Interview Question Plan by Experience Level

Four question banks, ~285 questions total, plus the Top 100 with answers. Questions test reasoning, not recall — no multiple choice, no "which of these is a valid S3 storage class."

1. How questions are distributed

Per teaching day: roughly 6 questions at 2–5 YOE and 5 at 6–10 YOE, always drawn from that day's material.

2. What each level is actually testing

flowchart LR
    A["2-3 YOE<br/>Do you know<br/>what it is?"] --> B["4-5 YOE<br/>Can you use it<br/>correctly?"]
    B --> C["6-7 YOE<br/>Can you design<br/>with it?"]
    C --> D["8-10 YOE<br/>Can you defend<br/>the decision and<br/>the failure modes?"]

The same topic, all four levels — SQS:

3. Coverage matrix — every day feeds every bank

4. Sample questions at each level

2–3 YOE (10 of 80)

1. What is the difference between a Region and an Availability Zone?
2. What is an IAM role, and how is it different from an IAM user?
3. What is S3 and what is it not good at?
4. What is a security group?
5. What does a load balancer do for a backend service?
6. What is a Lambda cold start?
7. What is the difference between SQS and SNS?
8. What is a partition key in DynamoDB?
9. How does your application get AWS credentials when running on EC2?
10. What is CloudWatch used for?

4–5 YOE (10 of 75)

1. When would you use S3 instead of EBS — and when is EBS the only right answer?
2. Your ALB health check passes but users get 502s. Where do you look first?
3. How do you size a connection pool for a service running 30 tasks against one database?
4. Explain visibility timeout and what happens when you get it wrong in both directions.
5. When is DynamoDB the wrong choice for a workload that "just needs key-value"?
6. How do you keep a secret out of your application config and out of your image?
7. Your Lambda gets throttled. Walk through what happens to a synchronous caller vs an asynchronous one.
8. How would you cache a read-heavy endpoint, and what breaks on a cache miss storm?
9. What is the difference between a read replica and a Multi-AZ standby, operationally?
10. Why might a service that only talks to S3 still generate a large NAT Gateway bill?

6–7 YOE (10 of 70)

1. Design a notification platform that sends email, SMS and push at 10M/day with per-channel failure isolation.
2. You must guarantee an accepted order is never lost, but the payment provider is unreliable. Where is your durability boundary?
3. How do you scale a service whose bottleneck is a single relational database? Give four options in the order you'd try them.
4. Design idempotency for a payment API. What do you store, keyed on what, for how long?
5. When would you choose Kinesis over SQS, and what do you lose?
6. How do you do a zero-downtime schema change on a live service?
7. Your event-driven system needs to replay three days of events for one consumer. How is that possible, and what did you have to build in advance?
8. Design the caching strategy for a product catalogue with 5M items and heavy read skew.
9. How do you structure IAM so 40 microservices in 3 accounts follow least privilege without blocking deploys?
10. A service must serve users in Europe and the US with <100ms reads. Walk me through the options.

8–10 YOE (10 of 60)

1. Your platform degrades and stays degraded after the triggering cause is fixed. Explain what happened and how the architecture caused it. (Metastable failure / retry storm.)
2. You are asked to make the system multi-Region. What do you tell the executive sponsor about cost, complexity, and what it does not buy?
3. A team proposes moving a 400 GB relational table to DynamoDB "for scale." Make the case both for and against, then decide.
4. Where exactly is the consistency boundary in the order system you just designed, and what does a customer observe inside it?
5. Your idempotency store and your business data live in different systems. Describe the failure window and three ways to close it.
6. The AWS bill grew 3× while traffic grew 1.2×. Describe your investigation and the three most likely causes.
7. You inherit a system with 200 Lambdas and no tracing. What do you do in the first two weeks, and what do you refuse to do?
8. Design the failure domains for a platform that must survive an AZ loss, a bad deploy, and a poisoned message, without a human in the loop for the first two.
9. When is "just use a bigger instance" the correct senior answer, and how do you defend it?
10. What would you have to see to conclude that your event-driven architecture was a mistake and a synchronous call was better?

5. How the course teaches answering, not memorizing

Every 6–10 YOE question ships with a reasoning trace, not an answer paragraph:

What I'd clarify first     the two or three questions that change the answer
The constraint that drives it  what actually forces the decision
Options considered         3, with the real trade-off of each
What I'd choose and why    a decision, stated plainly
What I'd give up           explicitly
How I'd know I was wrong   the metric or event that would change my mind

That last line is the one the course emphasizes most: a senior answer includes its own falsification condition.

The five-move senior pattern (taught on Day 14)

flowchart LR
    A[State assumptions<br/>out loud] --> B[Name the constraint<br/>that dominates]
    B --> C[Give 2-3 options,<br/>not a single answer]
    C --> D[Choose, and say<br/>what you traded]
    D --> E[Stress-test your<br/>own choice first]

6. Common interview traps (Day 14)

Nine statements that sound correct and are not. Each gets the wrong mental model, the correct one, and the one-sentence version to say in an interview.

7. Mock interview formats (Day 14)

Related

• 13 — Cheatsheet Structure
• 14 — Capstone & System Designs
• 11 — Troubleshooting Plan