Near-Right Architecture — AWS Machine Learning (MLS-C01)

Whether to replace a custom SageMaker ASR training pipeline with Amazon Transcribe for a commodity speech-to-text workload given a hard 40% cost-per-inference reduction mandate — i.e., apply the build-vs-buy threshold governing constraint.

Choose cost-effective schema-flexible object storage with a governed access layer (S3 + Lake Formation) over an analytics warehouse (Redshift) that forces structured ingestion and incurs compute-bound cost at training-data volume.

When text feature engineering requires domain-specific NLP extraction AND inference-time reproducibility inside a SageMaker pipeline, prefer a managed NLP service (Comprehend) invoked from a SageMaker Processing step over a general serverless ETL service that requires custom NLP scripting and cannot natively participate in SageMaker pipeline inference.

Whether the interactive latency and non-technical audience requirements are best satisfied by a fully serverless query-plus-BI stack (Athena + QuickSight SPICE) or by a provisioned columnar warehouse (Redshift) that meets query performance but introduces cluster management overhead the team cannot absorb.

Whether to run SageMaker managed training on CPU-optimized instances or provision GPU-backed EC2 instances with DLAMI, when the algorithm is CPU-optimized and a cost ceiling — not raw throughput — is the dominant constraint.

Whether the observed high-variance symptom (validation loss diverging from training loss) unambiguously maps to a targeted regularization hyperparameter adjustment on the existing training job, or whether launching a SageMaker Automatic Model Tuning job to search the broader hyperparameter space is the correct response given the time and infrastructure constraints.

Whether to evaluate model quality using offline accuracy on a held-out S3 test set, or to deploy via SageMaker shadow testing and measure precision/recall against delayed ground-truth labels captured from live traffic — when both severe class imbalance and an online-impact measurement requirement are simultaneously present.

When the dominant constraint is real-world business impact measured on live production traffic with minimal deployment risk, online evaluation via SageMaker production variants is required — offline batch evaluation with imbalance-aware metrics is disqualified regardless of metric sophistication, because it cannot satisfy the offline-online-evaluation-parity constraint.

Whether the demand forecasting business problem should be framed as a managed forecasting task delegated entirely to Amazon Forecast, or as a custom time-series model task built on SageMaker, given that team ML expertise and production timeline — not algorithmic flexibility — are the binding constraints.

Whether to use GPU-class instances (p3 family) or CPU-optimized instances (c5/m5 family) for a SageMaker XGBoost training job when a per-run cost ceiling is the binding constraint.

Whether to route Glue-to-S3 data transfer through a VPC gateway endpoint—keeping traffic on the AWS private backbone at zero per-GB cost—versus routing through a NAT gateway, which is internet-routed and incurs per-GB NAT processing charges that compound at multi-terabyte scale.

Enforce the data-access boundary with an S3 bucket policy containing an explicit Deny on all principals except the SageMaker execution role ARN, because only identity-layer controls can restrict which IAM principals call S3 APIs regardless of network path; a VPC endpoint policy alone cannot satisfy the 'only this role' constraint.