Exam Bundle

A rerun uses the same immutable S3 manifest and produces identical input checksums without public endpoint dependency. This would verify a different operation: Land a versioned, integrity-checked source copy in an authorized S3 data lake before launching repeatable processing jobs. That separate action addresses: A SageMaker processing job repeatedly downloads a multi-gigabyte model-input dataset from a distant public endpoint and fails on unstable external transfers. It cannot establish the required Keystone BioLabs correction for: A batch transformation joins slowly changing reference categories using today's labels for all historic examples, leaking future feature state. The decisive acceptance check is instead: A historical example contains the category active at its original event timestamp instead of a later corrected label.

The destination manifest reconciles object counts and checksums with the source snapshot and identifies any failed transfers. This would verify a different operation: Use an authorized transfer plan with manifest or checksum verification and record the source snapshot identifier for lineage. That separate action addresses: A sensitive ML corpus copied by DataSync into S3 has no reproducible record that the transferred objects match the approved source snapshot. It cannot establish the required Keystone BioLabs correction for: A batch transformation joins slowly changing reference categories using today's labels for all historic examples, leaking future feature state. The decisive acceptance check is instead: A historical example contains the category active at its original event timestamp instead of a later corrected label.

A scheduled retention test can still retrieve every manifest object needed to reproduce the archived model version. This would verify a different operation: Apply retention and lifecycle rules compatible with artifact lineage and archive immutable source objects required for approved reproducibility. That separate action addresses: A training manifest refers to S3 objects that were removed by an aggressive lifecycle rule, so reproducible model builds fail months later. It cannot establish the required Keystone BioLabs correction for: A batch transformation joins slowly changing reference categories using today's labels for all historic examples, leaking future feature state. The decisive acceptance check is instead: A historical example contains the category active at its original event timestamp instead of a later corrected label.

Use an authorized transfer plan with manifest or checksum verification and record the source snapshot identifier for lineage. It does not directly address the Timberline Foods case: A batch transformation joins slowly changing reference categories using today's labels for all historic examples, leaking future feature state. The relevant acceptance criterion for this case is: A historical example contains the category active at its original event timestamp instead of a later corrected label.

Separate and explicitly version S3 prefixes or manifests for train, validation and test data with role-based write control. It does not directly address the Timberline Foods case: A batch transformation joins slowly changing reference categories using today's labels for all historic examples, leaking future feature state. The relevant acceptance criterion for this case is: A historical example contains the category active at its original event timestamp instead of a later corrected label.

Define a validated incoming schema and reject malformed CSV records at ingestion before creating the curated training dataset. It does not directly address the Timberline Foods case: A batch transformation joins slowly changing reference categories using today's labels for all historic examples, leaking future feature state. The relevant acceptance criterion for this case is: A historical example contains the category active at its original event timestamp instead of a later corrected label.

Every required rare class appears at least the configured minimum count in evaluation and no training row is duplicated into test. This would verify a different operation: Use a stratified or otherwise class-aware splitting method with declared minimum representation checks for the holdout set. That separate action addresses: A multi-class label dataset is split without stratification and a rare class is missing entirely from the evaluation partition. It cannot establish the required Timberline Foods correction for: A batch transformation joins slowly changing reference categories using today's labels for all historic examples, leaking future feature state. The decisive acceptance check is instead: A historical example contains the category active at its original event timestamp instead of a later corrected label.

The final report records positive-class recall and precision on the original distribution, not only inflated aggregate accuracy. This would verify a different operation: Measure class balance and compare stratified resampling or class weighting on training data while preserving an untouched evaluation set. That separate action addresses: A marketing propensity model has a class distribution of 98 percent negative examples and selection by accuracy alone hides poor positive detection. It cannot establish the required Timberline Foods correction for: A batch transformation joins slowly changing reference categories using today's labels for all historic examples, leaking future feature state. The decisive acceptance check is instead: A historical example contains the category active at its original event timestamp instead of a later corrected label.

