TechPrufer
Sign inGet started

Blogs›Shopify›Multi-cloud GPU Orchestration

Multi-cloud GPU Orchestration

Multi-cloud GPU Orchestration

1
posts
2026

Shopify has integrated SkyPilot to manage multi-cloud GPU workloads, enabling engineers to define jobs in YAML and run them on available Kubernetes clusters across different cloud providers (Nebius, GCP). This includes a custom plugin for routing based on hardware requirements (H200s to Nebius, L4s/CPUs to GCP), fair-share scheduling via Kueue, and automatic injection of necessary configurations for specific hardware like InfiniBand. The system also supports development environments with dedicated labels and resource management, and automates storage provisioning and cleanup.

2026

SkyPilot at Shopify: Multi-cloud GPUs without the pain (2026) - Shopify

1/26/2026

This post details the integration of SkyPilot for multi-cloud GPU workload management at Shopify. Key contributions include building a SkyPilot plugin to intercept requests and route them to appropriate Kubernetes clusters based on hardware needs (e.g., H200s on Nebius, L4s/CPUs on GCP). It also covers the implementation of custom labels for cost allocation (`showback_cost_owner_ref`) and fair-share scheduling via Kueue (`ml.shopify.io/quota-group`, `ml.shopify.io/priority-class`). Specific Nebius integration details include automatic mounting of InfiniBand-related devices and shared caches, and dynamic storage provisioning. A development environment pattern is introduced with the `ml.shopify.io/dev: "true"` label, enabling interactive sessions with different scheduling and resource policies.