AWS Interview Questions: 16 Sample Answers on EC2, S3, IAM & VPC
16 AWS interview questions with sample answers: EC2, S3 storage classes, IAM roles, security groups vs NACLs, Lambda limits, Well-Architected and DR.
Last updated: November 2026
Quick Answer
AWS interview questions test three things: whether you know what the core services do (EC2, S3, IAM, VPC, RDS, DynamoDB, Lambda), whether you can choose between them for a real workload and explain the trade-off in cost, availability and security, and whether you can troubleshoot calmly when something breaks. The best answers define the service in one sentence, name the limit or setting that matters, then give one example from your own work. Rehearse them aloud in OphyAI Interview Practice, and on live video rounds OphyAI Interview Copilot suggests answers in real time.
| # | Question | What the interviewer is testing |
|---|---|---|
| 1 | What is the AWS shared responsibility model? | Security ownership |
| 2 | How do you choose an EC2 purchasing option? | Cost judgment |
| 3 | Which S3 storage class would you use, and why? | Matching storage to access patterns |
| 4 | What is the difference between an IAM role and an IAM user? | Credentials and least privilege |
| 5 | Security groups vs network ACLs: what is the difference? | Network security basics |
| 6 | How would you design a VPC for a three-tier web app? | Subnets, routing, isolation |
| 7 | When would you pick RDS over DynamoDB? | Data modeling trade-offs |
| 8 | How does RDS Multi-AZ work? | High availability |
| 9 | When is Lambda the wrong choice? | Serverless limits |
| 10 | How do Auto Scaling and Elastic Load Balancing work together? | Elasticity and health checks |
Most cloud loops mix these concept questions with one scenario (“the site is down, walk me through it”) and one behavioral question about an outage or a cost problem. Amazon’s own loops add Leadership Principles stories; see our Amazon interview guide. If the loop includes a design round, our system design interview questions list covers the common prompts.
The Questions, With Sample Answers
Sample answers are illustrative. Swap in your own systems and numbers; an interviewer will ask follow-ups about anything you claim.
1. “What is the AWS shared responsibility model?”
Why they ask it. It is the foundation of every security conversation on AWS, and a wrong answer suggests you would leave a gap nobody owns.
Sample answer. “AWS describes it as security of the cloud versus security in the cloud. AWS is responsible for the infrastructure that runs its services: the hardware, the software, the networking and the facilities. I am responsible for what I put in the cloud, and how much that is depends on the service. On EC2, that means the guest operating system and its patches, the applications I install and the security group rules. On a managed service like S3 or DynamoDB, AWS runs the operating system and platform, so my share shrinks to my data, its encryption and classification, and the IAM permissions. In my last role, that split was why we moved a self-managed database to RDS: patching the database engine became AWS’s job, and our team’s checklist got shorter.”
What weak answers look like. “AWS handles security.” Or a correct definition with no example of how the line moves between EC2 and a managed service.
2. “How do you choose an EC2 purchasing option?”
Why they ask it. Compute is often the biggest line on the bill, and they want to know whether you think about cost before you launch.
Sample answer. “I match the option to how predictable the workload is. On-Demand has no commitment, so I use it for new or spiky workloads while I learn the usage pattern. Once a baseline is steady, Savings Plans make sense: AWS says they can cut the bill by up to 72% against On-Demand in exchange for a usage commitment. For work that can be interrupted, such as batch jobs, CI runners or rendering, Spot Instances use spare capacity at up to 90% off On-Demand, but the instance can be reclaimed, so the job has to checkpoint or retry. Dedicated Hosts are for licensing or compliance needs. On one team we ran a steady baseline on a Savings Plan, kept On-Demand for peaks, and moved nightly data processing to Spot inside the same Auto Scaling group.”
What weak answers look like. Listing the options with no rule for when to use each, or recommending Spot for a stateful database.
3. “Which S3 storage class would you use, and why?”
