Live Scale Estimation Lab (Interactive)
Drive DAU, payloads, and burst multipliers; the board computes QPS, 5-year storage, bandwidth, and cache RAM. Run the 10^5-seconds-per-day mental math live: derive read/write QPS, replicated 5-year storage split into metadata versus blobs, ingress/egress bandwidth, and 80/20 cache sizing against hardware sanity limits.
Live Scale Estimation Board
Drive DAU, activity, and payload assumptions; the board computes QPS, 5-year storage, bandwidth, and cache RAM with the 10^5 mental-math trick.
Whiteboard assumptions
Derived numbers (show these on the board)
30-second sanity check vs hardware physics
Peak reads of 500.0k/s far exceed a single Postgres instance (~5-10k QPS): a Redis cluster and CDN edge cache are architecturally mandatory.
5-year replicated metadata is 274 TB: shard across roughly 46 nodes with 6 TB SSDs each.
Media blobs total 10.9 PB over 5 years - belongs in S3/GCS object storage behind a CDN, not in the database.
Read egress is 12.0 GB/s (96.0 Gbps): beyond a single Direct Connect link (10-100 Gbps) - offload 95 percent to CDN edge.
Mental shortcut used everywhere above: 86,400 s/day ~ 10^5 s (~15% error, well within engineering safety margins).
How It Works Under the Hood
Estimations dictate architectural physics. Rounding 86,400 seconds to 100,000 costs about 15% error but makes division instant. Peak multipliers of 2x-5x expose the difference between average load and the evening surge that crashes average-capacity systems. Separating metadata from media blobs decides whether data fits a sharded PostgreSQL cluster or demands S3 behind a CDN, and the Pareto 80/20 rule sizes the in-memory hot working set. This lab recomputes every number as you move the sliders and sanity-checks each result against real hardware ceilings so you learn to self-correct on the whiteboard.
Core Architectural Principles
- Average and peak QPS derived per path with a 2x-5x burst multiplier over the 10^5 seconds-per-day baseline.
- Five-year storage projections apply 3x replication and split relational metadata from unstructured media blobs.
- Sanity checks compare computed QPS, storage, RAM, and bandwidth against single-node physical limits to catch decimal errors.
State the rounding trick aloud, then write exactly four numbers on the board: read/write QPS, 5-year storage, egress bandwidth, cache RAM. If a number exceeds hardware reality - petabytes for text URLs - say so and fix it in public; interviewers reward self-correcting engineers. Keep the whole block inside five minutes, because precision is not the point, architectural deduction is.
Order-of-magnitude math converts vague designs into defensible capacity plans, but lingering over exact arithmetic steals time from the high-level design.