diff --git a/src/Storages/RabbitMQ/RabbitMQBlockOutputStream.cpp b/src/Storages/RabbitMQ/RabbitMQBlockOutputStream.cpp
index 17e4db3fb89..8e867db6de9 100644
--- a/src/Storages/RabbitMQ/RabbitMQBlockOutputStream.cpp
+++ b/src/Storages/RabbitMQ/RabbitMQBlockOutputStream.cpp
@@ -46,6 +46,8 @@ void RabbitMQBlockOutputStream::write(const Block & block)
 
     if (buffer)
         buffer->flush();
+
+    storage.pingConnection();
 }
 
 
diff --git a/src/Storages/RabbitMQ/StorageRabbitMQ.h b/src/Storages/RabbitMQ/StorageRabbitMQ.h
index 5aa77a9a732..635d53e6cf0 100644
--- a/src/Storages/RabbitMQ/StorageRabbitMQ.h
+++ b/src/Storages/RabbitMQ/StorageRabbitMQ.h
@@ -54,6 +54,8 @@ public:
     const String & getFormatName() const { return format_name; }
     NamesAndTypesList getVirtuals() const override;
 
+    const void pingConnection() { connection.heartbeat(); }
+
 protected:
     StorageRabbitMQ(
             const StorageID & table_id_,
@@ -88,7 +90,7 @@ private:
 
     event_base * evbase;
     RabbitMQHandler eventHandler;
-    AMQP::TcpConnection connection;
+    AMQP::TcpConnection connection; /// Connection for all consumers
 
     Poco::Semaphore semaphore;
     std::mutex mutex;
diff --git a/src/Storages/RabbitMQ/WriteBufferToRabbitMQProducer.cpp b/src/Storages/RabbitMQ/WriteBufferToRabbitMQProducer.cpp
index 86d3b32925a..e61a8e1ccd8 100644
--- a/src/Storages/RabbitMQ/WriteBufferToRabbitMQProducer.cpp
+++ b/src/Storages/RabbitMQ/WriteBufferToRabbitMQProducer.cpp
@@ -16,7 +16,7 @@ enum
 {
     Connection_setup_sleep = 200,
     Connection_setup_retries_max = 1000,
-    Buffer_limit_to_flush = 5000 /// It is important to keep it low in order not to kill consumers
+    Buffer_limit_to_flush = 10000 /// It is important to keep it low in order not to kill consumers
 };
 
 WriteBufferToRabbitMQProducer::WriteBufferToRabbitMQProducer(
@@ -113,38 +113,11 @@ void WriteBufferToRabbitMQProducer::flush()
     /* The AMQP::passive flag indicates that it should only be checked if there is a valid exchange with the given name
      * and makes it visible from current producer_channel.
      */
-
     producer_channel->declareExchange(exchange_name + "_direct", AMQP::direct, AMQP::passive)
     .onSuccess([&]()
     {
         exchange_declared = true;
 
-        /// The case that should not normally happen: message was not delivered to queue (queue ttl exceeded) / not forwareded to consumer
-        if (flush_returned)
-        {
-            /// Needed to avoid data race because two different threads may access this vector 
-            std::lock_guard lock(mutex);
-
-            LOG_TRACE(log, "Redelivering returned messages");
-            for (auto & payload : returned)
-            {
-                next_queue = next_queue % num_queues + 1;
-
-                if (bind_by_id || hash_exchange)
-                {
-                    producer_channel->publish(exchange_name, std::to_string(next_queue), payload);
-                }
-                else
-                {
-                    producer_channel->publish(exchange_name, routing_key, payload);
-                }
-
-                --message_counter;
-            }
-
-            returned.clear();
-        }
-
         /* The reason for accumulating payloads and not publishing each of them at once in count_row() is that publishing
          * needs to be wrapped inside declareExchange() callback and it is too expensive in terms of time to declare it
          * each time we publish. Declaring it once and then publishing without wrapping inside onSuccess callback leads to
@@ -159,27 +132,11 @@ void WriteBufferToRabbitMQProducer::flush()
 
             if (bind_by_id || hash_exchange)
             {
-                producer_channel->publish(exchange_name, std::to_string(next_queue), payload, AMQP::mandatory || AMQP::immediate)
-                .onReturned([&](const AMQP::Message & message, int16_t /* code */, const std::string & /* description */)
-                {
-                    flush_returned = true;
-
-                    /// Needed to avoid data race because two different threads may access this variable
-                    std::lock_guard lock(mutex);
-                    returned.emplace_back(std::string(message.body(), message.body() + message.bodySize()));
-                });
+                producer_channel->publish(exchange_name, std::to_string(next_queue), payload);
             }
             else
             {
-                producer_channel->publish(exchange_name, routing_key, payload, AMQP::mandatory || AMQP::immediate)
-                .onReturned([&](const AMQP::Message & message, int16_t /* code */, const std::string & /* description */)
-                {
-                    flush_returned = true;
-
-                    /// Needed to avoid data race because two different threads may access this vector 
-                    std::lock_guard lock(mutex);
-                    returned.emplace_back(std::string(message.body(), message.body() + message.bodySize()));
-                });
+                producer_channel->publish(exchange_name, routing_key, payload);
             }
 
             --message_counter;
diff --git a/src/Storages/RabbitMQ/WriteBufferToRabbitMQProducer.h b/src/Storages/RabbitMQ/WriteBufferToRabbitMQProducer.h
index 146be0c5796..c61a76a3e74 100644
--- a/src/Storages/RabbitMQ/WriteBufferToRabbitMQProducer.h
+++ b/src/Storages/RabbitMQ/WriteBufferToRabbitMQProducer.h
@@ -54,11 +54,8 @@ private:
     size_t next_queue = 0;
     UInt64 message_counter = 0;
     String channel_id;
-    std::atomic<bool> flush_returned = false;
-    std::mutex mutex;
 
     Messages messages;
-    Messages returned;
 
     Poco::Logger * log;
     const std::optional<char> delim;