Hands-on AWS knowledge base for developers Managed Ops
Talk to Us

MLS-C01 Exam Guide and Study Plan: 65 Questions, 36% Modeling

Export PDF Print-friendly layout with a copyright watermark; choose Save as PDF in the print dialog
MLS-C01 Exam Guide and Study Plan: 65 Questions, 36% Modeling

TL;DR

  • MLS-C01 is 65 questions in 180 minutes, scored on a 100–1000 scale with 750 to pass, $300 USD per attempt in the US, and valid for 3 years.
  • Modeling carries 36% of the weight — roughly 23 of the 65 questions — so if you only have 40 study hours, spend 15 of them there, not on EDA.
  • The questions test SageMaker operations (spot training, checkpointing, tuning jobs, endpoint variants) more than modeling skill; you cannot reason your way to the answer from ML theory alone.

Who MLS-C01 filters out, and who should sit a different exam first

The official guide describes a candidate with a year or two of hands-on experience building and running ML workloads on AWS, plus working knowledge of the modeling lifecycle. In practice the exam is stricter than that sentence suggests. I watched a strong data analyst with five years of SQL and pandas fail the first attempt by 40 scaled points because she had never created an endpoint, a tuning job, or a batch transform job in her life. The theory was fine; the service surface was not.

AWS now has three overlapping ML credentials, and picking the wrong one wastes six weeks.

Exam Level Best fit Time to prep
AIF-C01 (AI Practitioner) Foundational PMs, analysts, anyone new to ML on AWS 15–25 hours
MLA-C01 (ML Engineer Associate) Associate Engineers who deploy and operate models, light modeling 40–60 hours
MLS-C01 (ML Specialty) Specialty People who both build and operate models on SageMaker 80–120 hours

If your job is "train the model," go Specialty. If it is "ship the model someone else trained," MLA-C01 is a better return on your evenings — it assumes less algorithm trivia and more pipeline work. Don't stack all three; the badge overlap confuses recruiters more than it impresses them.

The four domains, and what 36% on Modeling actually costs you

The published weights have been stable across guide revisions. Multiply them by 65 and you get a rough question budget, which is the single most useful planning number I've found.

Domain Weight ~Questions High-yield content
Data Engineering 20% ~13 S3 layout and partitioning, Glue/Athena, Kinesis Data Streams vs Firehose, RecordIO-Protobuf and Parquet, Feature Store ingestion
Exploratory Data Analysis 24% ~16 Missing-value handling, scaling, class imbalance, SMOTE, PCA, TF-IDF, Data Wrangler, correlation vs causation
Modeling 36% ~23 Built-in algorithm selection (XGBoost, Linear Learner, DeepAR, BlazingText, Factorization Machines), metrics, regularization, hyperparameter tuning strategy, transfer learning
ML Implementation and Operations 20% ~13 Real-time vs serverless vs async vs batch inference, A/B and shadow variants, auto scaling, Model Monitor, managed spot training, IAM execution roles

The Modeling domain is where people over-study and under-score. The exam rarely asks you to derive a gradient; it asks which built-in algorithm fits a sparse high-dimensional click log, or whether AUC or F1 is the defensible metric when the positive class is 0.4% of the data. Knowing the difference between ml.m5.xlarge and ml.c5.2xlarge matters more than knowing why XGBoost uses second-order gradients.

Registration, cost, and the format details that bite

You register through your AWS Certification account and schedule on Pearson VUE. Both test center and online proctored slots were available within 48 hours the last time I booked, though I'd give yourself two weeks if you want a specific evening slot. The fee is $300 USD for Specialty exams in the US; regional pricing differs. Retakes cost full price and require a 14-day wait after a failed attempt, which is the real reason to book only after you're scoring consistently on practice sets.

The format is multiple choice (one correct) and multiple response (two or three correct out of five or more). There is no penalty for a wrong answer, and flagged questions can be revisited until the clock hits zero. The pass mark is a scaled 750 — AWS does not publish the raw-to-scaled mapping, so anyone telling you "you can miss 16" is guessing.

A study plan that survives a full-time job

Realistic hours depend on your starting point more than your IQ. These are the ranges I've seen hold up, not marketing claims.

Starting point Hours Calendar at 12 h/week
Ships models, 1+ year of SageMaker 50–70 4–5 weeks
Analyst: strong Python/SQL, no AWS ML 90–120 8–10 weeks
AWS ops or solutions architect, light ML 100–140 10–12 weeks
No ML fundamentals at all Stop — do AIF-C01 first —

The plan below assumes the middle row. Run it in order; the lab work must come before the practice exams or you'll memorize answers without context.

  1. Weeks 1–2, Data Engineering + EDA. Read the SageMaker Developer Guide sections on data prep and the built-in algorithm pages, one algorithm per evening. Build one Feature Store group and one Data Wrangler flow.
  2. Weeks 3–5, Modeling. Pick five algorithms and actually train them on the same dataset. Compare metrics in a notebook. Do not read about hyperparameter tuning — run a tuning job with three static ranges and watch the console.
  3. Weeks 6–7, Operations. Deploy one model four ways: real-time endpoint, serverless endpoint, batch transform, async inference. Break each one deliberately (send the wrong content type) and read the error.
  4. Week 8, practice and gaps. Take the free 20-question Official Practice Question Set cold, then re-read only the domains where you missed more than two.

