Quick Answer

The answer is to optimize the model by pruning or quantizing to reduce size and to deploy the model to multiple Google Cloud regions. These two strategies directly address the core requirements of high availability and low latency for Vertex AI deployments in real-time fraud detection. Pruning or quantizing shrinks the model’s footprint, which reduces inference time and helps achieve the sub-100ms latency target by minimizing computational overhead. Deploying across multiple regions ensures that if one region fails, traffic is seamlessly rerouted via regional load balancing and Cloud DNS, maintaining both availability and low latency. On the Google Professional Data Engineer exam, this question tests your understanding of operationalizing ML models under strict SLOs—a common trap is focusing only on scaling up a single region or over-optimizing for latency at the expense of redundancy. Remember the mnemonic “Prune for speed, spread for need” to recall that model compression handles latency while multi-region deployment ensures availability.