Day 1 surveyed 4 inference options. Days 2–3 covered real-time and advanced patterns. Today we dig into the other two sides of synchronous endpoints: batch transform and serverless inference. Both share "always-on cost → 0," but their mechanics and fit scenarios differ completely. Batch handles bulk data at once; serverless handles sporadic individual requests.
Exams test tuning parameters here directly. Batch transform: how to boost throughput (instance count, mini-batch, split). Serverless: how to handle concurrency and cold starts (MaxConcurrency, Provisioned Concurrency). Build numeric intuition for cost tradeoffs.