Self-Assessment Quiz: CI/CD for Data Pipelines

Twenty questions. Aim for 16 or more. Questions 8, 13, and 18 are the three this chapter's two incidents turn on.

Question 1

Which is not one of the four ways data CI/CD is harder?

• A. You cannot easily test against realistic data
• B. The build is slow and expensive
• C. Data code is harder to write
• D. There is no rollback for data

Question 2

Three of those four resolve to:

• A. A tool purchase
• B. "Accept something" — they are trades, not best practices
• C. More testing
• D. A bigger staging environment

Question 3

CI stages should be ordered by:

• A. Alphabetically
• B. What they catch per second
• C. Cost
• D. Team preference

Question 4

Which stages need no warehouse at all?

• A. None
• B. Lint, parse, policy lints, unit tests
• C. Only lint
• D. All of them

Question 5

A 1% random sample of Kestrel's orders contains at least one of the twelve wholesale distributors with probability:

• A. 12%
• B. 11.4%
• C. 1%
• D. 88%

Question 6

Which clause of a deliberate sample should you build first?

• A. The random sample
• B. Every row that has ever caused a test to fail
• C. The extremes of every measure
• D. Structurally unusual entities

Question 7

Slim CI against a missing manifest:

• A. Fails loudly
• B. Selects nothing, builds nothing, and exits 0
• C. Falls back automatically
• D. Rebuilds everything

Question 8

Which is the primary fix for that?

• A. Fail the job if the manifest is missing
• B. Assert that the build produced more than zero models
• C. Retry the download
• D. Cache the manifest longer

Question 9

Slim CI cannot see:

• A. Nothing; it is complete
• B. A hardcoded reference, a shared target, some macro changes, and anything outside dbt
• C. Only macro changes
• D. Test failures

Question 10

A data staging environment holding a 1% sample:

• A. Is representative
• B. Does not contain the row that breaks the model
• C. Is the industry standard
• D. Costs nothing

Question 11

Kestrel deleted its staging environment because:

• A. Budget cuts
• B. In eighteen months it never once caught something the slim build did not
• C. It was unmaintained
• D. Compliance required it

Question 12

The three deploy shapes are additive, structural, and:

• A. Incremental
• B. Definitional
• C. Emergency
• D. Rolling

Question 13

Treating a definitional change as additive produces:

• A. A slower build
• B. A table whose meaning changes at a date nobody recorded
• C. A test failure
• D. Duplicate rows

Question 14

When you fix a definition, the three options are rebuild, do not rebuild and record the boundary, and:

• A. Roll back
• B. Do not rebuild and do not record it — which is the absence of a decision
• C. Delete the column
• D. Version the model

Question 15

Which cannot be rolled back at all?

• A. Code
• B. Table contents inside the retention window
• C. A schema migration
• D. Data sent downstream

Question 16

A shadow deploy's gate should be:

• A. Zero rows different
• B. Only the differences you intended, and no others
• C. Under 1% different
• D. The same row count

Question 17

pr_report.py classifies conservatively because:

• A. It is easier to implement
• B. A false definitional costs thirty seconds; a false additive costs eleven weeks
• C. dbt requires it
• D. It reduces build time

Question 18

"Do not X; instead Y" is dangerous as a requirement because:

• A. It is ambiguous
• B. It is two requirements, and the negative one is easier and is what gets implemented alone
• C. It cannot be tested
• D. It is too long

Question 19

A reviewer's attention should go to:

• A. Hardcoded references
• B. The deploy shape, whether the definition is right, whether the grain changed, and who downstream should be told
• C. Formatting
• D. Test coverage counts

Question 20

Renaming an Airflow task:

• A. Is a cosmetic change
• B. Is a structural change to a historical record — history is keyed on task_id
• C. Requires a migration
• D. Is prevented by the scheduler

Answer Key

1. C — §27.1. Data code is not intrinsically harder to write; the difficulty is entirely in the other three plus the staging problem.

2. B — §27.1's 📐 callout. A team that has not named what it gave up gave up something anyway.

3. B — §27.2. So a bad change fails at second three, not minute eleven.

4. B — §27.2. They run on a laptop, in a container, in a fork, with no credentials.

5. B — §27.3. $1 - 0.99^{12} = 11.4\%$ — so the shape you tested against is missing the entity that causes Chapter 21's skew nine times out of ten.

6. B — §27.3. Every row that has failed a test has proven it can break something. A fixture set that grows by one row per incident is the cheapest regression suite in this book.

7. B — §27.4 and Case Study 2. PASS=0 WARN=0 ERROR=0 SKIP=0 TOTAL=0, exit 0, for nine weeks.

8. B — Case Study 2. Assert that the check RAN, not that it passed. Failing on a missing manifest was rejected — it punishes an uninvolved person and gets an exception carved into it.

9. B — §27.4's ⚠️ callout. Slim CI is a latency optimization, not a coverage one.

10. B — §27.1, §27.5.

11. B — §27.5. A 5% sample, eighteen months, $340 a month.

12. B — §27.7.

13. B — Case Study 1. Chapter 20 Case Study 1's "a trap with a timestamp on it," created by a deploy.

14. B — Case Study 1's ⚠️ callout. It is the default because shipping the code is the half with a button.

15. D — §27.8. An export, an email, a vendor API call. So make the irreversible step last, after the assertions.

16. B — §27.9. A definitional change should produce differences.

17. B — Case Study 1's 📐 callout. And the false positives are not waste: each is a recorded claim that history is consistent.

18. B — Case Study 2's 🔎 callout. Write the positive clause first and test the negative path.

19. B — §27.10. If a defect is invisible in the artifact under review, no amount of diligence is the fix.

20. B — §27.11's ⚠️ callout. Keep task_id stable and rename the function.

Topic map