Deployment and Orchestration of ML Workflows
Free AWS Certified Machine Learning Engineer - Associate practice — 6 questions on Deployment and Orchestration of ML Workflows, with explanations. No sign-up. Full 12-question mixed test →
An ML team serves a fraud-detection model where individual inference requests carry payloads up to 500 MB (batches of transaction images) and can take up to 20 minutes to process. Traffic arrives in unpredictable bursts with long idle periods overnight, and the team wants to avoid paying for idle compute during those periods. Which deployment option BEST satisfies these requirements?
Asynchronous inference supports payloads up to 1 GB and processing times up to one hour, queues requests in S3, and its auto scaling can be configured with MinCapacity=0 so no instances run (and no cost is incurred) during idle periods, resuming automatically when a request arrives.
A company trains a separate XGBoost model per customer (currently 400 customers, growing weekly). All requests use the same container image and framework, but each request must be routed to that specific customer's model artifact, and idle models should not consume dedicated compute. Which hosting pattern minimizes cost and operational overhead?
Multi-model endpoints are purpose-built for hosting many models sharing the same framework/container behind a single endpoint; artifacts are stored in S3, loaded on demand, and cached, with the specific model selected per request via the TargetModel parameter -- this scales to thousands of models cost-effectively.
A data science lead must manually review and approve each new model version before it can be promoted to production. The team wants deployment to CodePipeline to trigger automatically and ONLY when that manual approval occurs in SageMaker Model Registry, with no other trigger causing a production deployment. Which configuration correctly implements this?
SageMaker emits a native "SageMaker Model Package State Change" event to EventBridge whenever ModelApprovalStatus changes; filtering that rule specifically on Approved and targeting CodePipeline ensures the pipeline runs only in response to the intended manual approval action, with no polling or unrelated triggers.
A team has compiled a transformer model with the AWS Neuron SDK and needs to host it for real-time inference, prioritizing the lowest cost per inference while still meeting a strict low-latency SLA. Which instance family should they select for the SageMaker endpoint?
Inferentia2-based ml.inf2 instances are purpose-built accelerators for inference and are used specifically with models compiled via the Neuron SDK, delivering the lowest cost-per-inference for supported deep learning models such as transformers while meeting low-latency requirements.
A SageMaker real-time endpoint hosts a GPU-backed model. The team observes that GPU utilization spikes and saturates well before the SageMakerVariantInvocationsPerInstance metric reaches levels that would trigger the existing target-tracking auto scaling policy, causing latency SLA breaches during bursts. What is the MOST effective fix?
When the default invocations-per-instance metric doesn't correlate well with the actual resource bottleneck (GPU saturation), SageMaker auto scaling supports target tracking on a custom CloudWatch metric, allowing scaling decisions to be driven by the metric that actually reflects load, resolving the mismatch.
A team wants to safely roll out a new model version to a production SageMaker real-time endpoint by shifting live traffic gradually in fixed percentage increments over time (e.g., 10% every 5 minutes) and automatically rolling back if a CloudWatch alarm on error rate fires, all without introducing an external deployment service. Which approach correctly achieves this using native SageMaker capabilities?
SageMaker endpoints support native deployment guardrails via UpdateEndpoint's DeploymentConfig: BlueGreenUpdatePolicy with TrafficRoutingConfiguration Type=LINEAR shifts traffic in equal fixed percentage steps at a configured interval, and AutoRollbackConfiguration tied to CloudWatch alarms triggers an automatic rollback -- exactly matching the requirement without any external deployment tooling.
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 →