What Is Data Engineering?

A normal loop is a recruiter screen, a technical phone screen, then an onsite of three to five rounds. Almost every loop has a SQL round somewhere. After that the composition depends a lot on the company.

SQL Round: The bedrock. Around 69% of DE loops include one. Almost always window functions (LAG, LEAD, RANK), a CTE or two, an aggregation against a fact table, sometimes a self join over events. You type into a shared editor like CoderPad. Some interviewers run your output against hidden test cases, others just read your code and ask why.

Python or Coding Round: Shows up in just over half of loops. The bar is lower than a software engineering interview, but the problem space is different. Expect to parse a messy log, deduplicate records with a fuzzy key, or build a small generator pipeline. Forty-five minutes is typical. Pandas is fine. NumPy is overkill and signals you did not read the room.

Data Modeling: About a third of loops have an explicit modeling round. The prompt is usually a real product (Lyft rides, a payments ledger, a content feed) and you're asked to model it well enough to answer two or three analytical questions. SCD2 comes up. Star schema comes up less often than you'd think, in around 4.7% of loops, because most interviewers want to see how you reason about grain rather than which formal pattern you name.

Behavioral: Nobody loves these but everybody runs them. Have two stories ready about a data quality incident you owned end to end, and one story about a disagreement with a stakeholder. The number one failure mode is candidates who frame everything in terms of what their team did instead of what they personally did.

Reasons it fits: The pay band starts above the tech median and the senior bands compress less than software engineering, which is good for the middle of your career. Your work is hard to fake. Either the table is fresh and correct or it is not. Performance reviews are easier to argue when the answer is in the warehouse. Most teams are still understaffed, so even mid-level ICs end up owning systems end-to-end. That accelerates a resume. The skillset transfers. SQL and dbt look basically identical at Stripe, at a 40-person Series B, and at a hospital system.

Honest downsides: On-call. Pipelines fail at 2am because Snowflake had a regional incident, or because someone renamed a Salesforce field. A lot of the job is data quality archaeology. Why this number does not match that number. Hours of work, one Slack message of reward. Nobody notices when the warehouse is healthy. They notice loudly when it is not. The vendor churn is real. dbt, then dbt Cloud, then Coalesce, then SDF, then a Rust rewrite. Half of what you learn this year is a line item on a resume in three years. At some companies the data team is a service desk for whoever yells loudest. Read the org chart before signing.