Every number wears a label: measured, documented, derived or assumption.
Interviewers don’t expect a drive’s exact IOPS. They expect the right power of ten and an honest answer to “where does that come from?”
Has new hardware made these numbers obsolete?
Bad: Everything got faster
One NVMe drive does 2,000× a hard disk’s random reads.
Good: Round trips didn’t
Memory is still ~100 ns and light no faster. Batching ages well.
Light in fibre: 1 ms of round trip per 100 km.
That’s the floor. Measured round trips run 1.5–2× it, so Pune to Frankfurt, 6,679 km, takes 140 ms. No code makes it shorter.
7/s
synchronous commits a second, India to Europe
At 140 ms a round trip, a single writer waiting on each manages 1,000 ÷ 140 ≈ 7. Across zones, at 1–2 ms, sync replication is cheap.
A DynamoDB table provisioned for 40,000 writes a second. Most traffic hits one key.
The catch
Limits are per partition: that key throttles at 1,000 a second. Size from the hottest key, not the total.
Throughput is per unit, one server, primary or node.
Assume ~2,000 requests/s per app server, ~10,000 small writes/s per SQL primary and ~100,000 ops/s per Redis node. Then multiply by units, minus headroom.
25 million phones poll a cricket score every 10 seconds.
2.5 million requests a second means ~2,100 servers. Every answer is identical, so a 1-second CDN cache leaves the origin ~300 a second.
How to say a number so it’s trusted
- Say the label: measured, documented or assumption
- Check which threshold it crosses
- Round capacities down and demands up
- Challenged? Defend the decision, not the digit
Remember
Carry three ladders: latency, throughput per unit and sizes. Label every number, ask which threshold it crosses, and remember round trips don’t get faster. Defend the decision, not the digit.