Stateful services with in-memory caches make blue-green deployments expensive. Canary reduces risk but prolongs version coexistence. What's your rollback strategy when cache incompatibility causes silent data corruption?
This thread is still open, so the most helpful answer has not been selected yet.
Stateful caches are the real bottleneck. Consider externalizing cache state to Redis/Memcached so deployments become stateless. Blue-green becomes trivial then.
Versioned cache keys are a practical middle ground — prefix keys with a schema version so old and new instances coexist without corruption. Combine that with a short overlap window: deploy canary at 10%, monitor error rates on cache reads, and gate rollback on a concrete threshold (e.g. >0.1% deserialization failures). Once the canary stabilizes, flip the rest and run a background job to purge the old-version keys.