Deployment and Orchestration of ML Workflows – MLA-C01 Study Guide
This domain accounts for 22% of the exam and covers getting a trained model into production and automating the pipeline around it. The two most-tested skills: choosing the right SageMaker inference option for a workload's latency/throughput/cost profile, and recognizing deployment strategies (blue/green, canary, shadow). Learn the four inference options and when each applies, and this domain is very winnable.
SageMaker Inference Options
The central decision — match the option to the workload:
| Option | Best for | Key trait |
|---|---|---|
| Real-time endpoint | Low-latency, per-request, always-on traffic | Persistent endpoint |
| Serverless inference | Intermittent, unpredictable traffic with idle periods | Auto-scales to zero; pay per use |
| Asynchronous inference | Large payloads and long processing times | Queue-based, non-blocking |
| Batch transform | Offline scoring of a whole dataset | No persistent endpoint; shuts down after |
Decision cues:
- "Sub-second response per user request" → real-time endpoint.
- "Sporadic traffic, don't pay when idle" → serverless.
- "Large payloads / minutes-long processing, don't block the client" → asynchronous.
- "Score a big dataset once, cheaply" → batch transform.
A common combined pattern: real-time endpoint for live requests plus batch transform for periodic bulk backlogs — match each need separately rather than forcing one option to do both.
Scaling and Cost
- Endpoint auto scaling — adjusts instance count based on load metrics (e.g., invocations per instance) to match capacity to demand and control cost.
- Right-size instances; don't run maximum GPU capacity 24/7 for variable load.
Unlock with a paid plan
4 more sections (about 276 words) continue below. Every paid plan unlocks all MLA-C01 study guides, the full question bank and mock exams, from $4.97/month.