Project context is here.

This PR tackles the workflow of uploading connect to a designed S3 repo, which acts as the "rpk plugin" repo, used only by connect for now.

plugin_uploader.py
To be run after goreleaser has built the binaries. upload-archives creates and uploads the archives. upload-manifest creates and uploads the final manifest.json file. RPK uses manifest.json to know what plugins are available and how to download them (download URL, etc.). Comes with unit tests, basic linters, type checkers, setup in CI.

Additional release workflow
.github/workflows/upload_plugin.yml. High level:

• goreleaser build
• Run plugin_uploader.py upload-archives ...
• Run plugin_uploader.py upload-manifest ...

Only uploads for real when triggered from a release tag (inc. RC). Otherwise, runs in --dry-run mode - no writes to S3.

Why not use goreleaser built in blob publish support?

Goreleaser has "blob" publish support, which kind of can do upload-archives in spirit. However, our use case requires additional flexibility in how we set S3 object tags. Our manifest.json must contain SHA256 checksums of the binary payload WITHIN the tar.gz (rpk requirement). It would not work gracefully with said blob publishing within Goreleaser.

manifest.json construction logic (upload-archives) would be needed in any case and most cleanly implemented separately from goreleaser.

Generally the design of goreleaser is not conducive to running, testing, debugging, retrying specific parts of the goreleaser steps. E.g. Publishing steps are all or nothing. To test blob publishing, one would also be publishing to Github!

Overall the monolithic design of goreleaser's runtime makes things tricky for more complex use cases.

Thankfully goreleaser provides a clean interface for extension scripts (artifacts.json, metadata.json)

Why not use goreleaser custom publisher support?

Similar reasoning to above.

Note the calculus may be different for other use cases. E.g. next task is to build Linux packages for redpanda-connect (deb / rpm). The functionality may fit much more closely such that we can accept the drawbacks we describe here. We do that for redpanda-console (see nfpm usage there).

Why not extend existing release.yml Github action workflow?

This could be an option in the long term. However this functionality is new and may introduce instability in the main release.yml pipeline in the meantime. It will be much cleaner to test, debug, run (and rerun!) as an independent workflow. Some observations:

• The S3 upload flow does not need a goreleaser release, only goreleaser build!
• release.yml pushes real Github releases and may not be idempotent!

Once we gain confidence that both .github/workflows/upload_plugin.yml AND release.yml are very reliable, we can consider merging into release.yml.