Describe a Time You Made a Significant Technical Mistake
The question: "Describe a time when you made a significant mistake in your technical work. How did you identify it, what did you learn, and how did you apply that learning in the future?"
Also asked as: "Tell me about a technical decision you got wrong, and how you found out." · "What's the worst bug you've shipped? How long did it run before anyone knew?" · "What technical call would you take back, and what did you do differently the next time?"
Who's asking: an Engineering Director or Senior Engineering Manager, in a behavioral round.
Executive Summary: The director wants to know how long your mistake could live before anyone noticed, and whether what you learned has ever changed a later decision. What changes with level is what now catches it: your own check, your team's rule, a test before several teams commit, then a number executives see every month.
What the director is really asking
The question makes three checks:
- "How did you identify it?" Hearing it from a customer or an analyst isn't the failure. Not knowing why your own checks stayed quiet is. They want what you were watching, what it couldn't show you, and what should have told you sooner.
- "What did you learn?" They want a claim about how this kind of thing fails, such as "'no errors' only tells you the code ran". "Test more" gives them nothing to check.
- "How did you apply it?" This has to be a later event: the decision that went differently, and what happened.
The decision it feeds: "If I give you a system this size, how long will your next mistake run, and will I watch the same one twice?"
Pick the right kind of story. It should be a technical call you made: a bug, a design choice, an estimate. A story about owning up, or about a process or a person, belongs in "Tell Me About a Time You Made a Mistake at Work". And a people manager isn't grading your debugging, so each technical detail gets one clause a non-specialist can follow, plus what it meant for someone.
It maps loosely to Amazon's "Dive Deep", for noticing when metrics and anecdotes differ; approximate. Three ladder dimensions do the work: the failure story row, how impact is measured (the number you watched against the one that mattered), and changing a mechanism.
Level fit at a glance
| SDE | Senior SDE | Staff | Principal | |
|---|---|---|---|---|
| Asked at this level? | Core | Core | Common | Common |
| Story scope they expect | A bug or miss you caused in your own component | A technical decision on a project you led | A cross-team technical bet that failed | A technical direction you set, with a business cost |
| The bar in one line | Say how you found out, name the check you lacked, show one later use | Test your own diagnosis, change a team rule, show someone else using it | See what your own status hid, unwind it with each manager, run the next bet differently | Price the error in margin or revenue, change what executives watch, show it changing a decision |
| Instant down-level | "QA found it, I fixed it, and I test more now" | "We had a caching bug, so I added a test" | "My project slipped, and a neighboring team noticed before I did" | "Compute costs came in higher than anyone expected" |
Staff and Principal are Common because directors more often ask for a failed bet, or fold this into strategy questions. This wording tests your technical judgment, and whether your reporting can be trusted.
SDE: name the check you didn't have
What meets the bar
- The bug or miss is yours, in your own component, and it reached someone.
- You can say how you found out, when and from whom.
- You name the check that would have caught it sooner, and how much sooner.
- One later task went differently, and you can say what the habit caught.
- Level check: if nobody had to find it, there's nothing to detect. If your check needs a team process to work, you're reaching for Senior.
Model answer
"I paid about 1,900 airline loyalty members the same correction twice, and found out four days later from a finance analyst. I worked on the job that posts miles after each flight. A partner airline had changed its earning rates, and for 11 days our job credited too few miles. My ticket was to change the job and top up the members it had shorted. My change also taught the job to ignore corrections.
It was due to ship Friday. I ran the top-up on Thursday afternoon, because I wanted members paid before the weekend. My lead had reviewed the script; the timing I changed on my own. That night the old job treated each top-up as new activity and posted it again. My script's log said 2,300 flights corrected, no errors, so I closed the ticket.
On Monday the analyst, who reconciles miles weekly, asked why a correction I'd logged at 1.35 million miles had posted 2.7. It took me 20 minutes to see why, and I told the analyst and my lead it was my ordering before I had the full list. I'd checked that my script ran, and never that the miles added up. One total, compared with what I expected, would have shown it on Friday morning.
