Feature flags can target infrastructure, too
Pick a rollout unit that matches the thing that can fail.
I had an optimization to release where a failure would affect an entire pod. Targeting a few users or tenants wouldn’t contain it, and rolling out to a whole region would put too much at risk. The change had to happen at the pod level.
I added a pod context in LaunchDarkly and enabled the change for a percentage of pods. When the flagged pods started failing, traffic moved to healthy ones. I turned the flag off in seconds, before it became an incident.
Flags don’t have to be for users. They can control infrastructure as well. Pick the rollout unit that matches the thing that can fail.