fix: Avoid timeouts when pushing to a registry

[dragosgheorghioiu]

unikraft image build --output <ref> fails when pushing a large rootfs: a ~2 GB blob PUT dies 30s in.

This adds a dedicated HTTP client for container-registry traffic with no response-header timeout, so a registry may take as long as it needs to commit a large blob.

[jedevc]

🤔 what kind of ref?

This does feel odd to me, ResponseHeaderTimeout is after the request is fully written. It takes 30s to write after all the data is there?

See https://github.com/containerd/containerd/blob/574a60700ee870807a052e2a0d0c7eed08b7ce8f/core/remotes/docker/registry.go#L254-L269, for the containerd default transport... which defaults to 30s.

[craciunoiuc]

Just one more important comment

[dragosgheorghioiu]

🤔 what kind of ref?

This does feel odd to me, ResponseHeaderTimeout is after the request is fully written. It takes 30s to write after all the data is there?

See https://github.com/containerd/containerd/blob/574a60700ee870807a052e2a0d0c7eed08b7ce8f/core/remotes/docker/registry.go#L254-L269, for the containerd default transport... which defaults to 30s.

For a 2.5G image pushed to harbor:

FROM --platform=linux/x86_64 alpine:3.20 AS build

RUN dd if=/dev/urandom of=/big.bin bs=1M count=2560

FROM scratch

COPY --from=build /big.bin /big.bin
COPY --from=build /bin/busybox /bin/sh

You get this http2: timeout awaiting response headers so it seems it does take more than 30 seconds to respond after the data is there. It basically fails when committing the layers.
It may be the case that is just harbor being slow. Pushing to local registries doesn't have this problem.