Design Feature Flag System

Problem Statement: Runtime Feature Control

A feature flag platform lets product and engineering teams change application behavior at runtime without redeploying binaries. The system stores boolean, multivariate, and JSON configuration values, evaluates targeting rules against user and request context, and pushes updates to thousands of services within seconds. LaunchDarkly, Split, and Optimizely popularized this pattern because it decouples release risk from deployment cadence: you ship code dark, enable gradually, and kill instantly when metrics regress.

Interviewers probe whether you treat flags as configuration with governance rather than ad-hoc if-statements. Production platforms separate a control plane (admin UI, audit, RBAC, rule authoring) from a data plane (high-volume evaluation and config fan-out). Clients must keep evaluating flags in microseconds even when the control plane is unavailable—otherwise a flag outage becomes a product outage.

Design narrative

• Point 1: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 2: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 3: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 4: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 5: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 6: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 7: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 8: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 9: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 10: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 11: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 12: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 13: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 14: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 15: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 16: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 17: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 18: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 19: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 20: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 21: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 22: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 23: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 24: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 25: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 26: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 27: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.
• Point 28: Problem Statement: Runtime Feature Control — tie decisions to measurable SLOs (eval p99, propagation lag, audit completeness) and explicit failure ownership.

How to open this one

The framing that lands for a feature-flag system is decoupling deploy from release: code ships dark and a flag flips it on for a cohort, so a rollout becomes a runtime decision with an instant kill-switch instead of a redeploy. Lead with the low-latency evaluation path — flags resolved locally against a cached ruleset, not a network call per check — and the failure story that matters: the flag service goes down and clients fall back to the last-known-good ruleset rather than failing the request. That shows you understand why the evaluator's availability is non-negotiable.