AWS VI: Architecting on AWS & the exams: interview questions
The question you are most likely to get for each topic, a model answer, and what else comes up.
Walk me through how you would design an order-processing system on AWS that survives bursts and failing dependencies. Mid
The checkout writes the order to an SQS queue and returns, so a slow dependency becomes a backlog instead of an outage. A Lambda function or an ECS service consumes it with a visibility timeout longer than the work, partial batch responses and a DLQ with an alarm on it. A multi-step process - reserve, charge, ship, compensate - runs in Step Functions with retries on transient errors and a Catch for permanent ones. Everything spans at least two Availability Zones, roles are scoped per component, data is encrypted with KMS, and I size concurrency to what the database can take.
Also asked: When would you choose SNS over SQS, or EventBridge over both? · How do you pick a disaster recovery strategy for a workload? · How would you cut the cost of a steady containerized workload?
How does SQS make a system more resilient, and what settings matter? Mid
It turns a synchronous call into a hand-over: the producer only needs the queue, consumers process at their own pace, and a backlog absorbs bursts. The settings that matter are the visibility timeout - longer than the processing time, or messages are processed twice - a redrive policy with maxReceiveCount and a DLQ that someone alarms on, long polling to cut empty receives, and FIFO only when order or exactly-once processing is required.
Also asked: What is the difference between a standard and a FIFO queue? · How does SNS fan-out with SQS work? · What happens to a message that keeps failing?
How do you reduce Lambda cold starts, and how do you see them? Mid
A cold start creates a new execution environment and runs the module's top-level code; the logs show INIT_START and an Init Duration on that invocation's REPORT line. I keep the package and dependencies small, create SDK clients and connections at module level so warm invocations reuse them, give the function enough memory, and use provisioned concurrency on an alias for latency-critical paths. Then I compare Init Duration before and after.
Also asked: What happens when a Lambda function is throttled? · When would you set reserved concurrency to zero? · How does memory affect a function's cost and speed?
Learn it: 39.5 Lambda: functions, cold starts, the REPORT line and concurrency
A message keeps getting processed again and again by a Lambda consumer. What could cause it? Mid
Either it is never deleted or it reappears while still being processed. The function may throw, so the whole batch is redelivered - fixed with partial batch responses (ReportBatchItemFailures in the mapping and batchItemFailures in the code) and a DLQ on the queue. Or the visibility timeout is shorter than the processing time. I would compare the function's errors and duration with the queue's visibility timeout and make the handler idempotent.
Also asked: How do you stop one bad message from blocking a queue? · What is the difference between MaximumConcurrency and reserved concurrency? · What happens to a failed asynchronous invocation?
Learn it: 39.6 Lambda behind a queue: event source mappings, partial batches and destinations
Your API returns 502 under load but works when you test it by hand. Where do you look? Mid
A 502 comes from the integration, so I read the Lambda function's logs around the 5XXError spike for Invoke Error lines and REPORT lines with a timeout status. Under load the usual cause is an error path - a slow dependency, throttling - that returns something that is not a proxy response with statusCode and a string body. I would fix the error path, add timeouts to downstream calls below the integration timeout, and watch Throttles and Duration.
Also asked: What is the difference between a REST API and an HTTP API? · How do you throttle an API per client? · What does API Gateway need to invoke a Lambda function?
Learn it: 39.12 API Gateway: putting an API in front of functions
Your ECS tasks get AccessDenied when the application reads S3, but they start fine. Which role do you fix, and why? Mid
The task role. The task execution role is only used by ECS itself before and around the container: pulling the image from ECR, creating the CloudWatch Logs stream and fetching secrets - if that one were wrong, the task would not start. Once the container runs, the SDK in the application gets the task role's credentials, so the S3 permission belongs there, scoped to the bucket and prefix it needs.
Also asked: Why would a Fargate task fail with ResourceInitializationError? · How do you roll back a bad ECS deployment automatically? · When would you use Lambda instead of ECS?
Learn it: 39.15 Containers on AWS: ECR, ECS on Fargate, and when to use what
How would you implement an order process that must undo earlier steps when a later one fails? Mid
As a Step Functions state machine using the saga pattern. Each step is a Task with Retry on transient errors, with backoff and jitter. A Catch on the permanent errors, such as a declined card, goes to compensating steps that undo what already succeeded - releasing reserved stock, refunding a charge - and then to a Fail state with a clear error. Steps are idempotent, and the execution history shows support exactly where each order stopped.
Also asked: When would you choose Step Functions over SQS between functions? · What is the difference between Standard and Express workflows? · How do you add a human approval step to a workflow?
Learn it: 39.20 Step Functions: workflows with retries, catches and compensation
Kinesis Data Streams or Firehose - how do you choose? Mid
Kinesis Data Streams when consumers need records in real time, in order per partition key, several independent consumers, and replay within the retention period - I own the consumers and the shard capacity. Firehose when the goal is to land data somewhere such as S3, Redshift or OpenSearch with about a minute of buffering, optional Parquet conversion and no consumer code. They often combine: a stream for real-time consumers and Firehose as one consumer that archives it.
Also asked: How would you design a data lake on AWS? · What is a hot shard and how do you avoid it? · When would you use Athena instead of Redshift?
