Appearance
Stories That Work
Each model answer below uses an example project, a Node and PostgreSQL order platform for an online shop. Replace every detail with your own; interviewers can tell a borrowed story within two follow-up questions.
1. What is the STAR method, and how do you use it?
The short answer
STAR is a four-part shape for any "tell me about a time" question. Situation: the context in one or two sentences. Task: what you were responsible for. Action: what you did, step by step, with the reasoning behind each step; this should be most of the answer. Result: what changed, with a number if possible, and what you learned. Keep it to about two minutes, say "I" rather than "we" for your own actions, and stop at the result so the interviewer can ask follow-ups.
| Part | Share of the answer | Sounds like |
|---|---|---|
| Situation | 10% | "Our checkout API started timing out every evening at peak." |
| Task | 10% | "I was on call, and I owned the payments integration." |
| Action | 60% | "First I… because… then I… which showed…" |
| Result | 20% | "p99 fell from 4 seconds to 300 ms; we added an alert so it cannot recur silently." |
Prepare six stories, not sixty. Most behavioural questions can be answered with one of these: a production incident, a hard bug, a performance win, a disagreement, a mistake, and helping someone grow. Write each down once in STAR form, with the numbers, and practise saying it aloud.
Common mistakes:
- All situation, no action. Two minutes of background, thirty seconds of what you did.
- "We" throughout. The interviewer cannot tell what you personally did.
- No result. End with an outcome, even if it is "it failed, and here is what I changed".
- Too perfect. A story with no obstacles sounds rehearsed or invented.
2. Tell me about yourself and your backend engineering experience.
The short answer
Give a 90-second answer in three parts: present, what you do now and the scale you work at; past, the two or three experiences that led here and what each taught you; future, why this role is the next step. Pick details that match the job description, use concrete numbers, such as requests per second, team size or a measurable improvement, and end with a hook the interviewer will want to ask about. It is not your whole CV.
A model answer:
"I am a backend engineer, currently at a mid-sized online retailer, where I work on the order platform: Node and TypeScript services on PostgreSQL and Redis, handling about 2,000 requests per second at peak. Most recently I led moving our payment flow to idempotent, event-driven processing, which removed duplicate charges entirely.
Before that I spent two years on a smaller team doing everything from React front ends to deployments, which is where I learned to care about how code behaves in production, not just whether it works.
I am looking for a role where I can go deeper on distributed systems at larger scale, which is why this platform team appeals to me. I would be happy to talk about the payments migration, since it is the project I learned the most from."
Template:
- Now: "I am a … at …, working on … with … at a scale of …"
- Highlight: "Recently I … which led to …"
- Before: "Before that I … where I learned …"
- Next: "I am looking for … which is why this role …"
- Hook: "Happy to go into …"
3. Can you explain your current project architecture and your role in it?
The short answer
Draw the system at the right altitude: the main components, how a request flows through them, where data lives, and the scale, in about two minutes. Then say clearly which parts you own, the decisions you made or influenced, and one trade-off you would revisit. Be ready to go one level deeper on any box, such as why this database, how deploys work, or what happens when a dependency fails. Do not share confidential details; describe the shape and the reasoning.
A structure that works:
- Purpose and scale: "The order platform takes orders from web and mobile, about 150,000 orders a day."
- Request flow: "Requests go through an API gateway to the orders service in Node. It writes to PostgreSQL, publishes events through an outbox to Kafka, and payments, warehouse and email consume them."
- Data: "PostgreSQL for orders, Redis for carts and caching, S3 for invoices."
- Operations: "Kubernetes, canary deploys, OpenTelemetry tracing."
- Your part: "I own the orders and payments services, and I designed the outbox."
- A trade-off: "We chose one shared Kafka cluster for all teams; at our scale it is fine, but noisy neighbours are starting to appear, so I have proposed quotas per team."
Follow-ups to prepare: why microservices or not; how you handle a failed payment after the order is saved; how a new engineer deploys; the last incident and what changed afterwards.
4. Can you describe a production incident you investigated and resolved?
The short answer
Use STAR, and show the incident shape: you stabilised first, then found the root cause with evidence rather than guesses, then prevented it from recurring. Mention how you noticed, what you checked in which order, how you communicated during the incident, the specific fix, and the follow-up actions from the blameless review. Numbers help: duration, users affected, and the improvement afterwards.
A model answer:
Situation. "One Friday evening our checkout error rate jumped from 0.1% to 15%."
Task. "I was the on-call engineer for the orders service."
Action. "The alert fired on error rate. The dashboard showed errors began four minutes after a deploy, so I rolled back first, and errors fell to normal within five minutes. I posted updates in the incident channel every ten minutes. Then I compared traces from the bad version: every failing request spent 30 seconds waiting for a database connection. The new release had added a call to a reporting query inside the checkout transaction, which held a connection for over a second, and at peak the pool of 20 ran dry."
