Based on the MLA-C01 exam guideReport an error
It's the last day of the 10 weeks. Today we cover not new knowledge but the skill of taking the exam. Even at the same skill level, pass or fail is decided by time management and question decomposition strategy.
| Item | Details |
|---|---|
| Number of questions | 65 questions (50 scored + 15 unscored) |
| Time | 130 minutes |
| Question types | Multiple choice (single/multiple answer), case-based, ordering, etc. |
| Passing score | 720 / 1000 (scaled score) |
| Result | Calculated from the 50 scored questions (the 15 unscored are not reflected in the score) |
The domain weights are as follows.
Domain 1 Data Preparation ████████████████████████████ 28%
Domain 2 Model Development ██████████████████████████ 26%
Domain 4 Monitoring and Security ████████████████████████ 24%
Domain 3 Deployment and Orchestration ██████████████████████ 22%
└ 1+2 = 54% · 3+4 = 46% ┘You can't abandon any one domain, but since Domains 1 and 2 make up more than half, stability in the data and model areas is the foundation of passing.
The 15 unscored questions are trial questions for future exams, and you can't tell which ones they are. So don't guess "this one must be unscored"; treat every question equally. The passing mark of 720 is a scaled score adjusted for question difficulty, so it can't be converted directly into a percentage of correct answers. Therefore the goal is not perfection but "a solid majority + avoiding traps."
With 65 questions in 130 minutes, you have 2 minutes per question on average. We recommend the following pace.
This is a checkpoint table for mid-exam checks. By comparing it with the remaining time on the actual exam screen, you can tell right away whether you've fallen behind.
| Elapsed time | You should be done through here | Action if you're behind |
|---|---|---|
| 30 min | About 15 questions | Tighten the flag-and-move-on threshold from 2 minutes to 90 seconds |
| 60 min | About 30 questions | For long case-based questions, read only the question and qualifiers and win by elimination |
| 90 min | About 45 questions | Be prepared to cut review time down to 15 minutes |
| 100 min | First pass of 65 questions complete | Start reviewing in order of flagged questions |
| 125 min | Review ends | Final check of blanks and multiple-answer counts |
[130-minute allocation]
├──────────── First pass 90-100 min ────────────┤── Second review 25-30 min ──┤─ Third 5 min ─┤
All questions once · guess and flag if stuck Re-review flagged questions Blanks · counts⚠️ Trap: The most common failure is "pouring 10 minutes into one hard question and then not being able to solve the easy ones later." Every question is worth the same. The discipline of guessing, flagging, and moving on immediately when stuck protects your score. Also, always guess on a question you have no answer for — there is no penalty for wrong answers, so a blank is a pure loss.
Decompose long case-based questions in the following order.
[4-step decomposition]
① Last question ─▶ "What am I being asked to choose" + qualifier (MOST/LEAST)
▼
② Scan the body ─▶ Number the condition keywords (conditions ①②③)
▼
③ Answer before options ─▶ Write the feature name in your head using the translation table
▼
④ Eliminate options ─▶ Remove options that break any condition → the one remaining| Type | Characteristics | Approach |
|---|---|---|
| Single-answer multiple choice | 1 of 4 options | Recall the answer first with the translation table, then find it among the options |
| Multiple-answer (TWO, etc.) | You must pick the exact number and there is no partial credit | Check the count first, lock in the sure ones, then compare the rest |
| Long case-based | 2-3 condition keywords are scattered | Question and qualifier first → number the conditions → the option that satisfies all conditions |
| Matching/ordering | Pair items or arrange an order | Fix the pairs you're confident about and fill in the rest by elimination |
| "Which is incorrect?" | Negative question | Always check whether the question contains a negative. Find the 3 correct ones and eliminate them |
⚠️ Trap: If you miss a negative ("NOT", "which is incorrect"), you pick the opposite even though you know the answer. It's a mistake that is easier to make the more hurriedly you read. The habit of giving a second look to negatives and qualifiers (MOST, LEAST, TWO) when reading the question sentence prevents this loss.
Multiple-answer questions (pick 2 or more) need a separate strategy. When a question explicitly says "TWO" or "TWO actions," you must pick exactly that number, and getting only one right earns no partial credit. Also, superlative qualifiers such as "MOST cost-effective," "LEAST operational overhead," and "MOST secure" are the key. Even if several options technically work, only the single best one by the qualifier's criterion (cost, operational burden, security) is the answer. If you miss the qualifier, you pick a trap option that "works but isn't the best."
