I Am an Infrastructure Guy With Zero ML Background
I have been doing IT infrastructure for many years, and before this exam I had literally zero machine learning background. The most "ML" thing I had ever done was ask ChatGPT to write me a joke. The AWS Certified Machine Learning Engineer Associate is my 8th AWS certification, and it is the toughest associate exam I have taken so far.
If you are an ML engineer with years of experience, this exam will probably feel a lot easier for you than it did for me. But if you are like me — a cloud, infra, or DevOps person who wants to step into the AI world — this post is for you. Here is everything I learned the hard way, so you do not have to.
Heads-Up: MLA-C01 Is Being Replaced by MLA-C02
I took MLA-C01, the version AWS retired in English on September 28, 2026. It stays available in Japanese, Korean, and Simplified Chinese until its successor reaches general availability.
If you are starting now, you will most likely sit MLA-C02. Beta registration opened on September 1, 2026: 85 questions in 170 minutes, $75 USD at beta pricing, English only. General availability is January 14, 2027. The new version widens the exam into generative AI — foundation models, agentic workflows, and Amazon Bedrock now sit alongside SageMaker.
The good news: the core of this post still applies. ML fundamentals, metrics, SageMaker built-in algorithms, deployment options, MLOps, and monitoring are the backbone of both versions. Just expect more Bedrock and generative AI than I got.
What the Exam Actually Expects
According to the official exam guide, MLA-C01 validates your ability to build, operationalize, deploy, and maintain machine learning solutions and pipelines on AWS. The target candidate has at least one year of experience with Amazon SageMaker and other AWS ML services, plus a year in a related role such as backend development, DevOps, data engineering, or data science.
That part is a bit scary. Good thing I had not read it before I started — back then I thought SageMaker was something that makes sages.
You can still pass without prior ML experience. I am living proof of that. But I have to be honest: unlike Solutions Architect Associate, where you can partly get by on general AWS knowledge, this exam really expects you to understand ML concepts. You will have to learn them from scratch during your prep, and that is absolutely doable. One important thing: you do not have to know everything. You have to know enough.
Exam Format: MLA-C01 vs MLA-C02
MLA-C01 has a passing score of 720 out of 1000. Besides classic multiple choice and multiple response, I also got the newer ordering and matching question types, which take longer to answer.
| Detail | MLA-C01 (what I took) | MLA-C02 (beta) |
|---|---|---|
| Questions | 65 | 85 |
| Time limit | 130 minutes | 170 minutes |
| Cost | $150 USD | $75 USD (beta pricing) |
| Languages | English retired Sept 28, 2026; JA/KO/ZH until C02 GA | English only during beta |
| Delivery | Pearson VUE test center or online | Pearson VUE test center or online |
Should You Take It?
I do not work with machine learning day to day, and there is no direct need for it in my job right now. But AI is not hype — it is where IT is heading, and it is already here. I wanted to step into it, and the best way I know to learn something properly is to take a certification.
I would recommend it if you are an infra, DevOps, or cloud person who wants to stay relevant, if you already work with data pipelines and want to understand the AWS ML side properly, or if you have passed AI Practitioner and want to go deeper — this is the natural next step.
Be cautious if you have zero AWS experience. Do Cloud Practitioner first, and maybe even Solutions Architect Associate — that gives you solid ground on the core AWS services.
Be cautious, too, if you have zero AI exposure. I highly recommend taking AI Practitioner before this one: it gives you a soft introduction to ML terminology, and then this exam feels much less brutal. I nearly fell into this trap myself. Even with all my AWS experience, I thought I could skip AI Practitioner and started preparing for ML Engineer Associate directly. I quickly realised that was unrealistic, backed off, and did AI Practitioner first.
And a reality check for my fellow infra people: prepare for a different kind of pain. Solutions Architect Associate is broad but familiar — EC2, S3, VPC, RDS, you have touched most of it. This exam throws algorithms, statistics, metrics, and ML terminology at you that you may never have seen: inference, temperature, XGBoost, Random Cut Forest, AUC-ROC, the precision-recall trade-off. The first time I saw all of this, my brain froze. But once you push through the first week or two, things start clicking. Do not get discouraged — just know what you are signing up for.
How I Prepared: One Course + Plenty of Practice Tests
I used the same formula as for every AWS certification: one solid online course plus plenty of practice tests. It has never failed me, and it did not fail me this time either.
For the course I went with Stephane Maarek's Machine Learning Engineer Associate course on Udemy. For someone with zero ML background it was the right choice: he explains ML concepts from scratch without assuming you already know them, the pacing is excellent, and the depth is exactly right for the exam.
