CertiPrepare before you train
- Parquet for columns you will reread
- Feature Store for one definition online and offline
- Redact before you embed
- Chunk so retrieval returns a section
Study guide · Flashcards & memory notes
AWS ML Engineer Flashcards & Memory Notes
40 recall cards for the distinctions MLA-C02 keeps testing: how the data is prepared, when a foundation model is enough, how a release is versioned, and which control stops the bad output.
How to use these cards: Say the answer out loud before you flip the card, and mark it honestly. Revisit the cards you missed tomorrow rather than rereading them right away. Progress is kept only while this page is open. For the full explanations behind each card, see the core notes.
Study mode
Choose a deck, flip each card, and mark what you already know.
You marked every card in this deck as known. Shuffle and run through it again, or choose another deck.
All flashcards
Select a question to reveal its answer.
Deck 1 · Data preparation
Formats, features, retrieval prep, and data quality.
1.1What is the difference between a CSV landing file and Parquet for training?
Parquet is columnar, so a later job can read fewer columns. CSV stores whole rows and makes every reread scan more bytes.
1.2What does SageMaker Feature Store share between training and inference?
The same feature definition and values. A one-off file in S3 can drift from the live request.
1.3Which vector stores does the MLA-C02 data task name?
Amazon OpenSearch Service, Amazon RDS with pgvector, and Amazon S3.
1.4What is the difference between chunking and tokenization?
Chunking splits a document into spans a retriever can return. Tokenization splits text into the units a model consumes.
1.5What metadata helps a RAG index skip an obsolete policy?
Status and effective date, stored with the chunk, so retrieval can filter drafts out.
1.6When do you mask a value instead of deleting the row?
When the row is still useful and the sensitive value must not be seen. Deletion is for data that must not be retained.
1.7What is class imbalance?
One label or condition appears far more often than another. Accuracy can look high while the rare class is ignored.
1.8Which tools validate a data-quality rule before training?
AWS Glue Data Quality and AWS Glue DataBrew.
1.9What data shape does an instruction fine-tune need?
Prompt-and-response pairs you have checked for quality and safety.
1.10Why drop a column that exists only in the training file?
The live request cannot supply it, so the served model will not match the trained one.
Deck 2 · Model development
Bedrock, SageMaker AI training, and evaluation.
2.1When do you start with Amazon Bedrock instead of a training job?
When a foundation model, a prompt, or retrieval can meet the quality, latency, and cost bar.
2.2What is SageMaker AI automatic model tuning?
A search over hyperparameters. It is not a content filter and it is not a vector index.
2.3What does early stopping prevent?
More epochs after validation performance has stopped improving.
2.4What is catastrophic forgetting?
A fine-tune that learns the new examples and loses earlier useful behavior.
2.5What do BLEU and ROUGE measure?
Overlap between generated text and a reference. BERTScore is the semantic-similarity metric named beside them.
2.6What is a shadow variant?
A candidate that sees a copy of production traffic and does not replace the user-visible answer.
2.7Which service extracts text from a scanned form?
Amazon Textract. Amazon Transcribe is speech to text. Amazon Polly speaks.
2.8When is script mode the simpler path's alternative?
When a built-in SageMaker AI algorithm does not cover the framework and training loop you must run.
2.9What does retrieval accuracy tell you that BLEU on the final sentence mixes together?
Whether the right passage was retrieved, separate from how the generator wrote the answer.
2.10What is LLM-as-a-judge?
A model that scores another model's outputs. It still needs a check against a rubric or a human review.
Deck 3 · Deployment and pipelines
Endpoints, knowledge bases, agents, and release.
3.1When do you choose batch inference instead of a real-time endpoint?
When a stored dataset can be scored and written out, and nobody is waiting on a tap.
3.2What does a multi-model endpoint save?
Idle cost when several models are rarely called. A multi-container endpoint runs pieces of one request together.
3.3What does Amazon Bedrock Custom Model Import do?
Hosts a model that was built outside AWS. SageMaker Neo is not the MLA-C02 import path.
3.4What does a Bedrock knowledge base index?
Your documents, in a vector store, for retrieval. Uploading a file to S3 does not by itself refresh that index.
3.5What is agent state?
The information an agent must remember between steps, such as a completed tool call.
