Practice Tests: Check What You Actually Know

These are self-assessment questions, not exam preparation for anyone's certification. Answer before reading the answer — recognition is not knowledge, and reading a question then reading its answer feels like learning while producing almost none.

The questions are deliberately chosen where the intuitive answer is wrong. If you get most of them right immediately, the corresponding topic is probably solid; skip to the areas where you hesitate.

AI and retrieval#

1. Your RAG system answers 60% of questions correctly. You improve the prompt substantially and accuracy stays at 60%. What is most likely wrong?

Retrieval. If the correct passage is not being retrieved, no prompt can recover it — the retrieval hit rate is the ceiling. Measure retrieval separately from generation before touching anything else. (RAG)

2. When should you fine-tune instead of using retrieval?

When you need a behaviour — a format, tone, or task shape — rather than facts. Facts belong in documents you can edit; weights cannot cite a source and cannot be corrected without retraining.

3. Your agent occasionally runs up a large bill. You add a spend limit to its instructions. Is that sufficient?

No. An agent can reason past its own instructions. The cap must be enforced in the layer running the agent, which it cannot edit — and it must be tested by actually hitting it. (AI agents)

4. Semantic search returns nothing useful for the query "error PX-4471". Why?

Semantic search matches meaning, and an identifier has little. Exact strings need keyword search — this is the standard argument for hybrid retrieval, and identifiers are exactly what people search for when they need a real answer.

Engineering and testing#

5. Your test suite has 90% coverage. What does that tell you about quality?

That 90% of lines executed during the test run. It says nothing about whether behaviour was verified — coverage rises with execution, not with assertions. Better question: if this broke, which test fails? (testing)

6. A test fails intermittently. The pragmatic response is to re-run the build. What does this cost?

The suite's authority. Once re-running red builds is normal, the suite blocks nothing and you are paying runtime for no signal. Quarantine the flake immediately and fix or delete it on a deadline.

7. A pull request is 2,000 lines. You have an hour. What should you do?

Ask for it to be split. An hour on 2,000 lines produces an approval, not a review, and everyone involved knows it. (software engineering)

8. Which database migration is safe to deploy without downtime — adding a nullable column, or renaming one?

Adding. A rename breaks the currently running code, which makes both deploy and rollback an outage. Renames are done by expand-then-contract across several releases.

Containers, Kubernetes and Linux#

9. Your pod restarts every few minutes with no error in the application log. First suspicion?

The memory limit. The process is killed by the kernel, so the application never gets to log anything. Check whether the runtime knows its container limit — many size their heap against the host. (Kubernetes)

10. Your liveness probe checks that the database is reachable. What happens when the database is briefly slow?

Every pod fails liveness and restarts simultaneously, and the database then faces a reconnect storm. Liveness checks the process; readiness may check dependencies.

11. You deleted a large log file but the disk is still full. Why?

A process still holds the file open, so the space is not released until that process is restarted or the descriptor closed. (Linux)

12. Why is a secret unsafe even after you delete it in a later Dockerfile instruction?

Layers are additive — the file remains in the earlier layer and is retrievable from the image. The credential must be rotated; deleting the image is not enough. (Docker)

Cloud and data#

13. You move a virtual machine running at 8% utilisation to the cloud. What happens to cost?

It runs at 8% and now bills monthly and visibly. Cloud savings come from what the move enables — right-sizing, shutting non-production down, managed services, retiring things — not from the move. (cloud)

14. A nightly data job reports success but the dashboard shows zeroes. What was checked, and what was not?

That the code ran. Not that data arrived. Exit codes describe execution; only value checks — minimum row count, freshness, key uniqueness — describe the data. (data engineering)

15. What is wrong with an untested backup?

It is a belief. Backup job success measures that a job ran; only a restore proves the data is usable. Record how long the restore took, too — that number is your real recovery time.

Delivery and governance#

16. Fifteen small changes have each been approved individually. What is the standard failure here?

Nobody reported the cumulative effect. Individually reasonable changes have added weeks; report the running total at every review. (Change Management)

17. Your risk register has thirty risks, all scored medium. What is the problem?

It cannot prioritise, which is its only function. Undefined scales and reluctance to score anything high produce a register that says nothing. (Risk Register)

18. UAT passes with no defects raised, in three days, by two people with full workloads. What do you conclude?

Very likely that it was not done. UAT with no released time produces box-ticking. Absence of findings is not evidence of quality. (UAT)

What a test like this cannot tell you#

It measures recall of things somebody wrote down. It does not measure whether you can build a system, hold a design argument, notice the requirement nobody stated, or operate something at 3am.

Those are measured by Projects and by having shipped things. Treat a high score as evidence you have read carefully, and nothing more.

FAQ#

Are these questions from a real certification exam?#

No. They are written to test the ideas that most often catch people out in practice. Real certification exams have a different style and are usually broader and shallower.

How should I use these?#

Answer before reading. If you get one wrong, read the linked guide rather than memorising the answer — the answer is worth little without the reasoning behind it.

What is a good score?#

There is no threshold, because it is not an assessment. The questions you hesitated on are the result. Those are your topics.

Will there be more questions?#

They grow as topics are published. Every question here maps to a guide on this site, so the set expands with the material rather than separately from it.

Can I use these for interviewing candidates?#

They are better as conversation starters than as a scored test. Ask the question, then ask why — the reasoning distinguishes candidates far more than the answer does. See Interview Questions.

Is there a timed mode?#

No. Timing measures exam technique, which is only useful if you are sitting an exam. For that, use the practice exams sold alongside the certification you are taking.

Related Articles#

Read the material in Courses, practise in Exercises and Labs, and prepare for hiring with Interview Questions.

What else is coming for Practice Tests