreateClientSnafu)?;
+
+        Ok(Self::new_with_connection(
             connection,
-            compare_and_set_script: compare_and_set_script(),
+            factory,
             cluster_groups,
-        })
+        ))
     }
 }
 
 impl<R> Persistence for RedisPersistence<R>
 where
-    R: AsyncCommands + Clone,
+    R: AsyncCommands + Clone + Send + Sync + 'static,
 {
     #[instrument(skip(self, queued_query))]
     async fn store_queued_query(&self, queued_query: QueuedQuery) -> Result<(), super::Error> {
@@ -533,10 +662,55 @@ where
 
 impl<R> RedisPersistence<R>
 where
-    R: AsyncCommands + Clone,
+    R: AsyncCommands + Clone + Send + Sync + 'static,
 {
+    /// Wraps the initial connection in a [`Reconnectable`] and starts the background health check
+    /// that rebuilds it when it goes bad.
+    fn new_with_connection(
+        connection: R,
+        factory: ConnectionFactory<R>,
+        cluster_groups: Vec<String>,
+    ) -> Self {
+        let connection = Arc::new(Reconnectable::new(connection, factory));
+        Self::spawn_health_check(Arc::clone(&connection));
+
+        Self {
+            connection,
+            compare_and_set_script: compare_and_set_script(),
+            cluster_groups,
+        }
+    }
+
     fn connection(&self) -> R {
-        self.connection.clone()
+        (*self.connection.current()).clone()
+    }
+
+    /// Spawns a background task that periodically pings Redis and rebuilds the connection if the
+    /// ping fails.
+    ///
+    /// The redis crate's [`ConnectionManager`] reconnects on dropped-connection errors but not on
+    /// timeouts, so a black-holed connection (see the note on [`REDIS_TCP_KEEPALIVE_TIME`]) would
+    /// otherwise time out on every command indefinitely. A periodic ping detects this - and heals
+    /// it even while there is no other traffic - without having to wrap every individual operation.
+    /// [`Reconnectable::reconnect`] makes sure we only rebuild once per dead connection.
+    fn spawn_health_check(connection: Arc<Reconnectable<R>>) {
+        tokio::spawn(async move {
+            loop {
+                tokio::time::sleep(REDIS_HEALTH_CHECK_INTERVAL).await;
+
+                let current = connection.current();
+                let mut ping_connection = (*current).clone();
+                // The connection carries its own response timeout, so a black-holed socket makes
+                // this `PING` fail rather than hang forever.
+                let ping: Result<(), RedisError> =
+                    redis::cmd("PING").query_async(&mut ping_connection).await;
+
+                if let Err(error) = ping {
+                    warn!(?error, "Redis health check failed, rebuilding the connection");
+                    connection.reconnect(&current).await;
+                }
+            }
+        });
     }
 
     #[instrument(skip(self))]
@@ -610,3 +784,74 @@ fn compare_and_set_script() -> Script {
     ",
     )
 }
+
+#[cfg(test)]
+mod tests {
+    use std::sync::atomic::{AtomicU32, Ordering};
+
+    use super::*;
+
+    /// Returns a [`Reconnectable`] over a `u32` plus a counter of how many times it was rebuilt.
+    /// Every rebuild yields once (so concurrent rebuild attempts interleave deterministically) and
+    /// then returns the new rebuild count as the connection value.
+    fn counting_reconnectable() -> (Arc<AtomicU32>, Reconnectable<u32>) {
+        let rebuilds = Arc::new(AtomicU32::new(0));
+
+        let factory_rebuilds = Arc::clone(&rebuilds);
+        let factory: ConnectionFactory<u32> = Arc::new(move || {
+            let rebuilds = Arc::clone(&factory_rebuilds);
+            Box::pin(async move {
+                tokio::task::yield_now().await;
+                Ok(rebuilds.fetch_add(1, Ordering::SeqCst) + 1)
+            })
+        });
+
+        (rebuilds, Reconnectable::new(0, factory))
+    }
+
+    #[tokio::test]
+    async fn reconnect_replaces_the_connection() {
+        let (rebuilds, reconnectable) = counting_reconnectable();
+
+        let used = reconnectable.current();
+        assert_eq!(*used, 0);
+
+        reconnectable.reconnect(&used).await;
+
+        assert_eq!(rebuilds.load(Ordering::SeqCst), 1);
+        assert_eq!(*reconnectable.current(), 1);
+    }
+
+    #[tokio::test]
+    async fn reconnect_is_a_noop_for_an_already_replaced_connection() {
+        let (rebuilds, reconnectable) = counting_reconnectable();
+
+        let stale = reconnectable.current();
+        reconnectable.reconnect(&stale).await;
+        assert_eq!(*reconnectable.current(), 1);
+
+        // The handle is now stale (the connection was replaced); reconnecting with it must not
+        // rebuild again, otherwise repeated failure reports would rebuild over and over.
+        reconnectable.reconnect(&stale).await;
+        assert_eq!(rebuilds.load(Ordering::SeqCst), 1);
+        assert_eq!(*reconnectable.current(), 1);
+    }
+
+    #[tokio::test]
+    async fn concurrent_reconnects_rebuild_only_once() {
+        let (rebuilds, reconnectable) = counting_reconnectable();
+        let reconnectable = Arc::new(reconnectable);
+
+        let used = reconnectable.current();
+        let attempts = (0..8).map(|_| {
+            let reconnectable = Arc::clone(&reconnectable);
+            let used = Arc::clone(&used);
+            async move { reconnectable.reconnect(&used).await }
+        });
+        futures::future::join_all(attempts).await;
+
+        // Single-flight: many concurrent failures on the same connection rebuild it exactly once.
+        assert_eq!(rebuilds.load(Ordering::SeqCst), 1);
+        assert_eq!(*reconnectable.current(), 1);
+    }
+}