I started from DAU and worked toward QPS, which felt natural, but then I got a bit tangled trying to map per-request costs to actual instance types.
Start by clearly stating your assumptions about the workload (e.g., requests per user per day, peak-to-average ratio, payload sizes). Then break down each resource (CPU, memory, disk IOPS, network) with explicit formulas, using consistent units and rounding to order-of-magnitude. Finally, sanity-check the numbers against known benchmarks and mention the need for headroom and redundancy.
Pro tip: Always separate steady-state from peak load and apply a safety factor (e.g., 2x) to account for spikes and failures. Also, mention that real-world capacity planning is iterative and should be validated with load testing.
State user count, activity rate (e.g., requests per user per day), average request size, and peak-to-average ratio. Clarify whether the workload is read-heavy or write-heavy.
Convert daily activity to requests per second (RPS) and estimate concurrent requests using Little's Law (concurrency = RPS × average latency).
Use benchmarks or typical values (e.g., CPU time per request, memory per active connection) to compute total CPU cores and memory needed, then apply headroom.
Based on data read/written per request and payload sizes, calculate IOPS and bandwidth. Consider caching to reduce disk load.
Aggregate resources, add redundancy (e.g., N+1), and compare with known instance types. Mention that these are rough estimates to be refined with monitoring and load testing.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.