Question 230
An LLM returns variable JSON keys and occasionally prose, causing downstream schema errors. Which processing control is needed?
- Human review for low-confidence extraction
- Bedrock text classification request
- PII redaction before LLM call
- Schema validation of LLM output
Explanation
Correct Answer
D:Model responses require parsing and validation against an expected structured contract before ingestion; in this scenario, validate fields, types and allowed values; reject or retry malformed responses and retain the raw inference for traceability.
Incorrect Answers
A:A review queue can route uncertain automated extractions to humans instead of blindly publishing them; the current need is for preventing unstructured model responses from breaking ETL, not avoiding unchecked facts in ambiguous document extraction.
B:Amazon Bedrock can invoke foundation models to classify or summarize unstructured text in a pipeline; this can support deriving categories from free-form support notes, but would not address the need for preventing unstructured model responses from breaking ETL.
C:Sensitive fields should be removed or minimized before they are sent to a model where policy requires it; this helps with keeping customer identifiers out of model prompts, not preventing unstructured model responses from breaking ETL.
Question 231
A data lake has a shared Glue catalog and needs controlled SQL access to finance tables without duplicating files per team. What design is appropriate?
- DynamoDB access-pattern modeling
- Redshift star schema
- Lake Formation governed tables
- Surrogate warehouse dimension key
Explanation
Correct Answer
C:Lake Formation grants operate on registered data-lake catalog resources and can enforce fine-grained policies; in this scenario, register and govern catalog resources through Lake Formation and grant compatible readers only the required data permissions.
Incorrect Answers
A:DynamoDB tables are designed around known partition/sort key access patterns rather than normalized joins; this can support optimizing predictable NoSQL key queries, but would not address the need for controlling which analysts can read governed S3 tables.
B:A star schema links facts to dimensions for analytic joins and common reporting aggregations; this helps with modeling sales facts for BI measures by customer and product, not controlling which analysts can read governed S3 tables.
D:Surrogate keys can stabilize warehouse joins when operational business identifiers change; it would target preserving historical dimension references through source ID changes, while the question is about controlling which analysts can read governed S3 tables.
Question 232
A table originally partitioned by month now needs a finer strategy, but rewriting all older files would be expensive. What helps?
- Apache Iceberg snapshots
- Amazon S3 Tables table bucket
- Iceberg schema evolution
- Iceberg partition evolution
Explanation
Correct Answer
D:Iceberg can evolve partition specifications without rewriting historical data immediately into one fixed layout; in this scenario, evolve the table's partition specification so new data uses the changed strategy while metadata tracks earlier layouts.
Incorrect Answers
A:Iceberg tracks table snapshots and metadata so readers can use consistent committed table states; this helps with keeping readers consistent during an object-based table update, not changing partition strategy as data volumes grow.
B:Amazon S3 Tables provides purpose-built table buckets for Apache Iceberg with automatic file compaction and snapshot maintenance; although useful for hosting Iceberg tables with managed storage-level maintenance in Amazon S3, it leaves the need for changing partition strategy as data volumes grow unmet.
C:Iceberg supports schema evolution using field identities rather than relying only on physical column order; it would target adding columns without rewriting every historical file, while the question is about changing partition strategy as data volumes grow.
Question 233
A warehouse holds valid cardholder contact data, but ordinary report users should see only masked values while a trusted role sees originals. What feature helps?
- Redshift dynamic data masking
- One-way hashing of identifiers
- Deterministic PII tokenization
- Amazon Macie sensitive data discovery
- Data anonymization aggregation
Explanation
Correct Answer
A:Masking policies can return altered column representations to database users based on defined privileges; in this scenario, attach database masking policies to the relevant roles and test unmasking privileges across query interfaces.
Incorrect Answers
B:A cryptographic hash maps inputs to fixed outputs without keeping an ordinary reversible mapping; although useful for comparing known identifiers without storing plaintext values, it leaves the need for obscuring sensitive fields at query time for nonprivileged roles unmet.
C:Tokenization replaces sensitive identifiers with controlled surrogate values and protects the mapping separately; it would target linking customer events without exposing raw identifiers, while the question is about obscuring sensitive fields at query time for nonprivileged roles.
