Figure out docker hub caching

[ehuss]

In GitHub Actions we periodically have problems hitting the Docker Hub rate limit which was introduced in November 2020 (error is "429 Too Many Requests"). This hits any repo using Docker (such as rust-lang/rust, and rust-lang/cargo).

The anonymous Docker Hub rate limit is 100 pulls / 6 hours / IP. source

There have been a few solutions proposed:

1. Authenticate with Docker Hub. This changes the rate limit to 200 pulls / 6 hours / account. I do not know if that is sufficient for our needs across the org. I get the impression that infra team members do not like this option.
2. Mirror in GitHub Container registry (docs). From what I can tell, there doesn't seem to be a read limit. There is a 10GB/layer limit, which I think is fine. The main drawback is that it requires manually updating the images (like when new Ubuntu images are released).
   • I do not think we update the base images very often. I do not know if the infra team has a pre-existing mechanism for uploading, or how difficult that is to do manually.
3. Use Amazon ECR. I think the ECR Public Gallery has many of the images we typically use. I am uncertain, but I think the unauthenticated rate limit is 500GB/month (per IP?) source. We could authenticate using OIDC, which raises the cap to 5 TB / month.

I do not know what the performance and reliability compares between ghcr and amazon ecr.

Would the infra team have a preference here? I prefer whatever is easiest 😜. ghcr seems appealing to me if the infra team is ok with handling uploading new images.

[Mark-Simulacrum]

Authenticate with Docker Hub. This changes the rate limit to 200 pulls / 6 hours / account. I do not know if that is sufficient for our needs across the org. I get the impression that infra team members do not like this option.

Yeah, I'd prefer to avoid ~personal/team accounts on Docker Hub, seems like unnecessary hassle, and 200 pulls / 6 hours also doesn't feel that high that this fully solves the problem.

Use Amazon ECR. I think the ECR Public Gallery has many of the images we typically use. I am uncertain, but I think the unauthenticated rate limit is 500GB/month (per IP?) source. We could authenticate using OIDC, which raises the cap to 5 TB / month.

It is per IP: " *** Data transferred out from public repositories is limited by source IP when an AWS account is not used." (https://aws.amazon.com/ecr/pricing/)

5 TB isn't a cap, it's just the free tier. Past that we start paying, but I'd expect that in practice we wouldn't use much beyond 5 TB (if at all, that's a pretty large amount of data).

For ECR, if it's not in the existing public gallery, we could probably configure pull through caching (https://docs.aws.amazon.com/AmazonECR/latest/userguide/pull-through-cache.html), though it sounds like that would require authentication. I'm much more comfortable with not ending up needing multiple Docker hub accounts to distribute load (as seems likely if e.g. rust-lang/rust uses this).

I do not think we update the base images very often. I do not know if the infra team has a pre-existing mechanism for uploading, or how difficult that is to do manually.

I think base images get updated pretty regularly? At least I'd expect that e.g. ubuntu:22.04 is getting updates constantly -- it was updated just 8 days ago (way after initial release) https://hub.docker.com/layers/library/ubuntu/22.04/images/sha256-3d1556a8a18cf5307b121e0a98e93f1ddf1f3f8e092f1fddfd941254785b95d7?context=explore

[marcoieni]

This is on my todo list now 👍
I will switch from dockerhub to aws ecr as agreed here

[marcoieni]

ECR issues

I'm summarizing some of the discussions we had in Zulip.

The ECR unauthenticated rate limit of 500GB/month per IP is an issue for the following reasons:

• For contributors, authenticating to AWS ECR is more complicated than to DockerHub.
  I can imagine people would prefer to edit the base image in the Dockerfile rather than setting up AWS CLI and configuring it. This isn't a great developer experience.
• In PRs we can't authenticate to aws, so we are subject to the unauthenticated rate limit.
  This is because PR jobs from forks can't read GitHub secrets (see here), so they can't authenticate with aws.

An option is hardcoding the AWS key using to authenticate to ECR, but
this would allow external people to use the key to download from ECR public gallery for free by using our account. So we would pay for their traffic.
If they use the key in a private repo, we wouldn't notice.

GHCR

I suggest we use GHCR instead and forget about rate limits.

I found two solutions to periodically mirror the images from DockerHub to GHCR:

• https://github.com/getsentry/image-mirror — Solution by Sentry. It's a python script of ~200 lines of code (I excluded from this count the lines containing the images they mirror)
• https://github.com/rblaine95/dockerhub-mirror — Way simpler solution (the logic is 10 lines of GitHub Actions). I suggest we go with this one since it's very simple, pinning GitHub Actions and binary versions. Of course we need to test it first. I can create a POC.

[Mark-Simulacrum]

FYI -- https://github.com/orgs/community/discussions/45253#discussioncomment-4769997 suggests there are some hidden, internal rate limits for pushing to GHCR at least, though I somewhat doubt we'll hit them in practice. Generally speaking using GHCR seems pretty reasonable, if we can get the mirroring at least semi-automated -- that will take a bit of work but probably not too hard.

[marcoieni]

It was very easy to setup:

• repo: https://github.com/marcoieni/dockerhub-mirror

• images: https://github.com/marcoieni?ecosystem=container&tab=packages

• Should I add the dockerhub-mirror repo to https://github.com/rust-lang/team? If yes, do you have a better name in mind instead of dockerhub-mirror? Also, do I have to create the repo manually or will rust-lang/team create it for me when I add the toml file?

• Should we switch the license to MIT? The original project used the UNLICENSE, but I don't think we use it anywhere.

[marcoieni]

I created rust-lang/rust#135574 but it didn't work:

Error: PUT https://ghcr.io/v2/rust-lang/ubuntu/manifests/22.04: DENIED: permission_denied: write_package

Logs here.

I will investigate

[Kobzol]

I think that the issue might be that we're trying to push to the organization's image hub, rather than the repository's image hub, I think that more permissions are needed for that. But I suppose we could just push to rust-lang/rust/image instead, that should be fine.

[marcoieni]

Something was pushed
https://github.com/rust-lang/rust/pkgs/container/ubuntu

It just failed to write the manifest, which means the tags aren't there 🤔

[marcoieni]

So weird that the repository is listed twice. this is probably a gh bug

[Kobzol]

Isn't it r-l/rust and r-l-ci/rust?