Infrastructure & Operations › Working in Production · also in CI/CD & Deployment

Feature Flag Cleanup

Removing flags once rollout is done so they don't pile up as debt.

Backend Engineer · Mid-level Frontend Engineer · Mid-level Data Engineer · Mid-level

Also known as: flag cleanup, feature flag debt, stale flags

A feature flag is meant to be temporary. You add it to ship code dark, roll out gradually, or kill a feature in a hurry. Once the rollout is complete and the feature is fully on, the flag has done its job — and if you leave it in the code, it quietly becomes technical debt.

Stale flags cost more than they look:

Dead branches. The code still contains the “off” path, which may never run and is never tested, yet must be maintained.
Confusion. New engineers can’t tell which flags are load-bearing configuration and which are dead.
Risk. A stray flag set wrong in production changes behavior in ways nobody remembers.
Test matrix growth. Every flag can multiply the combinations worth testing.

flag enabled for 100% and stable for a release →
  remove the flag, keep the "on" code, delete the "off" branch, delete the flag definition

The classic mistakes:

Never removing them. “Temporary” flags from two years ago still gate production behavior. Treat every flag as having an expiry.
Using flags as configuration. If a setting is meant to stay forever and vary by customer, that’s configuration, not a rollout flag — model it as such.
No owner or ticket. A flag with nobody responsible for removing it is a flag that will live forever. Add the cleanup to the same work that introduced it.
Cleaning up without care. Removing the “off” path while it’s still reachable can break someone. Verify which side is live first.

The habit that works is small: when a flag is created, note what “done” looks like; when rollout finishes, delete it in a follow-up change, usually via the same continuous delivery flow that shipped it. This keeps the deploy-versus-release freedom that flags give you without letting them accumulate. Frontend and data pipelines have their own toggles and rollout switches, and they rot the same way.