Question 270
A compliance check has only a few thousand embeddings and must find the mathematically closest items without approximation error. What search method fits?
- HNSW approximate index
- Vector index recall tuning
- IVF vector clustering
- Exact vector similarity scan
Explanation
Correct Answer
D:Exact similarity evaluates distances against all relevant embeddings rather than using approximate pruning; in this scenario, compute exact distances and rank results, trading more work per query for exact nearest-neighbor results.
Incorrect Answers
A:HNSW organizes vectors in a navigable proximity graph for approximate nearest-neighbor retrieval; its benefit is favoring search speed and recall in large vector indexes; that is different from obtaining an exact ranking for a small embedding collection.
B:Recall measures how often approximate search finds neighbors present in an exact reference result set; the current need is for obtaining an exact ranking for a small embedding collection, not validating approximate-search quality against exact neighbors.
C:IVF partitions vector space into coarse clusters and searches selected candidate lists during retrieval; this can support limiting searches to coarse vector clusters, but would not address the need for obtaining an exact ranking for a small embedding collection.
Question 271
An EMR cluster was terminated after a nightly failure, but its Spark driver output must still be analyzed. What prior configuration helps?
- Glue job worker increase
- Glue job bookmark reset
- Glue job data partition tuning
- EMR managed scaling policies
- EMR step logs in S3
Explanation
Correct Answer
E:EMR cluster logging can preserve step and application output in an S3 log destination; in this scenario, enable appropriate log archiving so step output survives the cluster and can be inspected after termination.
Incorrect Answers
A:More appropriate Glue worker resources can help when Spark tasks are genuinely CPU or memory constrained; although useful for reducing a resource-limited Glue job runtime, it leaves the need for investigating a failed historical EMR processing step unmet.
B:Resetting a Glue bookmark permits an ETL job to reprocess previously tracked source data intentionally; it would target replaying a corrected historical data batch, while the question is about investigating a failed historical EMR processing step.
C:Spark partition sizing determines whether distributed ETL tasks can run evenly and make good use of available executor slots; its benefit is raising parallel task utilization when a Glue read produces too few partitions; that is different from investigating a failed historical EMR processing step.
D:EMR managed scaling adjusts available node capacity under configured bounds for supported clusters; the current need is for investigating a failed historical EMR processing step, not matching variable Spark compute demand without constant oversizing.
Question 272
A warehouse bill increased and engineers must identify which SQL statements scanned large tables. Which operational record is most relevant?
- S3 server access logging
- CloudTrail event history
- DMS task logs
- Redshift system query logs
Explanation
Correct Answer
D:Redshift query and system logging views provide evidence about SQL execution and warehouse operations; in this scenario, inspect Redshift query execution information and correlate users, durations and scan metrics with the cost incident.
Incorrect Answers
A:S3 server access logs can record requests made to a bucket and deliver log files to a chosen destination; its benefit is reviewing bucket request activity in a storage-focused audit; that is different from tracing an unexpectedly expensive warehouse query.
B:CloudTrail event history exposes recent management activity for an account within its documented retention; the current need is for tracing an unexpectedly expensive warehouse query, not tracing who changed a Glue job configuration recently.
C:DMS task logs record migration and change-capture operational events and diagnostic details; this can support diagnosing failed replication-table mappings, but would not address the need for tracing an unexpectedly expensive warehouse query.
Question 273
Hundreds of curated tables carry `classification=public` or `restricted`, and teams need rules that follow those labels. Which model fits?
- Lake Formation table SELECT
- Lake Formation column permissions
- Lake Formation data location permissions
- Lake Formation LF-Tag permissions
Explanation
Correct Answer
D:LF-Tags enable tag-based governance over supported catalog resources and scalable permission assignment; in this scenario, tag registered resources consistently and grant LF-Tag-based permissions to roles based on authorized classifications.
Incorrect Answers
A:Lake Formation grants can authorize SQL SELECT access on governed catalog tables for supported integrations; it would target granting analysts read access to approved lake tables, while the question is about granting access to datasets by business classification.
B:Lake Formation supports fine-grained grants selecting only approved columns of governed tables; its benefit is preventing access to sensitive columns without copying datasets; that is different from granting access to datasets by business classification.
