Capital One · Primly Community

Capital One data engineer interview: pipelines and SQL are basically the whole thing

de_derek · 4 replies

just finished the capital one data engineer loop for a senior DE role in mclean. sharing because i couldn't find many recent write-ups that were specific about what they actually test.

the loop is 4 rounds: one recruiter/intro call, one sql-and-data-modeling round, one system design for data (pipelines, not microservices), and one behavioral. they might add a hiring manager chat depending on the team.

sql round. this was the most intense one. they're not messing around with toy queries. expect: multi-level CTEs, window functions (lag/lead, dense_rank), and a question that requires you to detect gaps or duplicates in event data. i got a question that was basically: here's a table of transaction events, find customers who had two failed auth attempts followed by a successful one within 30 minutes. you need to think out loud the whole time because they care about your reasoning, not just the output.

pipeline design. they asked me to design a near-real-time pipeline for fraud signal aggregation. they wanted to talk trade-offs between kafka + flink vs. batch-on-spark-with-short-intervals. they push hard on: how do you handle late-arriving data? what's your dedup strategy? how do you manage schema evolution? i mentioned schema registry + avro and that seemed to land well. they use AWS heavily (kinesis, glue, redshift) so familiarity helps, but they're not asking you to recite service names, they want you to explain why.

behavioral. classic STAR, probably 4-5 questions. 'tell me about a time a pipeline you owned failed in prod' was the most memorable one. they really want to see ownership, not just what went wrong but how you communicated and what you changed afterward.

my offer was in the $165-185k TC range (base + bonus, no equity as a non-SWE track). took about 5 weeks from recruiter screen to offer call.

one thing i noticed: they seem to care a lot about financial domain knowledge. not deep quant stuff, but knowing what batch settlement means, or why exactly-once semantics matter for payment pipelines. brush up on that if you're coming from a non-fintech background.

4 replies

analyst_ana

thank you for this. do they test spark specifically or is it more general distributed processing concepts? i have pyspark experience but haven't touched flink.

de_derek

they went deeper on concepts than on specific frameworks honestly. i talked about flink's watermarking for event-time processing and they seemed fine with me saying 'in flink this works like X' even without having used it in prod. if you know spark streaming, explaining the trade-offs vs. micro-batch should be enough.

ds_dmitri

the gap-detection SQL question is real. i got something similar at another bank. the trick is usually rownumber() partitioned by userid ordered by event_time, then subtract that from a running sequence to find breaks. worth drilling that pattern.

market_realist

5 weeks is pretty fast for a big bank. i've heard capital one moves faster than most. did the recruiter give you a timeline upfront or did it just happen to be that fast?