Why they ask it. Storage classes are a common cost lever, and the answer shows whether you think about access patterns and minimum storage durations.
Sample answer. “It depends on how often the data is read and how fast I need it back. S3 Standard is for frequently accessed data. If I do not know the pattern, S3 Intelligent-Tiering moves objects between tiers for me. Standard-IA suits data that is read rarely but must come back in milliseconds, with a 30-day minimum storage duration. One Zone-IA keeps data in a single Availability Zone, so I only use it for data I can re-create. For archives, Glacier Instant Retrieval still returns in milliseconds, Glacier Flexible Retrieval takes minutes to hours, and Glacier Deep Archive is the cheapest for long-term retention, with retrieval within 12 hours and a 180-day minimum. For application logs, I would use a lifecycle rule: Standard for 30 days, then Standard-IA, then Deep Archive for the compliance retention period.”
What weak answers look like. Ignoring minimum storage durations and retrieval times, or putting irreplaceable data in One Zone-IA.
4. “What is the difference between an IAM role and an IAM user?”
Why they ask it. Leaked long-term access keys are a classic cause of breaches. They want to hear that you avoid them.
Sample answer. “Both are IAM identities with permission policies, but a user is tied to one person or workload and can have long-term credentials, such as a password or access keys. A role is meant to be assumed by whoever needs it, and it has no long-term credentials: assuming it returns temporary security credentials for that session. A role has two policies: a trust policy saying who can assume it, and a permissions policy saying what it can do. So I give an EC2 instance or a Lambda function a role instead of embedding keys, and I use roles for cross-account access. AWS recommends IAM users only for cases that federation cannot cover, such as some third-party tools or emergency access. I also scope permissions to the specific actions and resources a workload needs, not a wildcard.”
What weak answers look like. “A role is a group of users.” Or describing access keys hard-coded in an app as normal practice.
5. “Security groups vs network ACLs: what is the difference?”
Why they ask it. It is the most common networking question in cloud interviews, and the stateful versus stateless detail separates people who have debugged traffic from people who have read a summary.
Sample answer. “A security group works at the instance level, allows rules only, evaluates all its rules before deciding, and is stateful, so return traffic is allowed automatically. A network ACL works at the subnet level, supports both allow and deny rules, evaluates rules in ascending number order until one matches, and is stateless, so I have to allow return traffic explicitly, including the ephemeral port range. AWS recommends security groups as the primary control, because they are stateful and can reference other security groups, and network ACLs as a secondary layer, for example to block a specific IP range at the subnet. Once I fixed an outage where someone tightened a NACL and forgot the outbound ephemeral ports; the security group looked fine, which is why it took a while to find.”
What weak answers look like. Mixing up which one is stateful, or saying security groups can deny traffic.
6. “How would you design a VPC for a three-tier web app?”
Why they ask it. It tests subnets, routing and isolation together, and it is often a warm-up for a bigger design question.
Sample answer. “I would create one VPC for the workload and spread it across at least two Availability Zones. In each zone, a public subnet holds the Application Load Balancer and a NAT gateway; a private subnet holds the application servers; and a separate private subnet holds the database, which has no route to the internet. The public route table points to an internet gateway; the private ones send outbound traffic through the NAT gateway. Security groups chain the tiers: the load balancer accepts HTTPS from the internet, the app tier accepts traffic only from the load balancer’s security group, and the database accepts only the app tier’s security group on the database port. I would add VPC Flow Logs for troubleshooting and use a VPC endpoint for S3 so that traffic stays off the internet.”
What weak answers look like. Putting the database in a public subnet, or using one Availability Zone with no mention of failure.
7. “When would you pick RDS over DynamoDB?”
Why they ask it. Database choice is the decision most likely to hurt later. They want reasoning about access patterns, not a favorite tool.
