Based on the MLA-C01 exam guideReport an error
Today we cut back on concept explanations and solve integrated scenarios mixing the 4 domains at the same pace as the real exam. Solve the 8 questions below within 16 minutes if possible, and before looking at the answer, it is good to first say out loud "which domain is this question, and what is the keyword that decides the answer?"
| Step | What to do | Time |
|---|---|---|
| First pass | Go through all 8 questions without stopping | 16 min |
| Before self-grading | Write down each question's domain and deciding keyword | 4 min |
| Grading and review | For wrong answers, explain "why the other options were wrong" | No limit |
💡 Related theory: An integrated question usually has 2-3 "condition keywords" embedded in one sentence. Example: "hundreds of same-type models (→ MME candidate)" + "invoked occasionally (→ cost savings reinforced)" + "same framework (→ MME confirmed)". If you separate the keywords, map each to a feature, and then choose the option that satisfies all the keywords at the same time, you can filter out trap options.
As you read a question, first pinpoint "which cell of this map are we in now," and many of the options are eliminated immediately.
Ingestion Storage Preparation Training Tuning
───── ───── ───── ───── ─────
Kinesis ─▶ S3 ─▶ Data Wrangler ─▶ Training Job ─▶ AMT
Firehose Glue Processing Job Built-in/Custom Bayesian
Ground Truth Catalog Glue ETL/EMR File·Pipe Random·Grid
(labeling) Athena(query) FastFile Hyperband
Feature Store Spot+checkpoint
Clarify(pre-training bias) Debugger
│
┌──────────────────────────────────────────────┘
▼
Evaluation Deployment Monitoring Security
───── ───── ────────── ─────
Choose metric ─▶ Model Registry(version·approval) ─▶ Data capture ─▶ IAM (least privilege)
Accuracy/F1 Real-time · Serverless Model Monitor VPC (endpoints·isolation)
AUC · RMSE Async · Batch Transform ├ Data Quality KMS (encryption at rest)
Clarify MME · MCE · Inference Pipeline ├ Model Quality TLS (encryption in transit)
(post-training bias) Variant weights (A/B) ├ Bias Drift CloudTrail(audit)
Pipelines · EventBridge └ Attribution Drift CloudWatch(metrics)
│
CloudWatch alarm ─▶ retraining trigger| Keyword | Correct feature |
|---|---|
| Real-time stream ingestion | Kinesis Data Streams / Firehose |
| Bulk loading into a data lake | S3 |
| Serverless ETL, schema catalog | Glue / Glue Data Catalog |
| Visual cleaning without code | Data Wrangler |
| Run my own preprocessing script as a managed job | Processing Job |
| SQL queries without moving S3 data | Athena |
| Humans must attach labels | Ground Truth |
| Training-inference feature consistency and reuse | Feature Store |
| Bias check before deployment / explaining feature contribution | Clarify (SHAP) |
| Handling class imbalance | Resampling, class weights |
| Keyword | Correct feature |
|---|---|
| Demand for the next N days, multiple related time series | DeepAR |
| Unlabeled outliers, stream anomalies | Random Cut Forest |
| Structured table classification/regression | XGBoost, Linear Learner |
| Grouping without labels | K-Means |
| Too many features | PCA |
| Sparse user x item recommendation | Factorization Machines |
| Document classification, word embeddings | BlazingText |
| Efficient tuning with few trials | Bayesian |
| Early stopping of hopeless training | Hyperband |
| Training data larger than the disk | Pipe / Fast File mode |
| I/O bottleneck from repeated reads every epoch | FSx for Lustre |
| Reduce training cost (interruption allowed) | Managed Spot + checkpoints |
| Diagnosing bottlenecks and overfitting during training | Debugger |
| Keyword | Correct feature |
|---|---|
| Steady traffic, immediate response | Real-time endpoint |