Those four days cost something. About 140 members had spent the extra miles, and the loyalty team let them keep the miles, about $1,200 written off. I wrote the reversal for the other 1,760, and member services took about 90 calls.
What I learned is that 'no errors' only tells you the code ran. So before I change balances, I write down the total I expect, ship the code first and start small. Four months later I had to correct car-rental miles. I shipped the job change, watched one night's run, and corrected 100 members as a test. The total came in 8 percent high, because the rate table still held a partner promotion that had ended. The PM and I corrected it and those 100 balances. The other 3,000 members got their miles a week later than promised, and the total matched what I'd written down to within 400 miles."
Why this lands at SDE
- "I'd checked that my script ran, and never that the miles added up": the level beat. One script and one missing check, named exactly.
- "would have shown it on Friday morning": they answer "what should have told you sooner" before being asked.
- "The total came in 8 percent high": the second use is a real event, and it cost a week.
Follow-ups to expect
- "An analyst found this, not you. What would have made it you?": tests whether you own detection → the total, run before you close the ticket.
- "You ran it a day early. Who did you check with?": tests when you ask for help → your lead reviewed the script, not the timing, and what makes you ask now.
- "Has that check ever slowed you down for nothing?": tests whether it's real → a time it passed and cost you an hour, and why you kept it.
Red flags
- ❌ "QA should have caught it." It was your change.
- ❌ "I learned to test more." There's no check to name.
- ❌ "Next time I would…" If it's never come up, pick another story.
Senior SDE: test your theory before you fund it
What meets the bar
- It's a technical decision you made on a project you led, with an option you turned down.
- Identifying it includes your diagnosis: what you assumed, what you spent on that, and what the data would have shown.
- You changed a rule for your team, and someone else has used it, with a result.
- There's a system or product number, net: what you kept and what it cost.
- Level check: a slip anyone writing the code could have made reads as SDE. If other teams' managers were reading your status, it's becoming a Staff story.
Model answer
"A caching decision of mine caused about 18,000 duplicate refill requests at a 400-store pharmacy chain, and then I spent a month fixing the wrong thing. I was tech lead on the refill service, a team of seven.
The app kept timing out against the stores' slow systems, so I decided we'd keep a 15-minute copy of each prescription's status. One of our engineers proposed clearing it when a patient requested a refill: three more weeks, and I turned it down to hit our date. Timeouts fell from 2 percent to 0.3.
Two weeks after launch, the pharmacy operations lead told me technicians were clearing duplicates: 4 percent of about 60,000 requests a week, up from 0.6. The button gave no feedback for a moment, so I decided patients were double-tapping. I put two engineers on that, and a month on, duplicates went from 4 percent to 3.9.
Then I finally looked at the gap between each pair of requests, in our logs since launch. A double-tap is seconds; these were six to 14 minutes. Patients came back, still saw 'refill available' from my copy, and asked again. My theory protected my decision, and I'd never asked what it predicted.
I told the operations lead and my manager the cause was my decision. A shorter copy brought timeouts back, so duplicates ran three more weeks until we shipped the fix I'd turned down. They fell to 0.7 percent with the timeout gain intact. Those nine weeks cost about 600 technician hours at two minutes each, pharmacists stopped 31 filled duplicates before they reached a patient, and auto-refill slipped six weeks.
It taught me that a theory is cheap to test and expensive to fund. The rule I set for my team: before anyone builds a fix for a live problem, they write down what the data should show if their theory is right, and check it. In October, auto-refill reminders went to patients whose prescriptions were already filled. Our on-call engineer suspected the text vendor, and wrote down that vendor delays would show across all stores. The texts all came from 30 stores whose fills reached us a day late, and we found that in one afternoon, after about 300 wrong texts."
Why this lands at Senior SDE
- "The rule I set for my team": the level beat. It's a rule for the team they lead, with no reach beyond it.
- "duplicates went from 4 percent to 3.9": they state how little the wrong theory moved.
- "Our on-call engineer suspected the text vendor": the rule works without the candidate, and they still count the wrong texts.
Follow-ups to expect
- "An engineer proposed the fix before launch. What did you say to them when you went back to it?": tests how you treat people → what you actually said, and how you listen now.