Sample answer. “I start from the access patterns. RDS gives me a managed relational engine, such as PostgreSQL or MySQL, with joins, flexible ad hoc queries and a familiar schema, so I pick it when the data is relational, the queries change often or reporting matters. DynamoDB is a serverless NoSQL database that AWS describes as delivering single-digit millisecond performance at any scale; it supports key-value and document models but no JOIN operator, so I design the table around known access patterns and denormalize. I pick it for high-traffic key lookups like sessions, carts or user profiles, especially when traffic is spiky, because on-demand mode scales with no servers to manage. On a past project, orders went to PostgreSQL on RDS for reporting, while the shopping cart went to DynamoDB.”
What weak answers look like. “DynamoDB is faster, so always use it.” Or not knowing that DynamoDB has no joins.
8. “How does RDS Multi-AZ work?”
Why they ask it. It checks whether you understand the difference between high availability and read scaling, which candidates often confuse.
Sample answer. “In a Multi-AZ DB instance deployment, RDS provisions a standby replica in a different Availability Zone and replicates to it synchronously. If the primary fails, or the zone has a problem, RDS fails over to the standby, and the application reconnects through the same endpoint. It also helps during planned maintenance. The key point is that the standby does not serve read traffic, so Multi-AZ is for availability, not scaling. If I need to offload reads, I add read replicas, or I use a Multi-AZ DB cluster. One trade-off: synchronous replication can add some write and commit latency compared with a single-AZ instance, so I test write-heavy workloads before and after enabling it. I also make sure the app retries connections, because a failover drops existing ones.”
What weak answers look like. “The standby takes half the reads.” Or not knowing failover changes the instance behind the same endpoint.
9. “When is Lambda the wrong choice?”
Why they ask it. Knowing a service’s limits is a better signal than knowing its features.
Sample answer. “Lambda is built for short-lived, event-driven tasks that do not keep state between invocations. Standard functions can run for up to 15 minutes per invocation and use up to 10,240 MB of memory, with CPU allocated in proportion to memory. So it is the wrong choice for long-running jobs, such as a multi-hour ETL, unless I break the work into steps with Step Functions; for workloads that need a persistent connection or local state; and sometimes for steady, high-volume traffic where a container on ECS or a reserved fleet costs less. I also check the payload limit, 6 MB each for a synchronous request and response, and the default concurrency quota of 1,000 per Region, which can throttle a launch if API Gateway sends more traffic than Lambda is allowed to run.”
What weak answers look like. “Lambda scales infinitely, so it is always the best option.” Or not knowing the timeout.
10. “How do Auto Scaling and Elastic Load Balancing work together?”
Why they ask it. Elasticity is the point of the cloud, and they want to hear about health checks, not just “it adds servers.”
Sample answer. “An Auto Scaling group has a minimum, a maximum and a desired capacity. EC2 Auto Scaling never goes below the minimum or above the maximum, and scaling policies move the desired number up or down with demand, for example target tracking on average CPU or requests per target. It replaces impaired instances using health checks, and I can switch it to use the load balancer’s health checks, so an instance that is running but returning errors gets replaced. The group spreads instances across Availability Zones, and every launch or termination registers or deregisters the instance with the load balancer automatically. For deployments, an instance refresh rolls a new AMI or launch template across the group, and I can canary it on a small set first.”
What weak answers look like. No mention of health checks or Availability Zones, or confusing scaling with load balancing.
11. “What is CloudFormation, and why use infrastructure as code?”
Why they ask it. Manual console changes do not survive audits, disasters or new regions.
Sample answer. “CloudFormation lets me define AWS resources in a YAML or JSON template and deploy them as a stack, so the environment is versioned, reviewed and repeatable. If I need to rebuild in another Region or account, I deploy the same template instead of clicking through the console from memory. AWS’s own disaster recovery guidance says to deploy with infrastructure as code, such as CloudFormation or the CDK, because restoring without it is slow and error-prone and can blow the recovery time objective. StackSets extend this to many accounts and Regions in one operation. Before an update, I review a change set to see which resources will be replaced. On my last team, every production change went through a pull request on the template, which cut configuration drift to almost nothing.”
