AWS retired the MLS-C01 exam on March 31, 2026
The AWS Certified Machine Learning – Specialty (MLS-C01) is no longer available to take — March 31, 2026 was the last testing day. If you already hold it, your certification stays valid for three years from the date you earned it. If you’re looking to certify your AWS ML skills now, AWS points candidates to the successor associate-level exam, MLA-C01.
Go to the MLA-C01 Machine Learning Engineer Associate exam →AWS Machine Learning Specialty (MLS-C01) — Retired
MLS-C01 retired on March 31, 2026. This page stays up as a reference for existing certificate holders and for anyone learning AWS ML — the SageMaker practice questions below still teach current, exam-relevant concepts. To certify today, move to the successor exam, MLA-C01.
See the successor: MLA-C01 → Existing subscribers: open the MLS-C01 question bankMLS-C01 exam at a glance
- Vendor
- Amazon Web Services (AWS)
- Exam code
- MLS-C01
- Certification
- AWS Certified Machine Learning – Specialty
- Status
- Retired — last testing day March 31, 2026
- Successor
- MLA-C01 (AWS Certified Machine Learning Engineer – Associate)
- Level
- Specialty
- Format (while active)
- 65 questions (50 scored + 15 unscored); multiple choice and multiple response
- Duration (while active)
- 180 minutes
- Scoring
- Scaled 100–1000; passing score 750
- Cost (while active)
- $300 USD
- Validity of earned cert
- 3 years from the date earned
Sources: AWS — Machine Learning Specialty certification page (retirement notice). Facts marked “while active” describe the exam as it stood before retirement.
About the MLS-C01 certification (and what replaced it)
The AWS Certified Machine Learning – Specialty (MLS-C01) validated the ability to design, build, train, tune, deploy, and maintain machine-learning solutions on AWS. It was aimed at data scientists and ML practitioners with two or more years of hands-on ML or deep-learning experience on the AWS Cloud, and it was known as one of the harder AWS exams — roughly half the questions tested ML fundamentals (feature engineering, algorithm selection, evaluation metrics) rather than AWS service trivia.
AWS retired MLS-C01 on March 31, 2026. The retirement reflects how ML work has shifted: instead of training custom models from scratch, most teams now build on managed services and foundation models, so AWS refocused its ML credentialing on operating ML in production. The direct successor is the associate-level MLA-C01, AWS Certified Machine Learning Engineer – Associate, which emphasizes building, deploying, and operationalizing ML with SageMaker. If you already earned MLS-C01, it remains valid for three years from your earn date; if you’re certifying now, start with MLA-C01.
The practice questions below remain a useful way to learn AWS ML — the underlying SageMaker concepts (built-in algorithms, tuning, deployment options, monitoring) carry directly into MLA-C01 and real work. Every explanation links to the exact AWS documentation page it derives from. For the broader picture, see our AI & machine learning guide.
MLS-C01 exam domains and weights
These are AWS’s published domain weights for MLS-C01, from the official exam guide. They’re preserved here for reference and because the same topics carry into the successor exam. Modeling was the single heaviest area.
Algorithm selection across regression, classification, clustering, recommendation, and deep learning; hyperparameter tuning with SageMaker Automatic Model Tuning; evaluation metrics including precision, recall, F1, AUC-ROC, and RMSE.
Feature engineering, handling missing values and class imbalance, and visualization with SageMaker Studio notebooks and Amazon QuickSight.
Training-data pipelines on Amazon S3, AWS Glue, Amazon Kinesis Data Streams, and Amazon Data Firehose; fine-grained access with AWS Lake Formation.
Deploy real-time, batch transform, and asynchronous SageMaker endpoints; orchestrate retraining with AWS Lambda and Amazon EventBridge; monitor models in production.
Source: AWS Certified Machine Learning – Specialty (exam guide and certification page). Weights total 100%.
Who this material is for now
With the exam retired, this page serves two audiences:
- Existing MLS-C01 holders keeping the material current, or preparing to recertify by moving to the successor track.
- Learners and career-changers studying AWS ML fundamentals — SageMaker algorithms, tuning, deployment, and monitoring — much of which transfers directly to MLA-C01 and to day-to-day ML engineering.
- Data scientists and ML engineers who want scenario practice on SageMaker patterns regardless of a specific exam.
If your goal is a current AWS credential, the associate-level MLA-C01 is the right target, with the AIF-C01 AI Practitioner as a lighter on-ramp and DEA-C01 Data Engineer Associate covering the data-pipeline side. For the roles this skill set supports, with salary ranges and progression, see the AI engineer career path in our Career Hub.
What this practice material delivers
Learn mode
Get the correct answer, the reasoning, and why each other option is wrong — immediately after each question. Ideal for the algorithm-selection and metric-choice patterns that carry into MLA-C01.
Exam mode
A timed run that mirrors the MLS-C01 format — still useful pacing practice for anyone tackling AWS ML scenario questions.
