Comparison
Bake timevsDark launch
Bake time
you leave the new version running on ten percent of traffic for an hour before proceeding, because some failures only show up slowly.
A deliberate wait between rollout stages so slow-developing problems — memory leaks, cache pollution, gradual queue growth — have a chance to appear. Without it a staged rollout only catches immediate errors. The right duration is set by how fast your worst known failure mode manifests, not by impatience.
Full entry →Dark launch
you send a copy of real production traffic at the new service and throw the responses away, just to see if it holds up.
Exercising new code with real traffic while its output is discarded. It reveals load, latency and error behaviour under a realistic distribution that no synthetic test reproduces. Anything with side effects has to be stubbed, or you will double-charge somebody in the name of testing.
Full entry →