Netflix · Primly Community

Netflix data engineer interview, pipelines and SQL, here's what they asked

analyst_ana · 6 replies

i went through the Netflix data engineer loop earlier this year for a senior DE role. sharing because the internet has very little on what this loop actually looks like compared to their SWE process.

total rounds: phone screen + 4 technical rounds + culture contribution

the core technical areas:

SQL: two rounds had heavy SQL. and i mean heavy. Netflix uses Presto/Trino and Spark SQL. i got questions that required window functions, lateral joins, and handling NULL propagation across aggregations. one question was specifically about writing a query that identified users whose streaming quality degraded over a session - which required self-joins and window functions in the same query. i've been a DE for 7 years and it was hard.

pipeline design / architecture: i was asked to design a batch pipeline that ingests 10TB+ of daily event data, transforms it for analytics, and handles late-arriving records and schema evolution. the late-arriving records piece is where they went deep. how do you handle an event that shows up 48 hours late? does your pipeline reprocess? what's the cost? how do you make that decision? this is very real Netflix problem territory - they deal with data arriving from devices globally with wildly variable connectivity.

data modeling: one round was specifically about dimensional modeling and slowly changing dimensions. i got asked to design the data model for tracking a user's content preferences over time, including handling changes in their plan tier, geographic location, and device type. the interviewer wanted to know which SCD type i'd use for each dimension and why.

tools proficiency: Spark, Airflow, and dbt came up naturally in discussion. i was not asked to code in Spark but was asked about partitioning strategies, shuffle optimization, and how i'd tune a Spark job that was OOMing.

what i'd tell someone prepping: the SQL is harder than you expect. practice actual Presto-compatible window functions. the pipeline design needs to show operational maturity, not just architecture. data modeling seems old-fashioned but Netflix's analytics infra is deep and they care about it.

6 replies

infra_ines

the late-arriving records problem is a classic and genuinely hard. did they ask about watermarking / event time vs processing time at all, or was it more operational (reprocess windows, etc.)?

de_derek

it was framed operationally - when do you trigger a reprocess, how do you decide the cost vs accuracy trade-off. they mentioned watermarking when i brought it up but the focus was on the business decision, not the Flink/Kafka Streams mechanics.

analyst_ana

how different is the DE loop vs the DS loop? i'm a data analyst trying to figure out which track to target at Netflix

de_derek

very different focus. DE = pipelines, architecture, SQL for ETL purposes. DS = statistics, experiment design, SQL for analytical purposes. overlapping SQL skills but different depth. if you're an analyst i'd probably aim DS side rather than DE unless you're already writing production pipelines.

sre_sol

7TB daily event ingestion with late arrivals sounds like a Kafka/Flink problem but you said batch. do they not do streaming pipelines for that kind of volume?

de_derek

they do both. my scenario was framed as batch but i talked about hybrid approaches. i think the interviewer was specifically testing batch architecture knowledge since that's still the core of their analytics workflows. streaming came up as a follow-up.