Uber APM Interview — Surge Trip-Completion Drop in One City

It tests whether you can take a vague symptom, a marketplace metric that moved in one city, and run a structured root-cause investigation under interruption. The interviewer wants to see you confirm the move is real and not an instrumentation artifact, write the trip-completion ratio as a formula, segment it by geography, time window, rider versus driver side, and pricing tier, order your hypotheses by likelihood and impact, size the business impact of the leading one, and finish with a single concrete next data pull. The number you land on matters far less than the path you take to it.

Before any cause, confirm the drop is real: check the date range, the pipeline, and whether a logging change explains it. State the metric as completed trips divided by ride requests over a window. Then isolate one funnel stage rather than guessing globally, and segment by city, time, rider side, driver side, and price tier. Build a short ranked list of hypotheses, pick the most likely and highest impact one, quantify what it costs in rides or revenue, and propose the one data pull that would confirm or rule it out. Narrate the order out loud so the interviewer can follow your logic.

The biggest one is theorising causes before proving the metric move is real and not a logging or pipeline change. Close behind: never stating the metric definition or its denominator, listing ten possible causes without ranking them, jumping to a product feature fix, proposing a rider-side change that quietly starves driver supply, never sizing the business impact, and ending without a concrete next data pull. Treating it as a clean A and B test also fails, because riders and drivers spill across any boundary you draw.

It behaves like the analytical-round interviewer, not a friendlier version. It interrupts when you theorise too early, pushes for the funnel stage you would isolate first, and asks you to size impact before you escalate. It will not praise you mid-session or tell you whether you are doing well. It answers a precise clarifying question in one sentence and hands over a specific data point only when you ask a targeted question for it. What it cannot do is read body language, so everything is judged from what you actually say.

Your transcript is scored against the dimensions a real APM loop grades: confirming the anomaly is real, decomposing the metric, systematic segmentation, a ranked hypothesis tree, impact quantification, two-sided marketplace reasoning, and proposing a concrete next data pull. Each dimension has observable signals, for example whether you stated a denominator or named a control city. You receive a scorecard that quotes specific moments from your answers and names the exact step you skipped, so you can see where the investigation broke rather than just a number.

Do not start guessing causes. Use the opening to lock down what you are even looking at: ask what window the drop covers, restate the trip-completion ratio in your own words including its denominator, and ask whether anything changed in instrumentation or the data pipeline recently. Confirm whether other metros moved too. Those two minutes signal that you separate a real marketplace shift from a measurement artifact, which is the single behaviour this interviewer is watching for at the start.

Acknowledge why a clean experiment fails here: riders and drivers move across the boundary between treatment and control, so the control is contaminated by network spillover. Then offer evidence you can still get: a switchback design that alternates the change over time in the same market, a synthetic control built from comparable metros, or a before-and-after read on a tightly segmented slice with the leading confounders held constant. Name what each approach can and cannot prove. The interviewer is testing whether you reason about marketplace network effects, not whether you can recite an experiment design.

It sounds like someone slowing the interviewer down before speeding up. First, is this real, what window, did instrumentation change. Then, here is the ratio and its denominator, here is the one funnel stage I would isolate first and why. Then, I would cut this by city, time, rider versus driver, and price tier, and here is my ranked shortlist of causes. Then, my top hypothesis costs roughly this many rides, so here is the one data pull I want and exactly what it confirms or rules out. Specific, ordered, and two-sided.

It leans analytical and execution, but product judgment is woven in. You are graded on diagnostic structure, confirming the metric, decomposing it, segmenting, and quantifying, which is the analytical core. Product judgment shows up when you reason about both sides of the marketplace, when you weigh whether a rider-side fix starves driver supply, and when you size impact in the currency the business cares about. An answer that is rigorous on the data but ignores the two-sided consequences of a fix still falls short of the bar.

This practice round runs about twenty minutes, mirroring a working segment of the real forty-five minute APM analytical session. Spend the first stretch confirming the metric and stating its definition, the largest stretch on segmentation and your ranked hypotheses, a focused stretch sizing impact and proposing the next data pull, and a short close reflecting on what you would revisit. Do not burn early minutes brainstorming causes; the interviewer will cut a long unstructured list short and you will lose the time you needed for impact and the data pull.