Overview
Data Preparation for ML and AI has three task statements. Task 1.1 covers collecting and storing data: extracting from Amazon S3, EBS, EFS, RDS, DynamoDB and OpenSearch Service, streaming ingestion with Kinesis, Flink or Kafka, columnar formats such as Parquet and ORC, vector databases and SageMaker Feature Store. Task 1.2 covers transformation with AWS Glue, DataBrew, Spark on Amazon EMR and SageMaker Data Wrangler, feature engineering, embeddings, tokenization, RAG chunking, masking and data for FM fine-tuning or distillation.
Task 1.3 covers data quality, labeling, bias metrics, class imbalance, training data integrity and cleaning. MLA-C02 has been the exam in use since September 29, 2026, delivered as an English beta; MLA-C01 is no longer offered in English, and AWS lists general availability for January 14, 2027.
How it's tested
Domain 1 scenarios describe raw data and a target: a model to train, a feature store to populate or a knowledge base to fill. You choose the ingestion path, storage format, transformation tool or quality control. The best answer usually meets the stated constraint, such as cost, latency or compliance, with the least custom code.
-
Match store and format to access
A workload reads a few columns from huge tables or needs millisecond lookups, and you pick Parquet on S3, a feature store online store or a vector index.
-
Choose the transformation tool
A team needs ETL, visual wrangling or stream processing, and you choose among Glue, DataBrew, EMR, Data Wrangler and Lambda.
-
Fix data before training
Missing values, duplicates, skewed classes, PII or unsafe prompt-response pairs appear, and you pick the cleaning, masking or balancing step.
A fraud dataset has 0.5% positive cases and the model predicts every transaction as legitimate. Accuracy looks excellent, yet recall on fraud is near zero. The data problem comes first: rebalance with oversampling such as SMOTE, undersampling or class weights, and split with stratification so test data keeps the real ratio. Collecting more legitimate transactions would only deepen the imbalance.
Task statements 3
Official task statements from the Amazon Web Services outline, in outline order.
1.1 Collect and store dataSources, storage choices, streaming ingestion, formats, merging, vector stores and Feature Store ingestion.
- What it means
- This task statement covers pulling data from S3, EBS, EFS, RDS, DynamoDB and OpenSearch Service, choosing storage by cost, performance, structure and compliance, troubleshooting capacity issues, streaming with Kinesis, Flink and Kafka, picking Parquet, ORC, JSON or CSV, merging sources with Glue or Spark, configuring vector databases, storing text, image and audio data, and ingesting into SageMaker Feature Store.
- What to understand
- Amazon S3 is the default data lake for training data; EFS and FSx for Lustre suit shared file access for training clusters. Columnar Parquet and ORC cut scan cost when jobs read a subset of columns, while JSON and CSV suit small or row-wise exchange. Kinesis Data Streams captures real-time events, Amazon Data Firehose delivers them to S3 with optional format conversion, and Amazon Managed Service for Apache Flink processes streams. For vectors, OpenSearch Service, RDS or Aurora PostgreSQL with pgvector and Amazon S3 Vectors are listed options. Feature Store has an online store for low-latency inference and an offline store in S3 for training.
- How it shows up
- A data source, volume and access pattern are given, and you choose the ingestion service, storage target and file format.
- Common confusion
- Firehose delivers and lightly transforms streams; it is not a general stream processor like Flink.
1.2 Perform data transformation, feature engineering, and pre-processingAWS transformation tools, features, streaming transforms, embeddings, text pre-processing, RAG documents, masking and FM training data.
- What it means
- This task statement covers transforming data with AWS Glue, Glue DataBrew, Spark on EMR and Data Wrangler, managing features in Feature Store, transforming streams with Lambda or Spark, feature engineering techniques, embedding models for text and images, tokenization and augmentation, chunking and metadata extraction for RAG, masking and anonymization, and preparing data for fine-tuning, continued pre-training and distillation.
