Data Engineer Interview Questions

Data Engineer Interview Questions

Live-executed practice problems and rubric-scored design problems covering every round of the data engineer interview loop.

1,400+ data engineer interview questions covering all 5 rounds of a typical loop: SQL (95% of loops), Python (70%), data modeling (55%), pipeline architecture (52% of senior-and-above), and PySpark (30% at Spark-first companies). Every coding problem runs against a real grader; every modeling and design problem ships with a rubric-scored verdict.

A typical data engineer interview loop in 2026 has 5 rounds. SQL appears in 95 percent of loops as a 30-to-45 minute coding round on a Postgres-flavored editor, with patterns like top-N per group, dedup latest, sessionization, gap-and-island, SCD2 half-open joins, and idempotent MERGE. Python appears in 70 percent of loops as a 30-to-45 minute coding round on a sandbox, usually vanilla Python with pandas at most companies and polars at the leading edge. Data modeling appears in 55 percent of loops as a 45-minute whiteboard exercise on a marketplace, rideshare, payments ledger, or content platform domain. Pipeline architecture or system design appears in 52 percent of senior-and-above loops as a 45-to-60 minute end-to-end design with a concrete scenario (10 million events per day, 15-minute freshness SLA, downstream BI tool that cannot handle table swaps). PySpark appears in 30 percent of loops at Spark-first companies (Databricks, Netflix, Uber, Airbnb, DoorDash, Spotify) as a 45-to-60 minute coding round emphasizing DataFrame fluency, join strategy selection, skew diagnosis, and Spark UI reading.

What loses points across all 5 rounds is rarely syntax. In SQL it is a missed NULL case, a wrong grain, or a window definition that ties when the interviewer wanted a strict order. In Python it is catching bare Exception, not validating record-level errors, defaulting to O(n-squared) when a dict lookup is O(1). In modeling it is failing to state the grain in one sentence before drawing, or over-normalizing instead of picking a star with conformed dimensions. In design it is whiteboarding the right architecture without naming durability, replay, idempotency, and what happens when the upstream dies on a holiday. In Spark it is ignoring skew, broadcasting a table that will not fit in driver memory, or failing to read a Spark UI screenshot showing stage time imbalance.

Rubric weights are calibrated to data engineer interview write-ups from Meta, Amazon, Google, Netflix, Stripe, Databricks, Snowflake, Airbnb, Uber, DoorDash, and Pinterest. Mid-level (L4) rubrics weight correctness and clean code. Senior (L5) rubrics weight trade-off articulation and failure-mode naming (3 failure modes per component in design rounds). Staff (L6) rubrics weight adapt-on-the-fly behavior when the interviewer changes a requirement mid-round. The catalog is filterable down to seniority and to specific companies once you know your target.

How many rounds does a data engineer interview loop have?
5 typical rounds: SQL (95 percent of loops), Python (70 percent), data modeling (55 percent), pipeline architecture or system design (52 percent of senior-and-above loops), and PySpark (30 percent at Spark-first companies). Phone screens are usually SQL plus a take-home or a 60-minute design walkthrough. Onsites add the remaining rounds plus a behavioral round.
What is the difference between a mid-level and senior data engineer loop?
Mid-level (L4) loops weight correctness and clean code: did you produce a working solution that handles the obvious edge cases. Senior (L5) loops weight trade-off articulation and failure-mode naming: did you name 2 alternatives, defend the choice, and call out what happens when the upstream dies on a holiday. Staff (L6) loops weight adapt-on-the-fly behavior: when the interviewer changes the SLA from 15-minute freshness to 1-minute mid-round, did you pivot cleanly.
Do all data engineer interviews include a Spark round?
No. Spark-first companies (Databricks, Netflix, Uber, Airbnb, DoorDash, Spotify) test PySpark or Scala-Spark in a dedicated 45-to-60 minute round. Companies on Snowflake or BigQuery without Spark in production (Stripe, Block, Coinbase) generally skip the Spark round. Check the company-specific list for the round shape at your target.
How is the data modeling round graded if there is no 'correct' answer?
The rubric scores grain selection (did you state 'one row per X' before drawing), SCD type choice with justification (Type 1 vs Type 2 vs Type 3, why), dimension shape (conformed dimensions across fact tables), and trade-off articulation (why star over snowflake or OBT, why this grain). The interviewer accepts multiple valid designs; the score comes from defense, not from picking the canonical answer.
What is the format of the pipeline architecture round?
A concrete scenario (e.g. 40 million clickstream events per day, 15-minute dashboard freshness SLA, downstream BI tool that cannot handle table swaps), 45 to 60 minutes, end-to-end design on a whiteboard or canvas. The rubric weights SLA match, cost reasoning, failure modes (what happens when Kafka brokers fail, when the source schema changes, when a partition is late), tool fit (why Kafka here and not S3 with notifications), and mid-round adapt-on-fly when the interviewer flips a requirement.
Should I prep for behavioral rounds too?
Yes. Behavioral rounds appear in every loop and account for the hire/no-hire decision when technical rounds split. Use STAR-D format (Situation, Task, Action, Result, Decision postmortem). Specific numbers required (latency, dollars, hours, downstream consumers). The decision postmortem (what you would do differently and why) is the senior-vs-mid signal.
How many problems should a data engineer solve before a FAANG onsite?
Roughly 150 problems for FAANG-level loops, distributed across rounds: 60 SQL, 40 Python, 20 modeling whiteboards, 20 pipeline designs, 10 PySpark if Spark-first. Plus 3 timed mocks in the final 2 weeks. The bottleneck most candidates hit is not volume; it is narration: thinking out loud under pressure without going silent for 2 minutes.

1550 practice problems matching this filter. Domains: SQL (931), Pipeline Architecture (144), Data Modeling (63), Python (412). Difficulty: medium (741), hard (265), easy (544).

SQL (931)

Pipeline Architecture (144)

Data Modeling (63)

Python (412)