Interview questions

Interview questions (with the counter-questions you'll get)

Q1. What's the difference between a message queue and pub/sub?

They're testing: the basic distinction.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** A queue delivers each message to one consumer, so workers share the work; it suits commands like "send this email". Pub/sub delivers each message to every subscription, so many services react to one event; it suits facts like "order placed". Consumer groups combine them: pub/sub across groups, a queue within each group.
3
4**Counter-questions:**
5- *"How do you get both with AWS?"* → SNS topic fanning out to one SQS queue per subscriber.
6
7**Trap:** using a single queue for an event several services need, so they steal messages from each other.

Q2. Design a news feed: fan-out on write or on read?

They're testing: the classic trade-off.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** Fan-out on write precomputes each user's feed, so reads are cheap but posts cost one write per follower, which explodes for celebrities. Fan-out on read stores posts once and merges at read time, so posts are cheap but reads are expensive. Use a hybrid: push for normal accounts, pull for celebrities above a follower threshold, merged at read time.
3
4**Counter-questions:**
5- *"How do you pick the threshold?"* → From the follower distribution and read/write rates; measure the cost of both paths and pick where pulling becomes cheaper.
6
7**Trap:** choosing pure fan-out on write and ignoring celebrities.

Q3. What is the celebrity (hot key) problem in fan-out?

They're testing: awareness of skew.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** A few accounts have millions of followers, so pushing their posts creates huge write bursts and delays, and their content becomes a hot key on reads. The fix is to not fan them out on write: pull their recent posts at read time and cache them heavily, since everyone reads the same celebrity posts.
3
4**Counter-questions:**
5- *"What about a celebrity who suddenly goes viral?"* → Thresholds can be dynamic, based on current follower count or posting load.
6
7**Trap:** solving it only with more workers.

Q4. What is a consumer group?

They're testing: Kafka-style consumption.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** A set of consumers that share a subscription: each message goes to one member, and members split the partitions. Different groups each get every message. Scaling a service means adding members (up to the partition count); adding a service means adding a group.
3
4**Counter-questions:**
5- *"What happens when a member dies?"* → The group rebalances and its partitions are reassigned, resuming from the last committed offsets.
6
7**Trap:** thinking every consumer in a group gets every message.

Q5. What is event-driven architecture, and what are its downsides?

They're testing: architectural judgement.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** Services publish facts and others react, rather than calling each other directly. It decouples services, isolates failures and makes adding consumers easy. Downsides: flows are harder to trace and debug, data is eventually consistent, and multi-step processes need careful choreography or orchestration.
3
4**Counter-questions:**
5- *"How do you debug it?"* → Correlation IDs on every event, distributed tracing, and a catalogue of publishers and subscribers.
6
7**Trap:** presenting it as free.

Q6. Explain the transactional outbox pattern.

They're testing: reliable publishing.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** Writing to the database and publishing separately can lose events or publish ghost events. Instead, write the change and an outbox row in the same transaction; a relay or CDC reads outbox rows and publishes them, marking them sent. Delivery is at-least-once, so consumers de-duplicate by event ID.
3
4**Counter-questions:**
5- *"Polling relay or CDC?"* → Polling is simpler; CDC (reading the database log) has lower latency and less load on the table.
6
7**Trap:** "publish inside the transaction" as if the broker were part of it.

Q7. How do subscribers handle duplicate and out-of-order events?

They're testing: correctness under at-least-once.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** Every event has a unique ID; subscribers record applied IDs with a unique constraint and skip repeats. For disorder, events carry per-entity versions or sequence numbers and subscribers ignore older ones. Key by entity so related events share a partition, and prefer "state is now X" events over "change by Y".
3
4**Counter-questions:**
5- *"Where do you store applied IDs?"* → In the same database as the effect, updated in the same transaction.
6
7**Trap:** assuming the broker prevents duplicates.

Q8. How do you change an event's schema without breaking subscribers?

They're testing: evolution in practice.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** Make backward-compatible changes: add optional fields, never rename or remove fields that subscribers use, and never change a field's meaning. Use a schema registry with compatibility checks. For breaking changes, publish a new version or topic alongside the old one until all subscribers move.
3
4**Counter-questions:**
5- *"How do you know who uses a field?"* → A catalogue of subscribers and contract tests.
6
7**Trap:** renaming a field and redeploying the publisher.

Q9. A new analytics service needs all past orders as events. How?

They're testing: replay and backfill.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** If the topic is a log with enough retention, start a new consumer group from the earliest offset. If not, backfill from the database (a snapshot export, or CDC's initial snapshot), then switch to the live topic from a known position, de-duplicating the overlap.
3
4**Counter-questions:**
5- *"Why not keep events forever?"* → Some teams do (compacted topics or long retention), at a storage cost and with data-retention rules to respect.
6
7**Trap:** asking the order service to republish everything.

Q10. Twitter-style timelines at scale: walk me through posting and reading.

They're testing: putting it together.

Think about your answer first, then open it.

reveal

1Show the strong answer, counter-questions and trap
2**Strong answer:** Posting writes the tweet once, then a fan-out service pushes its ID into followers' timeline caches, skipping very high-follower accounts. Reading loads the user's cached timeline, merges recent posts from the celebrities they follow, hydrates IDs into full posts from a cache, and returns. Inactive users' timelines are rebuilt on demand rather than kept up to date.
3
4**Counter-questions:**
5- *"What if the timeline cache is lost?"* → Rebuild it on read with fan-out on read for that user, then cache it.
6
7**Trap:** storing full post content in every follower's timeline instead of IDs.
8
9---
10
11## Practice coverage
12
13Every section in this chapter is tested at least once.
14
15| Section | Tested by |
16|---|---|
17| 23.1 Queue vs pub/sub | Level 1A, Level 3.1, Q1 |
18| 23.2 Topics and subscribers | Level 1A, Q8 |
19| 23.3 Consumer groups | Level 1B.1, Level 2.4, Level 3.1, Q4 |
20| 23.4.1 Fan-out on write | Level 1B.2, Level 2.1, Level 2.2, Level 3.2, Q2 |
21| 23.4.2 Fan-out on read | Level 1B.3, Level 2.3, Level 3.2, Q2 |
22| 23.4.3 Hybrid | Level 3.2, Level 5.1, Q3, Q10 |
23| 23.5 Event-driven architecture | Level 5.5, Q5, Q9 |
24| 23.6 Duplicates and ordering | Level 1B.5, Level 3.3, Level 5.4, Q7 |
25| 23.7 Outbox | Level 1B.4, Level 2.5, Level 3.4, Level 3.5, Level 4, Level 5.3, Q6 |
26
27---

23. Pub/Sub and Fan-Out