AWS Certified Machine Learning Engineer - Associate (MLA-C01) Flashcards¶

170 minutes (2 hours 50 minutes)

65 questions

720/1000 (scaled scoring, approximately 72%)

$150 USD (50% discount for previous AWS certification holders)

3 years

Pearson VUE (online proctored or testing center)

columnar, compressed, schema-embedded, best for SageMaker

columnar, optimized for Hive/Spark

simple, widely supported, larger files

nested data, text/NLP workloads

SageMaker native format, optimized for streaming

optimized for SageMaker built-in algorithms

TensorFlow native format

stream data directly from S3 (faster than File mode)

distributed processing for large datasets (PySpark/Scala)

lightweight Python scripts for smaller datasets

distributed Python for ML preprocessing

real-time ETL with micro-batches

automated data preparation in pipeline

standalone data processing

ingest features for sharing and reuse

export as code for customization

scale to [0,1] range - good for neural networks

reduce skewness in distributions

convert continuous to categorical

binary column per category (sparse, no ordinality)

integer per category (implies ordinality)

replace with target mean (risk of leakage)

learned dense representations (deep learning)

term frequency-inverse document frequency

dense word embeddings (use BlazingText)

split text into tokens

count-based representation

remove highly correlated features

remove low-variance features

from tree-based models

iteratively remove least important

dimensionality reduction (SageMaker built-in algorithm)

single record ingestion (real-time)

using SageMaker Processing or Spark

from Kinesis Data Streams

select based on data size

built-in (scikit-learn, Spark) or custom

S3 data mounted to processing container

results written to S3

VPC configuration for security

general-purpose ML preprocessing

bring your own Docker container

built-in algorithm, framework container, or custom

CPU or GPU based on algorithm requirements

single or distributed training

S3 location and data channel configuration

S3 location for model artifacts

algorithm-specific settings

execution role with S3 and other service access

correct predictions / total predictions

true positives / predicted positives (avoid false positives)

true positives / actual positives (avoid false negatives)

harmonic mean of precision and recall

area under receiver operating characteristic curve

visualize true/false positives/negatives

root mean squared error (penalizes large errors)

mean absolute error (robust to outliers)

proportion of variance explained

mean absolute percentage error

collection of model versions

specific model version with metadata

Pending, Approved, Rejected

accuracy, latency, and custom metrics

trained model artifacts from S3

instance type, count, variants

hosted inference service

reduce precision (FP32 to INT8)

remove unimportant weights

train smaller model from larger one

process multiple requests together for GPU efficiency

DAG of ML workflow steps

individual tasks in the pipeline

configurable inputs at execution time

branch logic based on step outputs

skip unchanged steps for faster execution

data preparation with SageMaker Processing

model training with SageMaker Training

hyperparameter tuning

batch inference with Batch Transform

branch based on conditions (e.g., accuracy > threshold)

register model in Model Registry

create SageMaker model from artifacts

run Lambda function for custom logic

data/model quality check

bias and explainability check

explicitly fail pipeline with message

wait for external process (human approval)

text generation, analysis, coding

text generation, embeddings, image generation

open-source text generation

text generation, embeddings, reranking

image generation (Stable Diffusion)

text generation, coding

managed vector search

with pgvector extension

third-party vector database

in-memory vector search

document database with vector search

split by character/token count

parent-child chunk relationships

split by meaning/topic boundaries

each document is one chunk