Back-of-the-Envelope Estimation Lab (Interactive)
Turn requests-per-day into average and peak QPS, MB/s versus Mbps, and replicated storage using interview-grade mental math. Round 86,400 to 100,000 seconds, apply 2x-5x peak multipliers, and compare Jeff Dean latency numbers at the click of a selector.
100,000-Seconds-a-Day Mental Math Trainer
Convert business scale into QPS, bandwidth, and storage the way interviewers expect.
The slower op is 750,000×× slower.
Whiteboard output
How It Works Under the Hood
Back-of-the-envelope estimation bounds order of magnitude, not decimals: 1M requests/day is 10 QPS, 1B is 10,000 QPS once you round a day to 100,000 seconds. Peak sizing multiplies average by 2-3x (or 3-5x for flash events). Bytes and bits are conflated at your peril — storage is B/s, NICs are b/s, so multiply by 8. The latency hierarchy from 0.5 ns L1 cache to 75 ms transatlantic RTT explains why caching, indexing, and locality exist at all.
Core Architectural Principles
- Average QPS = daily requests ÷ 100,000; peak = 2x-5x average depending on the product.
- Jeff Dean's hierarchy: DRAM 100 ns, NVMe random 20-50 µs, LAN RTT 0.5 ms, NYC-London 75 ms.
- Physical sizing multipliers: 3x replication for cross-AZ durability plus 30-50% index overhead.
Say the constants out loud: "There are 86,400 seconds a day, I will round to 100,000." State peak multipliers before anyone asks, and never mix Bytes and bits — 25 MB/s is 200 Mbps. Bounding the order of magnitude in three minutes and then discarding unviable architectures is what interviewers actually score.
Aggressive rounding risks precision you do not need for architecture but must never use for billing; estimating average load while provisioning for peak is the classic failure mode.