Kacam.net

Observe the world

Interview drills: one stock question, five layers down

The prompt is the exact sentence an interviewer opens with — "what's the difference between a process and a thread", "TCP versus UDP". The textbook definition isn't the finish line, it's the start: it records the version you already have memorised, drags you down five layers, then makes you say it again in sixty seconds.

These prompts only exist in 简体中文 for now.

How to use this

This set interrogates rather than teaches. It is a mock interview: running out of answers is a normal outcome, and it will not cover for you or finish your sentence when you trail off. If you want to understand a topic rather than rehearse for an interview, go to Computer systems, Databases or Algorithms — those guide you instead.

How one stock question fills forty minutes: the five-layer ladder

  • L1 · Definition — what each term means, list the differences. Can't answer: you never learned it.
  • L2 · Mechanism — which structure, which step actually produces that difference. Can't answer: you memorised the conclusion only.
  • L3 · Numbers — by how much? How many microseconds, how many times, how many I/Os. Can't answer: you've never done it hands-on.
  • L4 · Choice — given a concrete scenario, which one, at what cost, and when is that choice wrong. Can't answer: you can't apply it.
  • L5 · Limits — under what conditions does everything you just said stop holding. Can't answer: you know it all, but you don't really understand it.

The bar: L4 is what passes an interview, L5 is what counts as really understanding it, and anything that stops at L1 is recorded as memorised.

Six scenes — the first and last are unique to this set

  • In scene one it does not interrupt, comment or hint. You deliver the version you already have, start to finish, and it records that as your baseline. The silence is deliberate — it is listening, it just isn't correcting you yet.
  • The middle four scenes work down the ladder; scene three, mechanism and numbers, is the main battleground.
  • The last scene puts sixty seconds on the clock and makes you deliver it again, side by side with your opening version — because what an interview asks you to hand over isn't "I understand it", it's a sentence you can actually say out loud.

How to start

  1. Pick one, open it, copy the prompt.
  2. Paste it into a brand-new conversation, say nothing, and wait for its first question.
  3. Then answer like you would in an interview: deliver your memorised version in one go. It's fine if it's rough — that's the baseline it compares against at the end.
  4. After that it starts pressing. Running out of answers is a normal outcome: this isn't study material, it's a mock interview.

Five commands you can call out at any time

  • 跳过 move to the next scene now
  • 提示 give me a hint, not the answer
  • 直说吧 spell the full answer out
  • 重讲 jump straight to the last scene, re-record the sixty-second version and compare it with the opening one
  • 复盘 jump to the wrap-up and print the review card

Three traps

  • The prompt you copy has the answer key at the end — don't read it first. One of its columns is the "recital fingerprints": the exact sentences you're most likely to blurt out. Read them and there's nothing left to drill.
  • Don't hand over adjectives where a number belongs. "Much faster", "barely any overhead" — it stops you and asks for the figure. A conclusion you can't put an order of magnitude on is recorded as memorised.
  • The last two groups are language-bound: do the Java group if you write Java, the Go group if you write Go. They are not translations of each other — each picks its own questions — so there's no need to do both.

Interviewing for a backend role next week? Take process versus thread, TCP versus UDP, why B+ trees, isolation levels, the three cache failure modes, and idempotency — the six most frequent, with almost no overlap. Been hammered in interviews and want to find the gaps? Start with "from URL to rendered page": it exposes every weak spot horizontally, and you can pick your next ones from what it turns up.

Operating systems and concurrency

乐观锁和悲观锁
「乐观锁假设不会冲突」—— 那你说得出你们那张表的冲突率吗?
★★★☆☆ · about 35 min

Databases

四种隔离级别、幻读到底是什么
「可重复读解决了幻读」—— 真的解决干净了吗?给个反例。
★★★★☆ · about 40 min
MVCC 是怎么实现的
「多版本,读不加锁」—— 那「多版本」具体存在哪儿?
★★★★☆ · about 40 min
redo log、undo log、binlog 和两阶段提交
全系列最难的几道之一:崩溃恰好发生在那一瞬间,数据还对得上吗?
★★★★★ · about 45 min

Caching and messaging

缓存和数据库怎么保持一致
「先更新数据库再删缓存」—— 那这套做法的反例长什么样?
★★★★☆ · about 40 min
RDB、AOF 和主从复制
「有持久化也有主从,所以不会丢」—— 那主库挂掉那一刻呢?
★★★★☆ · about 40 min

Distributed systems

CAP 和 BASE
「三个里只能选两个」—— 那你真的有得选吗?
★★★★☆ · about 35 min
分布式事务几种方案怎么选
全系列最难的几道之一。先证明你真的需要它。
★★★★★ · about 45 min
幂等怎么设计
「加张去重表判断一下」—— 先说清:谁在重复,多久内算重复?
★★★★☆ · about 35 min
These prompts are open source They live in the public kacam-net/hero-prompts repo. If one of them doesn't press hard enough, or its answer key is missing a line, go and change it — whether it ships here is a separate call, made by the catalogue on this side.