Airbnb · Primly Community

Airbnb data engineer interview: pipelines and SQL were basically all of it

de_derek · 5 replies

just finished the airbnb data engineer loop last month. posting this because when i was prepping i could only find posts from 2022-2023 and things have shifted a bit.

there were 5 rounds total:

Phone screen (recruiter): pretty standard. 30 min, role expectations, visa situation, rough timeline. nothing technical.

SQL / data modeling round: this was the heaviest. 60 minutes, two problems on a shared coding environment. first was a multi-step aggregation query on a reservations dataset. think: compute rolling 7-day booking revenue per host category, handle nulls correctly, and then filter to top-performing markets. second was a simpler schema design question, they gave me a business problem ("design tables to track host payout disputes") and asked me to walk through the ERD and which columns i'd index. interviewer was engaged, pushed on my indexing choices pretty hard.

Pipeline architecture / system design: this is where airbnb shows its hand. they really care about batch vs streaming tradeoffs. i got a prompt about building a near-real-time pipeline for host quality signals. we talked about Spark, Flink, Kafka, and they specifically asked how i'd handle late-arriving data and backfill scenarios. if you have Airflow experience, mention it, they use it heavily internally.

Behavioral (values-heavy): two separate 45-min sessions. airbnb has 6 core values and they explicitly map questions to them. the one that tripped me up was "champion the mission" - they want real examples of connecting your day-to-day work to product outcomes, not just shipping clean pipelines. have a story ready where your data work directly influenced a product decision.

Hiring manager chat: last round, mostly conversational. they wanted to know how i handle ambiguity when upstream data quality is bad (which, every DE knows, is just... always).

overall the SQL bar felt mid-to-high senior level. i'd say LC medium-equivalent complexity but domain-specific. not algorithmic puzzles. they want you to think like a product data engineer, not a competitive programmer.

took about 3.5 weeks from phone screen to offer. offer came in L5 equivalent, SF-remote hybrid.

5 replies

analyst_ana

the rolling 7-day window question sounds brutal. did they let you use window functions freely or were there any constraints on the SQL dialect?

de_derek

full postgres dialect, window functions totally fine. honestly the window function part was the easy part. the tricky bit was they had some rows with null host_category and you had to decide whether to impute, exclude, or bucket them separately. i bucketed them as 'uncategorized' and they seemed happy with that.

ds_dmitri

the backfill question is airbnb's signature move. had it in my DS loop too. the right answer they're fishing for is idempotent pipeline design with a clear audit table so you can re-run any date range without double-counting. if you mention deduplication keys and incremental load strategies you'll get nods.

infra_ines

flink vs spark streaming for that near-real-time question: did they care which one you picked or just that you could articulate the tradeoffs?

de_derek

tradeoffs, not the choice. i went with spark structured streaming and justified it by saying our team already had spark expertise and the latency requirement (sub-5 min) didn't mandate true streaming. they pushed back gently with 'what if the requirement were sub-30 seconds' and that's when flink came up. they weren't testing brand loyalty to a tool.