- What to understand
- Scaling and standardization keep features with large ranges from dominating distance-based or gradient-based models; log transforms tame skew; binning turns continuous values into categories; one-hot encoding suits nominal categories. Embedding models such as Amazon Titan Text Embeddings or Cohere Embed on Bedrock turn text or images into vectors, and the same model must embed both documents and queries. RAG chunking can be fixed-size, hierarchical or semantic, with overlap and metadata for filtering. Amazon Comprehend and Macie help find PII before masking. Fine-tuning data on Bedrock is labeled prompt-completion JSONL, while continued pre-training uses unlabeled text.
- How it shows up
- A dataset or document set needs reshaping for a model, and you pick the transformation, encoding, chunking or masking approach and tool.
- Common confusion
- Fine-tuning needs labeled prompt and response pairs; continued pre-training uses large volumes of unlabeled domain text.
1.3 Validate data quality and manage biasData quality rules, labeling, bias sources and metrics, class imbalance, training data integrity and cleaning.
- What it means
- This task statement covers validating quality with DataBrew and AWS Glue Data Quality, labeling and annotation, finding and reducing bias with splitting, shuffling and augmentation, applying bias metrics across numeric, text and image data, resolving class imbalance, validating AI training data such as prompt-response pairs and content safety, and cleaning data by handling outliers, missing values and duplicates.
- What to understand
- Glue Data Quality evaluates rules written in Data Quality Definition Language and can stop a pipeline when checks fail. SageMaker Ground Truth and Ground Truth Plus produce labels with human workers and active learning. SageMaker Clarify computes pre-training bias metrics such as class imbalance and difference in proportions of labels before any model exists. Imbalance is handled with oversampling, SMOTE, undersampling, augmentation or class weights. Missing values can be imputed with mean, median or model-based methods; outliers can be capped or removed after investigation. For FM datasets, check prompt-response pairs for correctness, format and harmful content before training.
- How it shows up
- A dataset shows quality, labeling or fairness problems, and you choose the AWS tool or technique that detects or fixes them before training.
- Common confusion
- Pre-training bias metrics describe the dataset; post-training metrics in Clarify describe the model's predictions.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Apache Parquet and ORC
Columnar, compressed file formats.
- Remember
- Best when queries read a few columns.
- Tested as
- Format selection questions.
- Trap
- Choosing CSV for large analytical scans.
-
Amazon Kinesis Data Streams
Durable real-time stream with shards and consumers.
- Remember
- Captures events for custom or Flink processing.
- Tested as
- Streaming ingestion questions.
- Trap
- Confusing it with Firehose delivery.
-
Amazon Data Firehose
Managed delivery of streams to S3, Redshift or OpenSearch.
- Remember
- Can convert JSON to Parquet in flight.
- Tested as
- Low-effort streaming to storage.
- Trap
- Expecting complex stateful processing.
-
SageMaker Feature Store
Central store for features with online and offline stores.
- Remember
- Online for inference lookups, offline for training.
- Tested as
- Feature reuse and consistency questions.
- Trap
- Using only the offline store for real-time inference.
-
SageMaker Data Wrangler
Visual data preparation inside SageMaker Studio.
- Remember
- Over 300 built-in transforms and export to pipelines.
- Tested as
- Low-code preparation questions.
- Trap
- Picking it for petabyte-scale scheduled ETL.
-
AWS Glue and Glue Data Quality
Serverless Spark ETL with a crawler, catalog and quality rules.
- Remember
- Rules use DQDL and can fail a job.
- Tested as
- ETL and validation questions.
- Trap
- Writing custom scripts for checks Glue already runs.
-
Embeddings
Numeric vectors that capture meaning of text or images.
- Remember
- Index and query with the same embedding model.
- Tested as
- Vector search and RAG preparation questions.
- Trap
- Mixing models between indexing and querying.
-
Chunking strategy
How documents are split before embedding.
- Remember
- Fixed-size, hierarchical or semantic, with overlap and metadata.
- Tested as
- RAG preparation questions.
- Trap
- Chunks so large that retrieval returns noise.
