After AWS Machine Learning Specialty: A Modern ML Roadmap

The definitive route to the AWS Certified Machine Learning – Specialty examination now begins with an important correction: there is no current route to that exam. AWS ended MLS-C01 testing on March 31, 2026. Its original study domains remain useful as a historical map of machine learning engineering, but a learner in October 2026 should not buy an MLS-C01 voucher or plan to renew that retired exam.

This roadmap preserves the educational value of the old subjects—data preparation, model development, evaluation, deployment, and operations—while pointing toward current learning. For someone focused on implementing and operating ML workloads, AWS Certified Machine Learning Engineer – Associate is a role-aligned option. Its updated MLA-C02 beta is being delivered in English, with general-availability delivery scheduled for January 14, 2027.

Stage 1: decide whether the engineering path fits

Distinguish ML engineering from general AI literacy. AI Practitioner is a Foundational credential for understanding use cases and responsible AI concepts. Machine Learning Engineer – Associate is for practitioners who build, deploy, and operate models and supporting pipelines. A data engineer may prefer DEA-C01 if their work centers on ingestion, transformation, and governed storage rather than serving predictions.

Before committing to an exam, read the current AWS guide and compare each domain with your experience. List which tasks you have performed independently, which you have only seen in demonstrations, and which you cannot explain yet. That assessment is more useful than a schedule based solely on the number of chapters in an older Specialty book.

Use a single end-to-end project throughout this roadmap. An example might predict a service demand metric from non-sensitive historical events or classify support tickets. Keep the dataset modest enough that you can repeat each experiment and understand every failure without relying on a large cloud budget.

Stage 2: establish trustworthy data and features

Define what a correct input record looks like, how missing values are handled, how labels are generated, and why a given feature would be available at inference time. These decisions affect the validity of every later metric. A timestamp that leaks future information can make a model appear excellent in testing and fail in production.

Practise ingesting data into a governed storage location, validating its schema, and creating repeatable train/validation/test splits. Record the transformation version alongside the model version. If a pipeline fails halfway through, it should be possible to rerun without silently duplicating records or changing what a dataset means.

On AWS, S3, data-processing services, and SageMaker AI can support these steps. Select the tools that fit the workload rather than attempting to include every service in one demonstration. The proof of readiness is that you can show how the data was produced and whether it can be reproduced.

Stage 3: train, compare, and evaluate models

Start with a measurable baseline. For classification, use a simple model and report precision, recall, confusion-matrix behavior, and the consequences of false positives or missed cases. For regression, inspect error distribution and whether rare outliers affect decisions. Never choose a metric simply because it makes the result look attractive.

Test one improvement at a time: a different feature set, preprocessing method, or model family. Log experiment parameters, training data, artifacts, and evaluation outcomes. A candidate that performs better on one test sample may still be worse across time periods, underrepresented groups, or changing operating conditions.

Hyperparameter optimization is useful when it addresses a known limitation, but the final holdout should remain an independent check rather than a target for repeated tuning. Explain the trade-off between accuracy, inference latency, interpretability, and deployment cost.

Stage 4: build production deployment and monitoring

Choose the inference pattern by the application. Batch predictions, online inference, and asynchronous processing place different demands on latency, throughput, storage, and error handling. Define an availability or delivery target and design the system to handle a missing input, unavailable endpoint, or delayed upstream feed.

Implement model artifact versioning, scoped identity, protected secrets, and a way to promote and roll back a model. Record the code and configuration version that produced a prediction. An endpoint returning HTTP 200 is not enough: monitor data drift, inference latency, failure rates, and a business-relevant quality signal when it can be measured responsibly.

Run a recovery drill. Deploy a deliberately underperforming model to a controlled environment, detect a defined quality problem, revert to the prior candidate, and verify that requests are served correctly afterward. The exercise should demonstrate ownership and operational discipline rather than only a deployment command.

Stage 5: add modern foundation-model and agent workloads

AWS’s MLA-C02 update expands the application context beyond the historical Specialty. It covers implementation of generative AI, model selection, foundation models, retrieval-augmented generation, agentic workflows, and responsible AI. The established domains remain but task statements have changed; do not assume an older MLS-C01 guide covers those additions.

Build a small retrieval system using approved non-sensitive documents. Evaluate whether answers can be grounded in the retrieved text, whether citations refer to relevant passages, and how the system behaves when no trustworthy answer is available. Test prompt-injection attempts in a safe environment and ensure retrieved content cannot silently grant additional tool permissions.

If you experiment with agents, choose read-only tools first and require explicit authorization before actions with external effects. Document authentication boundaries, logging, retry behavior, and how an incomplete workflow can be rolled back. Measure safety and task completion separately: an agent that finishes a task by taking an unauthorized shortcut is not successful.

Plan a sustainable study cycle

A practical cycle is to review one objective group, complete a focused lab, explain a scenario without notes, and compare your reasoning with the current first-party guide. Keep a short error ledger that separates gaps in vocabulary, service behavior, architecture judgment, and operational troubleshooting. These need different remedies.

Prioritize lab evidence over a fixed calendar. A learner with production MLOps experience may progress through data-preparation material quickly but need more work on generative AI evaluation. Someone with data science expertise may know modeling well but need sustained practice with access controls, deployment, and incident response.

Use official exam-style practice resources when available and inspect the explanations for every distractor. A practice percentage does not guarantee an AWS scaled score. Do not use copied or confidential real exam items, and verify that any commercial preparation source is aligned to the current exam version.

Understand the 2026–2027 exam transition

MLA-C01 English delivery ended September 28, 2026. The updated MLA-C02 beta began delivery September 29, 2026, in English, with broader language availability and the regular version scheduled for January 14, 2027. AWS has continued to offer MLA-C01 in certain other languages during the transition, subject to its published dates. These details matter when choosing study resources and a testing language.

Do not confuse MLA-C02 with the retired Machine Learning – Specialty. An existing Specialty credential remains a valid historical achievement until its individual expiry, but holding it does not automatically give someone the newer Engineer – Associate credential. Check the issuer’s registration page immediately before selecting a date because version availability can change.

What readiness should look like

At the end of this roadmap, be able to tell a complete story: the data problem, the chosen baseline, the experiment record, the deployment architecture, permission controls, evaluation methodology, observed failure, and recovery. Show a reproducible run and a concise description of what you would improve before production use.

The legacy MLS-C01 curriculum can still teach valuable fundamentals, but it is not a bookable destination. Anchor new preparation to a current role, an active or scheduled issuer exam guide, and hands-on evidence of the systems you can actually operate.