| Intermittent traffic, scale to zero, cold start acceptable | Serverless inference |
| Large payload, long processing, individual requests | Asynchronous inference |
| Bulk scoring of an entire dataset | Batch Transform |
| Hundreds to thousands of same-type models | MME |
| A few models with different frameworks | MCE |
| Apply the same preprocessing to inference as well | Inference pipeline |
| A/B testing, gradual migration | Production variant weights |
| ML workflow DAG, CI/CD | SageMaker Pipelines |
| Register only when criteria are met | ConditionStep + Model Registry |
| Retrain when new data arrives | EventBridge → Pipeline |
| Reduce deep learning inference unit cost | Inferentia |
| Keyword | Correct feature |
|---|---|
| Whether the input distribution changed | Model Monitor (Data Quality) |
| Whether performance against the truth dropped | Model Monitor (Model Quality) + ground truth |
| Bias change in production | Bias Drift (Clarify integration) |
| Feature contribution change in production | Feature Attribution Drift |
| Access S3 and ECR without the internet | VPC endpoints (Gateway/Interface) |
| Block container outbound communication | EnableNetworkIsolation |
| Encrypt data at rest | KMS |
| Encrypt data in transit | TLS |
| Who called which API | CloudTrail |
| Latency and error-rate alarms | CloudWatch |
| Model documentation and governance | Model Cards |
| Data → training → model lineage | Lineage Tracking |
| A role with excessive permissions | IAM least privilege |
[1) Inference options]
Need an immediate response? ──No──▶ Bulk over the whole dataset? ──Yes──▶ Batch Transform
│Yes └No (individual, large, long) ──▶ Asynchronous
▼
Steady traffic? ──Yes──▶ Real-time (minimum 1 instance, cannot scale to 0)
└No ──▶ Cold start acceptable? ──Yes──▶ Serverless
└No ──▶ Real-time + auto scaling
[2) Algorithm]
Are there labels? ──No──▶ Looking for outliers? ──Yes──▶ RCF
│Yes └No ──▶ Grouping K-Means / dimensionality reduction PCA
▼
Is the time axis central? ──Yes──▶ DeepAR
└No ──▶ What is the data?
├ Table ──▶ XGBoost / Linear Learner
├ Text ──▶ BlazingText
├ Image ──▶ Image Classification / Object Detection
└ Sparse interactions ──▶ Factorization Machines
[3) Evaluation metric]
Classification? ──No──▶ Regression: RMSE (sensitive to large errors) / MAE (dampens outliers) / R²
│Yes
▼
Imbalanced? ──No──▶ Accuracy is OK
│Yes
▼
Is the cost of a miss high? ──Yes──▶ Recall
└No ──▶ Is the cost of a false alarm high? ──Yes──▶ Precision
└No ──▶ F1 / AUC-ROCWhen a sentence has several conditions, what decides at the end is usually the qualifier.
| Qualifier | Comparison axis | Answer direction |
|---|---|---|
| MOST cost-effective | Cost | MME, serverless, batch, Spot, Inferentia |
| LEAST operational overhead | Operational burden | Managed and serverless services (Glue, Pipelines, serverless inference) |
| MOST secure | Security | VPC endpoints + network isolation + KMS + least privilege |
| LEAST latency | Latency | Real-time endpoint, Feature Store online, caching |
| FASTEST to implement | Implementation speed | Built-in algorithms, Data Wrangler, managed features |
| MINIMUM code changes | Code changes | Keep the existing setup + solve with configuration changes |
⚠️ Trap: Several options technically all work. The superlative qualifier is the only criterion that narrows the answer to one. An option that breaks the performance or availability requirement while "reducing cost" (scaling real-time to zero, discarding data, using a GPU for XGBoost) is wrong even if the cost goes down. Conversely, for "reduce operational burden," an option that builds directly on EC2 is wrong even if it works.
This is an example not of the question format but of the habit of decomposing conditions. Follow with your eyes which cell each condition goes to.
