learn.smmeyer.devmanual
menu
reference
manual
Everything worth memorizing, on one page. Print it before the interview — this page, the glossary, and chapter pages all render print-clean.
scale formulas
Say the factor you assume out loud. Round numbers, shown units.
- average RPS
- daily actions ÷ 86,400
- peak RPS
- average RPS × explicit peak factor
- annual storage
- writes/s × bytes/write × 31.5M × replicas
- bandwidth
- requests/s × bytes/request
- cache source load
- traffic × (1 − hit rate)
- availability (serial)
- A₁ × A₂ × … × Aₙ
- quorum rule
- R + W > N for intersection
- Little's Law
- concurrency = throughput × latency
availability budgets
Downtime allowed per 30-day month. Pick the SLO before the architecture.
- 99%
- 7h 18m / month
- 99.9%
- 43m 48s / month
- 99.95%
- 21m 54s / month
- 99.99%
- 4m 23s / month
- 99.999%
- 26s / month
AI-system metrics
Separate serving health from retrieval health from task health.
- TTFT
- queue + prefill responsiveness
- ITL
- time between output tokens
- TPS
- decode throughput
- recall@k
- retrieval coverage
- groundedness
- claims supported by context
- task success
- end-to-end user outcome
- cost/success
- spend per successful task
- tool accuracy
- correct tool + arguments + effect
50-minute interview checklist
In order. If time collapses, protect 1–5 and close with 10.
- 01
- requirements + out of scope
- 02
- scale and SLOs
- 03
- API + data model
- 04
- critical read/write path
- 05
- bottleneck + scaling trigger
- 06
- failure + retry semantics
- 07
- security + tenant boundary
- 08
- observability + rollout
- 09
- cost + degradation
- 10
- risks + recap
fast tradeoff frame: FACT
Sixty seconds per comparison. No position without a reversal trigger.
- F — frame
- the deciding constraints
- A — alternatives
- worth comparing (usually two)
- C — consequences
- correctness, latency, scale, complexity, cost, operations
- T — take a position
- and name the reversal trigger