cert-bank

Guides

이 글을 한국어로 읽기

DVA-C03 Study Guide: Follow One Request Through a Serverless App, From Sign-In to Trace

The AWS Certified Developer – Associate exam (DVA-C03) is the exam of the small serverless application. A user signs in, calls an API, a function runs, data is read and written, some work is handed to a queue, and when something breaks a trace shows where. Almost every question sits at one point on that path. So this guide follows a single request from the browser to the logs, and at each hop lists what the exam expects you to know and the setting people confuse.

Study guide · AWS Certified Developer – Associate (DVA-C03)

DVA-C03 Study Guide: Follow One Request Through a Serverless App, From Sign-In to Trace

The exam

DVA-C03 replaces DVA-C02 from 1 December 2026 (the last day for DVA-C02 is 30 November 2026). 65 questions, 50 of them scored and 15 unscored; the minimum passing score is 720 on a scale of 100–1,000 (the exam guide does not state the exam time). The four domains stay, with new weights: Development with AWS Services 30%, Security 26%, Testing and Deployment 22%, Troubleshooting and Optimization 22%. Troubleshooting gained the most weight. New content covers AI-assisted development (Kiro), calling managed AI services such as Amazon Bedrock from applications, securing those AI services, and container builds and deployments.

Of the 470 DVA-C03 questions in this question bank, 140 (30%) list AWS Lambda among their services. DynamoDB comes next at 13%, then API Gateway at 10%. Lambda is the spine of the exam; learn it first and learn it deeply.

DVA-C03: one request's path, hop by hop
1
The user signs in
Cognito user pool for sign-in and JWTs; identity pool for temporary AWS credentials
2
API Gateway
REST vs HTTP APIs; Cognito, Lambda or IAM authorizers; read 429, 502 and 504
3
The function runs
Memory, timeout, reserved vs provisioned concurrency; VPC access via NAT or endpoints
4
Data is read and written
DynamoDB keys for your queries, Query not Scan, GSI vs LSI; DAX or ElastiCache
5
Slow work goes asynchronous
SQS visibility timeout and DLQ, SNS fan-out, EventBridge routing; idempotent consumers
6
Something breaks
CloudWatch Logs for what one function did; X-Ray for which hop failed or was slow
The request path this guide follows. Hops 6 (secrets) and 7 (shipping the code) have their own sections, so card 6 here is hop 8, "something breaks".

Hop 1: the user signs in

Amazon Cognito has two parts that look alike. A user pool is a user directory: sign-up, password, MFA, social login, and the JWTs your API checks. An identity pool exchanges a token for temporary AWS credentials, so a mobile app can upload straight to S3 under a role scoped to that user's prefix. "Sign-in and tokens for my API" means user pool; "the app needs to call AWS directly" means identity pool, usually behind a user pool.

Hop 2: the request hits API Gateway

  • REST APIs carry the richer features (request validation, mapping templates, caching, usage plans and API keys); HTTP APIs are simpler and cheaper.
  • Authorization: a Cognito authorizer for user pool tokens, a Lambda authorizer for custom logic, IAM for AWS callers.
  • A mock integration returns fixed responses with no backend, which lets a frontend team start before the functions exist.
  • Read the status code: 429 is throttling, 504 is an integration that ran past the API's timeout (29 seconds by default for REST APIs), 502 often means the Lambda function returned a response in the wrong shape for a proxy integration.

Hop 3: the function runs

Know the Lambda settings and what each one does not do. Memory also sets CPU. The timeout only limits how long an invocation may run. Create SDK clients and load reference data outside the handler so warm invocations reuse them. Reserved concurrency caps a function and guarantees it that capacity; provisioned concurrency keeps environments initialized to remove cold starts. A function attached to a VPC reaches the internet, and public AWS endpoints, only through a NAT gateway or VPC endpoints.

