skypilot-multi-cloud-orchestration

1# SkyPilot Multi-Cloud Orchestration
2 
3Comprehensive guide to running ML workloads across clouds with automatic cost optimization using SkyPilot.
4 
5## When to use SkyPilot
6 
7**Use SkyPilot when:**
8- Running ML workloads across multiple clouds (AWS, GCP, Azure, etc.)
9- Need cost optimization with automatic cloud/region selection
10- Running long jobs on spot instances with auto-recovery
11- Managing distributed multi-node training
12- Want unified interface for 20+ cloud providers
13- Need to avoid vendor lock-in
14 
15**Key features:**
16- **Multi-cloud**: AWS, GCP, Azure, Kubernetes, Lambda, RunPod, 20+ providers
17- **Cost optimization**: Automatic cheapest cloud/region selection
18- **Spot instances**: 3-6x cost savings with automatic recovery
19- **Distributed training**: Multi-node jobs with gang scheduling
20- **Managed jobs**: Auto-recovery, checkpointing, fault tolerance
21- **Sky Serve**: Model serving with autoscaling
22 
23**Use alternatives instead:**
24- **Modal**: For simpler serverless GPU with Python-native API
25- **RunPod**: For single-cloud persistent pods
26- **Kubernetes**: For existing K8s infrastructure
27- **Ray**: For pure Ray-based orchestration
28 
29## Quick start
30 
31### Installation
32 
33```bash
34pip install "skypilot[aws,gcp,azure,kubernetes]"
35 
36# Verify cloud credentials
37sky check
38```
39 
40### Hello World
41 
42Create `hello.yaml`:
43```yaml
44resources:
45  accelerators: T4:1
46 
47run: |
48  nvidia-smi
49  echo "Hello from SkyPilot!"
50```
51 
52Launch:
53```bash
54sky launch -c hello hello.yaml
55 
56# SSH to cluster
57ssh hello
58…