AWS ML Engineer Associate flashcards

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

SageMaker Processing job paths

Input data is mounted under /opt/ml/processing/input and outputs are written under /opt/ml/processing/output. The job's OutputPath in S3 receives whatever the container writes to the output directory.

Feature Store event time

The event time feature records when an observation occurred. It enables point-in-time joins so training datasets use only feature values known before the label's timestamp.

Ground Truth augmented manifest

Labeling job output includes a manifest where each JSON line contains the source object plus the label attribute added by workers. It can be fed directly into training jobs.

Clarify pre-training bias metrics

Computed on the dataset alone, before a model exists. Examples include Class Imbalance and Difference in Proportions of Labels, which compare label distributions across sensitive facets.

Data Wrangler export targets

A Data Wrangler flow can be exported as a SageMaker Pipeline processing step, as Python/Pandas code, or as a PySpark job for large-scale execution.

Offline versus online feature store

Offline store keeps historical features in S3 for batch training and Athena queries. Online store keeps the latest values for low-latency real-time retrieval during inference.

SageMaker Autopilot output

Autopilot analyzes data, selects algorithms, and produces candidate models plus a notebook containing the generated preprocessing and training code, which you can inspect and modify.

Hyperparameter tuning objective

Each tuning job needs an objective metric name and type. Maximize is used for accuracy or AUC; Minimize is used for loss or RMSE.

SageMaker Debugger rules

Built-in rules such as overfit, vanishing_gradient, and loss_not_decreasing analyze tensors captured during training and emit findings, optionally stopping the job early.

Built-in XGBoost use case

SageMaker's XGBoost container handles tabular regression and classification. It supports CSV or libsvm input and exposes hyperparameters like max_depth, eta, and num_round.

DeepAR algorithm purpose

A built-in algorithm for time-series forecasting using recurrent neural networks. It trains on multiple related series and produces probabilistic forecasts with confidence intervals.

SageMaker Experiments trial component

A trial component records one run's inputs, hyperparameters, metrics, and artifacts. Trials group components, and experiments group trials for comparison.

Estimator fit and deploy

Calling fit() on a SageMaker Estimator launches a training job that writes model.tar.gz to S3. Calling deploy() creates an endpoint backed by that artifact.

BlazingText algorithm modes

BlazingText supports text classification and word2vec embeddings. Supervised mode classifies sentences; unsupervised mode learns word vectors from unlabeled corpora.

JumpStart foundation models

A catalog of pre-trained models deployable or fine-tunable with a few clicks or SDK calls. Models include text generation, summarization, and image classification options.

Serverless Inference configuration

You set MemorySizeInMB and MaxConcurrency. The endpoint scales to zero when idle, making it suitable for sporadic traffic but not sustained high-throughput workloads.

Asynchronous Inference behavior

Requests are queued to S3 and results written back to S3. It supports long processing times and can scale to zero instances when no requests are pending.

Batch Transform characteristics

Runs predictions over an S3 dataset, writes output to S3, and releases the instances afterward. No persistent endpoint is created, so it suits offline scoring.

Model Registry approval status

Model packages carry a status of PendingManualApproval, Approved, or Rejected. Pipelines can gate deployment on the Approved status of a specific version.

Multi-model endpoint

One endpoint hosts many models that share instances. Models are loaded on demand from S3 into memory and evicted when unused, reducing hosting cost.

Multi-container endpoint limit

A single endpoint can host up to 15 containers that run together and are invoked through a shared invocation path, useful for preprocessing plus inference.

SageMaker Pipelines step types

Supported steps include Processing, Training, Tuning, Transform, Condition, and RegisterModel. They are arranged as a directed acyclic graph executed by the service.

Endpoint auto scaling metric

Application Auto Scaling targets a metric such as InvocationsPerInstance or ModelLatency. A target-tracking policy adjusts instance count to hold the metric near target.

SageMaker Projects MLOps template

A project template provisions a CodeCommit repository, CodePipeline pipeline, and CodeBuild project to automate build, train, and deploy stages for a model.

Model Monitor baseline

A baseline captures statistics and constraints from a reference dataset. Monitoring jobs compare captured endpoint data against it and report violations.

Model quality monitoring labels

Requires ground-truth labels merged with captured predictions. It computes metrics like accuracy, precision, recall, F1, or RMSE and compares them to baseline thresholds.

Data capture configuration

DataCaptureConfig on an endpoint writes a sampled percentage of request and response payloads to S3, providing input for monitoring and retraining.

Bias drift detection

SageMaker Clarify works with Model Monitor to compare bias metrics on captured data against a baseline, flagging shifts in model fairness over time.

Network isolation mode

EnableNetworkIsolation blocks outbound internet access for training and hosting containers. VPC endpoints for S3 and ECR are then required for data and image access.

Execution role permissions

The IAM role passed to training or hosting needs S3 read/write on relevant buckets and ECR permissions to pull images, plus CloudWatch Logs write access.

Feature attribution drift

A Model Monitor type that uses SHAP values from Clarify to detect changes in which features drive predictions, comparing current attributions to the baseline.

Encryption in transit option

Endpoints can enable inter-container traffic encryption so that calls between containers on the same instance are encrypted, complementing TLS on the endpoint itself.