Deployment and Orchestration of ML and AI Workflows

MLA-C02 · 48 questions

• A city permitting-chatbot team must serve sub-second interactive responses to residents typing permit questions throughout business hours, with continuous, predictable traffic. Which inference option best fits this requirement?
• A county 311 call-center intent-classification model receives sporadic, unpredictable bursts of traffic with long idle gaps between citizen contacts. The team wants to avoid paying for capacity that sits idle. Which inference option fits best?
• A public-works department must run inference on scanned permit application PDFs that are individually large and can take several minutes to process, exceeding the payload size and timeout limits of a real-time endpoint. Which inference option should the team use?
• A water utility wants to score its entire customer base overnight for usage-anomaly detection, with no requirement for a persistent endpoint or an immediate response. Which inference approach fits this need?
• A county's data-science team fine-tuned an open-weights foundation model outside AWS and needs a way to bring those model weights into Bedrock for managed inference, without retraining inside SageMaker. What should they use?
• A school district trains a small eligibility-scoring model per individual school — dozens of near-identical models — and wants to avoid provisioning a separate costly endpoint for each one. What deployment pattern fits best?
• A transit authority's ridership-prediction workload requires a preprocessing container to run ahead of the model-inference container on every request, with both steps served from one endpoint invocation. Which deployment pattern fits this requirement?
• A court clerk's office trained a case-category classification model on a data scientist's laptop entirely outside any AWS service, and now needs it hosted in production for clerks to use. What is the right way to get it into service?
• A 311 resident-services assistant must call several municipal backend APIs — pothole reporting and utility billing lookup among them — as part of answering a citizen's question. What should the team configure to support this?
• A city is deploying a permitting agent and a separate records agent that must hand off work to each other mid-conversation as a citizen's request moves between the two domains. What does the team need to configure to support this?
• A public-health department needs high-volume, steady translation of health-education materials via a Bedrock foundation model, running continuously at a predictable rate. Which throughput option best fits this workload?
• A municipal-code Q&A assistant's answers are missing relevant regulation text because its retrieval step only performs keyword search. What retrieval strategy should the team adopt to also capture semantically related matches?
• A permitting-regulations assistant retrieves too many marginally relevant chunks from its knowledge base, diluting answer quality. The team wants to improve relevance ordering without changing the underlying embedding model. What should they add?
• An emergency-management office needs to classify incoming alert reports. Traffic is near-zero on most days but spikes sharply during storm events. Which inference option handles this pattern most cost-effectively?
• An emergency-dispatch office is deciding whether to deploy a domain-fine-tuned foundation model for triage classification or invoke a Bedrock base foundation model directly, weighing accuracy against deployment complexity. How should the team approach this decision?
• A housing-authority eligibility system requires a feature-transformation step and a scoring step to run in strict sequence for every request. What deployment architecture should the team use to chain these two steps reliably behind one prediction call?
• A water-quality inference workload runs at a steady, predictable rate around the clock, and the finance office wants the lowest-cost compute-purchasing approach for that steady baseline versus occasional demand spikes. What should the team do?
• A city IT department needs every department's SageMaker endpoint stack — permitting, 311, transit — deployed identically and reproducibly across dev, test, and prod accounts, without manual console configuration. What should the team adopt?
• A transit authority built a custom inference container for predictive-maintenance scoring and needs to run it at scale outside SageMaker, as a long-lived, always-on ML workload. What kind of service should the team use to run it?
• A public-health department's risk-scoring endpoint processes sensitive patient-adjacent records and must be reachable only from the department's private network, with no path to or from the public internet. How should the team configure the endpoint?
• A county's MLOps engineer wants a model's endpoint automatically redeployed the moment a new version is approved in the model registry, without any manual console step, driven by code invoked from a pipeline. How should the team implement this?
• A permitting office's chatbot runs on a real-time SageMaker endpoint, and the platform team must pick a CloudWatch metric to drive an Auto Scaling target-tracking policy for the endpoint's instance count. Which metric best reflects the actual request load each instance is handling?
• A county already runs its citizen-services data in an Amazon RDS for PostgreSQL database and is adding a Bedrock knowledge base for a resident-facing RAG assistant, with a document corpus in the low tens of thousands of pages. The team is weighing a dedicated managed search service against extending their existing relational database. Which choice best fits this situation?
• A courts department's Bedrock knowledge base keeps returning outdated guidance because newly filed procedural documents aren't reflected in retrieval results. Before launch, the team is designing the chunking and embedding configuration for the ingestion pipeline. Which approach best supports timely, accurate coverage of new documents?
• Residents chatting with a multi-turn 311 assistant find that the agent forgets what they said just a few messages earlier, forcing them to repeat themselves within the same conversation. The team traces this to how the agent handles context between invocations. Which fix directly addresses the problem?
