Choose batch or online inference
Start with when a consumer needs the result. A nightly forecast often fits batch inference; an interactive request may require online inference. Specify end-to-end latency, throughput, data freshness, and availability. Include feature lookup and preprocessing in the latency budget, not just model execution.
Choose an inference pattern for weekly product recommendations and for a live payment risk check. State the constraint driving each choice.
Make releases reversible
Store versioned model artifacts alongside their evaluation evidence and preprocessing dependencies. A canary release sends a small portion of traffic to a new version while you observe behavior. Define rollback thresholds before deployment. Compare business and reliability metrics as well as model quality, because a more accurate model can still be too slow or expensive.
Write a rollout plan with an initial traffic allocation, observation window, and concrete rollback condition.
Scale against measured demand
Measure tail latency under representative load and consider cold starts, batching, concurrency, and minimum capacity. Use the smallest hardware configuration that meets the objective with appropriate headroom. Protect access to endpoints and avoid placing sensitive input or output in application logs.
Your average latency is acceptable but the 99th percentile is too high. List three measurements you need before changing the infrastructure.
A retailer needs predictions for every product before its warehouse opens each morning. Which approach is the best starting point?
Keep exploring
Use the official certification page and its linked resources to validate your coverage. For guided labs, explore the Google Skills learning path. Revisit the practical exercise and explain your decision in your own words before moving on.