- "You put two engineers on your theory. How did your manager see that?": tests self-awareness → what they said, and what they asked to see afterward.
- "What did you trade off to hit the date?": tests judgment → three weeks saved and nine paid, and what you ask now before cutting that kind of work.
- "Did any of those duplicates reach a patient?": tests whether you looked → none did, how you know, and what you told the pharmacy quality lead.
Red flags
- ❌ "Patients used it in a way we didn't expect." It makes your cache the patients' fault.
- ❌ "We had a caching bug." No decision and no owner, so it reads as SDE.
Quick check before moving on:
Staff: see what your own status was hiding
What meets the bar
- It's a technical bet you pushed across several teams, whose managers spent something on your word.
- You can say what your reporting hid: the number you showed, the number that was true, and how long the gap ran.
- You led the unwinding with each manager, and you say what wasn't recovered.
- A later shared bet went differently, with other teams involved and a net result.
- Level check: if only your team and one neighbor changed plans, it reads as Senior. Executives funding it on your number read as Principal.
Model answer
"At a photo-sharing network, I pushed a technical bet across four teams, and for three quarters I reported it green while it was failing. To cut our $60 million photo-delivery bill by $9 million a year, I proposed one shared pipeline compressing photos harder on the phone. The managers of feed, stories, messages and creator tools agreed: about 110 engineer-weeks with my team's build. I'd measured quality, never time-to-post on older phones, and I'd assumed those teams wanted the saving, when their launches were judged on posts per user.
I found out in month 11, when finance asked why the bill had barely moved with four teams live. 'Four teams live' was my own status line. I pulled the share of photos using the pipeline: 11 percent. On older phones, compressing delayed posts by over a second. In experiments, feed, stories and creator tools each saw about half a percent fewer posts, and each had held at around 10 percent. Those write-ups had been in a shared folder since month four.
Finance and I priced it: those posts were worth about $14 million a year in ad revenue, so full rollout would have lost money. Leaving older phones out kept only about $2 million. I told my director that my green had been wrong. I took each manager their own experiment, and withdrew the bet. Messages kept the pipeline, because failed sends fell 18 percent. The other three spent 22 engineer-weeks removing it, and my team about 30 recompressing older photos on our servers. Net of server costs: about $3 million a year, mostly from work needing none of those teams, for about 160 engineer-weeks, and a creator tool that missed its launch season.
A team's launch test, I learned, tells you more than its agreement. Before any team integrates a shared bet, I ask each manager what would stop their launch, and I run that test myself, at my team's cost. A year later, on a shared service for video previews, that took three engineer-weeks and showed the same older-phone drop, so I cut those phones before anyone integrated. It saves $1.8 million a year, against a $2.2 million forecast. Feed's manager joined on one condition: their experiment, not my dashboard, decides whether they stay."
Why this lands at Staff
- "Before any team integrates a shared bet, I ask each manager what would stop their launch, and I run that test myself, at my team's cost": the level beat. It changes how several teams get asked, and the candidate pays to find out.
- "'Four teams live' was my own status line": they own the detection failure.
- "their experiment, not my dashboard": a manager sets the terms, so the second use isn't self-graded.
Follow-ups to expect
- "You told your director it was green for three quarters, and the write-ups were in a shared folder from month four. Why weren't you reading them?": the most dangerous one → it's built to make you defend the gap. Don't. Your number looked good, so you stopped looking, and integration work ran on for seven more months. At least one manager had mentioned the hold; say you filed it as scheduling.
- "If those managers didn't care about the bill, why did they agree at all?": tests whether you understand the people → the bill sat in a central budget, so saying yes cost them little until their own test came back, and how you'd read a yes now.
- "$3 million a year for 160 engineer-weeks sounds like a good project. What was the mistake?": tests whether you hide behind the net → most of it needed none of those teams, and the seven months at 10 percent lost about $0.8 million in ad revenue.
Red flags
- ❌ "The teams never really committed." They ran the test you never read.
- ❌ "Adoption was slower than expected." It hides the 11 percent behind a schedule word.
Principal: change what the executives see
What meets the bar
- It's a technical direction you set and were accountable for, with executives funding it on your estimate.
