Feature flags can make production changes easier, but the service behind them can become another dependency. How do you keep a flag outage from turning into an application outage, and what other simple change risks are easy to overlook?
What happens to your app when the feature flag service goes down?
Answers
Start by deciding what each flag should do when it can't be evaluated. A default isn't automatically safe. For a new feature, off might be sensible. For a control that protects access or prevents costly operations, the right fallback needs a closer look.
Check what happens both at startup and after the app has been running. A local cache can keep the last known values available during a service outage, but you also need a plan for stale values and a cold start with no cache. Test those paths rather than assuming the SDK handles them the way you expect.
Reply to this answer
Flags don't remove the risk of changing production. They just change how you make the change. Limit the initial rollout, watch the relevant metrics, and make sure the person making the change knows how to reverse it.
And try the rollback path before you need it. A switch that can be turned on but can't be turned off cleanly is a pretty nasty surprise during an incident.
Reply to this answer
Configuration mistakes deserve as much attention as fancy architecture failures. Time formats, units, environment selection, and the scope of a policy can all turn a small edit into a broad outage.
Use a small test scope first, then check the actual production values and target before applying them widely. A 24-hour clock can also make scheduled changes less ambiguous.
Reply to this answer
Don't stop at software dependencies. Certificates, connectivity, physical power, and access to equipment can all take a service down too. Put ownership and reminders around renewals, and make sure critical equipment isn't easy to disconnect or shut off accidentally.
Reply to this answer
Add your answer
Answers are reviewed before publication.
4v+orbitnotch_826:
Also check whether evaluation happens locally or requires a network call on the request path. Those have very different failure modes.