D:Macie can examine supported S3 data for sensitive information and report discovery findings; its benefit is detecting unexpected sensitive content in S3 objects; that is different from obscuring sensitive fields at query time for nonprivileged roles.
E:Aggregation can report measurements across groups while excluding direct person-level rows from published outputs; the current need is for obscuring sensitive fields at query time for nonprivileged roles, not publishing population statistics without individual record detail.
Question 234
A JDBC URL and password are correct, yet the Glue job times out only for a database behind private networking. What should be investigated?
- JDBC driver compatibility
- ODBC connector configuration
- Secrets-based JDBC authentication
- Glue VPC routing and DNS
Explanation
Correct Answer
D:Glue network connections need paths and DNS resolution to reach databases inside a VPC; in this scenario, check subnet routes, security groups and name resolution from the Glue connection rather than altering SQL transformations.
Incorrect Answers
A:JDBC requires a driver compatible with the source database and its supported protocol; this helps with making a Glue job connect to a vendor database, not resolving a JDBC timeout to an internal data source.
B:ODBC connection settings rely on a driver and data-source configuration compatible with the application runtime; although useful for integrating a tool that expects an ODBC data source, it leaves the need for resolving a JDBC timeout to an internal data source unmet.
C:ETL connections can retrieve database credentials from a secret rather than embedding plaintext; it would target separating sensitive credentials from ETL source code, while the question is about resolving a JDBC timeout to an internal data source.
Question 235
A source delivers records at least once and the destination cannot contain duplicate transaction IDs. Which sink behavior is necessary?
- Durable consumer checkpoints
- Kinesis retention adjustment
- Idempotent output writes
- Managed Service for Apache Flink keyed windows
- Stateless transformation
Explanation
Correct Answer
C:Idempotent writes make replaying the same logical input safe without duplicating downstream effects; in this scenario, use a stable event identifier and conditional upsert or deduplication policy; retries can then preserve a single logical result.
Incorrect Answers
A:Checkpoints track a stream application's committed read position separately from the stream itself; this can support resuming from the last successfully processed record, but would not address the need for preventing duplicate target rows after event redelivery.
B:Stream retention controls how long published records remain available for replay in Kinesis Data Streams; this helps with extending the recovery window for an offline consumer, not preventing duplicate target rows after event redelivery.
D:Managed Service for Apache Flink can maintain keyed window aggregates with checkpoint-based state recovery across streaming records; it would target running recoverable per-customer rolling-window totals in managed Flink, while the question is about preventing duplicate target rows after event redelivery.
E:Stateless processors derive each output solely from their current input and need no cross-event accumulated state; its benefit is applying independent per-record normalization rules; that is different from preventing duplicate target rows after event redelivery.
Question 236
Tenants share a Redshift fact table but must see only rows belonging to their authorized tenant. What database control fits?
- Redshift database role
- Redshift row-level security policy
- Redshift GRANT on schema objects
- Redshift dynamic data masking
Explanation
Correct Answer
B:Redshift row-level security can restrict table rows returned for principals under applied policies; in this scenario, define and attach an RLS policy with the appropriate predicate and validate policy coverage for each access path.
Incorrect Answers
A:A Redshift database role groups database privileges that can be granted to multiple database users; this helps with administering common warehouse table permissions centrally, not limiting warehouse query results to a user's assigned tenant.
C:Redshift SQL grants determine which database principals can access schemas, tables and other objects; it would target giving a reporting role SELECT on one curated table, while the question is about limiting warehouse query results to a user's assigned tenant.
D:Redshift dynamic data masking can change the presentation of sensitive columns according to database policy; its benefit is hiding customer email values from ordinary warehouse readers; that is different from limiting warehouse query results to a user's assigned tenant.
Question 237
A database extract takes hours because one Glue reader processes a huge table sequentially. Which ingestion configuration should be evaluated?
- DMS table mapping rules
- S3 prefix-based ingestion
- Glue job bookmarks
- Glue JDBC partition predicates
Explanation
Correct Answer
D:JDBC partitioning lets Glue spread relational reads over partitions instead of one serial scan; in this scenario, partition the supported JDBC read by a suitable column and bounds; review database load and skew before increasing parallelism.
