Implementing a Feature Flag System

Feature flags decouple deployment from release — you deploy code to all instances but only enable it for a controlled subset of users, which means you can release to 0.1% of users, validate in production, and roll out gradually without repeated deployments.

Why feature flags matter

• Without feature flags, deploying a new feature is the same event as releasing it — all users see it immediately; if the feature has a bug, every user is affected and the only fix is a rollback deployment
• With feature flags, you deploy the code (zero risk — the flag is off), then enable it for 1% of users (minimal risk), expand to 10% (validate with more data), then 100% — each step is a configuration change, not a deployment

Implementation options

• LaunchDarkly and Unleash are dedicated flag services — they provide a UI for managing flags, targeting rules (user segment, user attribute, random percentage), A/B testing, and an SDK that evaluates flags at runtime
• Simple in-house implementation: store flag names and percentages in a database or Redis; evaluate on each request by hashing the user ID with the flag name to get a consistent 0-100 value; enable if below the threshold — this gives consistent user experience (the same user always sees the same variant)

Implementation in Spring Boot

• A service class that evaluates flags: flagService.isEnabled("checkout-v2", userId) returns true/false; the calling code wraps the new path in an if-block
• Avoid flag proliferation — each flag is technical debt; plan to remove flags after full rollout; a flag that persists for more than 3 months is a smell that the rollout is stalled or the flag was never intended to be temporary

Operational flags vs user-facing flags

• Operational flags (circuit breakers, kill switches): "disable the recommendation engine if it's slow" — these need instant evaluation, often without user context; store in Redis for sub-millisecond reads
• User-facing flags: A/B testing, gradual rollouts — these require consistent user targeting and analytics on which variant users saw