Senior Data Engineer Interview Experience

I treated it like a multi-table join across the books, users, and checkout tables, then counted overdue unreturned books per user and returned the username with the count. The main thing was getting the joins and filtering around the expiry date right under time pressure.

I grouped the result by user age and counted users who had checked out geography-category books. This one was more straightforward, but in that round even the simpler SQLs mattered because the clock was so tight.

Find the number of people who checked out a book on the same day that another person returned that same book.

I handled it as a same-book, same-day condition between a return event and another user's checkout event, then counted the people matching that pattern. It was one of those questions where you had to be very careful with the event logic and not overthink it.

I used a hashmap or array approach. The input was a list of strings, and I returned the character and the number of occurrences. I also accounted for the edge case around adjacent characters and cases where the second character logic could fail, so I would skip or handle those cleanly instead of forcing it.

Write a function to read and process a CSV data file, and handle the case where the file is not found.

This one was really about structured error handling. I used try/except while reading the CSV and processing it, and if the file was missing I returned or printed 'file not found.' They seemed to care that I handled it in a clean Pythonic way, not just that I could read the file.

I grouped salaries by department and then found the second highest salary for each one. It was a typical data manipulation problem, but again the challenge was doing it fast and correctly because there were a lot of test cases.

I laid out a relational schema with the core entities like users, drivers, rides, payments, and locations, and defined the PK/FK relationships. Then they pushed on scaling and asked how I'd optimize the transactional table past 1 TB. I said I'd use bucketing based on a date timestamp, and when they got into partitioning tradeoffs I said hash-based partitioning made sense because it's easier for lookup at volume. That part felt like a pure technical depth check.

I treated it as a streaming aggregation problem and computed ride-request counts in 15-minute tumbling windows. The question itself was pretty direct, but it matched the rest of the loop in that they wanted something practical and data-engineering oriented, not just generic Python syntax.