-
Class imbalance remedies
Oversampling, SMOTE, undersampling, augmentation and class weights.
- Remember
- Stratify splits to keep ratios.
- Tested as
- Skewed dataset questions.
- Trap
- Judging the model by accuracy alone.
-
SageMaker Clarify pre-training bias
Bias metrics computed on data before training.
- Remember
- Class imbalance and DPL are examples.
- Tested as
- Bias detection questions.
- Trap
- Waiting for a deployed model to look for data bias.
-
PII masking and redaction
Removing or obscuring personal data before use.
- Remember
- Comprehend and Macie can detect PII.
- Tested as
- Compliance data questions.
- Trap
- Training an FM on raw customer records.
-
Fine-tuning vs. continued pre-training data
Labeled prompt-completion pairs vs. unlabeled domain text.
- Remember
- Both use JSONL in S3 for Bedrock.
- Tested as
- FM data preparation questions.
- Trap
- Supplying unlabeled text for fine-tuning.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Streaming ingestion services
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Kinesis Data Streams | Real-time stream with retention and replay | You build or attach consumers | Expecting automatic delivery to S3 |
| Amazon Data Firehose | Managed loading into storage targets | Buffers and converts, no custom consumers | Using it for stateful analytics |
| Managed Service for Apache Flink | Stateful stream processing | Windows, joins and aggregations in real time | Choosing it just to land files in S3 |
| Amazon MSK | Managed Apache Kafka | Fits teams already on Kafka | Assuming it is serverless by default |
Data transformation tools
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| AWS Glue | Serverless Spark ETL and Data Catalog | Scheduled, code or visual ETL at scale | Using it for interactive notebook exploration only |
| AWS Glue DataBrew | No-code cleaning with recipes and profiles | Analysts without Spark skills | Expecting custom model training |
| SageMaker Data Wrangler | Visual ML prep in Studio | Feature-focused, exports to Pipelines or Feature Store | Treating it as a streaming tool |
| Spark on Amazon EMR | Managed big data clusters | Full control over large custom jobs | Picking it when serverless Glue is enough |
Imbalance and bias fixes
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Oversampling or SMOTE | Adds or synthesizes minority examples | Keeps all majority data | Applying it before the train-test split |
| Undersampling | Removes majority examples | Faster but loses information | Using it on small datasets |
| Class weights | Penalizes minority errors more | No change to the data itself | Forgetting it is an option |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Firehose for analytics
- What candidates pick
- Use Amazon Data Firehose to compute rolling fraud scores on a stream.
- Why it looks right
- Firehose is the easiest streaming service to set up.
- Why it's wrong
- Firehose buffers and delivers data; windowed, stateful calculations need Managed Service for Apache Flink or a custom consumer on Kinesis Data Streams.
Clue to watch for The scenario mentions aggregations, windows or joins in real time.
-
Offline store for live lookups
- What candidates pick
- Read features for a real-time endpoint from the Feature Store offline store.
- Why it looks right
- The offline store holds the full history.
- Why it's wrong
- The offline store lives in S3 and is meant for training and batch use; low-latency inference reads from the online store.
Clue to watch for Millisecond latency at inference time.
-
Oversample before splitting
- What candidates pick
- Apply SMOTE to the whole dataset, then split into training and test sets.
- Why it looks right
- It seems simpler to balance once.
- Why it's wrong
- Synthetic points leak into the test set and inflate results; resample only the training split after a stratified split.
Clue to watch for The question asks for a reliable evaluation.
-
Bigger chunks are safer
- What candidates pick
- Embed whole documents as single chunks so no context is lost.
- Why it looks right
- Fewer chunks seems to preserve meaning.
- Why it's wrong
- Very large chunks dilute embeddings and exceed context budgets, so retrieval returns vague matches; sized chunks with overlap and metadata retrieve better.
Clue to watch for Users report irrelevant answers from a knowledge base.
-
Drop every outlier
- What candidates pick
- Delete all values outside three standard deviations.
- Why it looks right
- Outliers look like noise.
