Self-Assessment Quiz: The Lakehouse

Twenty questions. Aim for 16 or more.

Question 1

The root cause of every data lake failure in Chapter 9 is:

• A. Object storage is too slow
• B. There is no agreement about which files constitute the table at any moment
• C. Parquet lacks transactions
• D. Listing is paginated

Question 2

Which is NOT one of the five guarantees a table format adds?

• A. Atomic commits
• B. Snapshot isolation
• C. Faster queries by itself
• D. Efficient row-level deletes

Question 3

A Delta commit contains actions of which two kinds?

• A. insert and update
• B. add and remove
• C. begin and commit
• D. read and write

Question 4

File statistics are stored in the transaction log because:

• A. Parquet footers cannot hold them
• B. The engine can prune files without opening any of them
• C. They compress better there
• D. It is required for time travel

Question 5

dataChange: false on a compaction commit exists so that:

• A. The commit is faster
• B. A streaming reader can ignore it rather than reprocessing reorganized data
• C. Vacuum skips those files
• D. The commit is not counted toward the checkpoint interval

Question 6

Atomic commits rest on which object-storage primitive?

• A. Atomic rename
• B. Put-if-absent (create only if the object does not exist)
• C. Multipart upload
• D. Versioning

Question 7

What changed about S3 in 2024?

• A. It became strongly consistent
• B. It added conditional writes, removing the need for an external commit coordinator
• C. Pricing changed
• D. It added native Delta support

Question 8

Two writers both at version 47 attempt to commit. The loser:

• A. Always fails
• B. Always retries successfully
• C. Reads the winner's commit and retries if the changes are disjoint, fails if they overlap
• D. Waits for a lock

Question 9

Optimistic concurrency is a poor fit when:

• A. Writes are large
• B. Writers collide constantly — the fix is disjoint partition ownership, not retry tuning
• C. The table has many partitions
• D. Readers are numerous

Question 10

A commit request times out and the writer retries. If the first commit succeeded:

• A. The retry is rejected automatically
• B. The same rows are appended twice — a table format gives atomicity, not idempotency
• C. The log detects and merges the duplicate
• D. Nothing happens

Question 11

Which schema change is safe without column mapping?

• A. Renaming a column
• B. Dropping a column
• C. Adding a nullable column
• D. Narrowing a type from long to int

Question 12

Column mapping should be enabled:

• A. Only when you need a rename
• B. At table creation, because enabling it later is a protocol upgrade older readers cannot handle
• C. Never — it slows reads
• D. Only on partitioned tables

Question 13

mergeSchema = true is recommended:

• A. Everywhere, for convenience
• B. In bronze only — it disables the guarantee you adopted the format for
• C. In gold only
• D. Never

Question 14

At Kestrel, annual maintenance cost compared to annual storage cost for the bronze events table is approximately:

• A. One tenth
• B. Equal
• C. Six times higher
• D. A hundred times higher

Question 15

Lowering VACUUM retention during a storage incident is dangerous because:

• A. It is slow
• B. It deletes files that a long-running query has pinned, failing it with an unhelpful error
• C. It invalidates the checkpoint
• D. It requires an exclusive lock

Question 16

Time travel is NOT:

• A. Useful for reproducing a report
• B. Useful for restoring from a bad write
• C. A backup — it is bounded by retention and shares the same bucket and blast radius
• D. Useful for debugging a pipeline change

Question 17

Merge-on-read is the right choice for:

• A. Read-heavy tables with infrequent updates
• B. Write-heavy tables with frequent small updates, such as CDC targets
• C. Tables with no updates
• D. Tables smaller than 1 GB

Question 18

After DELETE on a merge-on-read table, the deleted rows are:

• A. Physically removed immediately
• B. Logically gone but physically present until OPTIMIZE, and still present until VACUUM
• C. Moved to a quarantine table
• D. Removed at the next checkpoint

Question 19

The one genuine capability gap between Delta and Iceberg is:

• A. Schema enforcement
• B. Time travel
• C. Partition evolution — Iceberg can change the partitioning scheme without rewriting
• D. Deletion vectors

Question 20

Which does a lakehouse still NOT provide?

• A. Snapshot isolation
• B. Multi-table transactions
• C. Schema enforcement
• D. Row-level deletes

Answer Key

1. B — §10.1 overview. Object storage is doing exactly what it promises; the problem is using a file listing as a table definition.

2. C — §10.1. The format gives you the tools for layout. Adopting it and skipping maintenance makes things slower.

3. B — §10.2. A compaction is "remove these 34,560, add these 4," in one atomic transaction.

4. B — §10.2. Pruning from one log read instead of 14.7 million footer reads.

5. B — §10.2. Otherwise every compaction looks like new data downstream.

6. B — §10.3. A single-object atomic operation, which object stores can provide.

7. B — §10.3, 🧭 Version Note. It obsoleted a large body of advice about DynamoDB log stores and commit services.

8. C — §10.3. Appends to different partitions do not conflict; overlapping modifications do.

9. B — §10.3. Twenty jobs merging into one table spend most of their time retrying.

10. B — §10.3, ⚠️ callout. Chapter 4 §4.5's rule stands: at-least-once plus idempotent writes.

11. C — §10.4. Old files return null for it. Renames and drops need column mapping; narrowing is never safe.

12. B — §10.4. Exactly the kind of change you do not want to make under time pressure.

13. B — §10.4. Bronze accepts what arrives; silver enforces a contract.

14. C — §10.5, 💸 callout. Roughly $600/year against $94/year — 6.4×.

15. B — §10.5, ⚠️ callout. And an incident is when you have the least information about what is running.

16. C — §10.6. Same bucket, same permissions, same blast radius, bounded by retention. Teams reliably conflate the two.

17. B — §10.7. Low delete latency at the cost of read latency that degrades until compaction.

18. B — §10.7. The trap that catches teams during their first erasure request. If your obligation is physical erasure by a deadline, DELETE alone does not satisfy it.

19. C — §10.8. Relevant if you might get partitioning wrong — and Chapter 4's Case Study 1 suggests you might.

20. B — §10.10. Design so consumers tolerate momentary inconsistency, or serialize and accept a window.

Topic map