Going deeper: what a cold start actually is

Lambda runs your code in an execution environment. The first request to a new environment pays for creating it, downloading your package or image, starting the runtime and running everything outside the handler (the init phase). Later requests to the same environment skip all of that and only run the handler. That is why moving client creation out of the handler helps, and why it does nothing for the first request. When traffic grows faster than warm environments are available, Lambda creates new ones and those requests pay the init cost. Provisioned concurrency runs the init phase ahead of time for a fixed number of environments, which is the only setting in the list that removes the cost. Reserved concurrency only changes how many environments may exist, not whether they are warm.

Hop 4: data is read and written

DynamoDB questions are about access patterns. Design the partition key (and sort key) for the queries you run; use Query, not Scan. A global secondary index can be added any time but is only eventually consistent; a local secondary index must be created with the table. Read-after-write needs ConsistentRead on the base table, at twice the read cost. Throttling (ProvisionedThroughputExceededException) is handled with exponential backoff and jitter, which the SDKs already apply. For caching, know DAX for DynamoDB and ElastiCache with lazy loading or write-through for everything else.

Hop 5: slow work goes asynchronous

If the caller only needs an acknowledgement, put the job on a queue and return a job ID. SQS hides a received message for the visibility timeout; if processing takes longer, the message comes back and is processed twice. Failed messages go to a dead-letter queue after maxReceiveCount. For SQS and stream event sources, ReportBatchItemFailures lets Lambda retry only the failed records. SNS fans one message out to many subscribers; EventBridge routes events by content. Consumers must be idempotent, because delivery is at least once.

Hop 6: secrets and keys

KMS Encrypt takes at most 4 KB of plaintext, so files use envelope encryption: GenerateDataKey, encrypt locally, store the encrypted data key with the file. Secrets Manager stores and rotates secrets; Parameter Store SecureString holds configuration more cheaply without built-in rotation. Users upload large files with S3 presigned URLs, never with keys shipped in the app.

Beginner tip: deploy one AWS SAM "hello world" API that writes to a DynamoDB table, then break it three ways, one at a time: set the timeout to 1 second and add a 2-second sleep; remove the table permission from the function's role; attach the function to a private subnet with no NAT gateway. Look at each failure in CloudWatch Logs. An immediate AccessDeniedException, a Task timed out line and a hang until timeout look very different, and many troubleshooting questions are just asking you to recognize which one you are seeing.

Hop 7: the code is shipped

SAM templates expand into CloudFormation; sam deploy deploys through a CloudFormation change set, and --confirm-changeset makes it wait for your approval before applying it. Lambda versions are immutable, aliases point to them, and a weighted alias sends a share of traffic to a new version. CodeDeploy shifts that traffic by canary (a small share, then the rest) or linear (equal steps), and rolls back on a CloudWatch alarm. For containers, tag images with a commit ID in ECR with tag immutability on, never only latest. AWS AppConfig changes feature flags gradually without redeploying code.

Hop 8: something breaks

CloudWatch Logs and Logs Insights show what one function did; AWS X-Ray traces show which hop of the request failed or was slow. The Embedded Metric Format writes custom metrics as log lines, so the function makes no PutMetricData calls. Read the signal before you choose a fix: errors that appear instantly are usually permissions or validation, errors that take exactly the timeout are usually networking or a slow dependency.

The AI layer on top

AI questions are spread across all four domains, and one rule settles most of them: AI generates, summarizes and explains, while deterministic controls decide. Tests decide whether a build passes, IAM decides what an agent may call, alarms decide a rollback. Bedrock Guardrails filter the content of prompts and responses; they do not authorize API calls, and a system prompt enforces nothing. In Kiro, specs make the plan reviewable before code is written and steering files carry team conventions into every session.

Study order

  1. Lambda end to end, then API Gateway and Cognito.
  2. DynamoDB and caching, then SQS, SNS, EventBridge and Kinesis.
  3. IAM, KMS, Secrets Manager; then SAM, CloudFormation, CodeDeploy and containers.
  4. Logs, metrics and X-Ray; Bedrock and Kiro; then timed mixed sets.

Practise

Practise in the question bank: every question explains each option, right and wrong. Start with the domain you just studied, then take a timed mock exam.

Exam facts as published for DVA-C03 at the time of writing. Service limits and defaults change; check the official AWS exam guide and the service documentation.