Tier 1 (必须会)

1. Why Kafka? When should you choose Kafka?

面试官真正想问：为什么不用 RabbitMQ？

Kafka 适用于：Event-driven / High throughput / Multiple downstream consumers / Replay / Decoupling services

Trigger: Trade / Price Update / Order Event / Analytics / Audit

2. Partition

Why partition? Horizontal scalability
How many partitions? 看 consumer 数量和 producer throughput
How to choose partition key? Preserve ordering for the same entity

例如 Partition Key = Asset，BTC 永远一个 Partition，保证顺序。只保证同一个 Partition 内有序。

3. Consumer Group

一个 Group 中：一条 Message 只消费一次。多个 Group：每个不同的 group 都消费一次。

4. Rebalance

什么时候发生：Consumer Join / Leave / Heartbeat Timeout / Partition Changed

为什么不好：Temporary pause / Latency spike

5. Offset

Kafka 最重要的概念。

Message → Offset
Consumer 记录 Offset
Replay: 修改 Offset

Replay 怎么实现？ 新创建一个 consumer group 并指定初始 offset。

Tier 2 (高频)

6. Why is Kafka fast?

四个原因：Append-only Log / Partition / Batch / Zero Copy

7. Ordering Guarantee

**同一个 Partition 内保证顺序，**跨 Partition 不保证。

如果 BTC 必须有序 → Partition Key 用 BTC。

8. Delivery Semantics

At Most Once  - 可能丢，但绝不重复 (poll → commit → process)
At Least Once - 不允许丢，但可能重复 (poll → process → commit)  ← 最常见
Exactly Once  - 最难实现

At Least Once + Idempotent Processing 是实际系统最常用的模式：

I would use at-least-once delivery together with idempotent business logic. Every transfer would have a unique transfer ID. Before processing, I’d check whether that ID has already been processed. If it has, I’d skip it.

Kafka 可以为 Kafka-to-Kafka 提供 exactly-once (idempotent producer + Kafka transactions)，但 end-to-end exactly-once 还需要 database transaction、idempotency key、outbox 等机制。

Tier 3 (了解即可)

9. Retention

Kafka Message 不删，有 Retention（例如 7 Days / 100 GB）。方便 Replay / Backfill / Audit。

10. Producer ACK

acks=0   → 不等 Broker (最快，可能丢)
acks=1   → Leader 收到就 ACK
acks=all → 所有 ISR 写成功才 ACK (金融交易推荐)

Trade-off: Latency VS Reliability

Replication Factor 决定每个 Partition 保存几份数据：

                Topic: Trade
           P0            P1             P2
Leader    Broker1       Broker2       Broker3
Follower  Broker2       Broker3       Broker1
Follower  Broker3       Broker1       Broker2

Producer → Leader (简化 consistency 和 ordering)
Follower 从 Leader 异步复制
Leader 挂了 → 从 ISR (In-Sync Replicas) 中选新 Leader

高频 Follow-up

#	Question	重要度
Q1	Why Kafka instead of RabbitMQ?	⭐⭐⭐⭐⭐
Q2	How do you preserve ordering?	⭐⭐⭐⭐⭐
Q3	What if consumers are slower than producers? (Backpressure)	⭐⭐⭐⭐
Q4	What happens if a consumer crashes? (Offset + Replay)	⭐⭐⭐⭐⭐
Q5	How would you scale consumers? (Partition + Consumer Group)	⭐⭐⭐⭐⭐
Q6	What causes rebalance?	⭐⭐⭐⭐⭐
Q7	How would you replay yesterday’s events?	⭐⭐⭐⭐⭐

Q8: Can Kafka guarantee global ordering?

No. Kafka only guarantees ordering within a partition. If global ordering is required, I would use a single partition, although that limits throughput.

Q9: Why doesn’t acks=all wait for every replica?

Because some replicas may be temporarily out of sync. Waiting for every replica would significantly reduce availability. Kafka only waits for the in-sync replicas (ISR), which provide a good balance between reliability and performance.