Skip to main content
VirtualInterview.ai
Sign InSign Up

Mock interviews by role · Data, AI & Cloud

Data Engineer mock interview

Data engineer interviews look for pipelines that fail loudly and can be rerun. A diagram of tools with no mention of late data, duplicates, or who owns the contract is incomplete.

Talk through one table from source event to the column an analyst trusts, including what you do when yesterday’s job did not finish.

Practice this role with an AI interviewer

What the loop usually looks like

  1. Screen

    30 min

    A pipeline you own and a failure you repaired.

  2. Design

    45 min

    Ingest, model, and serve a dataset with freshness needs.

  3. Collaboration

    30 min

    Contracts with analysts or software engineers.

What interviewers score

  • Idempotent pipelines
  • Freshness and lineage
  • Backfills
  • Contracts with consumers

10 data engineer interview questions and model answers

  1. Technical

    1. How do you make a daily job safe to rerun?

    What they are testing: Idempotency.

    The job writes a partition for that day and replaces it, or it upserts on a stable key. A second run does not append a second copy. I test the rerun on a small date before I trust the scheduler.

  2. Technical

    2. A source system changes a column meaning without notice. How do you catch it?

    What they are testing: Contracts.

    I check row counts, null rates, and a few known totals against the previous day. A break fails the job and pages the owner. I do not let a green job publish a silently redefined metric.

  3. Domain

    3. How do you decide batch versus stream?

    What they are testing: Fit for purpose.

    If the decision can wait an hour, batch is simpler and easier to backfill. Stream is for a product path that must react in seconds, and it costs more to operate. I ask what decision is late if the data is late.

  4. Situational

    4. An analyst says the dashboard is wrong. Where do you start?

    What they are testing: Debugging with a consumer.

    I ask which number, which filter, and which morning it changed. Then I compare the source count, the warehouse count, and the dashboard query. The first mismatch is the layer I fix.

  5. Behavioral

    5. Tell me about a backfill you ran on a large table.

    What they are testing: Operational care.

    I ran it in date chunks, watched warehouse load, and kept the old column until the new one matched on a sample of days. Consumers switched after the check, not when the job merely finished.

  6. Technical

    6. Where do you handle late-arriving events?

    What they are testing: Correctness.

    I define a lateness window and restate recent partitions inside that window. Events older than the window go to a quarantine I can inspect. I say how long a number can change after the day ends.

  7. Behavioral

    7. Describe a pipeline you deleted.

    What they are testing: Simplicity.

    It fed a report nobody had opened in a quarter. I confirmed with the owner, turned it off, and kept the code for a month. One less failure page is a real result.

  8. Domain

    8. What belongs in a modeled table versus a raw table?

    What they are testing: Modeling.

    Raw keeps the source shape for replay. Modeled tables name the business grain and the filters analysts should not reinvent. I do not let every dashboard join raw logs differently.

  9. Situational

    9. Finance and product disagree on revenue. What do you do?

    What they are testing: Definitions.

    I put both definitions on one page with the filters each uses. If they need different numbers, they get different names. I do not pick a silent winner in the warehouse.

  10. Technical

    10. How do you limit the blast radius of a bad deploy of a transformation?

    What they are testing: Safety.

    I run the new model in parallel and compare totals before I point dashboards at it. The swap is a view change I can reverse. I do not overwrite history as the first step.

Frequently asked questions

How much coding is in the loop?

Expect SQL and a design conversation. Some companies add a Python exercise for a transform. Practice narrating the grain of the data while you write.

Should I name specific warehouses and orchestrators?

Name the ones you have operated. The concepts they will test are reruns, freshness, and contracts, which survive a tool change.

What is a strong project story?

A pipeline that was wrong or late, what you changed, and how a consumer knew it was fixed. Skip the tour of every service in the company.

Can I practice the design round by voice?

Yes. Describe the source, the partition, the check that fails the job, and the backfill. A data engineer mock will ask what happens on a rerun.

Keep practicing