The course is around 24 hours long, but it includes sections on standard AWS services I already knew from Solutions Architect Associate — S3, EC2, IAM, CloudWatch and so on — so I could cut some corners. I also watched at 1.25x speed and skipped some labs and walkthroughs, so it took me around 15 hours. For ML concepts, I would keep it at 1.25x maximum: your brain needs time to digest things like overfitting, regularization, and what on earth a hyperparameter is.
If Stephane's style does not click with you, that is fine. AWS Skill Builder and Pluralsight are other options. I actually started with a different Udemy course and it did not click for me — that is purely subjective. Most courses have free preview lessons, so try a few and pick the one you like.
What I would not recommend is reading the AWS documentation cover to cover. It is excellent, and in theory it has everything you need, but unless you have the superpower of learning from walls of text, stick with a structured course.
My Study Method
Part of my approach goes against the standard advice, so bear with me.
Go through the course slowly. For ML concepts, slow is fast. Do not try to binge eight hours of new terminology in one weekend — your brain will simply refuse. When my head felt heavy, I stopped, closed the laptop, and went for a walk. I may have taken this a bit too far, because my preparation stretched over several months. But the result is a pass, so the strategy works.
Take notes like your life depends on it. Notes mattered far more for me on this exam than on any previous one. When you are new to ML you are constantly hit with new algorithms, metrics, and services, and by lesson 50 you will have forgotten what XGBoost was for. I kept one short note per algorithm (what problem it solves), one per metric (when to use it), and one per AI service (what it is for) — one or two lines each, in my own words, no copy-paste from slides. By the end I had my own ML cheat sheet, and it became my main revision tool in the final days.
Hands-On… or Not? My Controversial Take
I did zero hands-on practice for this exam. None. And I still passed.
I have done AWS infrastructure for many years, but SageMaker was completely new to me. Getting comfortable with it hands-on would have cost me weeks I did not have — I have a full-time job, a family, and a YouTube channel to run. So I made a deliberate choice to focus on theory and practice questions instead of clicking around the console.
My honest advice: if you can do hands-on, do it. Concepts stick much better once you have launched a training job, deployed an endpoint, and seen SageMaker Studio with your own eyes. Looking back, even a few key labs would have saved me some pain on exam day. But if you are short on time, do not panic if you skip it — you can still pass with strong theory and lots of practice tests.
One warning if you do go hands-on: SageMaker endpoints, training jobs, and notebook instances can burn through money frighteningly fast. This is not an idle t2.micro. Always delete your endpoints, stop your notebooks, clean up after training jobs, and set up a billing alarm before you start.
And if you want some free money from Jeff to practise with, my other video shows how to get $200 in free AWS credits.
Practice Tests: Survival Equipment for This Exam
Once you have finished the course, this is the step that made the difference for me. ML questions use very specific terminology and love trap answers that sound technically correct but do not quite fit. Often three of the four options use real, valid ML terms, and you have to know which one fits this exact scenario. The only way to train for that is a lot of practice questions.
I used Stephane Maarek's practice exams on Udemy: three full exams, close to 200 questions, and very affordable. Use Practice Mode while studying — no time pressure, you see immediately whether you were right, and you get a detailed explanation for every answer. That is where the real learning happens: every wrong answer becomes a lesson, and every explanation goes into your notes. When you feel more confident, switch to Exam Mode with the timer running to simulate the real thing.
Those practice tests are slightly harder than the real exam. For me, scores of 55-65% were a sign I had a good chance of passing — but only because I learned from every mistake. Calibrate against your own results rather than treating that range as a rule.
If you are on a tight budget or want to test the waters first, CloudNinja has a free MLA practice exam. To be fully transparent, it is not enough on its own to prepare you. But it is great for getting familiar with the question style, understanding how questions are structured, and testing your current level — completely free.
Exam Day: Book Early, Go to a Test Center, Manage Your Time
Book the exam early. Your brain is incredibly creative at inventing reasons not to take the exam this month — you never feel fully ready, and next month always looks better. Once there is a date in the calendar, your focus sharpens. AWS lets you reschedule or cancel without a fee up to 24 hours before your appointment, so there is very little risk.
For this exam I went to a test center, and I would 100% recommend it. For an exam this hard, the last thing you want is your Wi-Fi dropping at question 47 — I have had that happen on another exam, and it is not fun. Test centers are quiet and reliable. The computers are not the latest, but you are not there to play games.
MLA-C01 gives you 130 minutes for 65 questions — about two minutes each. That is enough for most questions, but the long scenario questions and the ordering and matching formats eat time fast. Answer everything you know on the first pass, flag anything that stumps you and move on, then come back to the flagged questions at the end. Do not spend five minutes on one question: better to lose one point than run out of time for ten more.
