AWS Machine Learning Specialty (AWS) flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

S3 multipart upload threshold

AWS recommends multipart upload for objects larger than 100 MB and requires it for objects larger than 5 GB; parts can be uploaded in parallel and retried independently.

Athena cost model

Amazon Athena charges based on data scanned per query; using columnar formats such as Parquet or ORC with compression reduces bytes scanned and therefore cost.

Glue Data Catalog

A managed, Hive-compatible metadata repository that stores table definitions and schemas; AWS analytics services such as Athena, Redshift Spectrum, and EMR can query it.

Kinesis Data Streams shard limits

One shard supports 1 MB/s or 1,000 records/s ingress and 2 MB/s egress; enhanced fan-out gives each consumer up to 2 MB/s dedicated throughput.

Kinesis Firehose delivery

A fully managed service that buffers streaming data and delivers it to destinations such as S3, Redshift, and OpenSearch; it does not support replay of already delivered records.

Redshift COPY command

Loads data into Redshift in parallel from S3, DynamoDB, or EMR; it is much faster than row-by-row INSERT and supports compressed columnar source files.

Glue crawler purpose

Scans data stores, infers schemas and partitions, and writes table metadata to the Glue Data Catalog so query engines can discover the data.

Lake Formation permissions

Provides centralized, fine-grained access control over S3 data and Glue Catalog resources, including table-level, column-level, and row-level permissions.

IQR outlier rule

A value is flagged as an outlier if it lies below Q1 minus 1.5 times the interquartile range or above Q3 plus 1.5 times the IQR.

Pearson vs Spearman

Pearson correlation measures linear association between continuous variables; Spearman correlation measures monotonic association using ranks and is robust to nonlinear monotonic relationships and outliers.

MCAR MAR MNAR

Missing completely at random, missing at random conditional on observed data, and missing not at random. The pattern determines whether deletion or imputation is unbiased.

Data leakage in preprocessing

Fitting scalers, imputers, or encoders on the full dataset leaks test information; fit them only on training folds inside cross-validation.

Class imbalance metric choice

With rare positive classes, accuracy is misleading; use precision, recall, F1, or PR-AUC, and consider resampling or class weights.

One-hot high cardinality

One-hot encoding a feature with many categories creates a very wide sparse matrix; target encoding or feature hashing can reduce dimensionality but require leakage-safe fitting.

Log transform right skew

For positive, right-skewed features, log or Box-Cox transforms can reduce skew and stabilize variance, often improving linear and distance-based models.

SageMaker Clarify bias

Detects potential bias in data and models using metrics such as disparate impact, and provides feature attribution explanations for predictions.

XGBoost tabular default

Gradient-boosted trees often perform strongly on structured tabular data, handle nonlinearity and interactions, and require relatively little feature scaling.

SageMaker Linear Learner

A built-in algorithm for regression and classification that trains linear models with L1 or L2 regularization and supports large sparse inputs.

Random Cut Forest

An unsupervised SageMaker algorithm for anomaly detection that assigns higher anomaly scores to points that are easier to isolate in random tree partitions.

DeepAR forecasting

A SageMaker built-in algorithm for probabilistic time-series forecasting that learns across many related series and supports covariates.

BlazingText word embeddings

A SageMaker built-in algorithm for text classification and word2vec-style embeddings; it is efficient for large text corpora.

Silhouette score clustering

Measures how similar a point is to its own cluster compared with other clusters; values range from -1 to 1, with higher values indicating better-separated clusters.

L1 vs L2 regularization

L1 adds absolute weights and can drive coefficients to zero for feature selection; L2 adds squared weights and shrinks coefficients smoothly.

Time series validation split

Use chronological splits or rolling-origin backtesting rather than random shuffling, because random splits leak future information into training.

SageMaker batch transform

Runs inference on large datasets without persisting a real-time endpoint; it reads input objects from S3 and writes predictions to S3.

SageMaker serverless inference

A deployment option that automatically scales to zero when idle and charges only for compute used during requests; suited to intermittent traffic.

SageMaker Model Monitor

Detects data drift, model quality drift, bias drift, and feature attribution drift by comparing endpoint traffic with a baseline.

SageMaker Pipelines

A managed CI/CD orchestration service for ML workflows; steps can include processing, training, tuning, evaluation, conditional logic, and model registration.

SageMaker Spot training

Uses spare EC2 capacity for training at lower cost; enable checkpointing so interrupted jobs can resume from the last checkpoint.

SageMaker multi-model endpoint

Hosts many models behind one endpoint and loads them on demand, reducing hosting cost when there are many rarely used models.

SageMaker endpoint auto scaling

Adjusts instance count based on metrics such as invocations per instance or CPU utilization, using a target-tracking scaling policy.

SageMaker VPC isolation

Configure training jobs and endpoints with VPC subnets and security groups to keep traffic private; use VPC endpoints for S3 and other services.