OCI Data Science and AI Infrastructure
OCI Data Science
A managed workspace for the ML lifecycle: projects, notebook sessions running in conda environments, jobs for scheduled or long-running batch work, the model catalog for versioned artifacts, model deployments that expose an HTTP endpoint with autoscaling, and ML pipelines for orchestration. Notebook sessions are interactive and can be idle-stopped; jobs are non-interactive and terminate when finished. Deployments persist and bill while running, so delete or scale down unused endpoints.
Supporting data services
Object Storage holds datasets and model artifacts; Block Volume backs notebook storage; Data Flow runs Spark workloads; Data Catalog tracks metadata; Streaming ingests events. Access is governed by compartments, IAM policies and dynamic groups for resource principals.
AI infrastructure
GPU options include flexible VM and bare-metal shapes for training and inference; cluster networking (RDMA, high-bandwidth low-latency fabrics, superclusters) matters for distributed multi-node training, where interconnect often dominates. Choose shapes by workload: small inference can sit on a single GPU or CPU; large training needs multi-GPU with fast interconnect and local NVMe. Flex shapes let you size OCPUs and memory.
Connectivity and security
FastConnect or VPN for private connectivity, private endpoints so traffic never crosses the public internet, Vault-managed keys for encryption, and least-privilege policies.
Traps
Know job vs deployment vs notebook session. The Data Science service is not the Generative AI service. Questions about lowest latency between training nodes point to cluster networking, not storage. Autoscaling applies to deployments, not notebooks.
Rule of thumb
Interactive exploration → notebook session. Repeatable batch scoring → job. Real-time predictions for an application → deployment endpoint.