I'm curious: I would expect that this approach would give rise to more coding errors (because the consequences of coding errors are smaller). Is that the case?

I'm curious: I would expect that this approach would give rise to more coding errors (because the consequences of coding errors are smaller). Is that the case? If you have a "cattle" setup, how often do your server instances die and get recreated? Is this something you measure and try to minimise? Do you work out why any particular instance died?