We built our own feature flag system three years ago. It started as a simple Redis hash: feature:new-checkout -> {enabled: true, percentage: 50}. A Rake task to flip it. A dashboard built in ActiveAdmin. It worked for the first two features.
Then came the edge cases.
What the homegrown system missed
No audit trail. When someone flipped a flag at 2pm on a Friday, we had logs but no context. Why 50%? Why not 25%? Who approved it? The git history on the Rake task told us what changed, not why.
No segmentation. We could do percentage rollouts. We couldn't do "enable for enterprise customers only" or "enable for users on the new pricing plan" without writing custom middleware each time.
No kill switch that actually killed. Our "off" button set enabled: false but in-flight requests still hit the new code path because the middleware cached the flag value for 30 seconds. In a high-throughput endpoint, that's thousands of requests on the old behavior.
No experimentation framework. We wanted to A/B test a new onboarding flow. Our system could split traffic. It couldn't track conversion events, compute statistical significance, or auto-ramp the winner.
The migration
We evaluated LaunchDarkly, Unleash, and GrowthBook. LaunchDarkly won on three factors:
- SDK maturity. The Ruby SDK handles caching, streaming updates, and offline mode correctly. We didn't have to debug flag delivery.
- Segmentation UI. Non-technical PMs can create targeting rules without opening a PR. "Enable for users where
plan = 'enterprise' AND created_at > '2024-01-01'" takes 30 seconds. - Experimentation built in. The same flag that controls rollout can run an experiment. No separate tool.
Migration took two sprints:
- Sprint 1: Run both systems in parallel. New flags in LaunchDarkly. Old flags stay in Redis. Compare behavior.
- Sprint 2: Migrate the 12 active flags. Deprecate the Redis system. Update runbooks.
What changed day-to-day
Faster iterations. A PM wants to test a new pricing page variant. They create the flag, define the 50/50 split, add the conversion metric. Engineer adds the flag check. Deploy. Done in an afternoon. Previously: two days of planning, a PR for the targeting logic, a deploy, a separate analytics setup.
Safer rollbacks. Last month a new search algorithm increased latency 3x for 5% of users. The on-call engineer killed the flag in the LaunchDarkly dashboard. Latency normalized in seconds. No deploy. No cache expiry wait.
Better conversations. Flag discussions now happen in the LaunchDarkly UI with comments and approvals. The context is attached to the flag, not scattered across Slack, Notion, and GitHub.
The cost question
LaunchDarkly isn't cheap. For our volume (~2M flag evaluations/day), it's roughly $3,000/month.
Our homegrown system "cost" zero dollars but burned roughly 0.5 FTE of engineering time per month on maintenance, debugging, and feature requests. At senior engineer rates, that's $10k+/month in opportunity cost.
The math works if you value engineering time on product over engineering time on internal tooling.
What we'd do differently
Start with a managed service. Even Unleash self-hosted would have saved us two years of technical debt. The feature flag surface area looks small. It isn't. It touches targeting, experimentation, audit, compliance, and reliability.
Build your product. Buy the flag infrastructure.
Need help setting up feature flags, experimentation, or release infrastructure? We do this for startups and scale-ups at developerz.ai

