All should handle the dispatch and all should be available for serving API. If you're using Kubernetes, you can configure them to self cluster