MLA-C02 cheat sheet
Updated September 20, 2026
Last-minute reference for MLA-C02. Remember the shape: three of four domains are operations, not modelling.
Exam facts
| |
|---|
| Passing score | 720 / 1000 |
| Level | Associate |
| Validity | 3 years |
| Price | 75 USD while in beta |
Domain weights
| Domain | Weight |
|---|
| Data Preparation for ML and AI | 28% |
| ML Model and Foundation Model Development | 24% |
| Deployment and Orchestration | 24% |
| Operating, Monitoring, and Securing | 24% |
Failure modes
| Symptom | Name | Fix |
|---|
| Great offline, poor in production | Target leakage | Remove features unavailable at inference |
| Predictions differ from validation | Training-serving skew | One shared feature definition |
| Accuracy declines over months | Data or concept drift | Monitor against the training baseline |
| 99.7% accuracy, never predicts positive | Class imbalance | Precision, recall, AUC-PR; resample or weight |
| Wide variance across CV folds | Unstable estimate | Repeated or stratified CV |
Requirement to answer
| Requirement | Answer |
|---|
| 5% of traffic, instant rollback | Canary with weighted routing |
| 40M records overnight, no latency need | Batch inference job |
| Retrain when new labels land, with lineage | Event-triggered pipeline with artefact tracking |
| Traffic spikes at month end | Autoscaling with min and max |
| Training job reads one prefix | Execution role scoped to that prefix |
| Endpoint not on the public internet | VPC-private endpoint |
| Labels arrive two weeks late | Proxy monitoring now, true metrics later |
| Missing a positive case is costly | Recall |
| Large errors disproportionately costly | RMSE, not MAE |
Traps
- Random splits leak the future on time series. Split chronologically.
- Fit encoders in training and persist them; re-fitting at serving changes the mapping.
- Equal performance, simpler model wins. Marginal metric gains rarely justify operational complexity.
- Interruptible capacity with checkpointing cuts cost without risking quality; halving the dataset does not.
Night-before checklist
- Four domains, near-equal weight
- Skew versus drift versus leakage, cold
- Batch versus real-time serving decision
- Check ID and proctoring rules — see exam day
Take the 20-question practice test.