C:Data location permissions can authorize principals to create or alter governed catalog resources at registered storage locations; the current need is for granting access to datasets by business classification, not allowing a controlled role to create tables in one registered S3 location.
Question 274
A source's null rate increases slowly and operations needs to compare quality evaluation outcomes by day. What output should be retained?
- Glue Data Quality ruleset
- Glue Data Quality evaluation results
- DataBrew validation recipe
- DataBrew profile statistics
- Quality quarantine S3 prefix
Explanation
Correct Answer
B:Quality evaluation results record which configured checks pass and fail for a tested dataset; in this scenario, store evaluation scores and failed-rule details with batch identifiers, then alert on meaningful changes from baseline.
Incorrect Answers
A:Glue Data Quality supports rulesets for evaluating data quality against supported datasets and ETL flows; although useful for applying repeatable quality checks in a Glue pipeline, it leaves the need for tracking rule failures across successive daily runs unmet.
C:A DataBrew recipe records transformations that can normalize inconsistent values and produce an approved output; its benefit is standardizing known fields before final quality evaluation; that is different from tracking rule failures across successive daily runs.
D:DataBrew profile jobs summarize null values, distributions and other properties of tabular input; the current need is for tracking rule failures across successive daily runs, not measuring baseline data characteristics before quality rule design.
E:A separate S3 prefix can preserve rejected records for controlled remediation without polluting curated queries; this can support keeping invalid records available for debugging but out of production analytics, but would not address the need for tracking rule failures across successive daily runs.
Question 275
A Spark job joins a billion events with a 30 MB country-code table and spends time reshuffling both inputs. What join optimization helps?
- Incremental ETL merge
- Redshift materialized view refresh
- JDBC predicate pushdown
- Spark broadcast join
Explanation
Correct Answer
D:A broadcast join sends a small relation to worker nodes to avoid shuffling both sides of a join; in this scenario, broadcast the small lookup relation when it fits worker memory, avoiding a large shuffle on the event side.
Incorrect Answers
A:Incremental merges apply inserts and updates from new source changes instead of rebuilding the entire destination; it would target processing only changed business records after an initial load, while the question is about reducing shuffle overhead when one lookup table is tiny.
B:A materialized view stores precomputed query results that can be refreshed as base data changes; its benefit is accelerating repeated joins and aggregations for warehouse users; that is different from reducing shuffle overhead when one lookup table is tiny.
C:Source-side predicates can reduce rows transferred from a JDBC database into a distributed ETL job; the current need is for reducing shuffle overhead when one lookup table is tiny, not filtering historical relational data before network transfer.
Question 276
Engineers want shared abstractions for a data pipeline's buckets, IAM roles and Glue resources in TypeScript. Which IaC tool suits that approach?
- CloudFormation change set
- AWS CloudFormation stack
- AWS SAM template
- AWS CDK construct
Explanation
Correct Answer
D:AWS CDK expresses cloud infrastructure using supported programming languages that synthesize CloudFormation; in this scenario, build reusable CDK constructs and synthesize them into reviewed CloudFormation stacks before deployment.
Incorrect Answers
A:A change set previews proposed modifications to a stack before they are executed; this helps with reviewing potentially destructive infrastructure updates, not reusing programmable infrastructure definitions across teams.
B:CloudFormation provisions and updates AWS resources through declarative templates and stack change tracking; although useful for recreating the same data pipeline infrastructure in different accounts, it leaves the need for reusing programmable infrastructure definitions across teams unmet.
C:AWS SAM extends CloudFormation with serverless resource declarations for applications; it would target packaging Lambda and Step Functions assets as serverless infrastructure, while the question is about reusing programmable infrastructure definitions across teams.
Question 277
A platform team allows project owners to attach policies but must prevent those roles from exceeding approved S3 access. Which IAM mechanism fits?
- IAM policy condition key
- IAM permissions boundary
- IAM resource-scoped S3 policy
- IAM explicit deny precedence
Explanation
Correct Answer
B:A permissions boundary limits the maximum effective permissions of a principal even when identity policies allow more; in this scenario, apply a permissions boundary to created roles and require boundary use through delegated creation permissions.
