Create a new cluster

Segmind cluster is a set of computation resources and cloud services that your team can use to run their workloads and other MLOps related services. Clusters are EKS (on AWS) or GKE (on GCP) Kubernetes clusters and are managed by Segmind for you. You can configure the cluster by creating, editing and deleting node pools.

Node pools

A cluster contains one or more node groups. Each node group contains a group of machines of single instance type eg. c5.2xlarge or p3.2xlarge on AWS. Creating a new cluster creates a default CPU node group.

Primary node group

We currently use 2 x t3.small as primary node group to keep basic services running. You can choose to add as many secondary node groups to the cluster depending on your compute requirements.

Create additional node groups by specifying:

  1. Node group name. Eg. t4-gpu
  2. Instance type. Eg. g4dn.xlarge (on AWS)
  3. Spot true or false. Setting Spot to true will run machines in Spot mode (aka preemptible on GCP). This can help cut compute costs substantially.
  4. GPU* (for GCP, since GPU as accelerator concept to add GPUs)
  5. max nodes (auto-scaling). Maximum no. of nodes to spin up.
  6. min nodes (auto-scaling). Minimum no. of nodes to spin up.