Requirements and estimation
Turn a vague prompt into numbers that decide the design: peak load, storage growth and fan-out, each with its uncertainty.
Estimation is not a ritual. Do it only to answer a question the design depends on: one database or many? Does it fit in memory? Can one region serve it? Stop as soon as the answer is clear.
Numbers worth knowing
| Quantity | Rule of thumb |
|---|---|
| Seconds in a day | ~86,400; use 10⁵ and say it is 16% high |
| Peak to average | 2 to 10x depending on how diurnal the traffic is; consumer apps often ~3x |
| Memory read / SSD read / same-region round trip / cross-continent round trip | ~100 ns / ~100 µs / ~0.5 ms / ~100 ms+ |
| One modern server | Tens of thousands of simple requests per second; a single relational primary handles thousands to tens of thousands of writes per second depending on row size and indexes |
| LLM output speed | Tens to low hundreds of tokens per second per stream; varies by model and provider, so ask or state the assumption |
These are order-of-magnitude anchors, not benchmarks. Say "I'll assume" and move on; the interviewer cares that the assumption is explicit and that you know which way it would move the design.
Requirements that change architectures
- Read/write ratio. Read-heavy points to caches and replicas; write-heavy to partitioning and log-structured storage.
- Consistency need. "A user must see their own write" is cheaper than global linearisability. Ask which one is meant.
- Latency target at a percentile. p99 drives fan-out limits, timeouts and hedging.
- Durability and RPO/RTO. Decides replication mode, backups and multi-region.
- Tenancy and compliance. Data residency or customer-managed keys can force a cell architecture on their own.
Capacity estimator
Change any input. Outputs are simple products; the point is to see which input dominates.
Decimal units (1 GB = 10⁹ bytes). Add headroom for failover: N + 1 per zone at minimum.
Practice
Written for this site to drill the lesson. Not reported interview questions.
- A chat app has 50M DAU sending 40 messages a day of 300 bytes. Estimate peak writes per second and a year of storage at RF 3.
- An internal RAG assistant serves 20,000 employees asking 15 questions a day. Estimate peak queries per second and the daily token volume at an assumed 3,000 input and 400 output tokens per query.
- For one of the two above, name the single input whose uncertainty matters most and why.
Check yourself
When should you stop estimating?
As soon as the numbers answer the design question you were estimating for.
Name three requirements that can change the architecture on their own.
Any of: read/write ratio, consistency need, latency percentile target, durability (RPO/RTO), tenancy or data residency.