Self-Assessment Quiz: Data Architecture Principles

Twenty questions. Aim for 16 or more.

Question 1

The chapter defines an architectural decision as one that is:

• A. Made by an architect
• B. Documented in a diagram
• C. Expensive to reverse
• D. About infrastructure rather than code

Question 2

Which of these has the highest reversal cost at Kestrel?

• A. Which BI tool
• B. The compression codec
• C. The partitioning scheme on the largest tables
• D. Which cloud provider

Question 3

Storing bronze data in Parquet rather than loading it straight into a proprietary warehouse is recommended primarily because it:

• A. Compresses better
• B. Moves "which warehouse" down the reversal-cost ranking
• C. Is faster to query
• D. Is required for streaming

Question 4

The question that actually decides batch versus streaming is:

• A. How much data is there?
• B. How fresh does the data need to be?
• C. What decision or action changes if the data is an hour old instead of a minute old?
• D. What does the team already know how to operate?

Question 5

Someone says they need real-time data so they will "catch problems sooner." The chapter's recommended response is to build:

• A. A streaming pipeline
• B. Monitoring and alerting
• C. An hourly batch job
• D. A materialized view

Question 6

Which is NOT one of the costs of streaming named in §3.2?

• A. Backfills become genuinely hard
• B. Event time and processing time diverge
• C. Data volume increases substantially
• D. State must survive restarts and rebalances

Question 7

Roughly what was the metered monthly cost ratio between the streaming and hourly micro-batch options for Kestrel's clickstream?

• A. About 2×
• B. About 9×
• C. About 40×
• D. They were roughly equal

Question 8

ELT displaced ETL primarily because:

• A. SQL became more expressive
• B. Storage became cheap and compute became elastic
• C. GUI ETL tools were discontinued
• D. Regulators required raw data retention

Question 9

Which is a legitimate reason to use ETL rather than ELT?

• A. The raw data is messy
• B. We don't want junk in the warehouse
• C. We legally may not store the raw data
• D. Our warehouse is expensive

Question 10

Which does a plain data lake (Parquet files on object storage, no table format) guarantee?

• A. ACID transactions
• B. Schema enforcement on write
• C. Time travel
• D. None of the above

Question 11

Two Spark jobs write the same S3 prefix concurrently. What happens?

• A. The second job blocks until the first finishes
• B. S3 rejects the second writer
• C. Both succeed; the prefix contains interleaved output and nothing indicates a problem
• D. The files are automatically merged

Question 12

What does a lakehouse table format add over plain files?

• A. Columnar storage
• B. A transaction log providing ACID, schema evolution, and time travel
• C. Compression
• D. SQL query capability

Question 13

Which has the highest operational burden?

• A. Data warehouse
• B. Data lake
• C. Lakehouse
• D. All are equivalent

Question 14

Kestrel's chosen architecture is:

• A. Everything in the warehouse
• B. Everything in a plain data lake
• C. Lakehouse for bronze and silver, warehouse for gold
• D. Warehouse for bronze, lakehouse for gold

Question 15

In §3.5, elastic sizing cost what percentage of fixed-at-peak sizing?

• A. 4.2%
• B. 19.0%
• C. 48.0%
• D. 83.3%

Question 16

The economic argument for cloud infrastructure, stated precisely, is that it is cheaper than:

• A. Owning hardware
• B. Owning hardware sized for your peak
• C. Any on-premises deployment
• D. Managed services from smaller vendors

Question 17

Little's Law states:

• A. $L = \lambda / W$
• B. $L = \lambda W$
• C. $W = L \lambda$
• D. $\lambda = L / W$

Question 18

Sizing a consumer on median rather than tail handling time fails because:

• A. Medians are harder to measure
• B. The queue builds during slow requests and never drains
• C. Kafka requires p99-based sizing
• D. Median latency is not correlated with throughput

Question 19

In data platforms, coupling most often hides in:

• A. Shared object storage buckets
• B. Downstream jobs reading another job's unpublished intermediate tables
• C. Common compression codecs
• D. Shared authentication systems

Question 20

The most valuable section of an ADR, according to §3.7, is:

• A. Context
• B. Decision
• C. Consequences
• D. What would reverse this

Answer Key

1. C — §3.1. It is a spectrum, not a category, and the ranking is more useful than the label.

2. D — §3.1. Cloud provider is 1–2 years: every service, credential, network assumption, plus an egress bill.

3. B — §3.1, 📐 callout. Real architectural work is converting expensive decisions into cheap ones. This is the strongest practical argument for open formats.

4. C — §3.2. Asked as "how fresh," the answer is always "as fresh as possible," which decides nothing.

5. B — §3.2. Someone worried about stockouts wants a stockout alert, not a stream.

6. C — §3.2. Streaming does not increase data volume; it changes when you process the same volume. The listed costs are always-on infrastructure, hard backfills, state management, event-time ambiguity, watermark policy, and debugging against a moving offset.

7. B — §3.2, 💸 callout. Roughly $900/month streaming against $96/month micro-batch — and the metered cost is the smaller consideration next to needing someone who understands consumer groups and state stores on call.

8. B — §3.3. An economic change, not a practice improvement. This matters because when prices change again, the right answer changes again.

9. C — §3.3. The three legitimate cases are legal prohibition on landing raw, enormous volume reduction where raw is genuinely worthless, and a source that can only be read once.

10. D — §3.4. And that is the point — a lake trades all guarantees for cost and engine independence.

11. C — §3.4, ⚠️ callout. Object storage is doing exactly what it promises. A directory of Parquet files is not a table.

12. B — §3.4. Columnar storage and compression come from Parquet, not from the table format.

13. C — §3.4. Compaction and vacuum become standing duties, plus the youngest ecosystem.

14. C — §3.4, §3.7. Each layer using the technology that fits its access pattern.

15. B — §3.5. $11,952 against $63,072. It also meets the SLA every day, which fixed-at-average does not.

16. B — §3.5. A business with flat, predictable load may genuinely be better off owning hardware, and some are.

17. B — §3.5. Items in system = arrival rate × time in system. Exact, and nearly assumption-free.

18. B — §3.5, 🧪 Try It. This is why the gap between p50 and p99 sizing is the interesting number.

19. B — §3.6, 🏭 callout. An intermediate table that quietly became an interface, with nothing in the codebase indicating the dependency.

20. D — §3.7. It converts a permanent commitment into a conditional one and tells your successor what evidence to look for.

Topic map