Skip to navigation
ReferenceDeployment & Architecture

Deployment Modes

View as Markdown

The Kumo Predict Platform ships as software you deploy and operate inside your own cloud environment. There are two deployment paths:

  • The Virtual Private Cloud deployment runs Kumo on a customer-managed Kubernetes cluster (EKS, AKS, GKE, or OpenShift) using the kumo-helm chart. This is the production path, supporting all data connectors and admin-configurable job concurrency.

  • The Simplified VPC Deployment runs all Kumo services on a single GPU VM using Docker Compose. It is designed for proof-of-concept evaluations: fast to stand up, limited to CSV and Parquet data sources, and capped at one training job at a time.

In both cases, all data and model artifacts remain in your cloud environment.

The following table compares the two deployment options:

VPC - Kubernetes (Helm)Simplified VPC (Docker Compose)
InfrastructureCustomer-managed Kubernetes cluster (EKS, AKS, GKE, OpenShift)Single GPU VM (AWS AMI, Azure VM image, or GCP VM image)
Data connectorsSnowflake, Databricks, BigQuery, S3, ADLS, GCS, file uploadCSV and Parquet files from cloud storage (S3, ADLS, GCS)
Data storageCustomer-owned object storage (S3/GCS/ADLS), Postgres, TemporalCustomer-owned cloud storage bucket; ephemeral local NVMe for DiskGraph index
ML computeCustomer-managed GPU and high-memory nodesSingle GPU on host VM
AuthBYO OIDC or password (configured through Helm values)Password or OIDC (disabled by default)
Job concurrencyAdmin-configurable1
Clouds supportedAWS, Azure, GCP, on-premises (OpenShift)AWS, Azure, GCP
Intended useProduction workloadsProof of concept and evaluation