Learn it: 39.23 Streaming and data ingestion: Kinesis, Firehose and the analytics map
Your RDS database is at 90% CPU from reads. What do you do? Mid
First the queries: find the top statements and missing indexes. Then caching: cache-aside in ElastiCache for the hot, repeated reads with a TTL the business accepts, and read replicas for reads that need fresher data. CloudFront in front of cacheable API responses removes load before it reaches the application. I would watch the hit ratio and make sure the database can survive a cold cache after a flush or failover.
Also asked: What is the difference between CloudFront and Global Accelerator? · When would you use DAX? · How do you keep cached data from going stale?
Learn it: 39.26 Caching layers: where to cache and what it costs you
What is the difference between high availability and disaster recovery? Mid
High availability keeps a service running through expected failures - an instance, an Availability Zone - automatically, with redundancy inside one Region. Disaster recovery is how you get back after losing a Region or after a mistake that replication copies, using another Region or backups. DR strategies range from backup and restore through pilot light and warm standby to multi-site, chosen by the RPO and RTO the business accepts and what it will pay - and rehearsed.
Also asked: How do you choose between pilot light and warm standby? · What must exist in the recovery Region besides the data? · How would you test a disaster recovery plan?
Learn it: 39.27 Designing for resilience: AZs, Regions and DR strategies
You turned on KMS encryption on an SQS queue and an SNS topic stopped delivering to it. Why? Mid
SNS now has to encrypt messages with the queue's key, so the key policy must allow sns.amazonaws.com to use kms:GenerateDataKey and kms:Decrypt - the queue policy alone is no longer enough. I would add that statement to the key policy, scoped to the topic with aws:SourceArn, and check that the topic's failed deliveries return to zero. Consumers of the queue need kms:Decrypt too.
Also asked: How do you apply least privilege to a Lambda function? · What is the confused deputy problem? · How is responsibility split between you and AWS for RDS?
Learn it: 39.29 Designing for security: least privilege across services, encryption, segmentation, identity
Our AWS bill doubled last month. Where do you start? Mid
Cost Explorer grouped by service, then usage type, tag and day, to find when and what changed - often NAT gateway bytes, inter-AZ traffic or a forgotten cluster. Then remove waste: turn off idle resources, right-size with Compute Optimizer, add lifecycle rules for old data, move S3 traffic to a gateway endpoint. After that, commit to the new baseline with a Savings Plan, move interruptible work to Spot, and add budgets and anomaly detection so the next jump is caught early.
Also asked: When would you use Spot Instances? · How do Reserved Instances differ from Savings Plans? · How would you choose an S3 storage class?
Learn it: 39.31 Cost-optimized design: purchase options, right-sizing and the shape of the bill
How would you review an existing AWS architecture you just inherited? Mid
Pillar by pillar with the team, using the Well-Architected questions but with evidence from the account: CloudTrail and Config for how things change, IAM Access Analyzer for who can do what, the alarms that exist, backup and restore history, Compute Optimizer and Cost Explorer for waste. I would record each high risk issue, rank them by impact and effort, fix the cheap high risks first, write down the trade-offs we accept on purpose, and repeat the review as a milestone.
Also asked: What are the six pillars of the Well-Architected Framework? · What makes a finding a high risk issue? · How do you keep a review from becoming a one-off exercise?
How would you plan the migration of a few hundred applications to AWS? Mid
Assess first: an inventory with owners, usage and dependencies, and a business case. Then one of the 7 Rs per application: retire what nobody uses and retain what cannot move yet, rehost most of the rest in waves with Application Migration Service, replatform where one change pays off at once - the database to RDS with DMS - and refactor only what the business needs rebuilt, usually after the move. Before the first wave: a landing zone, skills and a pilot.
Also asked: What is the difference between rehost, replatform and refactor? · How do you migrate a database with minimal downtime? · What does the AWS Cloud Adoption Framework cover?
Learn it: 39.36 Migrating to AWS: the Cloud Adoption Framework, the 7 Rs and the tools
Why the Solutions Architect certification, and how did preparing for it change how you design? Mid
It made me reason from requirements: the RPO and RTO, the access pattern, the traffic shape, who operates the system, and then pick the simplest managed design that meets all of them - the exam's least operational overhead and most cost-effective qualifiers in practice. Because I built and broke each pattern hands-on, I can also explain how each design fails, such as a queue without a DLQ or a task that cannot reach ECR.
Also asked: How would you approach a design question with several valid answers? · What are the four domains of the Solutions Architect exam? · Which AWS whitepapers would you read before the exam?
Learn it: 39.38 The SAA-C03 exam: AWS Certified Solutions Architect - Associate
What is the difference between monitoring and observability, in AWS terms? Mid
Monitoring answers the known questions: CloudWatch metrics and alarms on what we expect to break, such as queue age, 5XX rates and throttles. Observability lets me answer new questions from outside: structured logs I can query with Logs Insights, traces, CloudTrail for what changed, and metrics with useful dimensions. I want both, with alarms on symptoms users feel and enough logs to find the cause - the skill behind the symptom-first questions of the CloudOps exam.
Also asked: What would you check first when an EC2 instance does not appear in Systems Manager? · How would you automatically remediate a non-compliant resource? · What are the domains of the CloudOps Engineer exam?
Learn it: 39.42 The SOA-C03 exam: AWS Certified CloudOps Engineer - Associate
Practise these answers with flashcards and labs Free, in your browser - a real Ubuntu terminal to try it in, with missions that check your work.