What weak answers look like. Describing infrastructure as code with no benefit beyond “automation,” or never having reviewed what an update would replace. If your loop names Terraform instead, see our Terraform interview questions page.
12. “How do you monitor an application on AWS?”
Why they ask it. They want to know you would find a problem before customers do.
Sample answer. “I use CloudWatch as the base: metrics for each service, logs from the application and Lambda in CloudWatch Logs, and alarms on the signals that customers feel, such as error rate, latency at p99 and queue depth, not just CPU. Alarms notify the on-call through SNS or the paging tool. I add dashboards per service so the on-call can see the whole request path. For tracing across services I use X-Ray or OpenTelemetry, and VPC Flow Logs when it might be a network problem. I also alarm on cost anomalies, because a runaway loop shows up on the bill as well as in the logs. After each incident, I ask whether an alarm should have fired earlier and add it.”
What weak answers look like. Monitoring only CPU and memory, or alarms with no owner.
13. “What are the pillars of the AWS Well-Architected Framework?”
Why they ask it. Solutions architect and consulting interviews use it as a shared vocabulary for trade-offs.
Sample answer. “There are six: operational excellence, security, reliability, performance efficiency, cost optimization and sustainability. I use them as a checklist when I review a design. For example, on a recent review the reliability pillar flagged a single NAT gateway in one Availability Zone; adding one per zone cost more, so we weighed it under cost optimization and decided the outage risk was worth paying for. Sustainability came up when we moved a batch job to Graviton-based instances and right-sized it. The value is that the framework makes us discuss trade-offs explicitly instead of optimizing one thing and forgetting the others.”
What weak answers look like. Naming only five pillars (sustainability is the one people forget), or reciting the list with no example of a trade-off.
14. “How would you cut an AWS bill by 30%?”
Why they ask it. Cost is a constant pressure, and the answer shows whether you start with data or guesses.
Sample answer. “I would start with data: Cost Explorer grouped by service and by tag, to find the top three cost drivers, and check for anything untagged. Then the usual levers, in order of effort. First, delete waste: idle instances, unattached EBS volumes, old snapshots and forgotten environments. Second, right-size instances using utilization metrics. Third, commit for the steady baseline with Savings Plans, which AWS says can save up to 72% against On-Demand. Fourth, move interruptible work to Spot at up to 90% off. Fifth, storage: S3 lifecycle rules to cheaper classes, and check data transfer, especially NAT gateway charges. I would set a budget alarm so the savings stick. I would not promise 30% before looking at the bill, but most accounts I have seen had obvious waste in the first week.”
What weak answers look like. Jumping straight to “buy reservations” with no analysis, or cutting redundancy that protects availability without saying so.
15. “What disaster recovery strategies does AWS describe, and how do you choose?”
Why they ask it. It tests whether you can tie architecture to business targets: recovery time objective (RTO) and recovery point objective (RPO).
Sample answer. “AWS groups them into four strategies, from cheapest and simplest to most complex. Backup and restore: back up data, and redeploy infrastructure from code in the recovery Region when needed. Pilot light: data replicates continuously and core infrastructure exists, but app servers are switched off until failover. Warm standby: a scaled-down but fully working copy runs in another Region and can take traffic immediately, then scales up. Multi-site active/active: the workload runs in several Regions at once, which can bring recovery time close to zero but costs the most. I choose from the RTO and RPO the business needs and what it will pay. Whatever the choice, backups must be tested, because replication alone does not protect against corrupted or deleted data.”
What weak answers look like. “We have backups” with no restore test, or picking active/active for everything.
16. “Tell me about an outage you handled on AWS. What did you learn?”
Why they ask it. Every cloud loop has a behavioral question, and incidents show ownership, communication and calm. Answer with the STAR method.
