Data Engineer Interview Guide: The Questions Behind SQL, Pipelines, and Reliability

数据工程师面试指南：SQL、管道和稳定性背后真正会被问什么

9 January 2026

12 min read

Data Engineering Specialist

Principal Data Engineer at Amazon, author of 'Data Engineering Handbook'

摘要 Summary

Data engineer interviews often look like a wall of unrelated questions until you group them by the decisions engineers actually make. This guide focuses on data modelling, pipeline design, operational reliability, and the judgement signals that separate a merely technical answer from a good engineering answer.

数据工程师面试看上去像是一堆互不相关的问题，其实背后绕来绕去就那么几类决策。这篇文章会把它们收成数据建模、管道设计、运行稳定性和工程判断四块，帮你把答案从会写，推进到更像真的做过。

Strong data engineering interviews are rarely about showing the most complicated architecture you can imagine. They are about making sensible trade-offs under constraints: data quality, latency, cost, scale, ownership, and how much operational pain the team can absorb.

数据工程师的强答案，很少靠炫最复杂的架构取胜。真正重要的是，你能不能在约束下做合理取舍：数据质量、延迟、成本、规模、归属关系，以及团队到底承受得起多少运维负担。

That is why fragmented prep does not work well. A warehouse question, a streaming question, and a data quality question often collapse into the same underlying judgement once you look closely.

所以碎片化准备通常效果一般。仓库设计题、流式处理题、数据质量题，看起来不同，往深里看，底层考的往往还是同一种判断。

The Four Areas Interviewers Usually Care About| 面试官通常最在意的四块

Modelling: how you define grain, keys, history, and whether the dataset can support the business question cleanly.
建模：你怎么定义粒度、主键、历史保留，以及数据集是否能干净地支撑业务问题。
Pipeline design: batch versus stream, orchestration, dependencies, and recovery behaviour.
管道设计：batch 还是 stream、调度方式、依赖关系和失败恢复。
Data quality: validation rules, alerting, reconciliation, and ownership when something breaks.
数据质量：校验规则、告警、对账，以及出问题时谁负责。
Operational judgement: cost, observability, backfills, and how maintainable the whole thing is for the team.
运行判断：成本、可观测性、回灌、以及团队是否真的维护得动。

SQL and Modelling: Start With Grain| SQL 和建模题，先抓粒度

Candidates often jump straight into table names and joins. A better answer starts one step earlier: what is a row supposed to represent, how does that change over time, and which keys make duplicates impossible rather than merely unlikely.

很多候选人一上来就开始说表名和 join。更好的答法会往前退一步：一行数据到底代表什么、时间变化怎么记录、哪些 key 能让重复从根上不成立，而不是只是碰巧不常发生。

sql

WITH ranked_events AS (
  SELECT
    order_id,
    status,
    updated_at,
    ROW_NUMBER() OVER (PARTITION BY order_id ORDER BY updated_at DESC) AS rn
  FROM raw_order_events
)
SELECT order_id, status, updated_at
FROM ranked_events
WHERE rn = 1

Pipeline Design: Explain the Trade-Off, Not Just the Diagram| 管道设计题：别只画图，先讲取舍

Why this needs to be real time or why batch is enough.
为什么这件事必须实时，或者为什么 batch 其实已经够用。
What happens if one upstream feed arrives late or arrives twice.
如果上游迟到一次，或者重复来一次，会发生什么。
How you would backfill safely without corrupting downstream consumers.
怎么安全回灌，而不把下游消费方的数据搞脏。
What you would monitor at ingestion, transformation, and delivery layers.
在 ingestion、transformation 和 delivery 三层分别监控什么。
Who owns a failure at 2 a.m. and how the team would know before a stakeholder does.
凌晨两点出故障到底谁接，团队又怎么做到比业务方更早知道。

Reliability Questions Are Usually About Engineering Maturity| 稳定性问题，本质上是在看工程成熟度

Do you define data contracts or at least expectations between producers and consumers?
你有没有数据契约，至少有没有明确的生产者和消费者预期？
Do you distinguish retryable failures from logic failures?
你会不会把可重试错误和逻辑错误分开？
Do you measure freshness and completeness, not just job success status?
你看的不只是任务成功失败，还会看 freshness 和 completeness 吗？
Do you think about the cost of operating your design six months later?
你有没有想过，这套设计六个月后维护成本会变成什么样。

A Practical Prep Plan| 一个更接近实战的准备计划

Take three pipelines you have built and rewrite the story for each around grain, dependencies, failure modes, and monitoring.
找 3 条你做过的管道，把每条都改写成围绕粒度、依赖、失败模式和监控来讲的版本。
Practise one warehouse design question and one streaming question on the same day so you learn the overlap.
同一天练一道仓库设计题和一道流处理题，逼自己看清它们的共通处。
Review SQL with a focus on window functions, deduping, and data quality checks.
SQL 重点复习 window function、去重和数据质量检查。
Do one mock where the interviewer is allowed to keep breaking your assumptions.
至少做一次 mock，让对方不停打断并挑战你的假设。

常见问题 FAQ

What does Data Engineer Interview usually test?

数据工程师面试通常会重点看什么？

Most rounds in this guide test a mix of role understanding, structured communication, and follow-up resilience. For technical or case-heavy roles, you also need to show how you break a problem down instead of jumping straight to a memorized answer.

从这篇文章覆盖的内容来看，这类面试通常会同时看岗位理解、表达结构和追问下的稳定性。技术或案例占比更高的岗位，还会额外看你能不能把问题拆开，而不是只会背现成答案。

How should I use this guide if I only have a few days before the interview?

如果距离面试只剩几天，这篇文章应该怎么用？

Use the opening sections to identify the main signals first, then focus on the recurring examples, frameworks, or technical topics that the article highlights. The FAQ and summary help you decide what deserves practice time and what can stay secondary.

先用开头部分抓住这场面试最核心的判断标准，再回头练文中反复出现的案例、框架或技术点。摘要和 FAQ 的作用，就是帮你判断哪些内容值得优先练，哪些可以先放一放。

What mistake causes candidates to underperform most often in Data Engineer Interview?

准备数据工程师面试时，最容易拉低表现的错误是什么？

The most common problem is giving answers that sound prepared but do not survive follow-up questions. Interviewers usually notice when the structure is there but the underlying judgment, numbers, or trade-offs are missing.

最常见的问题，是答案表面上很完整，但一到追问就露出底子不够。面试官通常很快就能听出来：你的结构在，判断、数据和取舍却没有真正想清楚。