- The error is in business terms: margin, revenue, what it would have cost at full scale.
- You can say why leaders' view stayed green: which number they watched, who chose it, and who warned you.
- The org kept a mechanism, and it has changed a decision someone else made.
- Level check: several teams and their managers, with no executive funding the bet on your number, reads as Staff.
Model answer
"At a consumer-genomics company, I set the direction to retire the chip test on every kit, which reads a fraction of the DNA, and sequence every customer's whole genome at low depth. I told the CEO and CFO: $70 million over three years, $44 a kit against $41, and $30 million a year in health revenue. My cost was wrong by $28 a kit, and for 19 months the number I'd told everyone to watch, lab cost, fell on plan.
I'd barely priced re-analysis. We launch about a dozen reports a year, each re-reads every customer's data, and a whole genome is hundreds of times bigger. I'd allowed $3 a kit over five years; the truth was about $31. The head of bioinformatics raised this at the funding review, and I called it a rounding error, because our pilot had never run a new report.
I found out when the reports team asked for $150,000 of compute for one launch, where my model said $15,000. Earlier launches, on fewer customers, had fitted in the shared budget. I rebuilt it with the head of bioinformatics: about $45 million a year at full rollout, about half our kit margin. I told the CEO and CFO the estimate was mine, and recommended keeping the chip test for the $99 kit, with whole genome only in a $249 health tier.
Finishing cost $81 million against my $70 million, because we'd rebuilt re-analysis and kept both labs running; $48 million was already spent when I told them. The tier brings in about $19 million a year, not $30 million, and nets about $8 million.
I now believe a pilot shows the first run, and the bill is in the hundredth. I asked the CFO to replace lab cost on the monthly margin review with all-in lifetime cost per kit, owned by the head of bioinformatics. Every new product now shows that number, from bioinformatics' figures, before launch. A blood-test subscription I'd backed showed $19 a member a year in compute, against my $6. The product head raised the price and halved the forecast, and it's within 8 percent of that plan. We still carry about 300,000 whole-genome customers at the old price, about $1 million a year with no end date."
Why this lands at Principal
- "I asked the CFO to replace lab cost on the monthly margin review with all-in lifetime cost per kit": the level beat. It changes what the company watches, and it gives the number to someone else.
- "the number I'd told everyone to watch, lab cost, fell on plan": the detection failure is about how the org saw the bet, and the candidate chose that view.
- "I called it a rounding error": the owner raised it, the candidate overrode it, and they say so.
- "The product head raised the price": another leader changed a decision because of the mechanism.
Follow-ups to expect
- "The head of bioinformatics told you at the funding review. What did you say to them afterward?": tests how you treat people → the conversation itself, and why they own the number now.
- "You'd said on plan for 19 months. How did you bring the CEO and CFO back?": tests executive trust → what you showed them, and what it cost your standing.
- "$81 million for $8 million a year is about a ten-year payback. Why not stop at $48 million?": tests judgment on sunk cost → what the last $33 million bought (re-analysis cut 60 percent, at the cost of two quarters of new reports), and what stopping would have cost. Spend was also ahead of a straight-line $37 million: lab cost on plan never meant spend on plan.
- "What does the mechanism miss?": tests lasting impact → a cost it can't see, such as customers staying longer than your five-year model.
Red flags
- ❌ "Compute costs came in higher than anyone expected." Someone did expect it, and you waved it off.
- ❌ "We learned to do more thorough modeling." No mechanism and no owner.
Quick check before moving on:
One story, four altitudes
| SDE | Senior SDE | Staff | Principal | |
|---|---|---|---|---|
| Situation | A correction run before the change that made it safe | A status copy, with its fix turned down | One pipeline, reported as "four teams live" | Retiring the chip test, with re-analysis barely priced |
| Your role | Wrote the script | Tech lead, made the call | Pushed the bet, with no authority | Set the direction, gave executives the number |
| The key move | Named the ordering, wrote the reversal | Checked the gaps, shipped the fix, set the rule | Pulled real traffic, took each manager their experiment, withdrew | Rebuilt the model, kept the chip test, changed the margin review |
| Result | $1,200 written off, 90 calls | 18,000 duplicates, 600 technician hours | $3M a year against $9M, 160 engineer-weeks | $81M; $19M a year against $30M, $8M net |
| Reflection | Write the expected total; caught 8% at 100 members | Write what the data should show; found in an afternoon | Run each team's launch test first; older phones cut early | A pilot hides the hundredth run; a subscription repriced |
Draft your answer
Step 1: Pick the right story
Run each candidate story through the test for your target level:
- SDE: Did a bug or miss you caused reach someone, and has a changed habit since altered a later task?
