Meta Machine Learning Engineer Interview Questions

I would structure the answer around the problem, the model, the deployment path, and the impact. I would explain why the project mattered, what data I used, how I trained or evaluated the model, and what changed after launch. The reason I choose this structure is that interviewers want to see that I can connect modeling work to real product outcomes. I would also mention one hard trade-off or failure I ran into and how I handled it. If the interviewer wants more detail, I would talk about monitoring and iteration after launch.

I would first check data quality to make sure the drop is real and not a logging or pipeline issue. Then I would segment by device, region, surface, and recent product changes to find where the drop is concentrated. The reason I choose this approach is that metric changes are often caused by a small slice of traffic rather than the whole population. I would also compare the affected period to previous weeks and look for launches, outages, or external events. If needed, I would mention that I would keep narrowing the problem until I found the most likely root cause.

I would first ask whether the missingness is random or correlated with the target or another important feature. Then I would compare options like imputation, adding a missing-value indicator, using a domain-specific default, or dropping rows if the business impact is acceptable. The reason I choose this approach is that the best fix depends on why the values are missing and how the feature is used. I would also think about whether the missingness itself contains signal. If the interviewer wants more detail, I would explain how I would validate the choice on a held-out set.

I would join the orders and customers tables, filter orders to the last 30 days, group by customer, sum revenue, order by that sum descending, and limit the result to 5. The reason I choose this pattern is that it is the standard way to answer a top-k revenue question cleanly. I would also think about whether refunds, canceled orders, or currency conversion should be included. If the interviewer asks, I would mention using a window function as an alternative when more detail is needed. I would keep the query simple unless they specifically ask for edge cases.

An INNER JOIN returns only rows that match in both tables, while a LEFT JOIN returns all rows from the left table and matching rows from the right table when they exist. I would explain it with a simple example like customers and orders, where a LEFT JOIN can show customers even if they have not placed an order. The reason I choose this example is that it makes the difference easy to visualize in product terms. I would also mention that NULLs appear on the right side of a LEFT JOIN when there is no match. If needed, I would show how this affects reporting queries.