Source-linked explanations
Every answer links to the exact AWS documentation page — so you learn SageMaker from AWS’s own docs, not just a memorized letter.
Score by domain
Results break down by the four MLS-C01 domains, so you can see which SageMaker areas to shore up before moving to the successor exam.
Sample AWS ML practice questions (still current concepts)
Ten free scenario questions across the four MLS-C01 domains. The exam retired, but these SageMaker concepts remain accurate and carry into MLA-C01 and real ML work. Each has a full explanation and a source link to the AWS documentation it derives from.
A team needs to ingest 200 GB/hour of clickstream data into S3 in Parquet format, partitioned by event date, with minimal code. Which service handles this most simply?
- AWS Snowball Edge
- EC2 fleet running custom Python scripts
- Amazon Data Firehose with Parquet conversion and dynamic partitioning
- Manual S3 multipart uploads from on-prem
Show answer & explanation
Correct: C. Amazon Data Firehose natively converts JSON to Parquet using a Glue schema and supports dynamic partitioning by event attributes — fully managed at hundreds of GB/hour with minimal code.
Why not the others: custom EC2 scripts (B) are high-operations; Snowball Edge (A) is for offline bulk transfer, not streaming; manual multipart uploads (D) aren’t a streaming pipeline.
Source: AWS — Firehose record format conversion → Further reading: PowerKram — AI & machine learning →An analyst sees a regression target with strong right-skew and a long tail of high values. Which transformation typically improves a linear model on this target?
- Min-max scaling only
- One-hot encoding the target
- Log transformation of the target
- Removing all rows above the median
Show answer & explanation
Correct: C. A log transform compresses the long right tail toward normality, which better satisfies linear-model assumptions and usually improves fit.
Why not the others: min-max scaling (A) rescales but doesn’t change skew; one-hot encoding (B) is for categorical inputs, not a continuous target; truncating at the median (D) throws away signal and biases the model.
Source: AWS — SageMaker data preparation →A binary classifier on imbalanced fraud data shows 99% accuracy but catches almost no fraud. Which evaluation metric should the team optimize instead?
- Accuracy
- R-squared
- Mean squared error
- F1 score or AUPRC (precision-recall AUC)
Show answer & explanation
Correct: D. On imbalanced classification, F1 and AUPRC reflect minority-class performance far better than accuracy, which is dominated by the majority class.
Why not the others: accuracy (A) is exactly what’s misleading here; R-squared (B) and MSE (C) are regression metrics, not classification metrics.
Source: AWS — SageMaker tuning metric considerations → Further reading: PowerKram — AI & machine learning →An engineer wants the SageMaker built-in algorithm best suited for time-series forecasting across many related items (for example, per-store, per-SKU sales). Which algorithm is the right choice?
- Linear Learner
- Random Cut Forest
- K-Means
- DeepAR
Show answer & explanation
Correct: D — DeepAR. DeepAR trains a single global RNN across many related time series and outperforms per-series classical methods when many series are available.
Why not the others: Linear Learner (A) is for regression/classification; K-Means (C) clusters; Random Cut Forest (B) is for anomaly detection, not forecasting.
Source: AWS — SageMaker DeepAR algorithm →A model in production must respond in under 100 ms to single requests, and traffic is steady. Which deployment option fits?
- SageMaker Real-time Inference endpoint
- SageMaker Batch Transform
- SageMaker Asynchronous Inference
- Save the model to S3 and load it ad hoc
Show answer & explanation
Correct: A. Real-time endpoints provide low-latency, single-request inference for steady traffic — the standard choice for a sub-100 ms SLA.
Why not the others: Batch Transform (B) is offline; Asynchronous Inference (C) suits long-running or large-payload requests; loading from S3 each call (D) is far too slow.
Source: AWS — SageMaker model deployment options →A team must label 100,000 images for object detection with a managed labeling workflow and optional human-in-the-loop verification. Which service is purpose-built?
- Amazon Translate
- Amazon Comprehend
- Amazon Polly
- Amazon SageMaker Ground Truth
Show answer & explanation
Correct: D — SageMaker Ground Truth. Ground Truth provides managed labeling workflows (private, vendor, or Mechanical Turk) with active learning to reduce labeling cost.
Why not the others: Translate (A), Comprehend (B), and Polly (C) are pre-built ML APIs for translation, NLP, and speech — none provide data-labeling workflows.
Source: AWS — SageMaker Ground Truth →Hyperparameter tuning is needed across 8 dimensions with non-trivial training time per trial. Which strategy is most efficient?
- Bayesian optimization (SageMaker Automatic Model Tuning)
- Grid search across all combinations
- A single manual trial
- Random guess and stop after one
Show answer & explanation
Correct: A. Bayesian optimization learns from prior trials to focus on promising regions of the search space, dramatically outperforming grid search in high dimensions with costly trials.
