Data engineering interview questions that actually come up.

SQL is the single biggest signal in a data engineering interview. Expect at least one live SQL problem — usually windows or self-joins.

• Find the 2nd highest salary per department. Tests window functions (DENSE_RANK) vs correlated subqueries. Say why you picked one.
• Compute a 7-day rolling average of daily active users. Tests RANGE BETWEEN INTERVAL '6' DAY PRECEDING AND CURRENT ROW and handling of days with zero events.
• Given events, find sessions (30-min inactivity gap). Classic gaps-and-islands with LAG + a running sum.
• Deduplicate a table keeping the latest row per key. QUALIFY ROW_NUMBER() OVER (PARTITION BY id ORDER BY updated_at DESC) = 1.
• Explain the difference between WHERE and HAVING. And where QUALIFY fits.

Not LeetCode. Real "can you write a small ETL script" questions.

• Parse a JSON file with nested arrays and flatten it into rows.
• Write a generator that streams lines from a large file without loading it into memory. They want to see yield and understand why.
• Retry an HTTP call with exponential backoff. Bonus points for mentioning tenacity or writing a clean decorator.
• Difference between list, tuple, set, dict — and when you'd use each in a pipeline.
• What does __init__.py do? What is if __name__ == "__main__"? Sanity checks for people who Copy-pasted their way in.

• Narrow vs wide transformations. Which ones cause shuffles?
• What is data skew and how do you fix it? Salting, broadcast joins, AQE.
• Broadcast join vs sort-merge join — when does each win?
• How do you tune the number of partitions? Talk about spark.sql.shuffle.partitions, target partition size (~128MB), and file size after write.
• Explain lazy evaluation and the DAG. Why .count() can be expensive on a chain of transformations.

• Star vs snowflake schema — trade-offs.
• SCD Type 1 vs Type 2 — when do you need history?
• How would you model an e-commerce dataset (orders, users, products) for analytics?
• Explain dbt sources, staging, marts. Where do tests live? What's a snapshot?
• Incremental models — when to use merge vs append vs delete+insert.

System design separates junior from mid, and mid from senior. There's rarely one right answer — they're grading how you reason.

• Design a pipeline for near real-time analytics on user events. Kafka → Spark Streaming / Flink → Iceberg → dbt → warehouse. Discuss latency SLAs and exactly-once.
• Design a data warehouse for an e-commerce company. Sources, ingestion pattern (CDC vs snapshot), staging, marts, orchestration, monitoring.
• How do you handle schema evolution? Avro/Protobuf, backward/forward compatibility, contract testing, migrations in the warehouse.
• Batch vs streaming — how do you choose? Latency, cost, complexity, freshness requirement.
• Idempotency and exactly-once — how do you actually guarantee them? Merge keys, dedupe windows, transactional sinks, Kafka + Iceberg transactions.

• "Tell me about a pipeline that broke in production. What did you do?"
• "How do you handle a stakeholder who wants a dashboard 'by tomorrow' on data that doesn't exist yet?"
• "Describe a technical decision you disagreed with. How did you handle it?"
• "How do you document a pipeline so the next person doesn't page you at 3am?"

Use the STAR framework (Situation, Task, Action, Result). Keep answers under 2 minutes.

The 14-course roadmap covers every fundamental behind these questions — SQL windows, Spark internals, dbt patterns, Kafka, Iceberg, orchestration. Bug Hunt drills you on real broken pipelines. Once you can complete the roadmap and hunt bugs unassisted, most of these questions become one-line answers instead of stumbling paragraphs.