DVA-C03 Mistakes: Look-Alike Settings, Misread Error Signals and the Limits Nobody Remembers
Wrong answers on the AWS Developer – Associate exam tend to come from three habits. People pick a setting whose name sounds right but does something else. They pick a fix before reading what the error is actually saying. And they forget that an API returns its results in pages or refuses a payload above a certain size. We went through the explanations of the 470 DVA-C03 questions we wrote for this question bank and sorted the most common traps into these three families, plus a fourth that is new in DVA-C03: AI controls that look like security but are not.
Common mistakes · AWS Certified Developer – Associate (DVA-C03)

Family 1: settings with look-alike names
These pairs appear side by side in the options because each one is a plausible reading of the stem. The fix is to know what each setting does not do. Eight of our explanations discuss reserved and provisioned concurrency together, and another eight compare Cognito user pools with identity pools.
| Pair | What separates them | Stem words that decide |
|---|---|---|
| Reserved vs provisioned concurrency | Reserved caps a function and guarantees it capacity, but keeps nothing warm. Provisioned keeps environments initialized, but on its own does not cap scaling. | "never more than", "protect the database" vs "cold start", "latency at 9 a.m." |
| User pool vs identity pool | A user pool signs users in and issues JWTs. An identity pool trades a token for temporary AWS credentials. | "sign-up, MFA, token for our API" vs "upload directly to S3" |
| GSI vs LSI | A GSI can be added later but reads are only eventually consistent. An LSI shares the partition key, must be created with the table, and supports strongly consistent reads. | "existing table", "different partition key" vs "strongly consistent" |
| Lambda alias vs API Gateway stage | An alias points to a function version and can split traffic by weight. A stage is a deployed snapshot of an API with its own URL and variables. | "10% to the new version" vs "dev and prod URLs" |
| Canary vs linear (CodeDeploy) | Canary shifts a small share, waits, then shifts the rest. Linear shifts equal steps at fixed intervals. | "two steps" vs "every N minutes" |
| S3 presigned URL vs CloudFront signed URL | A presigned URL works only while the credentials that signed it are valid. A CloudFront signed URL uses a key group and its own expiry. | "stops working after a few hours", "valid for days" |
Going deeper: why a presigned URL dies with its signer's session
A presigned URL is not a separate credential. It is an ordinary S3 request, signed in advance with the signer's access key, with the signature and (for temporary credentials) the session token placed in the query string. When S3 receives it, it checks the signature and then evaluates the request as if the signer had made it at that moment. If the signer was a Lambda function, its credentials come from the execution role and are short-lived; once that session expires, every URL it signed fails, whatever expiry was written into the URL. The same logic explains a related trap: if the object uses SSE-KMS, the signer also needs kms:Decrypt on the key, because the download is evaluated as the signer's request.
Family 2: reading the error before choosing the fix
About one DVA-C03 question in five in our bank (97 of 470) is a troubleshooting scenario, and most of them hand you the clue in the symptom. The wrong options are usually real fixes for a different symptom.
| You see | It usually means | Tempting wrong fix |
|---|---|---|
| Every call times out right after the function joined a VPC | No route out: add a NAT gateway or VPC endpoints | More memory, or IAM changes |
Immediate AccessDeniedException on KMS although IAM allows it | The key policy does not let IAM policies grant access | Adding more IAM permissions |
429 / throttles while the account limit is unused | The function's own reserved concurrency is the cap | Raising the account quota |
504 only for slow requests | API Gateway's integration timeout; make the work asynchronous | A longer Lambda timeout |
| Preflight succeeds, GET fails with a CORS error (proxy integration) | The function must return the CORS headers itself | Enabling CORS in the console again |
| SQS messages reappear and end in the DLQ; function duration near its timeout | Processing outlasts the visibility timeout, or a slow dependency (confirm with X-Ray) | A bigger maxReceiveCount |
Invocations and cost spike after an S3 trigger, or the RecursiveInvocationsDropped metric appears | The function writes back into the prefix that triggers it (Lambda recursive loop detection stops the chain after about 16 invocations, but the design is still wrong) | Retries, DLQs or versioning |
| "Too many connections" on RDS during spikes | Each Lambda environment holds its own connection: use RDS Proxy | Reconnecting on every call |
Going deeper: why a permissions problem fails fast and a network problem fails slowly
When a request reaches an AWS endpoint, the service authenticates it and evaluates the policies before doing any work, and answers at once with an error such as AccessDeniedException. That needs a network path: the request must arrive. When there is no path (a private subnet with no NAT gateway and no endpoint), the connection attempt simply gets no reply. The SDK waits for its connection timeout, retries, and waits again, until the function's own timeout ends the invocation. So the timing is itself a clue: an instant failure that names an action points to policies, while a failure that takes exactly the configured timeout points to routes, security groups or a slow dependency.
Beginner tip: make a one-page "signal card" from the table above and add a row every time a practice question fools you: the exact symptom text in the left column, the cause in the middle, and the option you wrongly chose on the right. Before answering any troubleshooting question, put your hand over the options and say the cause first. Only then read the choices.
No route out: add a NAT gateway or VPC endpoints
The key policy does not let IAM policies grant access
The function's own reserved concurrency is the cap
API Gateway's integration timeout: make it asynchronous
Each Lambda environment holds a connection: use RDS Proxy
Instant failure: check policies. Failure at exactly the timeout: routes or a slow dependency.
Family 3: limits and pages that change the design
Some questions turn on a limit that forces a different design rather than a bigger setting. KMS Encrypt accepts at most 4 KB of plaintext, so files need envelope encryption with GenerateDataKey. Large browser uploads go straight to S3 with presigned URLs instead of through API Gateway and Lambda, whose request payloads are capped well below the sizes in these stems. And list and query APIs page their results: a DynamoDB Query that returns a LastEvaluatedKey has more items to fetch, and ListObjectsV2 returns at most 1,000 keys per call. The fix is to loop with ExclusiveStartKey or the continuation token (or use an SDK paginator), not to add capacity or switch to Scan.
Family 4: AI controls mistaken for security
DVA-C03 adds questions about AI in the developer workflow, and one confusion runs through them. Bedrock Guardrails inspect the content of prompts and responses: denied topics, harmful content, personal data. They do not authorize or block API calls, so they do not limit what an agent's tools can do; least-privilege IAM roles and tool-level policies do. A system prompt that says "never change production" enforces nothing. AI-generated code, tests and IAM policies are reviewed like any other change: an AI reviewer's comments inform a human, and a generated policy with bedrock:* on every resource still needs to be narrowed to the actions and ARNs the function uses.
A five-second test for each family
- Look-alike names: can you say what the tempting option does not do?
- Error signals: did you name the cause before reading the options?
- Limits and pages: is there a size, a count of 1,000 or a token in the stem?
- AI controls: does the option filter content or actually restrict permissions?
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.
- DVA-C03 exam page · All 530 DVA-C03 questions · Timed mock exam · 30-day study plan
- By domain: Development with AWS Services · Security · Testing and Deployment · Troubleshooting and Optimization
- Family 2: Troubleshooting and Optimization
- Families 1 and 4: Security · Family 3: Development with AWS Services
Patterns drawn from the explanations in our own DVA-C03 question bank. Service limits change; check the AWS exam guide for DVA-C03 and the current service documentation.