Estimation for ML

How this question breaks down

1. 1Data sizeCore20 minTurn users and events into training rows, rows into bytes, and bytes into the three stores an ML system keeps (logs, features and embeddings, derived artefacts), and see what sampling and precision do to each.
2. 2Training computeCore21 minCount the FLOPs a training run needs, divide by what a GPU really sustains, and get GPU-hours, wall-clock time and memory, then learn when FLOPs are the wrong thing to count.
3. 3Serving budgetCore22 minFrom daily users to peak model calls, from a latency target to a per-stage budget, and from both to a count of machines, for a feed ranker on CPUs and an LLM on GPUs.

Work through the parts in order and you’ve covered the full design. Each part is one topic.