- Senior SDE: Did a technical decision on a project you led miss, with an option you turned down, and has someone else used the rule you set?
- Staff: Did several teams spend something on your technical bet, did you unwind it with their managers, and did a later shared bet go differently?
- Principal: Was it a technical direction you set, with a business cost, and has the mechanism changed a decision someone else made?
Then check that it passes all of these:
- It's a technical call, not a process or people miss.
- You can say how you found out, and what should have told you sooner.
- The second use has actually happened, recently, with a result. Staff and Principal stories can start further back.
- Someone paid a cost, and you can put a number on it.
If you're a new grad. Capstone, research or internship code counts if your technical call reached someone and you've applied the lesson since, even in a later project. The scope shrinks, not the questions: how you found out, how this kind of thing fails, and what it changed later. No second use yet? Pick a smaller story that has one. "In our capstone I cached the room schedule for a day, and forty students went to the wrong building."
Aiming a level above your current title? Tell the story at its true size, and let the detection part show what you can see. A Senior aiming at Staff might say: "I only watched my own service's numbers. Now I'd ask the two teams downstream what would make them stop." That beats a bigger story that falls apart at "Which managers, and what did each decide?"
Step 2: Fill the skeleton
SCOPE (10s) — Your role, the system, how many people it reached: ______
THE MISTAKE (15s) — One sentence starting "I", inside the first 30 seconds,
plus any option you turned down: ______
FOUND OUT (25s) — When, from whom, and what you were watching instead: ______
SOONER (10s) — The signal that existed, or the check you lacked: ______
COST + FIX (25s) — Who paid, how much, who you told, what the fix cost,
the net number, and what wasn't recovered: ______
LEARNED (10s) — A claim about how this kind of thing fails: ______
SECOND USE (25s) — The later decision it changed, and what happened: ______
LEVEL BEAT (inside SOONER or SECOND USE) — the line only your level could say:
SDE: the one check you lacked.
Senior: your team's rule, and who else has used it.
Staff: what you do before several teams commit.
Principal: what executives now see, and who owns it.
Aim for about two minutes, with a 30-second cut ready (the mistake, how you found out, the lesson in a sentence, the second use).
Step 3: Calibrate it
| If your answer says… | It reads as… | Fix |
|---|---|---|
| "QA found it and I fixed it" | No detection story | What you were watching, and what should have told you |
| A Senior answer: "I added a test" | SDE | The decision you made, and the rule someone else has used |
| A Staff answer: "My migration broke a neighbor's service" | Senior | Which managers spent something on your bet, and how you unwound it |
| A Principal answer: "Our cloud bill came in high" | Staff | The price in margin or revenue, and what executives now watch |
Step 4: Rehearse the follow-ups
Say each of these out loud once:
- "What was already telling you, before you found out, and why weren't you looking?" The most dangerous one. Name the signal, say why, and don't defend it.
- "What did you learn, in one sentence?" A claim about how things fail, with no virtues in it.
- "When did you last use that, and what did it cost you?" One dated event, including the time it took.
- "How did your manager see it?" What they said, and what they checked more closely afterward.
- "Could the same thing happen today?" Yes, and say what your check doesn't catch.
Common mistakes
- Telling a detective story. This round doesn't grade debugging.
- Offering "it never happened again" as the second use. Give a later event where the learning changed a decision, and what that cost.
Key takeaways
- How you found out and how you applied it are the parts most candidates skip.
- The level shows in what now catches it: your check, your team's rule, a test before teams commit, a number executives see.
Test yourself on this question's key ideas: