MLA-C02 sample questions with answers

1. A.

   Clarify is closed to new customers; AWS describes computing the standardised bias metrics directly with pandas and scikit-learn and using the SHAP library that Clarify itself is built on, operationalised in a pipeline with SageMaker AI MLflow, CloudWatch and QuickSight.

2. B.

   Clarify remains available but processing jobs have been replaced by a real-time API, so the team should call the bias and explainability endpoints directly instead of submitting jobs.

3. C.

   Clarify is closed to new customers and Amazon Bedrock Evaluations is the designated replacement for all of its capabilities, including pre-training bias metrics and SHAP attributions on tabular models. Bedrock Evaluations accepts tabular datasets directly and produces the same pre-training bias report that Clarify processing jobs used to emit.

4. D.

   Clarify is closed to new customers and the documented replacement is to enable SageMaker Model Monitor bias drift monitoring, which computes the same pre-training metrics without a Clarify job.

1. A.

   For CSV, TSV and JSON, Athena determines the compression type from the file extension and treats files with no extension as uncompressed plain text; uppercase extensions such as .GZ are not recognised, so both prefixes must use lowercase gz extensions.

2. B.

   The table's SerDe is missing the compression table property, so Athena defaults to no compression; adding write.compression = GZIP to TBLPROPERTIES makes existing files readable.

3. C.

   Athena cannot read gzip-compressed text files at all and requires that text data be converted to Parquet or ORC before it can be queried, regardless of how the files are named. The unreadable characters in the result are the compressed bytes being rendered as text, which is the expected outcome for an unsupported storage format.

4. D.

   Gzip files must be concatenated into a single object per partition because Athena cannot open multiple independently compressed members in one prefix, which produces the binary output observed.

1. A.

   S3 Express One Zone cannot be used as a training data source at all, because SageMaker AI input modes require a general purpose bucket; the team must fall back to FSx for Lustre for low-latency access. The team should accept the VPC configuration and single Availability Zone subnet that a Lustre file system requires.

2. B.

   The directory bucket can serve as the training input location, but SageMaker AI output data in a directory bucket can only be encrypted with SSE-S3, so the output must be written elsewhere to satisfy the KMS requirement.

3. C.

   The requirement is satisfied automatically, because objects in directory buckets inherit the bucket key configuration of the general purpose bucket that the directory bucket is linked to.

4. D.

   Directory buckets support SSE-KMS with customer managed keys for both input and output, so the team only needs to grant the training role kms:Decrypt and kms:GenerateDataKey on the key.

1. A.

   Detecting fields containing PII scans only a sample of rows and labels a column as containing an entity only when the proportion of sampled rows carrying it meets the detection threshold, so a sparse entity in a wide column is missed; detecting PII in each cell scans all rows and is the setting for this case.

2. B.

   Free-text columns are excluded from column-level detection because their inferred type is not a recognised sensitive data type, so a fine-grained action override is needed to include the column in the scan.

3. C.

   The transform does not include a UK national insurance number pattern among its managed entities, so the column was scanned correctly but the entity type was never looked for; a custom detection entity with a regular expression is required.

4. D.

   The global detection sensitivity was left at its default of Low, which reduces false positives by detecting fewer entities; raising it to High makes the sampled scan detect the sparse entity.

1. A.

   Vector collections support IVF but only through the model training API, and read-after-write is immediate; the flaky test indicates a client-side retry problem rather than a service behaviour.

2. B.

   Both requirements are met by enabling the Lucene ANN engine on the collection, which supports IVF and provides synchronous refresh on write for indexes below the serverless size threshold. Enabling the engine is a collection-level setting that can be applied without recreating any of the existing indexes.

3. C.

   Classic vector collections support only HNSW with Faiss and do not support IVF, and their index refresh interval means a write is not immediately searchable, so the test must tolerate the refresh delay.

4. D.

   Both limitations are removed by scaling the collection's OpenSearch Compute Units above the default maximum, because refresh behaviour and available algorithms are functions of allocated capacity.

1. A.

   Convert both DynamicFrames to Spark DataFrames with toDF before the union, because DataFrames infer a common supertype for conflicting columns automatically and never fail on this.

2. B.

   Apply the relationalize transform to both frames, which resolves choice types by pivoting each conflicting field into its own child table joined back by a generated key. The two typed representations then appear as sibling rows in the child table, which the merge logic can reconcile without losing anything.

3. C.

   Apply resolveChoice with cast:long, which converts every value to a single numeric type and is therefore the safest option when the column is an identifier.

4. D.

   Apply resolveChoice with make_cols, which splits the ambiguous field into one column per type so no value is lost, then reconcile the two columns explicitly in the merge logic.

Choose 2.

1. A.

   Hold the hourly counters in an Amazon ElastiCache cache and read them from the authorisation path.

2. B.

   Load the transaction history into Amazon DocumentDB and run the nightly aggregation with its MongoDB-compatible aggregation pipeline.

3. C.

   Run the overnight join and windowed aggregation in an Amazon Redshift data warehouse, then unload the resulting training set to Amazon S3.

4. D.

   Model transactions, cards and merchants in Amazon Neptune and compute the nightly aggregates with openCypher queries.

5. E.

   Register the transaction files with AWS Lake Formation and serve both the nightly aggregation and the authorisation-time lookups through its permissions layer.

1. A.

   Fast file mode requires that the manifest be stored in the same prefix as the objects it references; moving the augmented manifest alongside the images resolves the configuration error.

2. B.

   Fast file mode supports Amazon S3 prefixes only and does not support manifest or augmented manifest files; the team must keep file mode for the augmented manifest, or restructure the data so labels are not carried in a manifest.

3. C.

   Augmented manifest files are supported only in pipe mode, and fast file mode replaced pipe mode entirely, so the team must first convert the manifest to RecordIO-protobuf and then use fast file mode. Fast file mode then streams the converted records from the prefix on demand without needing the manifest at all.

4. D.

   Augmented manifests require the ShardedByS3Key distribution type, which is incompatible with fast file mode's on-demand streaming; setting FullyReplicated allows fast file mode to accept the manifest.

1. A.

   Kullback-Leibler divergence, because it is the only pre-training metric defined for multicategory outcomes and it handles empty categories by assigning them a divergence of zero. Its range starts at zero and is unbounded above, which makes it directly comparable between numeric, text and image modalities without rescaling.

2. B.

   Difference in Proportions of Labels, because it is symmetric by construction and its normalised range makes it directly comparable between numeric, text and image modalities.

3. C.

   Jensen-Shannon divergence, because it is a symmetrised and smoothed measure of how much the outcome distributions of different facets diverge, which makes it better behaved than the raw asymmetric divergence when a category is empty for one facet.

4. D.

   Class Imbalance, because with an empty outcome category the divergence metrics are undefined and only a count-based measure of facet representation remains valid.

1. A.

   Re-apply min-max scaling to a wider target range so that the compressed values are spread across a larger interval, which restores the resolution lost in the unit interval. Because the transform is linear, widening the range preserves the ordering while giving the model far more numeric separation to work with.

2. B.

   Remove the scaling entirely, because tree-based and linear models are both invariant to monotone rescaling of a single numeric feature and the scaling step added no value.

3. C.

   Scale the column using statistics that are robust to outliers, such as a robust scaler based on quantiles, or clip the extreme values to a quantile bound before scaling, so the bulk of the distribution retains resolution.

4. D.

   Replace min-max scaling with max absolute scaling, which divides each value by the maximum absolute value and is therefore unaffected by the presence of extreme values.