Autoscaling and Distribution

Kubetorch has comprehensive support for distributed workflows, with its support for autoscaling and range of distributed options. Parameters for setting these up can be specified for the compute, using compute.distribute(args) or compute.autoscale(args)

Autoscaling

Kubetorch supports automatic scaling of your services based on concurrency or request rate. You have options such as setting the min/max scale or controling the autoscale to zero grace period to ensure interactive development is not disrupted and keep your service warm for hot reloads.

Distributed

Building on top of scaling, Kubetorch also helps wire up pod-to-pod communication and environment setup for distributed workflows. Kubetorch provides built-in distributed frameworks like Ray and PyTorch, or lets you simply specify the number of workers and sets up a pool of parallel replicas for your distributed programs.