registry.k8s.io VS torrent

If you stop to think about, what AWS is selling here, it's quite funny. Most of the infrastructure behind registry.k8s.io is hosted on AWS (and also GCP). So AWS essentially tells you: Don't trust the upstream registry, it might go down, cache it on your own registry, hosted also by us.

For example: if you only allow cluster autoscaler and metrics server from registry.k8s.io you can pull those images through the cache as someone who has create repo IAM privileges. If someone without create repo privileges tries to pull a new image it will fail because they can't create the initial repo.

https://registry.k8s.io (the root url at /) redirects you to https://github.com/kubernetes/registry.k8s.io where we have an issue tracker.

❗ This container is having trouble accessing https://registry.k8s.io

Hosting a forwarding / redirect server instead of actually hosting images is probably a decent idea.

The K8s proxy is redirecting from only hosting on GCR to community-owned registries - https://kubernetes.io/blog/2023/03/10/image-registry-redirec...

You can view the code here - https://github.com/kubernetes/registry.k8s.io

But because everyone is already pointing at gcr.io (just like many openfaas users point at docker.io/) - they're having to do a huge campaign to announce the new URL - the same would apply with the author's solution here.

I wrote some automation for hosting (not redirects) in arkade with the OSS registry - Get a TLS-enabled Docker registry in 5 minutes - https://blog.alexellis.io/get-a-tls-enabled-docker-registry-...

The registry is also something you can run on a VM if you so wish, and have act as a pull through cache.

Apart from reliability - GitHub's container registry is the current next best option - but we have to ask ourselves, what happens when they start charging or the outages start to last longer or are more frequent than 1-2 times per week as we've seen in Q1 2023.

One annoyance with how docker images are specified is they include the location where they are stored. So if you want to change where you store you image you break everyone.

I wonder if what regsitry.k8s.io does could be generalized:

https://github.com/kubernetes/registry.k8s.io/blob/main/cmd/...

The idea is the depending on which cloud you are pulling the image from, they will use the closest blob store to service the request. This also has the effect that you could change the source of truth for the registry without breaking all Dockerfiles.

If you are using updated helm charts then most of them have already replaced with registry.k8s.io, for example nginx ingress so not really breaking change.

https://github.com/anacrolix/torrent has a fuse driver since 2013. I'm in the early stages of removing it. There are WebDAV, 3rd party FUSE, and HTTP wrappers of the client all doing similar things: serving magnet links, infohashes, and torrent files like an immutable filesystem. BitTorrent v2 support is currently in master.

I'm the author of https://github.com/anacrolix/torrent (started in 2013) and https://github.com/anacrolix/dht (started in 2015). I have a DHT indexer implementation I developed in 2021. It's currently closed source but available for use as part of https://www.coveapp.info/. I have found that after several hours the search is excellent and stays up to date with ease.

https://github.com/anacrolix/torrent/blob/master/cmd/torrent... does exactly that. Install with `go get github.com/anacrolix/torrent/cmd/torrent@latest`, and then run `torrent serve `.

Hit me up if you want to discuss using BitTorrent to back images. https://github.com/anacrolix/torrent

It has some small latency but only when resources are spread across many different infos. If you can constrain your resources to a single DHT traversal, it's pretty quick. I run several services that stream from BitTorrent on demand, using https://github.com/anacrolix/torrent which are surprisingly quick to start. However it does choke up when you try to start many different resources at once, which multiplies horizontally the number of DHT traversals, and per-torrent related overhead to get started.

It is solvable, but any solution that spreads resources out across many different targets in the DHT is slow. Basically anything that was inspired by BitTorrent, but isn't BitTorrent itself does this, because they get overly excited by deduplication of data.

For whatever reason the ARM version Transmission does not work well on my M2 laptop - it downloads quickly at first and then drops off to zero. I tried playing around with different settings, running their nightly builds, etc, and nothing fixed it for me. In the end I searched for other clients and found them all filled with ads and bloatware, and decided to use this excellent open source command line client instead:

https://github.com/anacrolix/torrent

It has a few frontends built on top of it (linked in the project readme), but I just run `torrent download ` and it downloads at full speed / with no issues.

https://tcloud-lunik.herokuapp.com/ https://btorrent.xyz/ https://github.com/pldubouilh/webtorrent-webui https://github.com/anacrolix/torrent

The use case is in the refactoring in https://github.com/anacrolix/torrent/compare/smartban...lazylog. A lot of the parameters moved around as part of a performance optimization in https://github.com/anacrolix/log/compare/lazylog.