Which server-side defence gets a service out of a retry storm? The retry storm again: three lockstep retries and a five-second spike. Switch on skipping expired work, shedding by queue age, or newest-first serving, and see which ones break the loop. The demo needs JavaScript.
The storm is a loop: a long queue makes requests time out, timeouts make retries, and retries make the queue longer. Shedding by age caps the queue; newest-first stops fresh work timing out. Either is enough here.
It saves the workers, but the client has already retried, and in a FIFO queue whatever survives has waited almost as long as the timeout. Paired with newest-first, it clears the pile that newest-first leaves behind.
A request turned away in a millisecond costs the client one retry. A request served after its timeout costs a worker the full service time, and the client retries anyway.
All demos