OCI Data Science Pro key terms

The top-level OCI container that groups related notebooks, jobs, models, pipelines, and access policies so that a team's work shares one workspace and one set of permissions.

A managed compute environment that runs a JupyterLab server, backed by a configurable shape, block storage, and a conda environment, and used for interactive exploration and model development.

A published, versioned set of Python and system packages that a notebook session or job uses; OCI provides service-managed environments and supports custom environments you publish yourself.

The operation that installs a conda environment into a local directory so it can be archived and uploaded, making a custom environment reproducible across notebook sessions and jobs.

A user-built conda environment stored in Object Storage and registered with the Data Science service, allowing jobs and notebook sessions to run with packages the service-managed environments do not include.

The states a session passes through, from creating and active to deactivating and inactive; an inactive session stops compute billing but keeps its block storage and files until deleted.

The persistent disk attached to a notebook session that holds the home directory and files across deactivation; it is billed separately and must be sized at session creation.

A compute shape whose OCPU and memory can be chosen independently within allowed limits, letting a notebook session or job match resources to workload cost.

A managed, non-interactive run of a script or notebook on dedicated compute, defined by an artifact, a compute shape, environment variables, and optional logging, suited to scheduled or long-running training.

One execution instance of a job, with its own lifecycle state, logs, and output location; a job definition can be run repeatedly with different overrides.

The code bundle a job executes, typically a compressed archive of source files stored in Object Storage and referenced by URI when the job is created.

Key-value settings passed into a job run at launch, commonly used to supply paths, hyperparameters, or credentials without changing the artifact.

The serialized output of training, such as a pickle, ONNX file, or saved model directory, stored in Object Storage and registered in the model catalog for deployment.

A project-scoped registry of model artifacts, versions, and metadata that supports lineage, sharing, and selection of a specific version for deployment.

Recorded information about how a model was produced, such as the training job, source data, and environment, stored with the catalog entry so results can be audited and reproduced.

A hosted HTTPS endpoint that loads a catalog model artifact onto managed compute and serves predictions, with configurable bandwidth, instance count, and logging.

Adjusting the number of instances or compute shape behind an endpoint so throughput and latency match demand, either manually or through autoscaling policies.

Optional capture of prediction requests and responses from a model deployment to Object Storage or a logging service, used for monitoring, auditing, and later retraining.

A deployment option where you supply a container image implementing the required inference API, used when the default model server cannot load your framework or preprocessing.

The container for objects that Data Science reads and writes, used for job artifacts, model artifacts, datasets, and logs, referenced by URI in most service configurations.

A private network access path that lets Data Science resources reach Object Storage and other services without traversing the public internet, improving security and isolation.

IAM policy statements granting groups permission to use data science resources, such as managing projects, notebook sessions, jobs, models, and deployments within a compartment.

A group whose membership is defined by matching rules on resource attributes, used so jobs and deployments can access Object Storage or other services without stored credentials.

An ordered set of steps, each running code on managed compute, that automates data preparation, training, evaluation, and registration so runs are repeatable and observable.

A single unit of work in a pipeline, defined by its code artifact, compute shape, inputs, and outputs, and connected to other steps through typed parameters.

One execution of a pipeline that records step status, parameters, and outputs, allowing comparison across runs and diagnosis of failures.

The practice of operating machine learning systems reliably, covering versioning, automated training and deployment, monitoring, and governance across the model lifecycle.

Comparing live prediction inputs or outputs against a training baseline to detect distribution change that degrades model quality, triggering review or retraining.

A compute shape with attached GPUs used for deep learning training or inference, selected for notebook sessions, jobs, or deployments when CPU throughput is insufficient.

A policy that adds or removes deployment instances in response to metrics such as CPU utilization or request count, balancing cost against latency under variable load.

A logical grouping of OCI resources used for access control and cost tracking; Data Science projects and their children are created inside a compartment.

A modeling approach that learns a mapping from input features to labeled targets, covering classification and regression tasks and evaluated on held-out data.

Quantities such as accuracy, precision, recall, F1, ROC AUC, RMSE, or MAE used to judge predictive quality and to compare candidate models before deployment.

Creating, transforming, or selecting input variables so a model can learn useful patterns, including encoding, scaling, and aggregation steps that must be reproduced at inference.

Designing and operating models with attention to fairness, explainability, privacy, and accountability, including documenting limitations and monitoring for harmful outcomes.

The Python library that wraps OCI Data Science APIs, letting you create projects, sessions, jobs, models, and pipelines programmatically instead of through the console.

An OCI library layered on the Data Science SDK that adds helpers for datasets, model evaluation, explainability, and artifact handling to speed common workflow tasks.