Self-Assessment Quiz: API Ingestion

Twenty questions. Aim for 16 or more.

Question 1

Which is NOT one of the five ways an API is harder than a database?

• A. You cannot see the schema
• B. You are a guest with an enforced quota
• C. The data volume is larger
• D. Every request can fail independently

Question 2

Because every request can fail independently, retry handling is:

• A. An edge case
• B. The main loop
• C. Optional at small scale
• D. The orchestrator's responsibility

Question 3

Offset pagination on a collection receiving inserts:

• A. Works correctly
• B. Systematically loses rows at page boundaries and duplicates others
• C. Only fails at very high volume
• D. Fails only on deletes

Question 4

The one-parameter fix for offset pagination's loss is:

• A. A larger page size
• B. An ascending sort on an immutable key, plus a range filter
• C. Retrying each page
• D. Reducing concurrency

Question 5

Cursor pagination's failure mode is:

• A. Duplicates
• B. Cursor expiry on a long or paused extraction
• C. Skew
• D. It cannot be resumed

Question 6

Page-token pagination should terminate on:

• A. An empty page
• B. The token's absence
• C. A fixed page count
• D. A 404

Question 7

Which rate-limit signal is authoritative over your local token bucket?

• A. Your configured requests-per-second
• B. The server's X-RateLimit-Remaining header — other systems share your quota
• C. The documentation
• D. The observed latency

Question 8

A client that runs at full speed until throttled:

• A. Maximizes throughput
• B. Spends its life in backoff and may get blocked
• C. Is the recommended approach
• D. Is fine if it honors Retry-After

Question 9

For a backfill, the constraint is usually:

• A. The data volume
• B. The rate limit
• C. Network bandwidth
• D. Parsing speed

Question 10

Which mitigation for a slow backfill has the highest payoff?

• A. Parallel requests
• B. A larger page size, or an undocumented bulk endpoint
• C. Compression
• D. Caching

Question 11

An auth token should be refreshed:

• A. Once at job start
• B. On a 401
• C. On a margin before expiry, thread-safely
• D. At exactly the expiry time

Question 12

Eight concurrent workers each refreshing a token can produce seven invalid tokens because:

• A. Rate limiting
• B. Some providers revoke the previous token when a new one is issued
• C. Clock skew
• D. The refresh endpoint is not idempotent

Question 13

Jitter in exponential backoff:

• A. Is a minor refinement
• B. Prevents backoff from merely synchronizing the herd at a longer interval
• C. Reduces total wait time
• D. Is only needed above 100 clients

Question 14

Which is retryable?

• A. 400
• B. 403
• C. 422
• D. 429, after honoring Retry-After

Question 15

Retrying 401 repeatedly is harmful because:

• A. It is slow
• B. Several providers treat it as a brute-force attempt and block the account
• C. It corrupts the token
• D. It is not — 401 is retryable

Question 16

The client in §16.6 was blocked for 24 hours because it:

• A. Exceeded the data volume limit
• B. Retried 429 without honoring Retry-After
• C. Used too many connections
• D. Sent malformed requests

Question 17

Webhooks require a reconciliation path because:

• A. They are slow
• B. They are lossy in practice regardless of what the documentation promises
• C. They cannot be authenticated
• D. They deliver out of order

Question 18

At Kestrel, a 30-day re-fetch window caught what share of retroactive changes?

• A. 71%
• B. 95%
• C. 99.6%
• D. 100%

Question 19

The long tail of retroactive changes is best handled by:

• A. A wider re-fetch window
• B. A different mechanism — reconciliation against the provider's own aggregate
• C. Content hashing alone
• D. Accepting the loss

Question 20

In a contract test, a new field in the response should:

• A. Fail the test
• B. Log, not fail
• C. Be added to the expected shape automatically
• D. Trigger a page

Answer Key

1. C — §16.1. API volumes are usually smaller than database ones; the difficulty is elsewhere.

2. B — §16.1. At a 0.1% failure rate a 40,000-page extract has forty failures per run.

3. B — §16.2. And the duplicates are invisible because the write is idempotent, which removes the only symptom anyone would notice.

4. B — §16.2. Paginate over something that does not move — new records arrive past the region you have read.

5. B — §16.2. Record the last record id too, so you can restart from a filter.

6. B — §16.2. Some APIs return a final page with items and no token; others an empty page with one.

7. B — §16.3. Your local bucket does not know about the other systems in your company.

8. B — §16.3. And with some providers, blocked rather than throttled.

9. B — §16.3, 💸 callout. It decides whether a backfill is an afternoon or a weekend.

10. B — §16.3. 50 → 500 per page is a 10× reduction in requests, usually one parameter. Bulk endpoints are frequently undocumented — ask.

11. C — §16.4. Clock skew and round-trip time mean a token valid when checked can be expired when evaluated.

12. B — §16.4. And the cascade of 401s looks exactly like a credential problem.

13. B — §16.5. Full jitter — a uniform draw from zero to the exponential.

14. D — §16.6. 408, 429, 5xx, and network errors are retryable.

15. B — §16.6. It hammers an auth endpoint with credentials that will not work.

16. B — §16.6, 🏭 callout. Not throttled — blocked, requiring a support ticket.

17. B — §16.7. The standard pattern is webhooks for latency plus a periodic pull for completeness.

18. C — §16.8. 7 days catches 95%; 90 days costs thirteen times the requests for a further 0.38%.

19. B — §16.8, 📐 callout. Widening is expensive and never complete; a reconciliation catches the tail at constant cost.

20. B — §16.9. Chapter 13 §13.7's schema-drift taxonomy applied to an API, and it is what keeps the check from crying wolf on every provider release.

Topic map