Self-Assessment Quiz: NoSQL and Specialized Stores

Twenty questions. Aim for 16 or more.

Question 1

The one honest reason to add a specialized store is:

• A. The team already knows it
• B. It will scale better
• C. An access pattern the existing store serves badly enough to matter, measured
• D. The existing store is relational

Question 2

Scanning a key-value store is:

• A. A normal operation
• B. A red flag — you have chosen the wrong store
• C. Only a problem at scale
• D. Impossible

Question 3

Someone asks for analytics on data that lives only in Redis. The right response is:

• A. Build a better scanner using SCAN with a cursor
• B. Increase Redis persistence settings
• C. Start emitting events to a durable log, and build the analysis on those
• D. Add a read replica

Question 4

What makes DynamoDB not quite a key-value store?

• A. It is managed
• B. A partition key plus a sort key gives efficient range queries within a partition
• C. It supports transactions
• D. It has a SQL interface

Question 5

MongoDB change streams are notable because they are:

• A. The only change feed in this chapter
• B. Ordered, resumable, and cover inserts, updates, and deletes — the best non-relational change feed here
• C. Available only in the managed service
• D. Lossy but fast

Question 6

When landing document-store data, the chapter recommends:

• A. Infer a schema from a sample and flatten
• B. Land whole in bronze, enforce an explicit schema in silver, quarantine what does not match
• C. Reject documents that do not match a schema at landing
• D. Store only the fields you currently need

Question 7

Without a quarantine, "enforce a schema in silver" actually means:

• A. Fail the load
• B. Silently drop what does not fit
• C. Store the extras in a JSON column
• D. Alert an operator

Question 8

In a wide-column store, denormalization is:

• A. A performance optimization
• B. The data model — you design tables around queries in advance
• C. Discouraged
• D. Handled automatically

Question 9

Cassandra's extraction story is weak, so most teams:

• A. Use its CDC commit logs directly
• B. Dual-write to a message bus, or full-scan with token-range pagination
• C. Query a read replica
• D. Use logical decoding

Question 10

Kestrel's clickstream is not time-series data because:

• A. It has no timestamps
• B. It has millions of distinct session ids, needs joining, and events get corrected
• C. It is too high volume
• D. It is stored as JSON

Question 11

What kills a time-series database?

• A. Total data volume
• B. Cardinality — the number of distinct label combinations
• C. Query complexity
• D. Retention length

Question 12

Adding customer_id (1.9M values) as a label to a metric with 4,800 series produces roughly:

• A. 1.9 million series
• B. 9.1 million series
• C. 9.1 billion series
• D. No change — labels are indexed separately

Question 13

The person who takes down a metrics stack by adding a high-cardinality label is usually:

• A. Being careless
• B. Trying to improve observability
• C. Running a load test
• D. Migrating data

Question 14

Search indexes should be extracted from:

• A. The index, using a scroll API
• B. The source of truth the index was built from
• C. The index's own logs
• D. A snapshot

Question 15

What data does a search engine hold that exists nowhere else?

• A. The documents
• B. Search analytics — queries typed, results returned, positions clicked
• C. The relevance model
• D. Nothing

Question 16

Vectors produced by two different embedding model versions are:

• A. Directly comparable
• B. Comparable after normalization
• C. Not comparable — they live in different spaces
• D. Comparable only for text

Question 17

Which defense would have caught 6,000 products missing from a vector index?

• A. A latency alert
• B. A coverage assertion: index row count equals embeddable source row count
• C. Increasing the ANN recall parameter
• D. A schema test

Question 18

pgvector is generally sufficient until roughly:

• A. 10,000 vectors
• B. 100,000 vectors
• C. 10 million vectors
• D. 1 billion vectors

Question 19

The question that tests whether a time-series database is the wrong choice is:

• A. How much data is there?
• B. Does anything get corrected, and do you need to join it to a customer or product?
• C. What is the retention requirement?
• D. Is the data numeric?

Question 20

Before adopting a new store, the chapter says to consult:

• A. The benchmark results
• B. The extraction table — what the change feed is, because it determines the work for the store's whole life
• C. The pricing page
• D. The team's existing skills

Answer Key

1. C — §12.1. The word doing the work is measured.

2. B — §12.2. The operation exists and is a red flag.

3. C — §12.2, ⚠️ callout. Two days of work and the problem is gone permanently, versus a scanner that is slow, inconsistent, and subject to a TTL nobody connected to analytics.

4. B — §12.2. Which puts it closer to wide-column in capability, and brings wide-column modelling discipline.

5. B — §12.3. The closest thing to Debezium-quality CDC outside the relational world.

6. B — §12.3. §11.5's layered answer, applied to a document source.

7. B — §12.3. The quarantine is what makes the third option honest.

8. B — §12.4. A Cassandra schema frequently holds the same facts in four tables because there are four access patterns.

9. B — §12.4. Cassandra's CDC writes per-node commit-log segments requiring cross-replica deduplication, which most teams decline.

10. B — §12.5. Corrections and joins are the two things time-series databases are worst at.

11. B — §12.5, 📏 callout. And it arrives suddenly.

12. C — §12.5. 4 × 6 × 200 × 1,900,000. An out-of-memory kill, not a slow query.

13. B — §12.5. Which is why the rule — never put an unbounded identifier in a label — has to be known in advance.

14. B — §12.6, §12.9. The source has better guarantees and a better change feed.

15. B — §12.6, 🏭 callout. And it usually exists only in logs with a short retention nobody chose deliberately. Emit it as events.

16. C — §12.7. Upgrading the model means re-embedding everything, and a partial re-embed produces silently wrong results rather than errors.

17. B — §12.7. Nine words of SQL, and it catches this class of problem regardless of cause.

18. C — §12.7. Kestrel's 47,000 products is four orders of magnitude below the threshold.

19. B — §12.10. If yes to either, it is not time-series data whatever its timestamp suggests.

20. B — §12.9. A store with no change feed will be full-scanned — expensive, disruptive, and it misses deletes.

Topic map