In a cluster network, an ingress component can serve as an entry point for incoming request traffic and route requests to a plurality of backend servers for processing. Each of the backend servers can handle multiple flows (e.g., TCP) established with the ingress component. When a backend server is detected to have been overloaded with requests, a different peer server is selected on the network to take over processing of the requests from the overloaded server. A set of state information for at least one connection is then transferred from the overloaded server to the selected peer server. Thereafter, the ingress component can update its routing information to begin routing new incoming requests or packets for that connection to the new peer server. The state information can be transferred upon reaching a logical boundary between requests, or within a request, received on the connection.