A single on-prem Kafka cluster in us-east ingests IoT device messages globally. Poor connectivity causes edge batching and ingestion spikes, driving cost and management issues. What Google-recommended cloud-native architecture should you adopt?

The correct answer leverages an IoT gateway to handle device connections and edge batching, then publishes messages to Cloud Pub/Sub. Pub/Sub is a globally distributed messaging service designed for high throughput and low latency, making it ideal for ingesting messages from around the world and buffering spikes. Cloud Dataflow then provides a fully managed, auto-scaling service to process these messages, addressing the management and cost issues of the on-prem Kafka. Edge TPUs are for machine learning inference at the edge, not for message storage and transmission. Cloud Dataflow directly connected to Kafka wouldn't solve the global ingestion, edge batching, or ingestion spike issues. A virtualized Kafka cluster on Compute Engine with Cloud Load Balancing still requires significant management and doesn't inherently solve the global ingestion or spike buffering as effectively as Pub/Sub.