Data Preparation for Machine Learning (ML)
Free AWS Certified Machine Learning Engineer - Associate practice — 6 questions on Data Preparation for Machine Learning (ML), with explanations. No sign-up. Full 12-question mixed test →
A company ingests clickstream data at a sustained 500,000 records/second, requiring a real-time fraud-detection feature pipeline with end-to-end processing latency under 70ms. Three independent downstream applications must each read the entire stream in parallel without impacting each other's throughput or latency. Which solution BEST meets these requirements?
Kinesis Data Streams with Enhanced Fan-Out gives each registered consumer its own dedicated 2MB/sec-per-shard throughput and delivers records in roughly 70ms, independent of other consumers reading the same stream — exactly matching the isolation and latency requirements.
An ML team needs feature retrieval under 10ms for real-time inference and also needs the full historical feature history with point-in-time correctness for model training. Which configuration of SageMaker Feature Store satisfies both requirements with the least operational overhead?
A single feature group with both stores enabled is the native Feature Store design: the online store (DynamoDB-backed) serves low-latency inference reads, while the offline store (S3/Parquet) retains full history for point-in-time correct training queries via Athena — no manual syncing required.
A training dataset has a facet 'age_group' with two values. The positive-outcome label rate is 65% for one group and 20% for the other, measured directly on the raw labeled dataset before any model is trained. Which SageMaker Clarify pre-training bias metric is specifically designed to detect this condition?
DPL directly quantifies the difference in the proportion of positive labels between facet groups on the raw training data, exactly matching a scenario describing unequal positive label rates (65% vs 20%) prior to training.
A team must label 2 million images on a constrained budget and wants to minimize the number of images requiring human review while maintaining labeling accuracy. Which SageMaker Ground Truth configuration BEST achieves this?
Ground Truth's automated data labeling feature trains an active-learning model on an initial human-labeled sample, then automatically labels new data above a confidence threshold, routing only uncertain examples to humans — directly minimizing human labeling volume while preserving accuracy.
A compliance team needs to continuously scan all objects across dozens of S3 buckets containing training data extracts to automatically discover sensitive data such as credit card numbers and SSNs at rest, with findings surfaced to Security Hub and EventBridge, without writing custom detection code. Which AWS service should be used?
Amazon Macie is purpose-built for automated, continuous, ML- and pattern-based sensitive data discovery across S3 at scale, generating findings that integrate natively with Security Hub and EventBridge without requiring custom orchestration code.
An engineer training an XGBoost model on 200GB of tabular data using File mode notices significant time spent downloading the full dataset before training starts. They want to switch to Pipe mode to stream data directly from S3 and reduce startup time and local disk usage. Which data format is BEST suited for efficient Pipe mode streaming?
RecordIO-protobuf is the natively recommended, binary-optimized format for SageMaker built-in algorithms using Pipe mode, minimizing parsing overhead and enabling efficient sequential streaming directly from S3 with reduced startup latency and disk footprint.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$109.99 $34.99 with code FREETEST33 — valid through September 7.
Get my $34.99 deal →