Sample answer. “Situation: our checkout API started returning 5xx errors at 7 p.m. on a Friday. Task: I was on call, so I owned getting it back and communicating. Action: I checked the load balancer metrics first; healthy host count had dropped in one Availability Zone. The instances were running but failing health checks because the database connection pool was exhausted after a deploy raised the app’s worker count. I rolled back the deploy, posted updates in the incident channel every 15 minutes, and confirmed recovery on the error-rate alarm. Result: about 25 minutes of degraded service. In the post-incident review I added an alarm on database connections and a load test to the deploy pipeline, and we have not seen the issue since.”
What weak answers look like. Blaming another team, no timeline, or no change made afterwards.
How to Practice These
Say each answer aloud until you can give the one-sentence definition, the key limit and your example in under two minutes. Run a mock cloud or DevOps interview in OphyAI Interview Practice, which asks follow-up questions and gives per-answer feedback with a transcript, in behavioral, technical and system design formats. On the live video round, OphyAI Interview Copilot gives real-time answer suggestions on Zoom, Microsoft Teams and Google Meet. For a hands-on coding or design round, OphyAI Coding Interview is a Premium product that helps with live coding and system design.
OphyAI runs on web, Mac, Windows, iPhone and Android. In the desktop app, the answer window is excluded from screen sharing and recording by the operating system, but that does not hide it from a camera, a person in the room or proctoring software. The free plan’s 5-credit grant covers one practice session capped at 15 minutes; paid plans start at Pro, $29 a month, and there is no $9 plan, because the Basic plan was retired in July 2026 (see pricing).
Frequently Asked Questions
What are the most common AWS interview questions?
The most common are the shared responsibility model, EC2 purchasing options, S3 storage classes, IAM roles versus users, security groups versus network ACLs, RDS versus DynamoDB, Multi-AZ, Lambda limits, Auto Scaling with load balancers, and disaster recovery strategies. Most loops add one troubleshooting scenario and one behavioral question about an outage or a cost problem.
Which AWS certification should I get before an interview?
For most cloud engineering roles, AWS Certified Solutions Architect - Associate is the usual starting point, because it covers the services in this list. Developers often take AWS Certified Developer - Associate, operations engineers the CloudOps Engineer - Associate, and data engineers the Data Engineer - Associate. Cloud Practitioner is the foundational level for non-technical or entry roles. A certification helps you get the interview; the scenario questions decide the offer.
How much do AWS cloud engineers make?
The US Bureau of Labor Statistics does not track AWS roles separately. Its closest groups give a guide: network and computer systems administrators had a median wage of $99,130 in May 2025. Pay for cloud engineers and solutions architects varies widely by level, city and employer, so compare the pay range posted in the job ad and ask the recruiter how bonus and equity work.
How do I answer AWS scenario questions?
Use a fixed order: clarify the symptom and its scope, check the obvious signals first (load balancer health, recent deploys, CloudWatch alarms), form one hypothesis at a time, and say what you would check to confirm it. Then explain the fix and what you would change to prevent a repeat. Interviewers grade the process as much as the answer.
Do I need to know every AWS service for an interview?
No. AWS has far more services than any one engineer uses, and interviewers focus on the core: compute, storage, networking, identity, databases, serverless, monitoring and infrastructure as code. Know those deeply, with limits and trade-offs, and be honest when a question is outside your experience; then describe how you would find the answer in the documentation.
What is the difference between availability zones and regions?
A Region is a separate geographic area, and each Region contains multiple Availability Zones. AWS describes an Availability Zone as one or more discrete data centers with redundant power, networking and connectivity, physically separated from other zones by many kilometers but within 100 km (60 miles) of each other. Spreading across zones protects against a data center failure; multiple Regions protect against a regional event.
How does OphyAI help with AWS interview questions?
Interview Practice runs mock technical and behavioral interviews by voice or text, asks follow-up questions such as “why not DynamoDB?”, and gives per-answer feedback and a transcript, so you find weak spots before the real round. During a live video interview, Interview Copilot suggests answers in real time on Zoom, Teams and Google Meet.
