在打印窗口把「目标」选为「另存为 PDF」即可保存,页面已带 awsgood.com 版权水印。 返回原文

www.awsgood.com

MLA-C01 Exam Guide and Study Plan: Domains, Cost, Timeline

2026-11-15 MLA-C01 Machine Learning Engineer - Associate 约 8 分钟

TL;DR

Who MLA-C01 fits, and when you should not book it

This is an associate exam for people who ship models, not people who explain gradient descent. The exam guide assumes roughly a year of hands-on work with SageMaker, S3, IAM and at least one CI/CD-ish workflow. I've watched strong data scientists fail their first attempt because they knew XGBoost cold but had never touched a Model Registry approval status.

The questions are written from the operations side. "A pipeline fails at 03:00 and the CloudWatch alarm doesn't fire — what do you check first?" is closer to the real feel than any question about loss functions.

When MLA-C01 is the wrong next step. Skip it if any of these describe you:

There's also a budget angle. Associates are $150 USD on the AWS certification page (local currency and tax differ), and a retake is another $150. That's not a rounding error for a one-person shop, so don't sit it "just to see".

The four domains and what the weights actually mean

Domain Weight ~Questions of 65 Hours of a 45-hour plan
1. Data Preparation for ML 28% 18 13
2. ML Model Development 26% 17 12
3. Deployment and Orchestration of ML Workflows 22% 14 10
4. ML Solution Monitoring, Maintenance, and Security 24% 16 10

Read that table twice. Data prep is the single biggest block, and it's mostly unglamorous: S3 partitioning, Parquet vs CSV, Glue Data Catalog, Processing jobs, encoding categoricals for XGBoost (which wants numbers, not strings), train/validation/test splits, and what to do about class imbalance.

Domain 2 is where Python people feel safe and then get caught on SageMaker specifics — Spot training with checkpoints to S3, managed warm pools, Automatic Model Tuning strategies (Bayesian by default, Hyperband for cheap early stopping), Debugger, and the difference between Experiments and the Model Registry.

Domain 3 is the one SMB ops folks underestimate. Real-time vs serverless vs asynchronous inference vs Batch Transform is a decision question that shows up in some form on nearly every attempt. Autoscaling policy choice, multi-model endpoints for many small models, and Step Functions vs Pipelines for orchestration all live here.

Domain 4 is the 24% that people study last and lose points on: Model Monitor (data quality, model quality, bias drift, feature attribution drift), CloudWatch metrics like ModelLatency and InvocationsPerInstance, KMS encryption at rest, VPC endpoints, and iam:PassRole failures.

Registration, cost, and which exam to take first

Format facts worth memorizing because they shape your timing strategy: 65 questions, 130 minutes, scaled score 100–1000 with 720 to pass, valid three years, delivered by Pearson VUE either at a test center or online proctored. Question types are multiple choice (four options, one answer) and multiple response (five or more options, two or three answers). There's no penalty for guessing, so nothing stays blank.

AIF-C01 (AI Practitioner) MLA-C01 (ML Engineer – Assoc.) MLS-C01 (ML – Specialty)
Level Foundational Associate Specialty
Time / questions 90 min / 65 130 min / 65 180 min / 65
List fee (USD) $100 $150 $300
Hands-on expected none ~1 year ~2 years
Take it when new to AWS AI services you deploy models you already run ML at scale

My ordering advice for a small team: AIF-C01 if anyone on the team is client-facing and needs vocabulary, MLA-C01 for the person who actually owns the pipeline. Do not go MLA-C01 → MLS-C01 back to back; take at least a quarter between them or the second one becomes expensive revision.

A six-week study plan that survives client work

Six weeks at 7–8 hours per week, split into weekday 45-minute blocks and one longer weekend lab. Every week ends with something deployed, because reading about CreateEndpoint is not the same as watching one fail.

Verify your account can actually do the labs before Monday. Two commands save an hour of confusion:

# what SageMaker endpoints already exist in this region
aws sagemaker list-endpoints --region us-east-1 
  --query 'Endpoints[].[EndpointName,EndpointStatus]' --output table

# check your training quota before you write the notebook
aws service-quotas list-service-quotas --service-code sagemaker 
  --query "Quotas[?contains(QuotaName,'ml.m5.xlarge for training job usage')].[QuotaName,Value]" 
  --output table


