Container registry authentication

repo2rlenv push (v0.8.2.post3+) uploads the bootstrap Docker image to an OCI registry alongside the HF Hub dataset so any consumer can harbor run the published tasks. This page covers how to log into each supported registry and verify the credentials work.

Run repo2rlenv push --check-auth to probe every detected registry and get a one-shot report. It pushes no image and leaves nothing behind.

Default flowCopy Anchor Link

You don't have to pre-configure anything. repo2rlenv push auto-detects the first verified registry from ~/.docker/config.json. If nothing is logged in, push falls back to inline-Dockerfile mode (recipe-level reproducibility) and tells you how to upgrade.

For launches and CI, use --require-registry to hard-fail instead of falling back silently.

GHCR (default recommendation)Copy Anchor Link

Free for public images, no anonymous-pull rate limit, and it maps naturally onto a GitHub org. Recommended default for new datasets.

gh auth refresh -h github.com -s write:packages
echo "$(gh auth token)" | docker login ghcr.io \
  -u "$(gh api user --jq .login)" --password-stdin

repo2rlenv push will pick ghcr.io/<hf-dataset-owner>/... automatically. If you lack write:packages on that GitHub org, push falls back to ghcr.io/<gh-login-user>/.... Repository visibility is flipped to public via the GitHub API (PATCH /user|orgs/packages/container/<name>) when the HF dataset is public.

AWS ECR PrivateCopy Anchor Link

Token TTL is 12h, so install amazon-ecr-credential-helper once and forget about it.

brew install docker-credential-helper-ecr  # or apt
# Add to ~/.docker/config.json:
#   "credHelpers": { "<acct>.dkr.ecr.<region>.amazonaws.com": "ecr-login" }

Or one-off (re-run every 12h):

aws ecr get-login-password --region us-east-1 \
  | docker login --username AWS --password-stdin \
    <acct>.dkr.ecr.us-east-1.amazonaws.com

ECR requires the repository to pre-exist; repo2rlenv push creates it automatically via aws ecr create-repository if the L4 probe returns 404. Make sure your IAM role has ecr:CreateRepository.

AWS ECR PublicCopy Anchor Link

A separate service from ECR Private. 50 GB free public storage, forever.

aws ecr-public get-login-password --region us-east-1 \
  | docker login --username AWS --password-stdin public.ecr.aws

Or use amazon-ecr-credential-helper against public.ecr.aws.

Azure Container RegistryCopy Anchor Link

Token TTL is 3h.

az login                              # one-time
az acr login --name <myregistry>      # per session

For anonymous public pulls (Standard tier or higher): az acr update -n <reg> --anonymous-pull-enabled true.

GCP Artifact RegistryCopy Anchor Link

gcloud registers itself as a Docker credential helper per region:

gcloud auth login
gcloud auth configure-docker us-central1-docker.pkg.dev

Repository must pre-exist; repo2rlenv push creates it via gcloud artifacts repositories create if the L4 probe returns 404. For public pulls:

gcloud artifacts repositories add-iam-policy-binding <repo> \
  --location=us-central1 --member=allUsers \
  --role=roles/artifactregistry.reader

Docker HubCopy Anchor Link

The default tier has a 100 pulls / 6h / IP anonymous limit. That's fine for small datasets, but a real problem for benchmarks pulled by many consumers or in CI. repo2rlenv push will NOT pick Docker Hub as the auto-default when other registries are available; pass --image-registry index.docker.io/<user> to force it.

docker login                          # interactive PAT prompt
# OR
echo "$DOCKER_TOKEN" | docker login -u "$DOCKER_USERNAME" --password-stdin

Credential resolution. For Docker Hub the probe prefers the explicit env vars DOCKER_USERNAME + DOCKER_TOKEN (a PAT) over the docker credstore. The credstore returns an OAuth identity token that is often pull-only at the token endpoint, while an explicit PAT reliably grants push. The push namespace is the authenticated Docker Hub user (DOCKER_USERNAME), not the HF dataset owner, so images land under index.docker.io/<DOCKER_USERNAME>/…. (GHCR analogously reads GHCR_TOKEN / GITHUB_TOKEN.)

Multi-repo datasets (e.g. pr_runtime across many repos, with one bootstrap image per repo) are fully supported: push pushes each distinct image and rewrites each task's environment/Dockerfile to its own registry digest. If any push fails (or no registry is verified), it falls back to inline mode: each task bakes its own rebuild recipe and stays reproducible with no registry at all.

Local / registry:2 (testing only)Copy Anchor Link

For development tests that need a real OCI registry but don't want any cloud signup:

docker run -d -p 5000:5000 --name r2e-test-reg registry:2
repo2rlenv push ./datasets/mydataset owner/repo \
  --image-registry localhost:5000

The probe correctly classifies localhost and skips L2 auth (the default registry:2 image runs without auth). For an authenticated local registry, see the upstream docs at https://distribution.github.io/distribution.

Anonymous-ephemeral: ttl.shCopy Anchor Link

For one-shot end-to-end smoke tests with no signup:

repo2rlenv push ./datasets/mydataset owner/repo \
  --image-registry ttl.sh

Images auto-delete after 1h (configurable in the tag).

What --check-auth actually probesCopy Anchor Link

For each registry the L1–L4 protocol runs in order:

L4 leaves no image, layer, or session artifact. Total wire time per registry: ~200–500 ms. --fast stops at L2 if you just want a quick sanity check.