Cheat sheet & self-check

Pick an answer to see if you got it, and why.

1. Q1. The customer says 'the tracking API must be fast'. What's a good NFR?

   1. AThe API is fast
   2. B99% of GET /track requests complete in under 300 ms, measured at the ingress over a rolling 30 days, at up to 3,000 requests/s
   3. CUse SSDs
   4. DLatency must be zero
   Show answer

   B. A good NFR states the metric, target, percentile, measurement point, window and load.

   From lesson 01 · Requirements → NFRs

2. Q2. Two components in series each have 99.95% availability. What's the best case for the combined system?

   1. A99.95%
   2. BAbout 99.90%
   3. C99.99%
   4. D100%
   Show answer

   B. 0.9995 × 0.9995 ≈ 0.9990. Every serial dependency lowers the ceiling.

   From lesson 01 · Requirements → NFRs

3. Q3. What's the difference between RPO and RTO?

   1. AThey're the same
   2. BRPO: how much data (time) you may lose; RTO: how long until service is restored
   3. CRPO is for networks, RTO for storage
   4. DRTO is always zero
   Show answer

   B. RPO drives replication/backup frequency; RTO drives automation and failover design.

   From lesson 01 · Requirements → NFRs

4. Q4. Why write PoC success criteria before building anything?

   1. APaperwork
   2. BOtherwise results get interpreted to fit a preferred option; fixed criteria make the outcome objective and comparable across options
   3. CVendors require it
   4. DTo slow the project
   Show answer

   B. A PoC is an experiment: define the hypothesis and measurements first.

   From lesson 02 · PoC from scratch

5. Q5. Why measure fsync latency on the disks planned for etcd?

   1. AFor fun
   2. Betcd writes its WAL with fsync on every commit; slow fsync (etcd guidance is p99 well under ~10 ms) causes leader elections and API slowness
   3. Cetcd doesn't use disks
   4. DOnly for backups
   Show answer

   B. Disk latency is the most common hidden cause of unstable control planes.

   From lesson 02 · PoC from scratch

6. Q6. What's the purpose of a weighted scoring matrix?

   1. ATo pick the cheapest option
   2. BTo compare options on agreed criteria whose weights reflect ParcelPath's priorities, making trade-offs explicit and reviewable
   3. CTo avoid testing
   4. DTo choose by popularity
   Show answer

   B. Weights are agreed with stakeholders; scores come from PoC evidence.

   From lesson 02 · PoC from scratch

7. Q7. What does a 72-hour soak test reveal that a 10-minute load test doesn't?

   1. ANothing
   2. BSlow problems: memory leaks, growing queues or disk usage, certificate or token expiries, log rotation issues, and gradual latency drift
   3. COnly CPU speed
   4. DNetwork cabling
   Show answer

   B. Many production incidents come from things that accumulate over hours or days.

   From lesson 03 · Proving stability

8. Q8. Why run an upgrade while the load test is running?

   1. ATo save time
   2. BUpgrades are the most common planned disruption; proving they don't fail requests under load is evidence for NFR-07
   3. CIt's required by Kubernetes
   4. DTo stress the CPU
   Show answer

   B. An upgrade that only works on an idle cluster hasn't been proven.

   From lesson 03 · Proving stability

9. Q9. What belongs on the executive one-pager?

   1. AAll raw metrics
   2. BThe recommendation, how confident you are and why, the top risks with mitigations, cost, and the decision needed from them
   3. COnly good news
   4. DKubernetes YAML
   Show answer

   B. Executives decide on trade-offs; give them the decision, the evidence summary and the risks.

   From lesson 03 · Proving stability

10. Q10. Why size from requests rather than observed average usage?

    1. ARequests are smaller
    2. BThe scheduler places pods by requests: a node is 'full' when requests fill allocatable, regardless of actual usage
    3. CUsage can't be measured
    4. DKubernetes ignores usage entirely
    Show answer

    B. Measure usage to set sensible requests; then size capacity from requests (plus headroom).

    From lesson 04 · Sizing & capacity math

11. Q11. You need 8 nodes' worth of capacity and must survive the loss of 1 of 3 racks. How many nodes?

    1. A8
    2. B12 (8 × 3 ÷ 2), four per rack
    3. C9
    4. D24
    Show answer

    B. After losing a rack, the remaining two racks must hold all 8 nodes' worth of work.

    From lesson 04 · Sizing & capacity math

12. Q12. Why not plan for 100% utilisation of allocatable?

    1. AKubernetes forbids it
    2. BBin-packing is never perfect, rolling updates need surge room, and bursts need headroom; 60–80% is a typical planning target
    3. CIt's slower
    4. DFor licensing
    Show answer

    B. Planning at 100% means the first deploy or node drain fails to schedule.

    From lesson 04 · Sizing & capacity math