We've noticed that there's a lot of contention on get/set of schemas by ID in CachedSchemaRegistryClient if that object is shared across threads. All of the methods on CachedSchemaRegistryClient are synchronized and getBySubjectAndId is called when deserializing a record: https://github.com/confluentinc/schema-registry/blob/master/avro-serializer/src/main/java/io/confluent/kafka/serializers/AbstractKafkaAvroDeserializer.java#L122
It seems like a good idea to make a thread-local variant of this data structure, or perhaps make the entire thing be backed by thread local storage. The cost would be memory but it seems safe to do and would remove contention on the registry object across threads. We've implemented a simple thread local schema registry to do this but wanted to get community feedback on the best way to factor this before submitting a PR.
