Real Interview Questions for the Data Engineer at Meta
NewOverview
How the loop runs
Candidates describe a first conversation with a recruiter (background, motivation, fit) followed by a live technical screen of about an hour, split roughly evenly between SQL and Python in a shared editor (CoderPad is the one named in reports). Most accounts say the difficulty concentrates in the SQL half: multi-table schemas, window functions, and ratios that have to be computed in one pass rather than three.
What follows the screen is where accounts diverge. One candidate counts five rounds in total, including the recruiter call and a 30-minute round they don't describe precisely; others describe an onsite of several 45-60 minute rounds spanning data modeling, Python, and a behavioral or product-sense conversation. A 2022 report describes a problem-solving round where the prompt arrived 24 hours ahead and the candidate drew the architecture in Excalidraw. Treat the onsite shape as variable across teams and years: the reports are thin here, and no single fixed sequence should be assumed.
What interviewers are evaluating
The judgment being bought is whether you can turn an ambiguous product question into a schema and a query someone else could run and defend, under time pressure, out loud. It is not a tour of your tool stack. Nobody in these reports is scored on having used Spark, dbt or Airflow by name; they are scored on whether the grain of the fact table is right, whether the join fans out, and whether the metric definition survives a follow-up.
Three focus areas
SQL under a clock. Nearly every reported loop has a query round, and the questions are business-shaped: per-user counts with a date predicate, percentage-of-total ratios, top-N by an average computed over a related entity, deduplication across locations before counting. Window functions, CTEs, and careful handling of NULLs and ties are the recurring mechanics. Speed matters: candidates describe the screen as the hardest part precisely because there is no time to iterate.
Dimensional modeling and pipeline architecture. The onsite modeling rounds ask for tables, column types, primary and foreign keys, and the one-to-many and many-to-many relationships spelled out: ride-sharing, engagement metrics for a video feature, referral trees, cellular telemetry joined to app logs. Follow-ups push into slowly changing dimensions versus bridge tables, late-arriving events, and batch backfill against streaming updates. Interviewers want the trade-off stated, not the diagram alone.
Python and product-sense communication. The Python side is small and practical: dictionary inversion, word-by-word string comparison, fixed-size buffers over a stream, running averages: often with a complexity question attached. Alongside it sit metric-definition and behavioral rounds: one metric up 10% week-over-week while another falls 5%, a pipeline that missed its SLA, a disagreement with a data scientist over what a number means. Scoring is on how you reason and communicate, not on the drama of the story.
How to prepare
Answer aloud and time yourself: a schema design in 25 minutes with the trade-offs spoken, a SQL query narrated as you write it. Work through this card's Interview Q&A to compare your structure against a model answer, then the Practice MCQ to catch the mechanics you would otherwise get wrong silently. The role's own prep covers the concepts behind the modeling rounds.
Sources
Feedback on this overview
Curated: · Written: · Reviewed: