AWS Certified Machine Learning – Specialty (MLS-C01) — Exam Survival Guide

Window: ML engineers / data scientists on AWS | Format: 65 questions (50 scored + 15 unscored), 170 minutes, pass/fail against a standard; retirement announced for 31 March 2026

(If you’re reading this after that date, check AWS’s replacement exam — likely the ML Engineer Associate.)

Must-do topics

MLS-C01 = “Can you design, build, train, optimise, and deploy ML on AWS in a sane, production-ish way?”

Four domains in the exam guide:

Data Engineering

Ingestion (Kinesis, DMS, Glue, DataSync), storage patterns (S3 layouts, partitioning, Lake Formation).

Feature storage (S3, DynamoDB, feature stores), data quality checks, schema design.

Exploratory Data Analysis

Handling imbalance, missing data, outliers; bias/leakage; basic visualisation/summary logic.

Choosing metrics for classification/regression/ranking, etc.

Modeling

Choosing algorithms: XGBoost vs linear vs deep, classical vs DL, tabular vs text vs images.

Hyperparameter tuning (SageMaker HPO), training at scale, distributed training basics.

Deployment & Operations

SageMaker endpoints, batch transform, pipelines, CI/CD, shadow/canary deployments.

Monitoring: data drift, concept drift, model performance, retraining triggers.

Plus a thick layer of AWS-specific decisions:

Where to host data (S3 vs RDS vs Redshift vs DynamoDB).

How to secure and cost-optimise pipelines (IAM, VPC, spot instances, managed services).

Top traps (avoid)

Treating the exam like a pure ML theory test; it’s very architecture + trade-offs.

Knowing algorithms in isolation but not how to plug them into AWS services.

Ignoring cost and operations (autoscaling, monitoring, rollback) in answers.

Over-engineering tiny problems with heavyweight architectures.

Time split

65 Q, 170 minutes → about 2.6 minutes per question.

Aim:

60–70 questions in 150 minutes, leaving 20 minutes for revisits / flagged items.

Last-48h checklist

One full-length mock OR 2 × 35-Q blocks under timed conditions.

Exam guide sweep:

For each domain, list the top 5–7 AWS services that show up there and what you’d use them for.

Rehearse “architecture trade-off” stories:

Real-time vs batch.

Serverless vs long-running training clusters.

Managed feature store vs DIY solution.

Quick frames

On each scenario:

What is the business goal? (latency, accuracy, cost, compliance).

What is the data situation? (volume, velocity, labelled/unlabelled, streaming vs offline).

Where are we in the lifecycle? (ingest → prep → train → tune → deploy → monitor).

Which AWS service combo hits goal with least pain?

Typical “good” answers:

Use managed services you’d actually trust in a production account: SageMaker, S3, Glue, Kinesis, etc.

Keep architecture clean: minimal moving parts that still respect security, scalability, and cost.

Speed tactics

If a solution can be done with one managed AWS service instead of three custom ones, that’s usually the right answer.

Prefer:

Monitoring + rollback plan over “deploy and hope.”

Autoscaling + spot awareness when heavy training loads are mentioned.

For evaluation/metrics:

Match the metric to the business risk (e.g., recall in fraud detection, precision in harmful false-positive domains).

Day-of mini-plan

Morning: 10–15 architecture-style practice questions; no fresh whitepapers.

In the exam:

If stuck, eliminate obviously over-complicated or obviously under-engineered designs first.

Mental stance:

“I’m the person asked to make this ML system behave in the real world, not to build a research lab.”