Self-Hosted Runners

Default: GitHub-hosted runners are clean VMs (Ubuntu, Windows, macOS), 7GB RAM, 14GB disk, 2 cores. Provisioned per job. Free tier: 2000 minutes/month.

Trade-offs of cloud runners:

• ✅ No infrastructure to manage
• ✅ Latest base images
• ✅ Free for public repos
• ❌ Slow (cold start each job)
• ❌ Limited compute (2 cores, 7GB)
• ❌ Cost scales with usage
• ❌ No GPU access (until recently)

Self-hosted runners are your own machines, registered with the CI provider:

# Download runner binary; register with token
./config.sh --url https://github.com/myorg/myrepo --token ABC123
./run.sh

The runner polls the API for jobs, executes them, reports back.

Pros:

• Faster (warm caches, persistent disk)
• More compute (whatever you provision)
• Custom hardware (GPUs, big RAM)
• Behind your firewall (access internal resources)
• Cheaper at high volume

Cons:

• You manage them (patching, scaling)
• Security: workflow code runs on YOUR machine. Untrusted PRs can be dangerous.
• Persistent state can leak between jobs

Best practices:

• Use ephemeral runners (one job, then destroy)
• Limit to private repos (no untrusted code)
• Network isolation (separate from production)
• Rotate runner tokens