www.awsgood.com
MaxWaitTimeInSeconds, and fuzzy thinking about when to use batch transform instead of real-time inference.The honest answer depends on whether you have ever stared at a Failed training job in the SageMaker console and had to reason your way out. If yes, MLS-C01 is a decent forcing function — it turns scattered console habits into a checklist you can apply on a client's account at 11pm.
If your AWS experience is EC2 and S3 and you have never launched a notebook instance, you will spend most of those 100 hours learning SageMaker vocabulary and ML vocabulary at the same time. That is a bad trade. Take MLA-C01 (the Machine Learning Engineer – Associate exam AWS added in 2024) first; it covers a lot of the same ground at a shallower depth and costs less.
The people who get the clearest return: ops engineers running SageMaker endpoints for small teams, data engineers who own the Glue/S3/Athena pipeline feeding a model, and consultants who need a credential on a proposal. In my experience the exam also maps well to what SMB clients actually ask for — "why is this endpoint costing $400 a month" is a Domain 4 question in disguise.
Sixty-five questions, 180 minutes, scaled score out of 1000 with 750 to pass. That's 2 minutes 46 seconds per question on average, and the multi-response items (pick three of five, no partial credit) eat more than their share. Some questions are unscored pilot items and you cannot tell which, so don't gamble on the weird one.
| Domain | Weight | What it actually tests |
|---|---|---|
| Data Engineering | 20% | S3 layout and partitioning, Glue crawlers, Kinesis/Firehose ingestion, RecordIO vs Parquet, Pipe vs File mode |
| Exploratory Data Analysis | 24% | Data Wrangler, imputation, class imbalance, feature engineering, statistics sanity checks, QuickSight |
| Modeling | 36% | Built-in algorithm choice, hyperparameter tuning, evaluation metrics, regularization, transfer learning |
| ML Implementation and Operations | 20% | Endpoint types, auto scaling, Model Monitor, Debugger, Pipelines, A2I, cost control |
Read that table twice. A third of the exam is modeling, but "modeling" here means which SageMaker tool and which metric, not deriving gradient descent by hand. The questions that killed me in practice were the metric-selection ones: knowing that a fraud model with 0.4% positives should be judged on recall/F1 or AUC rather than accuracy.
Registration goes through Pearson VUE, either at a test center or online-proctored. The online option needs a quiet room, a webcam sweep, and a closed door — the proctor will stop your clock if a roommate walks in.
| Line item | Typical cost (USD) | Notes |
|---|---|---|
| Exam fee | $300 | Region/currency varies; AWS localizes pricing |
| Official practice exam | ~$20 | A free 20-question sample set also exists |
| SageMaker labs | $20–60 | Dominated by notebook instances left running |
| Retake | Full fee again | 14-day waiting period after a fail |
That last row is the one to internalize. There is no cheap second attempt, so do not book the slot before you are scoring comfortably above 80% on practice sets. A single ml.t3.medium notebook left running for a week is about $8 in us-east-1 — the cheapest lesson available, and one I have paid for more than once.
# Cheap guardrail: list every notebook instance and its status, any region
for r in us-east-1 us-west-2 eu-west-1; do
echo "== $r =="
aws sagemaker list-notebook-instances --region "$r"
--query "NotebookInstances[].[NotebookInstanceName,NotebookInstanceStatus,InstanceType]"
--output table
done
## An 8-week study plan that survives a full-time job
Eight to ten hours a week is realistic if you have production SageMaker exposure; 120+ hours is realistic if you don't. Do not spend week one reading the exam guide end to end — it tells you *what* is on the exam and nothing about how it's phrased.
| Week | Focus | Artifact you should have by Sunday |
|---|---|---|
| 1 | Data engineering | Partitioned Parquet dataset in S3, queried with Athena |
| 2 | EDA and feature prep | Notebook handling missing values + class imbalance end to end |
| 3–4 | Built-in algorithms + HPO | Tuning job with 10–20 trials on ml.m5.xlarge |
| 5 | Evaluation and metrics | Confusion matrix + AUC written up, not just computed |
| 6 | Inference options | Same model deployed as real-time, serverless and batch |
| 7 | Operations | Model Monitor baseline + a Pipeline with two steps |
| 8 | Practice exams | Error log of every miss, grouped by domain |
The labs are where the exam becomes obvious. One concrete example: run the same training job with `TrainingInputMode` set to `Pipe` and then `File` against a multi-GB augmented manifest, and watch the startup time change. That difference is worth two or three exam questions and takes an afternoon.