A benchmark reads fewer files, preserves event totals and lowers processing overhead for the same approved date partition. This would verify a different operation: Introduce a compaction process that creates suitably sized columnar objects and retains partition and source lineage in the curated dataset. That separate action addresses: An ML data lake receives tiny telemetry objects every second and Spark processing time is dominated by listing and opening millions of files. It cannot establish the required Timberline Foods correction for: A batch transformation joins slowly changing reference categories using today's labels for all historic examples, leaking future feature state. The decisive acceptance check is instead: A historical example contains the category active at its original event timestamp instead of a later corrected label.

Use a designed DynamoDB export to S3 or appropriate controlled query pattern instead of repeated unrestricted production table scans. It does not directly address the Westhaven Retail case: A Spark feature preparation stage converts all encoded records to driver memory and fails for large training partitions. The relevant acceptance criterion for this case is: The large synthetic partition completes without driver out-of-memory errors and its aggregate counts match the reference run.

Correct the offline-store S3 and IAM configuration and verify feature group creation and data delivery permissions. It does not directly address the Westhaven Retail case: A Spark feature preparation stage converts all encoded records to driver memory and fails for large training partitions. The relevant acceptance criterion for this case is: The large synthetic partition completes without driver out-of-memory errors and its aggregate counts match the reference run.

Write immutable uniquely named S3 objects and publish a versioned manifest after all parallel data-part writes complete. It does not directly address the Westhaven Retail case: A Spark feature preparation stage converts all encoded records to driver memory and fails for large training partitions. The relevant acceptance criterion for this case is: The large synthetic partition completes without driver out-of-memory errors and its aggregate counts match the reference run.

A training role cannot access the mapping store while an explicitly authorized re-identification role can complete an audited lookup. This would verify a different operation: Protect the token mapping with separate least-privilege access and encryption controls and limit its retention to approved workflows. That separate action addresses: An ML pipeline stores masked user IDs but uses a reversible mapping file in a publicly accessible S3 location alongside training data. It cannot establish the required Westhaven Retail correction for: A Spark feature preparation stage converts all encoded records to driver memory and fails for large training partitions. The decisive acceptance check is instead: The large synthetic partition completes without driver out-of-memory errors and its aggregate counts match the reference run.

The final report records positive-class recall and precision on the original distribution, not only inflated aggregate accuracy. This would verify a different operation: Measure class balance and compare stratified resampling or class weighting on training data while preserving an untouched evaluation set. That separate action addresses: A marketing propensity model has a class distribution of 98 percent negative examples and selection by accuracy alone hides poor positive detection. It cannot establish the required Westhaven Retail correction for: A Spark feature preparation stage converts all encoded records to driver memory and fails for large training partitions. The decisive acceptance check is instead: The large synthetic partition completes without driver out-of-memory errors and its aggregate counts match the reference run.

A measured extract contains only authorized columns and the intended time window while transfer volume decreases. This would verify a different operation: Push filters and projections into an appropriate query or unload operation before delivering selected fields to the training pipeline. That separate action addresses: A distributed feature preparation process overfetches every Redshift table column even though the model uses just a few features and one time period. It cannot establish the required Westhaven Retail correction for: A Spark feature preparation stage converts all encoded records to driver memory and fails for large training partitions. The decisive acceptance check is instead: The large synthetic partition completes without driver out-of-memory errors and its aggregate counts match the reference run.

Write immutable uniquely named S3 objects and publish a versioned manifest after all parallel data-part writes complete. It does not directly address the Granite SaaS case: A Spark feature preparation stage converts all encoded records to driver memory and fails for large training partitions. The relevant acceptance criterion for this case is: The large synthetic partition completes without driver out-of-memory errors and its aggregate counts match the reference run.

Correct the offline-store S3 and IAM configuration and verify feature group creation and data delivery permissions. It does not directly address the Granite SaaS case: A Spark feature preparation stage converts all encoded records to driver memory and fails for large training partitions. The relevant acceptance criterion for this case is: The large synthetic partition completes without driver out-of-memory errors and its aggregate counts match the reference run.

