Deployment and Orchestration of ML Workflows

A media company deploys a document-processing model that accepts PDF uploads up to 500 MB and can take several minutes per request to run OCR and summarization. Requests arrive in unpredictable bursts, and during idle periods the team wants to avoid paying for running instances. Clients do not need an immediate synchronous response and can poll for results. Which SageMaker deployment option best fits these requirements?

A retail company retrains a churn-prediction model weekly. Every night, they need to generate churn scores for their entire customer base of 40 million records stored as Parquet files in S3. The scoring job must complete before business hours, but there is no requirement for individual real-time predictions during the day. The ML engineer wants the most cost-effective deployment approach that avoids paying for idle compute. Which SageMaker deployment option should the engineer choose?

A machine learning team runs a production SageMaker real-time endpoint serving a recommendation model. They need to deploy a new model version but are concerned that a faulty version could cause a spike in inference errors affecting customers. They want SageMaker to automatically shift traffic to the new version gradually, monitor CloudWatch metrics during the shift, and automatically roll back to the previous version if error thresholds are breached. Which approach should they configure?

An ML engineering team deploys SageMaker real-time endpoints across development, staging, and production accounts. Currently each environment is provisioned manually through the console, causing configuration drift and inconsistent instance types between accounts. The team wants a programmatic, version-controlled way to define the endpoint infrastructure using a familiar general-purpose programming language (TypeScript), reuse the same construct across all three accounts, and integrate the deployment into their existing CI/CD pipeline. Which approach best meets these requirements?

A machine learning platform team wants to define their SageMaker infrastructure (endpoints, endpoint configs, and IAM roles) as code. The team is composed of Python developers who want to use loops, conditionals, and reusable classes to generate the infrastructure definitions, while still deploying through AWS's native provisioning and rollback engine. Which approach best meets these requirements?

A team manages a SageMaker real-time endpoint entirely through a CloudFormation stack that is deployed identically across dev, staging, and production. During an incident, an on-call engineer manually changed the production endpoint's instance count and instance type through the SageMaker console to handle a traffic surge. Later, the team wants to programmatically detect whether the running production resources still match what the CloudFormation template declares, before running their next stack update. Which approach lets them identify these out-of-band changes with the least custom effort?

A team hosts 300 customer-specific fraud models behind a single SageMaker multi-model endpoint (MME) to save cost. Most models are invoked a few times per hour, but three high-traffic customers generate the majority of requests. Operators notice that these three heavily-used models occasionally see elevated p99 latency spikes that coincide with invocations of rarely-used models. What is the most likely cause and the best remedy?

A retail company has trained 400 separate XGBoost models, one per product category, to predict demand. Each model is small (under 50 MB) and receives sporadic, low-volume traffic throughout the day. The ML engineering team wants to serve all models for real-time inference while minimizing the number of endpoints and controlling infrastructure cost. Which SageMaker deployment approach best meets these requirements?

A media company has 40 genre-specific recommendation models, each trained on a different content category. Traffic to individual models is unpredictable and often sparse, but occasionally one genre spikes during a promotion. The team wants to minimize the number of endpoints they manage and reduce idle infrastructure cost, while still being able to invoke any single model on demand by specifying its identifier at request time. Which SageMaker deployment approach best fits these requirements?

A team deploys a lightweight fraud-scoring model on a SageMaker Serverless Inference endpoint. Traffic is intermittent, but a partner integration has a strict requirement: the first request after an idle period must respond within 300 ms. During testing, the team observes that cold starts add 2-4 seconds of latency to the first invocation. They want to keep the serverless model but eliminate the cold-start latency for the initial requests. What should they configure?

A logistics company runs a fraud-detection model that must score every incoming transaction within 100 milliseconds, 24/7, with steady and predictable traffic of roughly 500 requests per second. The ML engineering team needs to choose the SageMaker deployment option that meets the latency requirement while keeping the infrastructure fully managed. Which deployment approach best fits these requirements?

A retail company deploys a product-recommendation model to a SageMaker real-time endpoint. Traffic is steady during business hours but spikes 5x during flash sales that occur unpredictably. The team wants the endpoint to add instances automatically when load rises and remove them when traffic subsides, while keeping per-instance concurrency near a healthy level. What is the MOST appropriate way to configure this?

A machine learning team deploys a real-time SageMaker endpoint hosting a deep-learning image classification model built on a large convolutional neural network. During load testing on ml.c5.4xlarge CPU instances, per-request latency exceeds the 200 ms SLA, and CPU utilization is pinned at 100% even at moderate request volumes. The model uses matrix-heavy inference operations that parallelize well. What change should the team make to meet the latency requirement most effectively?

A retail company runs a real-time SageMaker endpoint that serves product recommendations during business hours. The endpoint is currently configured with a single ml.c5.xlarge instance in one production variant. During a recent Availability Zone disruption, the endpoint became completely unavailable, causing a customer-facing outage. The ML engineering team must improve availability with minimal configuration changes and without over-provisioning. What is the most effective change to make?

A machine learning team has a real-time SageMaker endpoint serving a fraud-detection model. They have trained a new model version and want to route only 10% of live traffic to it while keeping 90% on the existing model, so they can monitor CloudWatch metrics before fully committing. They want to accomplish this within a single endpoint without provisioning separate endpoints. Which approach meets these requirements?

A healthcare company deploys a SageMaker real-time endpoint that serves predictions to an internal application running in a private VPC. Compliance policy prohibits any inference traffic from traversing the public internet. The ML engineer must configure the endpoint so that the application can invoke it entirely over the AWS private network. What is the most appropriate approach?

A machine learning team has trained a new image classification model that will be deployed to a real-time SageMaker endpoint. They must handle a sustained throughput of about 200 requests per second while keeping p99 latency under 100 ms and minimizing cost. The team is unsure which instance type and count to provision and wants a data-driven recommendation before going to production. Which approach best meets these requirements with the least manual effort?

A machine learning team runs long deep-learning training jobs (6-10 hours each) on SageMaker using ml.p3.2xlarge instances. Finance has asked them to reduce training costs significantly, and the jobs are not time-critical as long as they eventually complete. The team is concerned about losing progress if the underlying capacity is reclaimed. Which approach BEST meets these requirements?

A machine learning team has a fraud detection workflow that requires three sequential processing steps on every real-time request: a scikit-learn preprocessing container that normalizes and encodes features, an XGBoost model container that produces a score, and a lightweight post-processing container that applies business thresholds and formats the output. All three must run on the same request with minimal added network latency, and the team wants to deploy them behind a single HTTPS endpoint without managing separate services or custom orchestration code. Which SageMaker deployment approach best satisfies these requirements?

An ML engineering team runs a SageMaker Pipeline daily that includes a data-processing step, a training step, and an evaluation step. The raw input data only changes about once a week, but the training code and hyperparameters change frequently as data scientists experiment. Each full pipeline run is expensive because the processing step reprocesses the same unchanged data every day. The team wants to reduce cost and run time by avoiding reprocessing when neither the processing inputs nor the step configuration have changed, while still guaranteeing that any change to code or parameters triggers a fresh run. What is the most appropriate approach?