```bash
aws sagemaker create-training-job
--training-job-name mls-spot-demo-$(date +%s)
--algorithm-specification '{"TrainingImage":"<ecr-image-uri>","TrainingInputMode":"File"}'
--role-arn arn:aws:iam::123456789012:role/SageMakerExecutionRole
--input-data-config '[{"ChannelName":"train","DataSource":{"S3DataSource":{"S3DataType":"S3Prefix","S3Uri":"s3://mls-prep-us-east-1/train/","S3DataDistributionType":"FullyReplicated"}}}]'
--output-data-config '{"S3OutputPath":"s3://mls-prep-us-east-1/output"}'
--resource-config '{"InstanceType":"ml.m5.xlarge","InstanceCount":1,"VolumeSizeInGB":30}'
--stopping-condition '{"MaxRuntimeInSeconds":1800,"MaxWaitTimeInSeconds":3600}'
--enable-managed-spot-training
--checkpoint-config '{"S3Uri":"s3://mls-prep-us-east-1/checkpoints"}'
## High-yield topics and the traps that cost points
Five things show up repeatedly in practice sets and in real accounts:
- **Algorithm selection by data shape.** Linear Learner for wide sparse data, XGBoost for tabular with mixed types, DeepAR for many related time series, BlazingText for text classification, Random Cut Forest for anomaly detection.
- **Inference mode.** Real-time endpoints for low-latency persistent traffic; Serverless Inference for spiky, low-volume calls; Batch Transform when nobody is waiting; Asynchronous Inference for large payloads and long processing.
- **Monitoring split.** Model Monitor watches data drift and bias drift; Clarify explains feature attribution and pre-training bias; Debugger watches gradients and loss during training. Mixing these up is a classic wrong answer.
- **Ground Truth and A2I.** Labeling workflows versus human review of live predictions — different services, often confused.
- **Cost questions.** Managed Spot Training plus checkpointing, and knowing that a batch job on a bigger instance can be cheaper than a small endpoint running 24/7.
## The troubleshooting lesson: `ResourceLimitExceeded` and the Spot timeout
Late on a Tuesday in 2025-11 I kicked off a lab training job on `ml.p3.2xlarge` in us-east-1 to practice transfer learning, and it failed in under 20 seconds:
ClientError: An error occurred (ResourceLimitExceeded) when calling the CreateTrainingJob operation:
The account-level service limit 'ml.p3.2xlarge for training job usage' is 0 Instances,
with current utilization of 0 Instances and a request delta of 1 Instances.
My first assumption was IAM — the execution role was new, so I spent twenty minutes rewriting the trust policy. Wrong. CloudTrail showed the `CreateTrainingJob` call was authorized and the service itself rejected it. The real cause was a fresh account with GPU quota at zero, which is normal.
```bash
aws service-quotas list-service-quotas --service-code sagemaker --region us-east-1
--query "Quotas[?contains(QuotaName,'p3.2xlarge')].[QuotaName,Value]" --output table
That returned `0`. I filed a quota increase request through Service Quotas with a one-line justification, and it landed inside a day; accelerator quotas can take longer and sometimes want a support case. Verification was simple: the same command returned `1` and the job started.
Then the second failure, which is the more instructive one. I switched to managed Spot training on `ml.m5.xlarge` and after about 50 minutes got:
Training job mls-spot-demo-1731... was stopped due to MaxWaitTimeInSeconds
I had set `MaxWaitTimeInSeconds` equal to `MaxRuntimeInSeconds`, so there was no headroom for Spot capacity to come back after an interruption. The fix was checkpointing to S3 plus a wait time of roughly double the runtime — that combination means an interrupted job resumes from the last checkpoint instead of restarting. I verified it by deliberately requesting an interruption and watching the job resume from iteration ~400 rather than step 0. Two exam domains, one evening.
## Exam technique when you have 2:46 per question