Reduce the costs of ML workflows with preemptible VMs and GPUs

Preemptible VMs

Compute Engine instances that last a maximum of 24 hours, provide no availability guarantees, and are priced lower than standard VMs.

Google Kubernetes Engine (GKE)

A service for managing Kubernetes clusters that can utilize preemptible VMs for ML workflows.

Kubeflow

An open-source project for deploying machine learning workflows on Kubernetes.

Kubeflow Pipelines

A feature of Kubeflow that allows users to build and deploy scalable ML workflows based on Docker containers.

Cost Reduction

The primary benefit of using preemptible VMs in ML workflows, especially for jobs with flexible completion times.

Node Pool

A group of preemptible, GPU-enabled instances in a GKE cluster used for running ML workloads.

Idempotency

A property that ensures preemptible steps can either be repeated without side effects or can checkpoint work to resume after interruption.

Tensor2Tensor

A model training framework mentioned as an example of using preemptible VMs in a Kubeflow pipeline.

Stackdriver Monitoring

A tool used to inspect logs for both current and terminated pipeline operations in Kubeflow.

Autoscale

A feature that allows node pools to automatically adjust the number of instances based on workload, helping to reduce costs.