Lifecycle of a feature flag experiment from creation to cleanup

WHAT THIS TESTS: This question tests whether you understand that a feature flag is production infrastructure with a full lifecycle, not a one-time switch. For experimentation specifically, the interviewer wants to see that you know an experiment flag requires instrumentation, monitoring, a decision framework, and mandatory cleanup. Senior engineers are expected to prevent flag debt and treat rollout safety as a first-class concern.

A GOOD ANSWER COVERS: Stage one is implementation, where engineering responsibility is to instrument the feature behind a flag using an SDK decide method and add event tracking for the experiment metrics without requiring extra deploys. Stage two is configuration and QA, where the flag is off in production but enabled for internal environments or targeted test segments to validate behavior. Stage three is rollout, where the engineer sets percentage-based or canary exposure, monitors error rates and latency, and keeps a kill switch ready. Stage four is the experiment state, where traffic splits run against the defined metrics and the team avoids peeking until statistical significance is reached. Stage five is decision, where the team ships the winning variant, rolls back the loser, or iterates; engineering must then move the winning behavior to the default code path. Stage six is cleanup, where the flag is archived and the conditional branches and dead code are removed in a follow-up deploy.

COMMON WRONG ANSWERS: Treating the flag as permanent configuration instead of temporary scaffolding. Failing to mention event tracking or success metrics, which turns the flag into a release toggle rather than an experiment. Leaving flags in the codebase after the experiment ends, which accumulates technical debt and makes the system harder to reason about. Ignoring rollback or kill switch procedures during rollout.

LIKELY FOLLOW-UPS: How do you prevent flag sprawl in a large codebase? What is the difference between a feature flag and an experiment flag? How do you handle dependencies between multiple experiment flags? What metrics do you monitor before ramping from 1 percent to 100 percent?

ONE CONCRETE EXAMPLE: Suppose you are testing a new checkout button color. You wrap the new component in a flag and instrument click events. You enable it for 5 percent of users in one region while watching conversion and error dashboards. After reaching significance you ship the winner, set the flag to 100 percent default, and schedule a ticket to remove the flag and old component code in the next sprint.

Source: optimizely.com