Traps, tiebreakers, and 'most cost-effective'
Lesson 3 of 4 in Exam Readiness: Reading Questions Like an Engineer.
Take one scenario — "train a large model every week" — and swing the qualifier: under MOST cost-effective, the key is managed Spot training with checkpointing; under FASTEST training time, it becomes distributed training across more powerful instances; under LEAST operational overhead, it drifts toward built-in algorithms with AMT — or out of training entirely, to a Bedrock model that needs no training. Same scenario, three different right answers. That is not a flaw; it is the exam. The qualifier is the question.
| Qualifier | What it really asks | Typical winners | Typical losers |
|---|---|---|---|
MOST cost-effective | Cheapest option that still meets every stated requirement | Spot training + checkpoints, batch transform, serverless and on-demand pricing, lifecycle rules to colder storage | Always-on endpoints for occasional traffic, provisioned clusters, fleets sized for peak |
LEAST operational overhead | Fewest things you patch, scale, and babysit | Serverless and managed: Bedrock, Athena, Glue, Firehose, Serverless Inference, Step Functions | Self-operated EC2/EMR/EKS, custom containers, self-hosted anything |
HIGHEST performance / LOWEST latency | Speed, even at a price | Real-time endpoints on larger or GPU instances, FSx for Lustre, distributed training, provisioned capacity | Batch and queued modes, archival storage classes, cold-start-prone options |
Interactive sorting exercise: Sort each stem phrase onto the property that wins its tiebreak.
Trap 1 — right service, wrong mode
A real-time endpoint for nightly bulk scoring (want batch transform); Kinesis Data Streams when the stem says "no consumer code" (want Firehose); a provisioned Redshift cluster for occasional queries (want Athena or serverless); fine-tuning when the need is grounding in documents (want RAG). The service family is right — the operating mode loses on the requirement.
Trap 2 — overkill architecture
An EMR cluster to convert one CSV (Glue or DataBrew does it serverlessly); EKS to serve one small model (an endpoint suffices); building a moderation model from scratch (Guardrails is configuration); custom OCR training (Textract is pre-trained). If a simpler option meets every stated requirement, the grander one is a distractor.
Trap 3 — underkill that misses a requirement
Lambda for GPU inference (no GPUs, 15-minute ceiling); Glacier tiers for data the pipeline trains on tomorrow (retrieval delays); single-AZ designs when the stem says highly available. Cheap and simple only wins when it still satisfies every constraint in the stem.
Trap 4 — the C01 ghost
Topics removed in the C02 revision still haunt old practice materials: SageMaker Neo edge compilation, model pruning/compression techniques, deep bring-your-own-container gymnastics, and the ordering/matching/case-study question formats. If your practice source tests these, it predates MLA-C02 — recalibrate against the current exam guide.
Interactive checkpoint quiz (3 questions) — open this page in a browser to take it.