American Express · Primly Community

American Express data engineer interview: pipelines, SQL, and what they really care about

de_derek · 4 replies

Finished the Amex DE loop about six weeks ago for a senior role on their payments data platform team. I'll share what happened, including what caught me off guard.

Recruiters said the process was 3-4 rounds. In practice it was five. Classic.

Round 1 (technical phone screen): SQL focused. Window functions, CTEs, a question about deduplication logic for transaction records. If you don't know how to think about idempotency in SQL cleanup, read up before you go in. They also asked about a time a pipeline I owned caused a data quality issue. Have that story ready.

Round 2 (system design for data): Design a real-time fraud signal ingestion pipeline. I walked through Kafka for streaming, Spark Structured Streaming for processing, and landing into a Delta Lake layer. They asked about exactly-once semantics. Know this cold if you're interviewing for payments infra: what does "exactly once" mean at the Kafka consumer level vs. the sink level? I fumbled it a bit and they were patient but noted it.

Round 3 (coding): Python, no tricks. Write a function to parse nested JSON from a payment webhook into a flat schema. Realistic problem, not LeetCode. They cared about handling null/missing fields gracefully.

Rounds 4-5 (behavioral + team fit): One with the hiring manager, one with a cross-functional partner (a risk DS). Both were STAR format. "Tell me about a time you had to rebuild trust with a stakeholder after a pipeline outage" was basically word for word asked.

Amex DE stack is mostly Spark, some Kafka, GCP BigQuery in some teams, legacy Oracle in others. Don't be surprised if they ask you about reading from and writing to Oracle in 2026. It's not glamorous but it's reality at a 175-year-old financial company.

Total comp offer I heard through the grapevine for a peer who accepted: around $175k base, 15-18% bonus target, strong 401k match. No equity, which is on brand for traditional financial services.

4 replies

sre_sol

The Oracle mention is so real. I interviewed for an infra-adjacent role there and they asked me how I'd handle incremental loads from an Oracle source system. I'd completely blanked on Oracle CDC. Not a disqualifier, but definitely prep for it.

backend_bekah

Exactly-once semantics in Kafka is a genuinely hard topic and a lot of people handwave it. The short version for anyone prepping: Kafka gives you idempotent producers and transactional APIs for exactly-once at the broker level, but end-to-end exactly-once depends on your consumer logic and whether your sink is idempotent. Know where the guarantee stops.

de_derek

Yes exactly. They wanted me to distinguish between "Kafka delivered it once" and "my BigQuery table has it once." Those are different problems.

visa_vik

Did the process move fast? I'm on a timeline and trying to gauge if Amex is worth pursuing right now.