Exercises: The Medallion Architecture

Solutions and grading notes are in the instructor companion. Exercises marked 🧪 use code/layer_check.py.

Warm-Up

Exercise 34.1 — State the guarantee

Difficulty: ★☆☆ · Time: 15 minutes

Without looking back, write the one-sentence guarantee for each layer. Then compare against §34.2–34.4.

Then: for each guarantee, name one thing you could check automatically that would detect a violation. If you cannot name one, the guarantee is decoration.

Exercise 34.2 — Which layer?

Difficulty: ★☆☆ · Time: 20 minutes

Assign each to bronze, silver, or gold, and give the reasoning in one line:

  1. Casting "142" to 142.
  2. Deciding a customer is active.
  3. Deduplicating three CDC rows for one order.
  4. Excluding orders flagged as tests.
  5. Renaming cust_id to customer_id.
  6. Converting EUR to USD at yesterday's rate.
  7. Storing a malformed JSON line exactly as received.
  8. Turning quantity = -1 into is_return = true.
  9. Splitting "Ada Lovelace" into first and last name.
  10. Assuming a timestamp with no offset is UTC.

Two of these are genuinely arguable. Say which, and argue both sides.

Exercise 34.3 — Read the violations

Difficulty: ★☆☆ · Time: 20 minutes · 🧪

python code/layer_check.py --check
  1. Eight findings across six models. Which two models produce two findings each, and why?
  2. Five are blocking and three are warnings. Apply §34.5's classification test to each of the three warnings and say whether you agree.
  3. gold.fct_session trips layer-skip and cast-in-gold. Which would you fix, and what happens to the other?

Core

Exercise 34.4 — Run it on your own project

Difficulty: ★★☆ · Time: 60 minutes · 🧪

  1. Express your own models in build_graph()'s format — layer, dependencies, materialization, unique key, grain, and operations. Twenty models is enough.
  2. Run --check. Report the violations honestly, including the count.
  3. For each violation: was it created deliberately, or did it accumulate?
  4. §34.13 claims drift is the normal end state. Does your project support that claim?

Exercise 34.5 — The boundary cases

Difficulty: ★★☆ · Time: 40 minutes

§34.3's test: if two competent people could disagree about the answer, it is a business rule.

  1. Find three transformations in your own silver (or staging) layer that fail this test.
  2. For each, apply Case Study 2's three markers: did the source specify it, did you use context outside the column, could a reasonable person land elsewhere?
  3. Move one of them to gold. What did it cost — a model, a grain statement, an owner, a test?
  4. §34.3 says "when genuinely unsure, push it up." What does that cost when you are wrong in that direction?

Exercise 34.6 — Tests at the right layer

Difficulty: ★★☆ · Time: 45 minutes

§34.6: a test at the wrong layer is a false alarm or a missed defect.

  1. Take Chapter 23's twenty-two assertions and assign each to a layer.
  2. Find the duplicates — assertions that exist at two layers and fail for the same reason. Kestrel found 31.
  3. Now find the row-count exception: the assertion that belongs at two layers because it fails for opposite reasons. Write both versions.
  4. Do you have a bronze assertion that asserts correctness? Weaken it or move it, and say which.

Exercise 34.7 — Replay cost

Difficulty: ★★☆ · Time: 40 minutes · 🧪

python code/layer_check.py --replay silver.stg_orders
python code/layer_check.py --replay bronze.customers_raw
python code/layer_check.py --replay gold.daily_revenue
  1. Three models, three costs: $25.92, $102.00, $0.96. Explain the pattern in one sentence.
  2. Compute the same for your own project, using your own node-hour rate.
  3. What is your full-rebuild cost? Kestrel's is $198.96.
  4. Have you ever attempted one? If not, what would break? Predict, then find out if you can.

Exercise 34.8 — Depth and blast radius

Difficulty: ★★☆ · Time: 40 minutes · 🧪

python code/layer_check.py --depth
  1. bronze.customers_raw reaches 47% of the graph. Compute the equivalent for your project — which model has the widest blast radius?
  2. How many assertions does it have? How many does your most business-critical model have?
  3. §34.10 proposes: test breadth by blast radius, test depth by visibility. Apply it and say what moves.
  4. Find any model whose depth contradicts its layer. If there is none, say how confident you are that your depth calculation would show one.

Exercise 34.9 — Design a quarantine

Difficulty: ★★☆ · Time: 45 minutes

§34.12's four properties: a paging size limit, the assertion that rejected the row, replayability, and a retention that forces the drain.