Related Guides
AWS questions show up in cloud, backend and infrastructure loops at these companies:
- Amazon interview guide: the AWS parent company, with Leadership Principles behind every behavioral answer.
- Amazon Leadership Principles interview guide: how to turn an outage story into Ownership and Dive Deep.
- Cloudflare interview guide: networking and infrastructure depth.
- Palo Alto Networks interview guide: cloud security roles where IAM and VPC questions are common.
- Datadog interview guide: monitoring and observability on cloud platforms.
- MongoDB interview guide: managed databases running on AWS and other clouds.
- Capital One interview guide: a large bank that runs on AWS.
- Microsoft interview guide: Azure roles ask the same concepts with different names.
For the wider question set, see our data engineer interview questions, technical interview prep for software engineers, system design interview guide and behavioral interview questions and answers.
Sources and verification notes
Checked November 2026. Each AWS page below was opened in this run. Sample answers are illustrative; the numbers in them are examples, not data.
- AWS Shared Responsibility Model: “security of the cloud” versus “security in the cloud”; customer duties on EC2 versus S3 and DynamoDB.
- Amazon EC2 pricing: On-Demand, Savings Plans (up to 72% off), Spot (up to 90% off), Dedicated Hosts.
- Amazon S3 storage classes and S3 data durability: classes, minimum storage durations, retrieval times; 99.999999999% durability and at least three Availability Zones for most classes; zone distances.
- IAM roles (IAM User Guide): no long-term credentials, temporary security credentials, trust and permissions policies, when to use IAM users.
- Infrastructure security in Amazon VPC: security groups versus network ACLs comparison table and AWS’s recommended use of each.
- What is Amazon DynamoDB?: serverless NoSQL, single-digit millisecond performance, no JOIN operator, key-value and document models.
- Multi-AZ DB instance deployments for Amazon RDS: synchronous standby in another zone, standby does not serve reads, possible added write latency.
- Lambda quotas: 15-minute timeout, 128 MB to 10,240 MB memory, 6 MB synchronous payloads, 1,000 default concurrent executions.
- What is Amazon EC2 Auto Scaling?: minimum, maximum and desired capacity, health checks, Availability Zone balancing, load balancer registration, instance refresh.
- AWS Well-Architected Framework: the pillars: the six pillars.
- Disaster recovery options in the cloud (AWS whitepaper): backup and restore, pilot light, warm standby, multi-site active/active; infrastructure as code with CloudFormation; StackSets.
- AWS Certification: current Foundational and Associate certification names.
- BLS Occupational Outlook Handbook: Network and Computer Systems Administrators: May 2025 median pay $99,130.
Tags:
Share this article:
Rehearse your answers, then get live help on the call.
Practice free for up to 15 minutes, then use Interview Copilot in the OphyAI desktop app: real-time answer suggestions on Zoom, Teams and Google Meet, in a window hidden from screen sharing.
On your phone? Practice interview answers in the OphyAI app.
Hidden from screen share and recording; a webcam or a person in the room can still see your screen.
Related Articles
Spring Boot Interview Questions: 23 Answers for Boot 4, JPA & Security
Interview Tips
23 Spring Boot interview questions with sample answers: auto-configuration, @Transactional pitfalls, JPA N+1, security filter chain, test slices and Boot 4.
Read more →
24 Python Interview Questions & Answers (2026): GIL, Decorators, Code
Interview Tips
24 Python interview questions with sample answers: dict hashing, is vs ==, decorators, generators, the GIL vs asyncio, MRO, plus coding tasks and pandas.
Read more →
Java Interview Questions: 18 Answers With Code-Level Samples (2026)
Interview Tips
18 Java interview questions with sample answers: equals and hashCode, HashMap internals, JVM memory, concurrency, streams, records and virtual threads.
Read more →