• After a hurricane, an emergency-management office needs to run image-based structural-damage assessment on thousands of drone photographs and must provision compute that handles this workload efficiently. Which choice best matches the job?
• A housing authority is preparing to launch a caseworker-assistant agent that looks up case records and needs somewhere to hold conversation state across a caseworker's multi-step session. Before the agent can go live, what supporting infrastructure must the team provision alongside the agent definition itself?
• A transit authority trains its ridership-demand model on GPU instances but only needs to serve predictions afterward. The platform team is picking compute for each phase separately. Which approach best matches cost to workload?
• An emergency chatbot deployed on a serverless SageMaker endpoint shows noticeable delay on the first request after periods of low traffic, and residents report the assistant feels unresponsive at the start of a conversation. Which endpoint configuration change best addresses this cold-start delay?
• A housing authority's case-management model endpoint must only be reachable from the internal subnets caseworkers use, not from the broader network. Which combination of controls best restricts access at the network level?
• A permitting office's asynchronous document-inference endpoint accumulates a growing backlog of queued requests during peak filing season, and the current Auto Scaling policy reacts to a per-instance invocation-rate metric borrowed from a real-time endpoint. Which metric should the scaling policy target instead for this asynchronous workload?
• A school district's multi-model endpoint hosts dozens of per-school models, and the team notices frequent model swapping and latency spikes because the endpoint is undersized. How should the team right-size the endpoint's instance type?
• A benefits-eligibility fraud-detection model needs to move automatically through dev, test, and production stages as code changes are merged, with each stage gated by an approval or automated test step. Which pipeline design best supports this?
• A permitting-classification endpoint is being updated to a new model version, and the rollout must shift traffic gradually while automatically reverting if the new version's error rate rises. Which deployment approach fits this requirement?
• A 311 assistant's new endpoint version, deployed as a canary, starts showing elevated latency shortly after rollout. The team wants the rollback to happen automatically the moment the problem is detected, rather than depending on someone noticing and intervening manually. Which mechanism achieves this?
• A public-health risk model must not reach production until a designated reviewer explicitly approves the newest trained version. The team wants this approval to be the actual gate that controls deployment, not just a policy people are expected to follow. Which mechanism enforces this?
• A transit authority retrains its ridership-demand model periodically and needs to track and compare metrics across the retrained candidates so the team can decide which one to promote. Which practice best supports this comparison?
• A water utility's anomaly-detection model needs to retrain automatically once monitoring detects that incoming usage data has drifted meaningfully from the training distribution, without anyone manually starting the job. Which design achieves this?
• A county's ML pipeline occasionally fails in production because malformed input data reaches the training job partway through a run, wasting compute time before the failure surfaces. The team wants automated checks that catch schema problems before a training job ever launches. Which pipeline design addresses this?
• A permitting chatbot's prompt template needs to be versioned and promoted from a test environment to production separately from the underlying foundation model, so prompt edits can be reviewed and rolled back on their own. Which approach fits this need?
• A 311 Bedrock agent needs a new version promoted to production behind an alias, while the previous version stays available so the team can roll back if needed, as part of an automated agent-release pipeline. Which mechanism supports this?
• Before promoting a revised prompt template for the permitting chatbot, the team wants an automated test suite that compares the new prompt's outputs against a set of known-good reference answers, catching regressions before the change reaches residents. Which practice best supports this?
• A county fine-tunes a foundation model periodically on updated case-notes data and wants each new fine-tuned version automatically brought into Bedrock with proper version tracking as part of the deployment pipeline, rather than handled as a one-off manual task each time. Which approach fits?
• A municipal-code knowledge base must be reindexed on a recurring schedule whenever new ordinances are published, coordinating document ingestion, chunking, embedding, and vector-store update as one automated workflow. Which approach best orchestrates this multi-step process?
• A school district wants its model-training pipeline to trigger automatically whenever a data scientist pushes updated training code to the district's source repository, with no manual pipeline start required. Which pipeline design achieves this?
• A county's data-science team wants the same training job definition reused across multiple model variants by parameterizing inputs like hyperparameters and dataset location, instead of hand-editing a separate script for each variant. Which approach fits this goal?
• Before fully cutting over an emergency-dispatch triage agent to a new version, an MLOps team wants the new version to process live dispatch traffic and have its outputs compared against the current version's outputs, without the new version's responses ever reaching dispatchers, until confidence is established. Which deployment pattern in the CI/CD pipeline best achieves this?
• A unified citizen-services chatbot release bundles both an updated traditional eligibility-scoring model and a revised foundation-model prompt template. The release must ship both artifacts together in one coherent rollout so they stay in version sync. How should the CI/CD pipeline handle this?