3.6What should a pipeline do when an evaluation gate fails?
Roll back to the last approved version. Do not leave the failed version in production overnight.
3.7Which services record a model version and a prompt version?
SageMaker Model Registry or MLflow for the model, and Amazon Bedrock Prompt Management for the prompt.
3.8What is reranking?
A second pass that reorders retrieved passages before the model sees them.
3.9Which service orchestrates build and deploy stages?
AWS CodePipeline. CodeBuild runs the build. CodeConnections links the repository.
3.10What is Amazon Bedrock AgentCore for on this exam?
Agentic workflow infrastructure, not a single stateless model call.
Deck 4 · Operations and security
Drift, cost, IAM, and guardrails.
4.1What is the difference between CloudTrail and CloudWatch?
CloudTrail records AWS API calls. CloudWatch holds metrics, logs, and dashboards from the workload.
4.2What does a data-drift alarm mean?
Live inputs no longer look like the training distribution, so quality may fall even if the endpoint is up.
4.3When is Spot a poor fit?
A user-facing endpoint that cannot be interrupted. Spot fits training that can resume.
4.4Which AI costs are not the instance hour?
Token usage, embedding computation, and vector-index storage.
4.5What does least privilege mean for an endpoint role?
The role can read the model artifact and nothing broader. A wildcard on every bucket fails that test.
4.6Which two credential types does the exam name for Bedrock access?
IAM credentials, usually from a role, and an Amazon Bedrock API key kept as a secret.
4.7What do Amazon Bedrock Guardrails enforce?
Content and sensitive-data policies on model input and output. They do not replace IAM.
4.8Which tools should fail a pipeline on a vulnerable image or risky code?
Amazon Inspector and Amazon CodeGuru.
4.9What is the difference between an A/B test and a shadow variant?
An A/B test shows variants to users. A shadow variant copies traffic and does not answer the user.
4.10Which service finds sensitive data already stored in Amazon S3?
Amazon Macie. A guardrail filters the live model path. It does not scan the bucket.
Memory notes
Groupings that make the highest-yield MLA-C02 distinctions easier to recall.
Prompt, retrieve, or train
- Prompt when the model already knows the task
- Retrieve when the answer is in your files
- Fine-tune when pairs exist and the prompt still misses
- SageMaker AI when you need your own training loop
Version, then shift
Model Registry for weights. Prompt Management for the prompt. Test, then shift traffic. Roll back to the last approved version.
See the user or hide the candidate
- A/B users receive one variant
- Shadow the candidate only watches
- Drift inputs moved
- Tool failure the agent called out and got nothing
Two bills
Instance hours are one bill. Tokens, embeddings, and the vector index are the other. Spot is for work that can stop.
Lock the path
IAM for the caller. Guardrails for the text. Macie for data already in S3. Inspector or CodeGuru before the image deploys.
Continue the AWS Machine Learning Engineer study path
Test what you have learned with a timed practice exam.
Overview
Exam format, the 720 passing score, domain weights, and a study plan.
Core Notes
Reference notes for data, models, deployment, and operations.
Flashcards & Memory Notes
40 recall cards in four decks, plus memory notes for models, retrieval, and controls.
Practice Exams
200 original questions with custom exams, explanations, and a score report by domain.
Related Tools
Useful companions while you study.
JSON YAML Converter
Convert structured data between JSON and YAML with formatting and validation.
CSV JSON Converter
Convert CSV spreadsheet data to JSON or turn JSON arrays into downloadable CSV files.
Base64 Converter
Encode text and files to Base64 or decode Base64 data with UTF-8 support.
Diff Checker
Compare two text blocks and highlight added, removed, and changed content.
Legal Disclaimer
Certicab provides general informational, educational, and convenience resources. Calculations, conversions, estimates, and learning materials may contain errors or become outdated. Financial, tax, medical, legal, and travel information is not professional advice. Verify important results and current requirements with qualified professionals or authoritative sources. Protect sensitive files and personal information, review each tool’s privacy limitations, and use only content you are authorized to process. You are responsible for how you use and share results. Study resources are independent and do not guarantee exam success or imply certification-provider endorsement. Amazon Web Services, AWS, and related marks are trademarks of Amazon.com, Inc. or its affiliates. Certicab is not affiliated with, endorsed by, or sponsored by Amazon.