Brief read.

This paper is about TCP failover in a web environment. It assumes that a web service is provided by one frontend and backends. The backend servers are identical and in a ring structure so that one covers each other for the failure. Any one of the backend alive can respond to the request. The frontend receives requests from the client and forward to the backends. All backend receives the same request but only one is really handling it as the frontend only allows the response packet from one of them to pass through.

According to TCP state machines, all backend receives the SYN packet from client but only one of them would have its SYN/ACK packet returned to the client. Upon client’s ACK packet arrives, those who are not active as determined by the frontend will find the ACK packet out of syn and dropped its connection.

The frontend keeps track on the healthiness of backends, e.g. by heartbeat. Failover is enabled by replay of data packets from client: When a backend fails, the frontend selects another one and resend all the client packets to the new backend. After the frontend rebuilds the connection, the data are forwarded to the client as if the connection is not broken. Therefore, before the TCP connection is finished, the frontend will hold the full transaction of packets.

Bibliographic data

   title = "Efficient TCP Connection Failover in Web Server Clusters",
   author = "Ronghua Zhang and Tarek E Abdelzaher and John A. Stankovic",
   booktitle = "Proc INFOCOM",
   year = "2004",