A data engineer can pass a difficult SQL interview and still struggle with the work that prompted the vacancy: numbers that change unexpectedly, pipelines nobody trusts and backfills that require the CTO to supervise. If your problem is dependable data delivery, assess ownership of a dataset through failure and recovery, not just the ability to produce a transformation.
The following interview is designed for a senior individual contributor joining a small or growing data team. It uses a fictional orders pipeline, so candidates do not need access to your customer data. The assessment format and hiring thresholds are recommendations to adapt to your environment.
Decide which part of the data journey needs an owner
Write down where work currently stalls. If reliable source tables exist but business definitions are inconsistent, an analytics engineer may be the more useful hire. If source ingestion, change handling and recovery are fragile, investigate data engineering experience. If several teams need shared infrastructure, the vacancy may belong in data platform engineering. These titles overlap; the responsibilities should not.
For a UK or Netherlands SaaS team, a practical brief might be: “Own the path from application order events to a reconciled daily revenue dataset, working with finance on definitions and platform engineers on infrastructure.” State who can change upstream schemas, who approves business logic and who responds when the daily dataset is late. Do not make the hire accountable for agreements leadership has not secured.
Give candidates a small, explicit evidence pack
Prepare a source schema, a transformation, a few synthetic records and a one-page incident timeline. The fictional system receives order-created and order-updated events, writes them to a warehouse and produces a daily aggregate. Include one repeated event, one late update and a currency field that appears partway through the history. Define whether the dataset measures bookings, payments or recognised revenue; otherwise you are testing guesses about accounting language.
Share the packet at the start of a 60-minute discussion. Allow documentation and explain the tool policy beforehand. Spend ten minutes clarifying the contract, twenty on diagnosis, twenty on recovery and ten on ownership. There is no requirement to build a working pipeline. Ask the candidate to sketch the smallest useful change and explain what evidence would justify it.
Ask them to make recovery safe
Introduce the incident: yesterday’s aggregate is wrong, and the team proposes replaying a week of events. Ask what must be known before approving that replay. Listen for questions about stable identifiers, update semantics, transformation versions, downstream consumers and whether repeating a write changes the result. “Rerun the job” is an action, not yet a recovery plan.
A good response should distinguish correcting stored data from correcting decisions already made from it. The candidate might propose an isolated reconstruction, a comparison against source totals and a controlled replacement once differences are understood. An alternative can be equally sound if they explain its assumptions. Require them to address how normal processing continues while the repair runs and how a failed repair would be stopped.
Apache Kafka’s design documentation explains that exactly-once processing involving an external destination requires coordination with that destination. Use that narrow point to challenge an unsupported “Kafka guarantees it” answer. Do not turn the interview into recall of broker settings: the hiring question is whether the candidate can define where the guarantee begins and ends.
Use a rubric tied to observable decisions
| Capability | Evidence that meets the role | Gap to investigate |
|---|---|---|
| Data contract | Clarifies keys, update rules and meaning with an accountable owner. | Assumes a timestamp uniquely identifies a business event. |
| Recovery | Explains replay effects, validation and a safe cutover. | Overwrites production outputs before checking differences. |
| Diagnosis | Separates missing data, late data and incorrect transformation logic. | Treats a successful job status as proof of correct data. |
| Operations | Names actionable checks and a response owner. | Adds alerts without defining what anyone should do. |
Keep a “not observed” option. If the interview never reaches recovery, do not silently mark the candidate weak at it. Use a focused follow-up. Calibrate the rubric with an engineer who knows your environment before interviewing applicants, and record evidence independently before comparing conclusions.
Test the handover between engineering and the business
Ask the candidate to explain the incident to a finance colleague in three sentences. They should separate confirmed facts from uncertainty, identify which outputs are affected and describe the next update. Do not reward confident reassurance when the scope is unknown. Then ask what they would document so another engineer could perform the next recovery.
For a team split between Britain and the Netherlands, make the operational window concrete: which local-time deadlines matter, who covers absence, and when should an issue be escalated? A one-hour time difference is less important than an unstated expectation that the same person always remains available. Clarify that expectation during the search.
Turn the interview into a credible first assignment
The first assignment could be to map one critical dataset, agree its freshness and correctness checks, and rehearse a recovery in a safe environment. It should have an upstream collaborator and a business owner, not just a deadline. This creates continuity between what you assessed and what you ask the person to do.
Use our technical interview scorecard guide to record the evidence. If you cannot identify the dataset, its owner or the decisions it supports, resolve those questions before adding more technologies to the job description.