Grant narrowly scoped source-object and KMS decrypt permissions to the authorized consuming role and confirm the key policy permits the operation. It does not directly address the Granite SaaS case: A Spark feature preparation stage converts all encoded records to driver memory and fails for large training partitions. The relevant acceptance criterion for this case is: The large synthetic partition completes without driver out-of-memory errors and its aggregate counts match the reference run.

A held-out adjudication sample meets the agreed inter-annotator agreement threshold and disputed items remain traceable. This would verify a different operation: Define annotation guidelines and use supported consensus or adjudication workflows before including labels in model training. That separate action addresses: SageMaker Ground Truth annotators disagree frequently on ambiguous classification labels yet the pipeline treats every raw label as ground truth. It cannot establish the required Granite SaaS correction for: A Spark feature preparation stage converts all encoded records to driver memory and fails for large training partitions. The decisive acceptance check is instead: The large synthetic partition completes without driver out-of-memory errors and its aggregate counts match the reference run.

An audit of transformed images preserves orientation-sensitive labels and prohibited flips are absent from model training. This would verify a different operation: Restrict image augmentation to clinically valid transformations and verify label invariance for each permitted operation. That separate action addresses: A clinical image dataset uses an augmentation pipeline that reverses left-right orientation despite labels depending on laterality. It cannot establish the required Granite SaaS correction for: A Spark feature preparation stage converts all encoded records to driver memory and fails for large training partitions. The decisive acceptance check is instead: The large synthetic partition completes without driver out-of-memory errors and its aggregate counts match the reference run.

An approved source update appears within the freshness objective and replayed change messages do not create duplicate feature rows. This would verify a different operation: Introduce an appropriate CDC or streaming ingestion design that preserves update ordering and deduplicates repeated change events. That separate action addresses: A team uses a daily full-table export for near-real-time ML fraud features even though the source produces frequent small changes. It cannot establish the required Granite SaaS correction for: A Spark feature preparation stage converts all encoded records to driver memory and fails for large training partitions. The decisive acceptance check is instead: The large synthetic partition completes without driver out-of-memory errors and its aggregate counts match the reference run.

Classify and remove unnecessary protected columns before the ML staging layer and enforce restricted access to retained originals. It does not directly address the Bluehaven Transit case: A streaming Lambda transformer retries failed events but appends duplicate feature writes each time its invocation is replayed. The relevant acceptance criterion for this case is: Two deliveries of one test event result in one accepted logical feature update and a failed payload remains retriable.

Define a validated incoming schema and reject malformed CSV records at ingestion before creating the curated training dataset. It does not directly address the Bluehaven Transit case: A streaming Lambda transformer retries failed events but appends duplicate feature writes each time its invocation is replayed. The relevant acceptance criterion for this case is: Two deliveries of one test event result in one accepted logical feature update and a failed payload remains retriable.

Push filters and projections into an appropriate query or unload operation before delivering selected fields to the training pipeline. It does not directly address the Bluehaven Transit case: A streaming Lambda transformer retries failed events but appends duplicate feature writes each time its invocation is replayed. The relevant acceptance criterion for this case is: Two deliveries of one test event result in one accepted logical feature update and a failed payload remains retriable.

A training-container read test reaches the specific feature files while a different identity cannot access the protected directory. This would verify a different operation: Configure authorized EFS access-point or POSIX permissions and ensure the job's VPC mount connectivity supports the intended path. That separate action addresses: A SageMaker training job reads from a shared EFS mount but fails because the container user lacks filesystem permissions on the feature directory. It cannot establish the required Bluehaven Transit correction for: A streaming Lambda transformer retries failed events but appends duplicate feature writes each time its invocation is replayed. The decisive acceptance check is instead: Two deliveries of one test event result in one accepted logical feature update and a failed payload remains retriable.

