BEHAVIORAL · DIVING DEEP
Tell me about a problem where the obvious cause was not the real one: sample answer
Whether you go to the data and the code yourself instead of accepting the first explanation, and whether you can explain a technical cause simply. They also want to see that you checked your theory before acting on it.
Also asked as
- Tell me about a problem where the obvious cause was not the real one.
- Describe a time you used data to change a decision.
The short version: Show you went to the data or the code to find the real root cause.
Sample STAR answer
An invented example from typical software engineering work, about 250 words, around two minutes spoken. Use the structure, not the story: yours should be real.
Situation Our product catalogue service had short latency spikes roughly every ten minutes, where the slowest requests jumped from about 200 milliseconds to four seconds. The team had blamed the database for weeks and was planning to move to larger database servers.
Task I was on call that week and wanted to be sure the expensive fix would actually work before we paid for it.
Action First I lined up the spikes against database metrics, and the database was not busy during them; queries were slow because many arrived at once. So I looked at what sent them. Our service cached product data for exactly ten minutes. When we deployed, every instance filled its cache at the same moment, so the entries also expired at the same moment, and thousands of requests went to the database together. I confirmed it by reading the cache code and by reproducing the pattern on staging with a load test: same ten-minute spikes. The fix was small. I added a random spread of plus or minus two minutes to each entry's expiry and made sure only one request per key reloaded data while others waited for it. I tested it on staging, then rolled it out to one instance before the rest.
Result The spikes disappeared; the slowest requests stayed under 300 milliseconds. We cancelled the database upgrade, which would have cost more than the rest of our yearly hosting for that service. I wrote a short note on synchronised cache expiry, and we added it to our design review checklist.
Weak answer vs strong answer
Weak
"We had latency problems, so I looked at the logs and found an issue with the cache. I fixed it and latency went back to normal."
Why it scores low: The interviewer cannot see how you reasoned, what the wrong explanation was, or how you proved the real one. "Looked at the logs" hides all the depth the question is asking about.
Strong
A specific situation, an action you took yourself, the reasoning behind it, and a result with a number. Show you went to the data or the code to find the real root cause.
How the story changes with level
At SDE 1, a careful debugging story in your own code, with a clear explanation of how you narrowed the cause, is ideal.
At senior or staff level, show that you questioned a decision others had already made, saved real money or time, and turned the finding into a check other teams use.
Follow-up questions interviewers ask
- Why did everyone believe it was the database?Show empathy: the symptom really looked like a database problem.
- How did you prove your theory before changing production?Describe the reproduction and the single-instance rollout.
- What metrics would have shown this sooner?Mention cache hit rate and request counts to the database over time.
- How do you know when to stop digging?When you can reproduce the cause and the fix removes it.
Amazon Leadership Principles that test this
Dive Deep · Tell me about a time the data did not match what people were saying.
Stay close to the details, check them often, and be sceptical when numbers and stories do not match.
Situation After we redesigned our product page, our dashboard showed that "add to cart" clicks had risen by 15 percent. But sales had not changed at all, and the support team said users found the new page confusing.
Task The team wanted to call the redesign a success. I thought the numbers and the stories did not fit, and I wanted to know which was true.
Action I went to the raw click events instead of the dashboard. In the new page, the add-to-cart button sent an event when it was pressed, and the page also sent the same event when the cart opened. So one click often counted as two. I checked by replaying a sample of real sessions and counting actual clicks by hand. After correcting for the double count, add-to-cart was actually down three percent. I also watched recordings and found people were missing the size selector, which now sat below the fold on smaller phones.
Result We fixed the event, moved the size selector up and re-ran the test. Real add-to-cart then rose by four percent and sales followed. We added an automated check comparing event counts with server-side cart changes.
Tip: Show that a mismatch between numbers and anecdotes made you go to the raw data yourself.
Companies that test this
Prepare your own
Write your story in the story bank and tag it "Diving deep"; it will show up against every company value that tests this theme.
Related: Ambiguity and judgment