Incorrect Answers
A:IAM policy conditions can limit allowed API use based on supported request attributes and context; its benefit is restricting object writes to approved encryption settings; that is different from constraining delegated ETL-role administration.
C:An IAM identity policy can scope supported S3 operations to specific bucket and object ARN patterns; this can support granting ETL read access to one landing prefix, but would not address the need for constraining delegated ETL-role administration.
D:An explicit deny in an applicable policy overrides otherwise allowing IAM statements; this helps with diagnosing access denied despite an Allow statement, not constraining delegated ETL-role administration.
Question 278
A partner frequently adds nested JSON keys and the raw zone must preserve all events before curated contracts are defined. Which architecture helps?
- Controlled schema migration version
- Schema-on-read normalization
- Schema registry compatibility check
- Backward-compatible optional column
Explanation
Correct Answer
B:Schema-on-read applies structure when data is queried or transformed rather than forcing one physical shape at ingestion; in this scenario, store raw records faithfully and validate/normalize them in a curated layer where expected schemas are explicit.
Incorrect Answers
A:Versioned schema migration scripts make database changes reproducible and reviewable across environments; this can support coordinating relational schema changes with ETL releases, but would not address the need for retaining unpredictable third-party fields for later curation.
C:A schema registry can enforce defined compatibility rules for producers and consumers sharing encoded records; although useful for blocking incompatible event schema changes during deployment, it leaves the need for retaining unpredictable third-party fields for later curation unmet.
D:Adding an optional column can allow older records and consumers to coexist with newer schema versions; it would target introducing a nullable attribute without breaking old data, while the question is about retaining unpredictable third-party fields for later curation.
Question 279
New S3 files arrive while a warehouse batch is loading, and operators must load only an approved file list. What COPY input feature helps?
- Redshift UNLOAD to S3
- Redshift COPY from S3
- Redshift staging table merge
- Redshift COPY manifest
Explanation
Correct Answer
D:A COPY manifest specifies the exact objects that a load should consume for reproducibility; in this scenario, use a manifest of explicit object references so the load's input membership is defined rather than a changing prefix.
Incorrect Answers
A:UNLOAD writes Redshift query results to S3 objects for exchange or downstream analysis; this helps with exporting a curated warehouse query result to the lake, not preventing partial object sets during a warehouse load.
B:COPY loads supported data files from S3 into Redshift tables using managed parallel reads; although useful for ingesting a large staged dataset into warehouse tables, it leaves the need for preventing partial object sets during a warehouse load unmet.
C:Loading into a staging table before a controlled merge separates ingestion from publication of target data; it would target atomically applying validated changes to warehouse tables, while the question is about preventing partial object sets during a warehouse load.
Question 280
A clickstream alternates between quiet periods and traffic spikes; manual shard allocation has become error-prone. Which stream mode is suitable?
- RDS Proxy connection pooling
- Exponential backoff with jitter
- DynamoDB adaptive capacity
- Kinesis on-demand capacity
- Client-side concurrency limit
Explanation
Correct Answer
D:Kinesis on-demand mode adjusts stream capacity without manually planning shard counts for typical demand patterns; in this scenario, on-demand mode automates capacity management within its scaling behavior; monitor observed throughput and applicable limits.
Incorrect Answers
A:RDS Proxy pools application database connections and reduces connection churn against supported databases; this helps with protecting a relational database from short-lived connection bursts, not absorbing variable streaming ingestion demand without shard planning.
B:Backoff and jitter distribute API retries in time after throttling responses or transient errors; although useful for preventing synchronized retries after a shared API quota is exhausted, it leaves the need for absorbing variable streaming ingestion demand without shard planning unmet.
C:DynamoDB adaptive capacity can redistribute available throughput to heavily accessed partitions within service limits; it would target mitigating uneven partition access within a DynamoDB table, while the question is about absorbing variable streaming ingestion demand without shard planning.
E:A semaphore or worker cap bounds simultaneous API requests sent by an ingestion job; the current need is for absorbing variable streaming ingestion demand without shard planning, not keeping API requests below a known in-flight limit.