- Why it's wrong
- In fraud or failure detection the outliers are often the signal; investigate first, then cap, transform or keep them.
Clue to watch for The target event is rare.
-
Custom quality scripts
- What candidates pick
- Write Lambda code to check null counts and value ranges after each ETL run.
- Why it looks right
- Custom code feels flexible.
- Why it's wrong
- AWS Glue Data Quality and DataBrew profile jobs already evaluate such rules with less code and can stop the pipeline on failure.
Clue to watch for Least operational overhead.
Answer framework
Use these steps for Domain 1 questions about ingestion, storage, transformation and data quality.
- Identify the data shape and arrival pattern: batch files, change data, real-time events, documents for retrieval or media files.
- Pick storage by access: Parquet on S3 for analytics, Feature Store online store for low-latency features, a vector index for semantic search.
- Choose the lightest transformation tool that fits: DataBrew or Data Wrangler for visual work, Glue for serverless ETL, EMR for heavy custom Spark.
- Check quality and fairness before training with Glue Data Quality rules, Clarify pre-training bias metrics and imbalance fixes on the training split.
- Protect and shape GenAI data: mask PII, chunk with metadata, embed consistently and format fine-tuning pairs as JSONL.
Practice question
Want more like this? The MLA-C02 practice test has questions for every domain.
Checklist
0 of 14
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Start with storage and formats, because almost every Domain 1 scenario begins there. Build a small table of S3, EFS, FSx for Lustre, RDS, DynamoDB and OpenSearch Service, with the access pattern each one serves. Add the streaming services next and practice saying in one line what Kinesis Data Streams, Firehose, Flink and MSK each do that the others do not.
Then work hands-on. Load a public dataset into S3, crawl it with Glue, convert it to Parquet and add a Glue Data Quality ruleset. Open Data Wrangler in SageMaker Studio, apply scaling and encoding, and export features to Feature Store. Run a Clarify pre-training bias report on the same data.
The GenAI additions deserve their own session. Create a small Bedrock knowledge base, try fixed-size and hierarchical chunking, and compare the retrieved passages. Write ten prompt-completion pairs in JSONL to see what a fine-tuning dataset looks like, and note how continued pre-training data differs.
Finish with imbalance and cleaning drills: given a skewed dataset, decide on resampling, class weights or new metrics, and explain where the split must happen.
FAQ
How much of MLA-C02 is Domain 1?
Data Preparation for ML and AI is 28% of scored content, roughly fourteen of fifty scored questions, across three task statements. It is the largest domain on the exam.
Which MLA exam version is current?
MLA-C02 has been in use since September 29, 2026 as an English beta. MLA-C01 ended in English on September 28, 2026 and stays in Japanese, Korean and Simplified Chinese until MLA-C02 general availability, which AWS lists for January 14, 2027.
What GenAI data skills did MLA-C02 add to Domain 1?
Vector database configuration, embeddings, advanced text pre-processing, RAG document preparation, masking and anonymization, FM fine-tuning data, training data integrity checks and data cleaning.
When should I use Data Wrangler instead of AWS Glue?
Use Data Wrangler for visual, feature-focused preparation inside SageMaker Studio. Use Glue for scheduled, serverless ETL over large datasets and for catalog-driven pipelines.
What is the difference between the Feature Store online and offline stores?
The online store serves the latest feature values with low latency for inference. The offline store keeps historical values in S3 for training and batch scoring.
How do I handle class imbalance on the exam?
Use oversampling, SMOTE, undersampling, augmentation or class weights on the training data only, keep a stratified test set and judge results with recall, precision or F1.
Which vector stores does the exam guide name?
OpenSearch Service, Amazon RDS with pgvector and Amazon S3 are given as examples. Bedrock knowledge bases also support other stores such as Aurora PostgreSQL.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 1 knowledge
You have covered ingestion, storage formats, feature engineering, RAG data preparation, data quality and bias. Check whether you can pick the right data tool under pressure.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the MLA-C02 AI tutor Get any concept explained your way