how do you handle feature flagging for large-scale applications?

we're starting to use feature flags more widely in our microservices. at first, it was only for A/B testing, but now we're considering them for phased rollouts and even kill switches. for those of you running feature flags in production, especially at scale, what's your approach? are you building your own system, using a managed service, or maybe something else? what are the biggest problems you've run into, particularly with managing flag states across different systems or making sure flag evaluation is super fast