Case A - "There are 2,000 per-store XGBoost models across the country, and each is invoked only a few times a day. Minimize cost."
| Condition fragment | Translation |
|---|---|
| 2,000 models | Many models → MME candidate |
| All XGBoost | Same framework → MME confirmed (MCE eliminated) |
| Invoked only a few times a day | Low frequency → sharing instances is advantageous |
| Minimize cost | Qualifier = cost |
Case B - "You train on regulated data. It must not pass through the internet, the container must not send data outside, and data at rest must be encrypted."
| Condition fragment | Translation |
|---|---|
| No internet traversal | VPC endpoints (Gateway for S3, Interface for ECR) |
| Block outbound communication | EnableNetworkIsolation |
| Encryption at rest | KMS |
| 3 conditions at once | Only an option containing all three is correct |
Case C - "A real-time fraud detection model has been deployed. Monitor whether accuracy drops over time, and retrain automatically if it does."
| Condition fragment | Translation |
|---|---|
| Whether accuracy drops | Model Quality monitor |
| Prerequisite for computing accuracy | Need to collect ground truth labels |
| Automatic retraining | CloudWatch alarm → EventBridge → Pipeline |
| Real-time fraud | Imbalanced, so the metric is not Accuracy |
When the conditions span two domains as monitoring + automation, as in Case C, one of the options usually does "monitoring only" and another does "automation only." The option that contains both is the answer. Conversely, an option that adds a lot of features when there is only one condition is over-engineered and loses under the "LEAST operational overhead" criterion. Matching the number of conditions against the number of components in each option is also a quick elimination method.
| Symptom (how people tend to get it wrong) | Cause | Fix |
|---|---|---|
| Chose Accuracy for imbalanced data | Reflexive reaction to the word "accuracy" | Build the habit of first finding the sentence with the data ratio |
| Chose the option that scales a real-time endpoint to 0 | Forgot the real-time minimum of 1 instance | Scaling to 0 is only for serverless, asynchronous, and batch |
| Chose Batch Transform for individual requests | Judged only by payload size | "Individual requests" means asynchronous, "entire dataset" means batch |
| Chose KMS for a network requirement | Simple association of security = encryption | For network use VPC endpoints and isolation; for storage use KMS |
| Chose CloudWatch for an audit question | Confused because both are "logs" | "Who" is CloudTrail; "how slow and when" is CloudWatch |
| Chose a regression algorithm for a time series | Association of forecasting = regression | If the time axis is central, DeepAR |
| Chose a technically workable option and it was wrong | Did not read the superlative qualifier | Underline the qualifier in the last line of the question |
| Chose an option that satisfies only 2 of 3 conditions | Did not count the conditions to the end | Count the conditions and check each option |
Now it is time for the real thing. Solve 8 questions. Tomorrow we reorganize these wrong-answer patterns by trap type.
Click a choice to reveal the answer and explanation.
Question 1
A fintech company wants to detect credit card fraud in real time. 99.7% of transactions are legitimate, and the cost of missing fraud is very high. Which evaluation metric should take priority?
Question 2
There are per-store demand forecasting XGBoost models for 2,000 stores nationwide. Each model is invoked only a few times a day, and the cost of 2,000 endpoints is a burden. What is the most cost-effective deployment?
Question 3
A medical imaging model receives a 250 MB image per individual request and runs inference for about 8 minutes, and the result is retrieved from S3 after completion. What is the most appropriate inference option?
Question 4
You train on regulated data, and the training container must access S3 and ECR without going through the internet and must block data exfiltration to the outside. What is the most appropriate setup?
Question 5
A data scientist wants to visually explore and clean an S3 dataset with almost no code, apply missing-value imputation and encoding, and then export that flow into a training pipeline. What is the most appropriate tool?
Question 6
You want to monitor whether the prediction accuracy of a production model drops over time (against the actual truth). What setup is needed?
Question 7
You want to forecast demand for the next 30 days from multivariate time series sales data. Which SageMaker built-in algorithm is best suited to learning the patterns of multiple related time series together?
Question 8
Whenever new training data arrives in S3, you want to automatically perform retrain → evaluate → (if criteria are met) register, and manage model versions and approval status. What is the most appropriate combination?