Result. "Twenty minutes of degraded checkout, roughly 900 failed attempts, most retried successfully. Afterwards I moved the reporting query to a replica outside the transaction, added an alert on pool wait time, and added a check in code review for queries inside transactions. We have not had a pool exhaustion since."
What interviewers check: that you rolled back before debugging, used data, communicated, and fixed the system, not just the symptom.
5. Can you describe a memory leak you investigated and how you fixed it?
The short answer
Explain how you knew it was a leak, memory rising after garbage collection over hours, not just high memory; how you found it, heap snapshots compared over time, reading the retainer chain; what it was, such as an unbounded cache, listeners never removed, or timers never cleared; the fix; and how you verified it, the memory floor stopped rising. Mention the safety steps, such as taking the instance out of the load balancer before snapshotting.
A model answer:
Situation. "Our notification workers were being killed for running out of memory roughly every six hours."
Task. "I was asked to find out why, because each restart dropped in-flight jobs."
Action. "The heap graph showed a sawtooth with a rising floor, so it was a real leak, not just load. I took one worker out of rotation and took two heap snapshots an hour apart. Comparing them in DevTools, the biggest growth was in closures retained by an
EventEmitter. The retainer chain led to our provider client: every job added a'rate-limit'listener to a shared emitter and never removed it, so each listener kept that job's whole payload alive."Result. "I changed it to
onceand removed the listener in afinally. The memory floor went flat, restarts stopped, and I added an alert on heap growth over 24 hours so a slow leak would show up before it crashed anything."
The technique is in diagnosing memory leaks.
6. Can you explain a performance optimization you implemented and how you measured its impact?
The short answer
Start with the user-facing problem and the baseline number, such as "order history p95 was 2.4 seconds". Show how you measured to find the cause before changing anything: a trace, a profile, an EXPLAIN ANALYZE. Explain the change and why it worked. Give the after number under comparable load, and any trade-off you accepted. Interviewers want to hear that you measured before and after, not that you guessed.
A model answer:
Situation. "The order history page had a p95 of 2.4 seconds and was the most complained-about page."
Task. "I picked it up as a performance improvement for the quarter."
Action. "Tracing showed 51 database queries per request: one for the orders and one per order for its items, an N+1 pattern. I replaced them with one query for orders and one
WHERE order_id = ANY($1)for all items.EXPLAIN ANALYZEon that second query then showed a sequential scan, so I added an index onorder_items(order_id)withCREATE INDEX CONCURRENTLY. Finally I added cursor pagination, so we load 20 orders, not all of them."Result. "p95 fell from 2.4 seconds to 180 ms, and database CPU at peak dropped by about 15%. The trade-off was that 'jump to page 30' was replaced by 'load more', which product agreed to."
7. Can you describe a difficult bug you fixed?
The short answer
Choose a bug that was hard for a reason: intermittent, only in production, a race condition, a time zone or encoding issue, or a dependency's behaviour. Explain how you narrowed it down: reproducing it, adding logging or tracing, forming and testing hypotheses one at a time, and what finally pointed to the cause. Give the fix, and what you changed so that this class of bug cannot come back. Show patience and method, not luck.
A model answer:
Situation. "About once a week a customer's order showed 'paid' but the warehouse never received it."
Task. "It had been open for a month; I took it on."
Action. "I could not reproduce it, so I added the order ID and trace ID to every log line on the path and waited for the next case. It showed the order was saved and the payment event published, but the publish happened inside our transaction, and the transaction then rolled back on a rare constraint error. The warehouse had received an event for an order that, in our database, failed. Then a retry created a second order, which was never published."
Result. "I moved publishing to a transactional outbox, so an event exists only if the order committed. The weekly mismatch disappeared, and I added a nightly reconciliation between orders and the warehouse that now catches any drift within a day."
8. Tell me about a mistake you made and what you learned.
The short answer
Pick a real mistake with real consequences, not a disguised strength such as "I work too hard". Own it plainly in one sentence, explain the impact, describe how you fixed the immediate problem and told the people affected, and spend most of the answer on what you changed in your habits or the team's process so it does not happen again. Interviewers are checking honesty, accountability and learning.
A model answer:
Situation. "I ran a database migration that added a
NOT NULLcolumn with a default to our largest table, during business hours."Task. "It was my change, and I had tested it on a small staging database."
Action. "On production it locked the table for nearly two minutes, and checkout failed for that time. I cancelled it, confirmed the service recovered, and told the incident channel exactly what I had done. That evening I rewrote it as an expand-and-contract migration: add a nullable column, backfill in batches, then add the constraint as
NOT VALIDand validate it separately."Result. "The safe version ran without any user impact. I then added
lock_timeoutto every migration, wrote a short guide on lock-safe migrations, and we now rehearse migrations on a production-sized copy. The lesson for me was that 'it worked in staging' means nothing when staging is a thousand times smaller."
The safe technique itself is in safe migrations.