The day before and the morning of the exam, quickly review the following.
| Area | One-line check |
|---|---|
| Inference options | Separate real-time/serverless/asynchronous/batch by immediacy, traffic, and payload |
| Advanced deployment | Many similar models = MME, a few models from different frameworks = MCE, preprocessing chain = inference pipeline |
| Algorithms | Time series = DeepAR, anomaly = RCF, tabular = XGBoost, recommendation = FM, text = BlazingText |
| Evaluation metrics | Imbalance = F1/AUC, missed cases = Recall, false alarms = Precision, regression = RMSE |
| Training infrastructure | Large data = Pipe/FastFile, repeated reads = FSx, cost = Spot + checkpoints |
| Tuning | Few trials = Bayesian, early stopping = Hyperband |
| Data preparation | Visual = Data Wrangler, script = Processing Job, ETL = Glue, querying = Athena |
| Features | Training-inference consistency = Feature Store (online single record/offline bulk) |
| Monitoring | Input drift = Data Quality, performance degradation = Model Quality (+labels) |
| Bias and explainability | Clarify (pre/post-training bias, SHAP), ongoing monitoring in production is Model Monitor |
| Security | IAM least privilege · VPC endpoints (S3 = Gateway, the rest = Interface) · isolation · KMS |
| Governance | Who = CloudTrail, metrics = CloudWatch, lineage = Lineage, documentation = Model Cards |
| MLOps | EventBridge → Pipelines (ConditionStep) → Model Registry |
| Cost | MME · Inferentia · serverless/asynchronous · Spot · batch transform |
Right before the exam, solidifying the retrieval paths of what you already know is more effective than cramming in new concepts. Retrieval practice, in which you map the table above out loud from keyword → feature, helps with recall in the exam room far more than reading the information anew. Sleep enough, arrive at the exam site early, and don't be shaken even if you get stuck on the first few questions — hard questions clustered at the front is just a common arrangement.
This is a check of the items to bring and the environment, so you don't lose points for reasons outside knowledge. The detailed rules differ depending on how you take the exam (test center / online proctored), so treat the instructions in your booking confirmation email as the final authority.
| Category | Check item |
|---|---|
| The day before | Get enough sleep. Don't start a new course or a new problem book |
| The day before | Recheck the booked time, exam delivery method, and location (or the online proctoring procedure) |
| What to bring on the day | The ID specified in the booking confirmation. Check that the name spelling matches the booking |
| Test center | Arrive with time to spare. Check the rules for storing belongings |
| Online exam | Run the system check beforehand, tidy your desk, and secure a quiet, undisturbed space |
| Right after starting | First check where the remaining time is displayed and where the flag button is |
| During the exam | Compare your progress at the 30/60/90-minute marks. If you're behind, adjust your pace immediately |
| Before finishing | Confirm 0 blanks and matching counts on multiple-answer questions, then submit |
Don't read; say the answer first, then check the right side. If any row stalls you, review only that week's content again.
| Question | Answer |
|---|---|
| Which inference option doesn't scale to 0? | Real-time (keeps at least 1 instance) |
| Individual requests that are large and take a long time? | Asynchronous inference |
| Scoring the whole dataset at once? | Batch transform |
| Thousands of models on the same framework? | Multi-model endpoint (MME) |
| To reuse the training-time preprocessing as is at inference? | Inference pipeline or Feature Store |
| Finding outliers without labels? | Random Cut Forest |
| Forecasting the future of multiple related time series? | DeepAR |
| Performance metric for 99% normal data? | F1, AUC, Recall (not Accuracy) |
| Metric for a problem you can't afford to miss? | Recall |
| Which monitor only runs when ground-truth labels exist? | Model Quality |
| Which monitor watches input distribution changes? | Data Quality |
| One-time bias check before deployment? | Clarify |
| To reach S3 without the internet? | Gateway VPC endpoint |
| To reach ECR without the internet? | Interface endpoint (PrivateLink) |
| To cut off a container's outbound communication? | EnableNetworkIsolation |
| To check "who deleted it"? | CloudTrail |
| To trigger retraining when new data arrives? | EventBridge → Pipelines |
| To block registration of a model that misses the bar? | ConditionStep |
| To cut training cost but still be able to resume? | Managed Spot + checkpoints |
| To find hyperparameters with few trials? | Bayesian |
The reason the cards above work is that their format is the same as the exam's. What you need in the exam room is not "knowledge you can explain" but the speed of pulling out a single name from a short cue. Rereading material makes you feel like you know it (the illusion of familiarity), but when you actually try to pull it out, the rows where you stall are revealed. The final review must be done by pulling things out.
MLA-C01 is ultimately an exam about translating requirements into SageMaker features. Over 10 weeks we've gone full circle, from collection, feature engineering, Feature Store, and Clarify through training, tuning, and evaluation, inference options and advanced deployment, Pipelines and MLOps, and Model Monitor, IAM, VPC, and KMS. With stability in data preparation (Domain 1) and model development (Domain 2) as your foundation, you just need to recall the keyword mappings for deployment (Domain 3) and monitoring and security (Domain 4) reflexively.
Click a choice to reveal the answer and explanation.
Question 1
Which statement about the MLA-C01 exam structure is most accurate?
Question 2
During the first pass, you can't be sure of the answer to a scenario question after 2 minutes. What is the most recommended action?
Question 3
Which is the most appropriate recommended order for solving a long case-based question efficiently?
Question 4
What is the key point of caution when solving a question that includes a superlative qualifier such as "Which of the following is the MOST cost-effective option?"
Question 5
What is the most effective final review method right before the exam?