If English is not your first language, you can request the ESL +30 accommodation for an extra 30 minutes on AWS exams. My AI Practitioner video shows how to request it, along with a few ways to save money on exam fees.
What to Actually Focus On
The exam covers four domains, but here is where I found the points actually hiding.
ML fundamentals — do not skip these even if they look basic. Supervised learning uses labelled data; unsupervised learning has no labels. Classification predicts a category (spam or not spam); regression predicts a number (what a tomato will cost in Alabama 100 days from now). Underfitting is when the model is too simple and errs on both training and real data — high bias. Overfitting is when it learns the training data too well, essentially memorising the answers, and fails on new data — high variance.
Metrics — and when to use which. For classification: accuracy, precision, recall, F1, and AUC-ROC. For regression: RMSE (root mean squared error) and MAE (mean absolute error). I lost points here through a brain glitch: I was weighing accuracy, precision, and recall on a question that was clearly about regression, where the answer was MAE. Read every question slowly, and check whether it is classification or regression before looking at the options.
Data preparation: SageMaker Data Wrangler and AWS Glue are usually the right answers.
MLOps: SageMaker Pipelines, Step Functions, Model Registry, Model Cards, and basic CI/CD concepts for ML workflows. When the question is about CI/CD for ML, SageMaker Pipelines is almost always the answer.
Monitoring and responsible AI: SageMaker Model Monitor and the four things it watches — data quality, model quality, bias drift, and feature attribution drift. For infrastructure metrics like CPU and memory, the answer is CloudWatch. SageMaker Clarify covers bias detection and explainability.
Security: IAM execution roles, VPC endpoints, network isolation, and KMS encryption — standard AWS security applied to SageMaker. If you have passed Solutions Architect Associate, these are free points.
Generative AI: Amazon Bedrock, foundation models, RAG (retrieval-augmented generation), fine-tuning versus prompt engineering, and Bedrock Guardrails. This was a growing part of MLA-C01 and is a much bigger part of MLA-C02.
SageMaker Built-In Algorithms: Know What Each One Solves
For each built-in algorithm you mostly need to know what problem it solves and whether it is supervised or unsupervised. There is no shortcut — you have to memorise them. And again, you do not have to be perfect: if one does not click, move on. Better to lose a few points than to panic and never book the exam.
| Algorithm | What it solves | Type |
|---|---|---|
| XGBoost | Classification, regression, ranking | Supervised |
| Random Cut Forest | Anomaly detection | Unsupervised |
| DeepAR | Time-series forecasting | Supervised |
| BlazingText | Text classification and word embeddings | Supervised (classification) / unsupervised (Word2Vec) |
Deployment: Heavily Tested
Deployment questions come up constantly, and the skill being tested is picking the right inference option for the scenario. Read carefully whether the question needs real-time answers or can tolerate a delay. Also know multi-model and multi-container endpoints, and the deployment strategies: A/B testing, shadow deployments, and blue/green.
| Inference option | Pick it when |
|---|---|
| Real-time endpoint | You need low-latency predictions for steady, ongoing traffic |
| Serverless inference | Traffic is intermittent or unpredictable and occasional cold starts are acceptable |
| Asynchronous inference | Payloads are large or processing is slow, and callers can wait for a queued result |
| Batch transform | You need predictions for a whole dataset offline, with no persistent endpoint |
The Other AWS AI Services: Quick Wins
A lot of points hide in simply knowing what each managed AI service does. You do not need deep knowledge of any of them — just what they are for.
| Service | What it does |
|---|---|
| Amazon Textract | Extracts text and data from documents |
| Amazon Comprehend | Natural language processing, such as sentiment analysis |
| Amazon Rekognition | Image and video analysis |
| Amazon Polly | Text to speech |
| Amazon Transcribe | Speech to text |
Your Exam-Morning Revision
The CloudNinja MLA cheat sheet summarises most of these topics — algorithms, metrics, inference options, Model Monitor, and the AI services — in one free PDF. It is worth a read on the morning of your exam. The certification guide covers the domains and study strategy in more depth.
Was It Worth It?
Was it tough? Yes — for me, the toughest AWS associate exam so far. Was it worth it? Absolutely. AI on AWS no longer feels scary, SageMaker no longer feels like a black box, and I can have meaningful conversations with ML engineers instead of nodding politely with no idea what they are saying.
If an infra guy with zero ML background can pass this, you can too. Good luck — and delete your SageMaker endpoints.