Skip to main content

Estimation and Framing Lab

Retrieval Prompts​

  1. Name the three separate lists the framing step produces, in order. Why are they three and not one?
  2. State the formula for QPS from DAU and actions-per-user.
  3. From memory: the powers-of-two values for 10, 20, 30, 40. The Jeff Dean latency numbers for memory, SSD random read, intra-DC round trip, and cross-continent round trip.
  4. What is the 80/20 heuristic for cache sizing, in one sentence?
  5. Define "hard part" in the sense used in Cluster 1 concept 3.

Compare and Distinguish​

Separate these cleanly:

  • functional requirement vs non-functional requirement vs constraint
  • latency vs throughput
  • read:write ratio vs cache hit ratio
  • bottleneck vs single point of failure
  • skew (in data distribution) vs hot key

For each pair, produce a one-sentence distinction you could state out loud on a whiteboard.

Common Mistake Check​

For each statement, identify the error:

  1. "The system must support 1 billion users." (Missing what?)
  2. "P99 latency under 200 ms." (Acceptable as stated, or incomplete?)
  3. "We need a cache." (What makes this wrong as framing?)
  4. "Read:write is 10:1, so we need three read replicas." (What hidden assumption is this making?)
  5. "At 100× scale, we will shard." (Shard on what, why, and at what partition?)

Mini Application​

Pick two prompts you have not used yet. For each, produce in 15 minutes:

  1. Three lists (functional, non-functional with numbers, constraints).
  2. Five estimation lines: QPS avg, QPS peak, storage/day, cache size, latency-budget breakdown.
  3. A ranked list of 2-3 hard parts.
  4. The single number that would most change the design if it were 10× larger.

Candidate prompts (or invent your own):

  1. Global distributed rate limiter for 100 K tenants.
  2. Photo sharing app with 500 M DAU and 3 photos/day uploaded.
  3. Real-time sports score feed for an app with 50 M DAU.
  4. Centralized API usage metering service for a multi-tenant cloud platform.

Numbers Drill (No Calculator)​

Do these in under 10 minutes. Write order of magnitude only.

  1. 300 M DAU × 20 actions/day -> average QPS?
  2. 1 B posts/day × 2 KB each -> storage/day?
  3. 10 K QPS × 512 B request + 8 KB response -> bandwidth in MB/s each direction?
  4. 2 TB hot set × 3× replication -> total cluster memory needed?
  5. Latency budget of 150 ms with 30 ms network each way: how much budget remains for compute?
  6. 99.95% SLO -> minutes of allowed downtime per 30-day month?
  7. If your cache hit rate drops from 95% to 90%, what is the relative increase in origin load?
  8. Typical cross-continent round trip (Jeff Dean table) -> how many serial cross-continent hops can fit in a 400 ms P99 budget?

Evidence Check​

This page is complete only if:

  • you wrote down five estimation lines for each of your two chosen prompts, without a calculator
  • you named 2-3 hard parts per prompt with one-sentence reasons
  • you can recite the powers-of-two row (10, 20, 30, 40) from memory
  • you checked your first-pass framing for an unsupported claim, corrected any actual error, and otherwise explained a labeled counterexample