Stand up the lab cheaply. A ml.t3.medium notebook is enough for every exercise except distributed training, and it bills by the second — so stop it the moment you close the laptop.

# Sandbox notebook — smallest instance that runs the SDK comfortably
aws sagemaker create-notebook-instance 
  --notebook-instance-name mls01-lab 
  --instance-type ml.t3.medium 
  --role-arn arn:aws:iam::123456789012:role/SageMakerExecutionRole 
  --volume-size-in-gb 5 
  --region us-east-1

# Stop it immediately after each session; running instances are the #1 surprise line item
aws sagemaker stop-notebook-instance --notebook-instance-name mls01-lab --region us-east-1


## When NOT to take MLS-C01

Skip it if you have never trained a model outside a tutorial. The exam assumes vocabulary like precision-recall tradeoff, regularization, and stratified splits; you'll burn 40 hours just reaching the starting line. Take AIF-C01, then decide. Skip it too if you need a credential before the end of next month for an RFP — MLA-C01 is a shorter ramp. And skip it if the $300 budget only covers one attempt and your practice scores are still below 70%; the 14-day retake lockout turns a near-miss into a two-month delay.

## The spot-training validation error that taught me the Operations domain

I was building a lab to demonstrate managed spot training, mostly to make the cost questions concrete. My first `create-training-job` call enabled spot and set `MaxWaitTimeInSeconds` to 7200 with `MaxRuntimeInSeconds` at 3600, but nothing else. The API rejected it outright:


ClientError: An error occurred (ValidationException) when calling the CreateTrainingJob
operation: Managed spot training requires checkpointing to be enabled.


My first assumption was a missing IAM permission, so I started reading CloudTrail. Wrong path. The real constraint is architectural: spot capacity can be reclaimed mid-job, and SageMaker restarts the job from the last checkpoint, so it refuses to launch without somewhere to write them. That means both a `CheckpointConfig.S3Uri` and a training script that writes to `/opt/ml/checkpoints`.

The fix was one extra flag:

```bash
aws sagemaker create-training-job 
  --training-job-name mls01-spot-demo 
  --algorithm-specification '{"TrainingImage":"683313688378.dkr.ecr.us-east-1.amazonaws.com/sagemaker-xgboost:1.7-1","TrainingInputMode":"File"}' 
  --role-arn arn:aws:iam::123456789012:role/SageMakerExecutionRole 
  --input-data-config '[{"ChannelName":"train","DataSource":{"S3DataSource":{"S3DataType":"S3Prefix","S3Uri":"s3://mls01-demo/train/","S3DataDistributionType":"FullyReplicated"}}}]' 
  --output-data-config '{"S3OutputPath":"s3://mls01-demo/output/"}' 
  --resource-config '{"InstanceType":"ml.m5.xlarge","InstanceCount":1,"VolumeSizeInGB":30}' 
  --stopping-condition '{"MaxRuntimeInSeconds":3600,"MaxWaitTimeInSeconds":7200}' 
  --enable-managed-spot-training 
  --checkpoint-config '{"S3Uri":"s3://mls01-demo/checkpoints/"}' 
  --region us-east-1


Verification is the part people skip. `describe-training-job` reports `TrainingTimeInSeconds` and `BillableTimeInSeconds` separately, and on a successful spot run the billable figure is the smaller of the two — that gap is the whole point of the feature.

```bash
aws sagemaker describe-training-job --training-job-name mls01-spot-demo 
  --query '{Status:TrainingJobStatus,Sub:SecondaryStatus,Reason:FailureReason,Billable:BillableTimeInSeconds,Training:TrainingTimeInSeconds}'


That error is also the shape of a dozen Operations questions: spot needs checkpointing, tuning jobs need a valid `ObjectiveMetric` regex that matches a line in your logs, and endpoints need an invocation path that matches the container's expectations. When a lab breaks, read the validation message literally before you go hunting in IAM. If babysitting quotas, checkpoint buckets, and endpoint scaling isn't how you want to spend Fridays, that's the sort of routine SageMaker plumbing we take over at [awsgood.com managed services](https://awsgood.com).

## Exam-day technique: 166 seconds per question

180 minutes over 65 questions is 2.77 minutes each, but that average lies. Multiple-response items cost three to four minutes because you must validate each option, and a scenario with a table of metrics can eat five. My split: 120 minutes for the first pass at roughly 110 seconds per question, then 60 minutes for flagged items and a full review of every multiple-response answer.

Decode the qualifiers, because they eliminate half the options. "MOST cost-effective" usually points to spot training, batch transform, or serverless inference. "Least operational overhead" points to a managed service over anything you'd orchestrate yourself. "Near real-time" rules out batch transform. And on multiple-response questions, the prompt states exactly how many to pick — selecting two when it asks for three scores zero, so never guess an extra checkbox to "cover" an option.

Book the exam date before you finish the Modeling

Leave a Reply

Your email address will not be published. Required fields are marked *