Why does the balancer say a back-end is dead when I think it's alive?

After the Load Balancer decides a back-end is dead, it will mark it dead for a fixed amount of time (by default, 60 seconds). During this time, no further work will be sent to the machine. So even if the machine starts working soon after, the balancer will not try it until this time is up. This delay is to prevent wasting time on machines that are long-dead. The delay can be set with the tunable tuning!dead_time (specified in seconds).

An important point about the operation of the balancer: it does not poll the back-end servers for failure by itself. Rather, it relies on incoming requests that it sends to the back-ends. So, after a machine has been marked as dead, the balancer will not consider the machine alive until a new request arrives, and the balancer successfully gets a response from the back-end.

This is illustrated in the GUI by the colouration of back-ends. A dead back-end is coloured red. If the back-end was dead last time it was used, but the balancer is ready to try the back-end again to see if it has recovered, then it is coloured faded-out red.

Content Manager [Administrator] 12 August 2005