For Domain 4 practice, pull a real metric instead of memorizing the list:

```bash
aws cloudwatch get-metric-statistics --namespace AWS/SageMaker 
  --metric-name ModelLatency --dimensions Name=EndpointName,Value=ml-demo 
  --start-time 2026-11-01T00:00:00Z --end-time 2026-11-02T00:00:00Z 
  --period 300 --statistics Average --region us-east-1


If that returns an empty `Datapoints` array, your endpoint has been idle or deleted — a five-second lesson in why CloudWatch alarms on `Invocations` matter.

## The lab that ate my Sunday: a quota wall at 21:40

I was prepping the Week 3 lab on a fresh sandbox account, 21:40 on a Sunday, with a GPU notebook ready to go. `CreateTrainingJob` came back in under two seconds:


ClientError: An error occurred (ResourceLimitExceeded) when calling the CreateTrainingJob operation:
Failed to create training job because the account-level service limit
'ml.g5.2xlarge for training job usage' is 0 Instances


My first configuration was wrong in a way that's easy to repeat: I assumed a new account inherits sensible GPU quota because the console happily let me pick `ml.g5.2xlarge` from the dropdown. New accounts often start GPU instance quotas at zero, and the console doesn't warn you until the API call fails.

Diagnosis took three steps. I ran the `list-service-quotas` command above against `ml.g5.2xlarge`, which confirmed `Value: 0.0`. Then I checked CloudTrail for the failed event to be sure it was quota and not an IAM deny — the `errorCode` was `ResourceLimitExceeded`, not `AccessDenied`, which ruled out the execution role. Then I filed a Service Quotas increase request.

The request wasn't instant. Rather than lose the evening, I re-ran the same script on `ml.m5.xlarge` (non-GPU, already at a usable quota) with `UseSpotInstances` and a checkpoint path. It trained in about 12 minutes on a 40k-row dataset, which was plenty for the exercise. The quota landed the next morning.

Two lessons I now apply to every new account: check quotas before writing the notebook, and keep a CPU-sized fallback for every lab so a quota approval never blocks a study session. The same pattern bites in production — a retraining pipeline that's fine for months dies when someone bumps the instance type.

## High-yield traps and the 2-minute clock

130 minutes divided by 65 questions is exactly 2.0 minutes each. Multiple-response items eat double, so you need slack. My routine: first pass takes 75–80 minutes answering only what's obvious, flagging everything else; second pass burns 35 minutes on flagged items; the last 10 minutes re-read the flagged ones and confirm nothing is blank.

| What the stem says | What it's actually testing |
|---|---|
| "LEAST operational overhead" | Managed service beats self-managed — usually Pipelines, Model Monitor, or serverless inference |
| "MOST cost-effective" for spiky traffic | Serverless inference or Batch Transform, not a 24/7 real-time endpoint |
| "Payloads up to 1 GB, long-running" | Asynchronous inference with an S3 output path |
| "Real-time features, single-digit ms" | Feature Store *online* store, not the offline S3 store |
| "Detect drift in predictions" | Model quality monitor, not data quality |
| "Pipeline fails with AccessDenied on S3" | Execution role trust policy or bucket policy, often `iam:PassRole` |

The traps are language traps more than knowledge traps. "Least operational overhead" and "most cost-effective" point at different answers for the same scenario, and AWS writes both into adjacent questions. Read the constraint clause before the scenario.

One more: don't over-study Amazon Forecast, Personalize, or Comprehend. They appear, but a handful of questions at most. Time spent on Model Monitor baselines pays better.

## Sources

- [AWS Certified Machine Learning Engineer – Associate exam guide (MLA-C01)](https://docs.aws.amazon.com/aws-certification/latest/machine-learning-engineer-associate-01/machine-learning-engineer-associate-01.html)
- [AWS Certification: Machine Learning Engineer – Associate](https://aws.amazon.com/certification/certified-machine-learning-engineer-associate/)
- [Amazon SageMaker Developer Guide](https://docs.aws.amazon.com/sagemaker/latest/dg/whatis.html)
- [Amazon SageMaker Model Monitor](https://docs.aws.amazon.com/sagemaker/latest/dg/model-monitor.html)

When we set up study accounts for the team, the routine that stuck was one shared sandbox with a hard budget alarm and a documented teardown command list — if you'd rather not maintain that yourself, awsgood.com runs managed AWS environments for small teams, including the teardown discipline that keeps a lab account from quietly billing you.

常见问题

How many questions and how long is the MLA-C01 exam?
MLA-C01 has 65 questions in 130 minutes, which works out to 2 minutes per question. Question types are multiple choice (one correct answer of four) and multiple response (two or three correct of five or more). The scaled score runs 100–1000 and you need 720 to pass. There is no penalty for guessing, so never leave blanks.
What are the MLA-C01 domain weights?
Four domains: Data Preparation for ML at 28%, ML Model Development at 26%, Deployment and Orchestration of ML Workflows at 22%, and ML Solution Monitoring, Maintenance, and Security at 24%. Data prep and monitoring