harbor-aws

AWS EKS/Fargate execution backend for Harbor benchmarks. Run thousands of sandbox trials in parallel with per-second billing and VM-level isolation.

How it works

Orchestrator (laptop / CI)
    |  HTTPS via NLB
    v
harbor-control pod (in-cluster gateway)
    |  TCP via /dev/tcp
    v
Trial pods (one per task, Fargate)

Each trial pod runs a small bash runner (runner.sh) as PID 1 — no Python needed in the trial image. The runner dials the in-cluster harbor-control gateway over plain TCP. The orchestrator talks to harbor-control over an NLB. The K8s API server is only used for pod create/delete, never in the exec data path.

Install

pip install "harbor-aws[cdk]"

Quick start

1. Deploy (~15 min, one-time)

python -m harbor_aws deploy --region us-east-1

Creates everything: VPC, EKS, harbor-control, NLB, Load Balancer Controller.

2. Run benchmarks

harbor jobs start -p ./task -a nop -n 2500 \
  --environment-import-path harbor_aws.adapter:AWSEnvironment \
  --ek stack_name=harbor-aws --ek ecr_cache=true

The adapter auto-discovers the NLB endpoint and admin token from the stack. No env vars needed.

3. Clean up

python -m harbor_aws stop      # delete trial pods, keep infra
python -m harbor_aws destroy   # tear down everything

Cost

Development

pip install -e ".[dev,cdk]"
ruff check src/
mypy src/

License

Apache License 2.0 — see LICENSE.