Why not the others: grid search (B) explodes combinatorially across 8 dimensions; a single manual trial (C) or one random guess (D) won’t find good configurations.
Source: AWS — SageMaker Automatic Model Tuning →A deployed model’s predictions are slowly degrading over weeks. Which AWS service detects feature and prediction drift over time?
- AWS CloudFront
- Amazon SageMaker Model Monitor
- Amazon Macie
- AWS Trusted Advisor
Show answer & explanation
Correct: B — SageMaker Model Monitor. Model Monitor schedules baseline-versus-production checks for data quality, drift, model quality, and bias — exactly the slow-degradation case.
Why not the others: CloudFront (A) is a CDN, Macie (C) is data-security/PII discovery, and Trusted Advisor (D) is account best-practice checks — none monitor model drift.
Source: AWS — SageMaker Model Monitor → Further reading: PowerKram — AI & machine learning →An engineer wants to use a pre-trained image classifier and adapt it to a new dataset of medical images. Which technique is the natural fit?
- Train from scratch on the small dataset
- Transfer learning by fine-tuning the pre-trained backbone
- Use only k-nearest neighbors on raw pixels
- Train a linear regression on image bytes
Show answer & explanation
Correct: B — transfer learning. Fine-tuning a backbone pre-trained on a large source dataset typically achieves much higher accuracy than training from scratch on a limited target set. SageMaker’s built-in image classification supports transfer learning.
Why not the others: training from scratch (A) needs far more data; KNN on raw pixels (C) and linear regression on image bytes (D) ignore image structure and perform poorly.
Source: AWS — SageMaker image classification →A workload spikes for about 30 minutes a day and is idle the rest of the time. Cost is the priority. Which inference option is most cost-effective?
- A 24/7 ml.p4d real-time endpoint
- SageMaker Serverless Inference
- Multi-Region active/active provisioned endpoints
- Always-on Multi-Model Endpoint at peak size
Show answer & explanation
Correct: B — Serverless Inference. It scales to zero between bursts and bills per request — ideal for short, infrequent spikes where paying for idle capacity would dominate cost.
Why not the others: a 24/7 GPU endpoint (A), multi-Region provisioned endpoints (C), and an always-on peak-sized Multi-Model Endpoint (D) all keep costly capacity running while idle.
Source: AWS — SageMaker Serverless Inference →Keep going: Learning & Career resources
AWS ML skills remain in high demand even as the specific credential changes. Two PowerKram hubs back this material.
Deep dive: the retirement, the successor exam, and how your prep transfers
What retired, and when
AWS retired the Certified Machine Learning – Specialty (MLS-C01) with a last testing day of March 31, 2026. Exam prep resources were available up to that date. Certifications already earned remain valid for three years from the date they were earned, and holders can continue to display the badge through Credly during that window. AWS is no longer offering the exam to new candidates. See the successor exam →
Why AWS retired it
When MLS-C01 launched, most teams built and tuned models from scratch, which demanded deep knowledge of algorithms and feature engineering. Today, organizations increasingly rely on managed services and foundation models (for example, Amazon Bedrock) that integrate into applications, while SageMaker continues to support training and deployment. AWS shifted its ML credentialing toward operating ML in production — identity, cost control, security boundaries, observability, and integration — which is the focus of the successor associate exam. See the AI Practitioner exam →
How your MLS-C01 prep transfers
Most of what you studied still applies. SageMaker built-in algorithms, Automatic Model Tuning, the deployment options (real-time, batch transform, asynchronous, serverless), Model Monitor, and Ground Truth all appear in the successor exam and in day-to-day ML engineering. The main shift is emphasis: the successor leans further into building, deploying, and operationalizing ML rather than model theory. Use the practice questions here to keep those SageMaker patterns sharp, then move to MLA-C01-specific prep. See the Data Engineer Associate exam →
Choosing your next AWS exam
If you build and deploy ML systems, the associate-level MLA-C01 (Machine Learning Engineer – Associate) is the direct successor. If you’re newer to AWS AI/ML, the AIF-C01 AI Practitioner is a foundational on-ramp. If your work centers on the data pipeline feeding ML, the DEA-C01 Data Engineer Associate is a strong complement. Match the exam to your role rather than chasing the hardest badge. AI & ML career paths →
Frequently asked questions about MLS-C01 retirement
Is the AWS Machine Learning Specialty (MLS-C01) exam still available?
I already passed MLS-C01. Is my certification still valid?
What replaced the Machine Learning Specialty certification?
Are these practice questions still useful now that the exam retired?
What were the MLS-C01 exam domains and weights?
Certifying your AWS ML skills? Move to MLA-C01
MLS-C01 has retired. The associate-level Machine Learning Engineer exam is the current path — and much of your SageMaker prep carries straight over.
Go to the MLA-C01 practice exam → Existing subscribers: open the MLS-C01 question bank