Incorrect Answers
A:DMS table mappings select source schemas and tables and can apply supported transformations; this can support selecting only permitted source tables for replication, but would not address the need for accelerating a large batch read of a numeric-keyed table.
B:Explicit S3 prefixes narrow the object keys examined by a batch ETL input; this helps with scoping a daily batch to the intended landing folder, not accelerating a large batch read of a numeric-keyed table.
C:Glue job bookmarks can track previously processed source data between supported job runs; although useful for avoiding the routine reread of already processed batches, it leaves the need for accelerating a large batch read of a numeric-keyed table unmet.
Question 238
A churn dashboard requires account age and region but receives raw birth dates and full addresses from ingestion. What privacy principle should guide redesign?
- SageMaker Catalog asset subscription approval
- Purpose-based data minimization
- Data governance stewardship roles
- Retention schedule governance
Explanation
Correct Answer
B:Data minimization limits collected and shared fields to those needed for the approved use case; in this scenario, publish only necessary derived fields and restrict the raw dataset so the dashboard does not expose excessive identifiers.
Incorrect Answers
A:SageMaker Unified Studio uses project-level data-asset subscription requests and approval workflows to grant governed Lake Formation table access; this helps with approving a project subscription to a restricted catalog table, not reducing unnecessary exposure in downstream analytics.
C:A governance framework assigns ownership and stewardship for datasets, definitions and access decisions; it would target establishing accountability for shared business datasets, while the question is about reducing unnecessary exposure in downstream analytics.
D:A retention schedule documents how long classes of data may be stored and when deletion or archiving is required; its benefit is aligning lake cleanup with legal data retention obligations; that is different from reducing unnecessary exposure in downstream analytics.
Question 239
A data lake creates millions of SSE-KMS objects and KMS request charges become material. What supported S3 feature may help?
- EBS volume encryption
- SSE-KMS on S3 objects
- RDS storage encryption
- S3 Bucket Key with SSE-KMS
- KMS key policy permissions
Explanation
Correct Answer
D:S3 Bucket Keys can reduce the number of KMS requests made for eligible SSE-KMS object operations; in this scenario, enable S3 Bucket Keys for eligible objects after reviewing their effect on encryption-context and audit requirements.
Incorrect Answers
A:EBS encryption protects supported block volumes at rest and in transit between instance and volume; although useful for protecting EC2-attached processing scratch volumes, it leaves the need for reducing KMS request costs for many encrypted objects unmet.
B:SSE-KMS uses an AWS KMS key for server-side encryption of objects stored in S3; it would target enforcing controlled key usage for stored sensitive lake objects, while the question is about reducing KMS request costs for many encrypted objects.
C:RDS storage encryption uses KMS integration to protect underlying database storage and supported related backups; its benefit is encrypting a managed relational source's stored data; that is different from reducing KMS request costs for many encrypted objects.
E:Key policies define key-level permission foundations for KMS operations, alongside other applicable access policies; this can support allowing authorized ETL roles to decrypt with an existing customer-managed key, but would not address the need for reducing KMS request costs for many encrypted objects.
Question 240
A compliance archive is almost never read and a retrieval wait of several hours is acceptable. Which storage class fits?
- S3 Glacier Deep Archive
- S3 Standard-IA
- S3 Intelligent-Tiering
- S3 Glacier Flexible Retrieval
Explanation
Correct Answer
D:Glacier Flexible Retrieval offers archive storage with asynchronous retrieval options for infrequently needed objects; in this scenario, transition data to archive storage under a lifecycle policy, documenting restoration times and minimum retention costs.
Incorrect Answers
A:Glacier Deep Archive targets very rarely accessed data with long retrieval expectations; the current need is for archiving files when retrieval can take hours, not storing long-term archives at low storage cost.
B:Standard-IA provides lower storage pricing for infrequently accessed objects with retrieval charges and minimum-duration constraints; this can support storing infrequent but quickly retrievable lake archives, but would not address the need for archiving files when retrieval can take hours.
C:Intelligent-Tiering monitors eligible object access and can move objects between access tiers automatically; this helps with handling unpredictable access frequency for retained objects, not archiving files when retrieval can take hours.