Self-Assessment Quiz: Distributed Systems Foundations

Twenty questions. Aim for 16 or more. On the streaming path, treat 18 as the bar — Chapters 14, 15, and 29 assume all of this.

Question 1

The third outcome of a remote request, which does not exist on a single machine, is:

• A. The request fails with an error
• B. The request succeeds
• C. You cannot tell whether it executed
• D. The request is queued

Question 2

Why does an operation you might retry need a client-generated key?

• A. To improve performance
• B. So the server can recognize a resubmission of the same logical work
• C. To satisfy audit requirements
• D. To route the request to the correct partition

Question 3

Which partition key would be catastrophic for Kestrel's clickstream?

• A. session_id
• B. event_type
• C. A hash of event_id
• D. session_id concatenated with a salt

Question 4

A job finishes when its slowest task finishes. The implication for partitioning is:

• A. Average partition size determines cost
• B. Maximum partition size determines cost
• C. Partition count should equal node count
• D. Smaller partitions are always better

Question 5

A customer_id partition key has millions of distinct values but one B2B customer places 8% of orders. This demonstrates that:

• A. Hash partitioning is unreliable
• B. High cardinality is not the same as even distribution
• C. Range partitioning would be better
• D. The partition count was too low

Question 6

Which is a reasonable default target size for a Parquet file in object storage?

• A. 1 MB
• B. 16 MB
• C. 256 MB
• D. 8 GB

Question 7

Three things commonly defeat partition pruning. Which is NOT one of them?

• A. A function applied to the partition column in the WHERE clause
• B. A type mismatch forcing a per-row cast
• C. A subquery in the partition predicate that cannot be pushed down
• D. Using a compressed file format

Question 8

Replication lag causes permanent row loss in a watermark-based extract because:

• A. The replica drops rows under load
• B. The watermark is recorded from the replica's view and rows arriving in the lag window fall below it
• C. Replicas apply writes out of order
• D. The extract times out during lag

Question 9

Which fix for the replication-lag watermark bug is exactly correct rather than defense in depth?

• A. Subtract a 15-minute safety margin
• B. Refuse to extract when lag exceeds a threshold
• C. Read the watermark from the primary
• D. Increase the extract frequency

Question 10

The quorum condition for guaranteeing a read sees the latest write is:

• A. $W + R > N$
• B. $W + R < N$
• C. $W = R = N$
• D. $W > R$

Question 11

The most common misreading of CAP is:

• A. That it applies only to relational databases
• B. That it is "pick two of three"
• C. That it was disproved
• D. That it applies only during network partitions

Question 12

PACELC's second clause says that in normal operation you choose between:

• A. Availability and consistency
• B. Latency and consistency
• C. Durability and throughput
• D. Partition tolerance and availability

Question 13

S3's consistency model:

• A. Is eventually consistent for overwrites
• B. Became strongly read-after-write consistent for all operations in 2020
• C. Requires a consistency layer such as EMRFS consistent view
• D. Depends on the region

Question 14

"Eventually consistent" promises convergence if writes stop. It does NOT promise:

• A. That replicas will agree
• B. When convergence happens, or that reads move forward in time
• C. That writes are durable
• D. That the system remains available

Question 15

Kafka's transactional producer provides exactly-once semantics:

• A. End to end, including writes to external databases
• B. From a Kafka topic through a stream processor to another Kafka topic
• C. Only for single-partition topics
• D. Only when combined with idempotent consumers

Question 16

The pattern this book uses throughout is:

• A. Exactly-once delivery
• B. At-most-once delivery with monitoring
• C. At-least-once delivery plus idempotent writes
• D. Two-phase commit across systems

Question 17

Which is NOT one of the four idempotent write strategies?

• A. Delete-insert
• B. Merge on a natural key
• C. Partition replacement
• D. Retry with exponential backoff

Question 18

For analytics, the right time basis is usually:

• A. Processing time, because it is always available
• B. Event time, because the business question is about when things happened
• C. Ingestion time, because it is monotonic
• D. Whichever the source system provides

Question 19

A watermark, in stream processing, is:

• A. A checksum on a batch of events
• B. A declaration that you no longer expect events older than a given event time
• C. The offset of the last processed message
• D. A per-partition sequence number

Question 20

Exponential backoff without jitter fails because:

• A. It retries too slowly
• B. It synchronizes the herd at a longer interval instead of spreading it
• C. It overflows the retry counter
• D. It cannot be combined with circuit breakers

Answer Key

1. C — §4.1. Retry and you may duplicate; do not and you may lose. There is no third option, and no engineering removes the uncertainty.

2. B — §4.1, 🏭 callout. The key identifies the logical work, not the attempt. Three concurrent Spark jobs ran because "submit a job" was treated as idempotent when it was not.

3. B — §4.2. Seven distinct values, and page_view dominates. Low cardinality plus natural popularity is the classic skew generator.

4. B — §4.2. You pay for the maximum partition times the number of nodes waiting on it.

5. B — §4.2, ⚠️ callout. A key can have millions of values and still be dominated by a few. The 22-minute job became 71 minutes about once a week.

6. C — §4.2. 128 MB – 1 GB is the working range; 256 MB is a good default.

7. D — §4.2, 💸 callout. Compression is orthogonal to pruning. The three are a function on the partition column, a type mismatch, and an unpushable subquery — all written by careful engineers producing correct results.

8. B — §4.3, ⚠️ callout. Rows committed on the primary during the lag window arrive on the replica with updated_at values below the stored watermark.

9. C — §4.3. Reading the watermark from the primary solves it exactly. The margin and the lag threshold are defense in depth, and Kestrel does all three because they do different jobs.

10. A — §4.3. With $N=3$, $W=R=2$ is the common default. $W=R=1$ gives fast operations and stale reads.

11. B — §4.4. Partition tolerance is not optional — networks partition whether you choose or not. The real choice is binary and applies only during a partition.

12. B — §4.4. And this is the choice you actually make every day, which is why PACELC is more useful than CAP.

13. B — §4.4, 🧭 Version Note. Pre-2021 advice about working around S3 eventual consistency is obsolete and still ranks well in search results.

14. B — §4.4. "Eventually" is unbounded, and reads may go backwards — a value can appear, disappear, and reappear.

15. B — §4.5. Real and well-engineered, and it is a claim about that hop. The write to Postgres or S3 sits outside the transaction boundary.

16. C — §4.5. At-least-once plus idempotent writes gives effectively-once processing, with any transport and any sink, and no distributed transaction coordinator.

17. D — §4.5. Backoff is a retry policy, not an idempotency strategy. The fourth is deduplicate-on-read.

18. B — §4.6. "What did customers do on Tuesday," not "what did our servers see on Tuesday." The cost is that you must handle late data.

19. B — §4.6. And what happens to later events is a stated policy — dropped, side-outputted, or triggering recomputation. It is a business decision in technical clothes.

20. B — §4.7. Without randomization, backoff merely re-synchronizes the herd. The jitter is the part people omit.

Topic map