A test input with NULL is marked missing and a true zero-dollar record remains distinguishable in the clean dataset. This would verify a different operation: Validate nullable field semantics and reject or impute explicit missing tokens without converting them to legitimate zero values. That separate action addresses: A source CSV sends the literal string NULL for missing prices, but the preparation job treats it as a valid zero-dollar transaction. It cannot establish the required Bluehaven Transit correction for: A streaming Lambda transformer retries failed events but appends duplicate feature writes each time its invocation is replayed. The decisive acceptance check is instead: Two deliveries of one test event result in one accepted logical feature update and a failed payload remains retriable.

A reproducible export contains the required records while production capacity and latency remain inside agreed limits. This would verify a different operation: Use a designed DynamoDB export to S3 or appropriate controlled query pattern instead of repeated unrestricted production table scans. That separate action addresses: A training pipeline needs recent transaction attributes from DynamoDB, but its ad hoc full table scan increases read capacity consumption during business hours. It cannot establish the required Bluehaven Transit correction for: A streaming Lambda transformer retries failed events but appends duplicate feature writes each time its invocation is replayed. The decisive acceptance check is instead: Two deliveries of one test event result in one accepted logical feature update and a failed payload remains retriable.

Measure partition-key distribution and provision or scale stream capacity using suitable shard or on-demand settings. It does not directly address the Brookside Manufacturing case: A streaming Lambda transformer retries failed events but appends duplicate feature writes each time its invocation is replayed. The relevant acceptance criterion for this case is: Two deliveries of one test event result in one accepted logical feature update and a failed payload remains retriable.

Split examples by patient identity and keep all records for one identity in only one dataset partition. It does not directly address the Brookside Manufacturing case: A streaming Lambda transformer retries failed events but appends duplicate feature writes each time its invocation is replayed. The relevant acceptance criterion for this case is: Two deliveries of one test event result in one accepted logical feature update and a failed payload remains retriable.

Attach source system, ingest time and immutable batch manifest metadata to accepted training records during data integration. It does not directly address the Brookside Manufacturing case: A streaming Lambda transformer retries failed events but appends duplicate feature writes each time its invocation is replayed. The relevant acceptance criterion for this case is: Two deliveries of one test event result in one accepted logical feature update and a failed payload remains retriable.

A manifest audit finds every record in exactly its intended split with no overlapping S3 object references. This would verify a different operation: Separate and explicitly version S3 prefixes or manifests for train, validation and test data with role-based write control. That separate action addresses: An S3 training inventory uses object prefixes that mix production and evaluation records, causing the preparation job to include the wrong dataset split. It cannot establish the required Brookside Manufacturing correction for: A streaming Lambda transformer retries failed events but appends duplicate feature writes each time its invocation is replayed. The decisive acceptance check is instead: Two deliveries of one test event result in one accepted logical feature update and a failed payload remains retriable.

The destination manifest reconciles object counts and checksums with the source snapshot and identifies any failed transfers. This would verify a different operation: Use an authorized transfer plan with manifest or checksum verification and record the source snapshot identifier for lineage. That separate action addresses: A sensitive ML corpus copied by DataSync into S3 has no reproducible record that the transferred objects match the approved source snapshot. It cannot establish the required Brookside Manufacturing correction for: A streaming Lambda transformer retries failed events but appends duplicate feature writes each time its invocation is replayed. The decisive acceptance check is instead: Two deliveries of one test event result in one accepted logical feature update and a failed payload remains retriable.

The benchmark query scans fewer bytes and still returns identical approved training rows and feature values. This would verify a different operation: Convert the curated training data to partitioned Parquet with column statistics and choose partitions aligned to actual predicates. That separate action addresses: A data engineering job stores a large column-oriented training dataset as thousands of verbose JSON files and suffers excessive scan costs in Athena. It cannot establish the required Brookside Manufacturing correction for: A streaming Lambda transformer retries failed events but appends duplicate feature writes each time its invocation is replayed. The decisive acceptance check is instead: Two deliveries of one test event result in one accepted logical feature update and a failed payload remains retriable.