  1. Design the quarantine table for one of your pipelines. Include every column.
  2. Set the size limit. Kestrel's is 0.1% of a load or 500 rows, whichever is smaller. Justify yours.
  3. Write the replay job. What does it do with a row that fails again?
  4. Case Study 2's analyst queried the quarantine and produced a number wrong by 1.8%. What would stop that in your design — and is it a technical control or a catalog one?

Exercise 34.10 — How many layers?

Difficulty: ★★☆ · Time: 30 minutes

§34.8: three is a convention.

  1. Argue for two layers for a specific system you know. What does silver buy there?
  2. Argue for five. Name the two extra layers and what each guarantees.
  3. §34.8 says a layer that exists because a different team owns it is a mesh domain in disguise. Find one in your organization, or say why there is none.
  4. What would make you remove a layer from an existing platform, and how would you do it safely?

Advanced

Exercise 34.11 — Extend the checker

Difficulty: ★★★ · Time: 90 minutes · 🧪

Add three rules to layer_check.py, each with a fixture that triggers it and one that does not:

  1. A silver model with a timezone-assuming parse (Case Study 2). What is your false-positive rate?
  2. A gold model with no consumer — no dashboard, no export, no downstream model. Is it a violation or a finding?
  3. A bronze table with no retention policy (Chapter 30 §30.7).
  4. Two silver models reading the same bronze column with different logic — the divergence check.
  5. A model whose depth exceeds the maximum for its layer.
  6. Expiring suppressions (§34.13), failing the build on an expired one.

Then: classify each new rule as blocking or warning using §34.5's test, and defend the classification.

Exercise 34.12 — Attempt the rebuild

Difficulty: ★★★ · Time: 2–3 hours

Case Study 1's rebuild failed three times in two years, each time revealing something no test could.

  1. Attempt a full rebuild from raw for a project you own, into a scratch schema.
  2. Record every failure and classify it: a missing input, a non-deterministic transformation, a changed source, a hidden dependency.
  3. If it succeeds on the first attempt, be suspicious: check that you rebuilt from raw rather than from an intermediate that was already materialized.
  4. What did it cost, in dollars and in hours?
  5. Put it on the calendar. §34.15 argues a capability you never exercise is one you do not have.

Exercise 34.13 — The ambiguity flag

Difficulty: ★★★ · Time: 60 minutes

Case Study 2's durable fix flags ambiguous timestamps at ingestion.

  1. Scan your bronze layer for timestamp columns arriving without an offset. Kestrel found 14 of 41.
  2. For each: has anybody parsed it yet? What did they assume?
  3. Extend the idea beyond timestamps. What else arrives ambiguous? Currency without a code, a quantity without a unit, an ID without a namespace, a name without an encoding.
  4. Design the catalog field that records "a decision is required here and nobody has made it."
  5. Case Study 2 leaves nine columns unresolved and calls that correct. Do you agree?

Exercise 34.14 — Argue against the pattern

Difficulty: ★★★ · Time: 45 minutes

§34.14 names three conditions under which the medallion architecture is wrong.

  1. Write the strongest case against adopting it for a system you know. Be specific about the costs.
  2. §34.11 lists storage, compute, latency, and arguments. Price the first three for your system.
  3. The chapter claims one well-tested transformation layer beats three untrusted ones. Test the claim: what would you lose, concretely?
  4. Now argue the other side, and say which argument you actually believe.

Project Milestone

Exercise 34.15 — Specify and enforce your layers

Difficulty: ★★★ · Time: 3–4 hours · 🧪

Formalize the Kestrel platform's layers, per §34.15.

  1. Write the three guarantees for your platform, in one sentence each, as a document in the repo.
  2. Add the four bronze envelope columns to every bronze table that lacks them.
  3. Implement layer_check.py's nine rules against your real model graph, in CI.
  4. Fix the blocking violations, or suppress them with an expiry and a reason.
  5. Assign every assertion to a layer (Exercise 34.6) and remove the duplicates.
  6. Build the quarantine for one silver model, with all four properties.
  7. Compute depth and blast radius, and reallocate assertions accordingly.
  8. Run one full rebuild and record what it cost and what it found.

Deliverable: the guarantees document, the CI check with its violation count before and after, the assertion-to-layer map, the quarantine, and the rebuild log.

The rebuild log is the deliverable that matters. Everything else states an intention; the rebuild is the only thing that demonstrates the layers do what they claim.