✈️ SkyShift

Run Any Workload on Any Compute 🌐

---

📰 News 📰

• [Dec 2024] Run LLM inference with vLLM and large scale data processing with Spark.

---

⚡Quick Install

Install with pip:

# Clone SkyShift to local repo
git clone https://github.com/sky-shift/skyshift.git && cd skyshift
# Install SkyShift dependencies for clients.
pip install -e .
# (Optional) SkyShift dependencies for admins.
pip install -e .[server]

🤔 What is SkyShift?

SkyShift is a unified framework for running workloads anywhere—from on-premises compute to edge devices and cloud instances.

SkyShift abstracts away many clusters into one contiuuum of compute, achieving the Sky:

• Unify Kubernetes, Slurm, and Ray clusters + any cloud instance into one singular cluster.
• Automatically manages job and services.
• Transparently handles networking and communication across clusters.

SkyShift simplifies workload and compute management:

• Supports one unified job/service API across clusters.
• Automatically autoscale cloud clusters and manually add/remove existing clusters.

🚀 Launch SkyShift

SkyShift requires at least one cluster to run. This example will use Kind, which provisions local Kubernetes (K8) clusters on your local laptop:

# Launch two K8 clusters.
for i in {1..2};
do
kind create cluster --name cluster$i
done

Once the K8 clusters are provisioned, launch SkyShift:

# In skyshift root path
./launch_skyshift.sh

Check SkyShift's status:

> skyctl get clusters
NAME               MANAGER    RESOURCES                          STATUS
cluster1             k8       cpus: 1.83/2.0                     READY
                              memory: 6035.6/7954.6 MiB
cluster2             k8       cpus: 1.83/2.0                     READY
                              memory: 6035.6/7954.6 MiB

Congrats! SkyShift is now ready to submit jobs with its simplified interface. Provided below is an example YAML for a simple Nginx deployment:

kind: Job

metadata:
  name: example-job
  labels:
    app: nginx

spec:
  replicas: 2
  image: nginx:1.14.2
  resources:
    cpus: 1
    memory: 128
  ports:
    - 80
  # Always restart a job's tasks, regardless of exit code.
  restartPolicy: Always

To create a SkyShift job, run the following command:

skyctl apply -f [PATH_TO_JOB].yaml

Since one cluster does not have sufficient resources, SkyShift automatically schedules and launches the job across two clusters:

> skyctl get jobs
NAME          CLUSTER    REPLICAS    RESOURCES               NAMESPACE    STATUS
example-job   cluster1   2/2         cpus: 1                 default      RUNNING
              cluster2               memory: 128.0 MiB

> skyctl get clusters
NAME               MANAGER    RESOURCES                          STATUS
cluster1             k8       cpus: 0.83/2.0                     READY
                              memory: 5907.6/7954.6 MiB
cluster2             k8       cpus: 0.83/2.0                     READY
                              memory: 5907.6/7954.6 MiB

Finally, clean up SkyShift and the two K8 clusters:

# Kill SkyShift service
./launch_skyshift.sh --kill
# Terminate Kind clusters
for i in {1..2};
do
kind delete cluster --name cluster$i
done

Refer to Quickstart to get started with SkyShift.

📖 Documentation

The above tutorial describes a small subset of features SkyShift supports. Refer to the documentation for a more complete set. For example, SkyShift supports:

• Slurm and Ray clusters.
• Automatic management of services via ClusterLink.
• Fine-grained Role-based Access Control (RBAC) for regular users.
• Advanced scheduling, including custom priorities, preferences, and filters.
• Advanced examples, incl. but not limited to LLM serving with vLLM, large-scale data processing with SparkSQL, and general microservices.

🙋 Contributions

As an open-source project, we gladly welcome all contributions, whether it be a new feature, better documentation, or simple bug fixes. For more information, see here.