Research exemplars and model answer outline

Two of the ten weekly questions are worked as calibration: S3 for a design/trade-off answer and S7 for an analysis/responsibility answer. The model outline is guidance, not a model answer to copy — it shows the shape and depth expected, not text to submit.

The full question set, the submission spec and the rubric are in the course assessment guidance. The citations are on the S3 week page and the S7 week page.

Exemplar 1 — S3, a design/trade-off question

Which mix of block and object storage, and which tiering and snapshot policy, keeps the 1 TB video archive and 200 GB asset library both cheap and restorable?

Why this one. The question has a defensible answer, but only after the student prices the options and chooses a restore objective. It rewards a decision rule, not a list of storage classes.

Model answer outline

  1. Frame the decision. 1.2 TB of media, 135 GB/month egress, 300 video views and 5,000 thumbnail requests a day. First choose the substrate: block volumes carry the live working set and its RAID layout, object storage carries the archive and its tiers. Then state when an asset moves and what restore time it will accept. Define the restore-time objective first; price second.
  2. Thesis. Adopt a three-tier lifecycle with versioning and a quarterly restore test: hot for the first 30 days, infrequent access from day 31 to day 90, cold after day 90; keep noncurrent versions for one day; target restore under four hours for the last 90 days and under 24 hours for the archive. The policy is cheap because the cold tier holds the long tail, and restorable because the restore test is part of the policy, not an afterthought.
  3. Evidence. Use Ghemawat et al. (2003) for the durability and chunk model, Rodeh et al. (2013) for snapshotting, and current provider prices for the three tiers. Build a small table: gigabytes per tier, monthly storage, egress, requests and the restore path. Compare it with a two-tier and a single-tier policy. State the assumptions (annual growth, average object size, deletion rate).
  4. Counterargument. Cold tiers have minimum storage durations and retrieval costs; a 1.2 TB archive may be too small to justify three tiers; snapshots are not backups if the bucket is deleted; restore time can blow the objective. Answer each: keep client masters in a second account or region; test a sample restore; use versioning plus a lifecycle rule; cap the cold tier at assets untouched for 90 days.
  5. Decision rule. A table by asset age and type — camera masters, finished clips, catalogue photos, 3D models, client legal holds — with the tier, the retention, the restore objective and the owner. The rule is what makes the answer operational.
  6. Risks and what would change the answer. Restore latency, deletion by lifecycle misconfiguration, egress on restore, price changes, and a client contract that requires a shorter restore. Name the evidence that would flip the decision (for example, a 10× growth in archive size or a legal hold).
  7. References. Five to eight, including the two curated papers, provider pricing and at least three independent sources.

Marking calibration for S3

Level What it looks like
4 A clear policy with a restore objective, a costed comparison, a counterargument about cold-tier retrieval, and a decision table. The arithmetic is visible and the assumptions are stated.
3 A sensible policy with some costing and a restore test, but one option is under-analysed or the counterargument is thin.
2 Describes the tiers and lifecycle features, but does not choose a policy or price it; restore is mentioned but not tested.
1 A summary of storage classes with no decision, no arithmetic and fewer than five sources.

Exemplar 2 — S7, an analysis/responsibility question

Where exactly does the provider's responsibility end for a media asset that is stored in DigitalOcean, cached by Cloudflare, and processed by an OpenRouter-routed model?

Why this one. It is the PLO3 pattern: the student must trace one asset through the trust boundary, name the control owner at each step, and defend a residual-risk judgement — not recite the shared-responsibility slogan.

What a strong answer does. Picks one concrete path (for example, a client clip that is stored on a DigitalOcean block volume, cached at Cloudflare's edge, and sent through OpenRouter to a model endpoint). At each step it names the provider control, the customer control and the data class crossing the boundary; it uses the lifecycle map from the security thread; it cites the providers' shared-responsibility documentation and the PDPA cross-border rule; it finishes with one risk-register row and a residual-risk judgement. A weak answer lists the model's bullet points without choosing an asset or a boundary.

How to use these