AWS Certified Machine Learning Engineer - Associate (MLA-C02) Practice Exam

A retailer stores transaction records in Amazon RDS for MySQL. An ML team needs an initial copy of the records in Amazon S3 and an ongoing feed of committed inserts, updates, and deletes for feature preparation. After the initial load, the feed must preserve successive changes to the same record, even when those changes occur between extraction runs. The application permanently deletes some records and cannot be modified to produce change events. Which solution meets these requirements with the least custom development?

A machine learning engineer runs a TensorFlow training job on an Amazon EC2 GPU instance using an AWS Deep Learning Container. Each epoch contains 200 batches, and the optimizer performs one update per batch. The engineer intends to halve the learning rate every 5 epochs during a 20-epoch run. The training script uses tf.keras.optimizers.schedules.ExponentialDecay(initial_learning_rate=0.001, decay_steps=5, decay_rate=0.5, staircase=False). Training loss improves during the first few batches but then stops improving. Logs show that the learning rate becomes nearly zero during the first epoch, while gradients remain finite and nonzero. Which change best addresses the convergence issue while implementing the intended decay interval?

A media company must run a trained video-classification model against 1,200 independent input shards in Amazon S3 each weekend. Each shard requires one NVIDIA GPU and approximately 75 minutes of processing. Results are needed the next day, not interactively. The existing container runs a command-line inference program that reads a shard and writes predictions to Amazon S3; it does not implement an HTTP inference server. The company cannot modify the container and wants AWS to provision and scale the compute capacity, with no running compute instances between workloads. Which solution meets these requirements?

A company is selecting an Amazon EC2 instance configuration for a synchronous image-classification API. Traffic alternates between sparse requests and sustained bursts. The API must return a complete response within 100 milliseconds for at least 99% of requests during both traffic patterns. An ML engineer tests four configurations using representative requests. Each configuration meets the required throughput and model accuracy. The client-observed latency includes network transport, preprocessing, batch waiting, model execution, and response delivery. The table shows the higher p99 latency observed across the sparse-traffic and burst-traffic tests. Prices are illustrative. | Configuration | Hourly cost per instance | p99 model execution latency | p99 client-observed latency | |---|---:|---:|---:| | Compute-optimized CPU, no batching | $0.60 | 85 ms | 125 ms | | Inference accelerator, batches of up to 32 requests | $0.85 | 18 ms | 145 ms | | Small GPU instance, no batching | $1.20 | 42 ms | 78 ms | | Large GPU instance, no batching | $2.40 | 24 ms | 58 ms | The company will use the same fixed instance count for every configuration. It must select one of the tested configurations without further serving changes. Which choice meets the response-time requirement at the lowest instance cost?