RFC: try free OSS tier of Aspect Workflows

[alexeagle]

CI for this project seems slower than would be desired: an Avg job run time 28m 23s is currently reported on https://github.com/buildbarn/bb-storage/actions/metrics/performance

Aspect hosts a "managed" Buildbarn instance and also runs a fleet of CI runners that connect to GitHub Actions, Buildkite, and some others. We now have a free tier for OSS projects, which runs co-tenant on our own cloud and we've been using ourselves (https://github.com/alexeagle/awk.bzl is my most recent hobby example)

We'd be happy to offer this to the Buildbarn community, in place of engineering work to optimize the existing GitHub Actions setup.

[EdSchouten]

Hey @alexeagle,

Yes, Buildbarn's CI is robust, but it is slow. The main reason for it is that GitHub Actions doesn't allow us to easily persist Bazel's cache directory. There is some step that permits it, but saving/restoring the cache directory is so incredibly slow. Is this something that Aspect's CI addresses? If so, I'd be more than happy to switch.

[alexeagle]

Good questions. To be transparent, there is a non-commercial option to make it faster as well:

• https://github.com/bazel-contrib/setup-bazel is maintained by Rules Authors SIG and it does allow persisting both the disk_cache of actions, the repository_cache of downloads, and the external/ folder of expanded repositories. It just uses GitHub Action's native cache, which is pretty slow as you say.
• The pipeline is currently serialized, but doesn't need to be IIUC - we could create parallel GHA jobs per platform, or use transitions to build&test all configurations in a single build
• Find low-hanging fruit to speed up steps, for example refactor: use pre-built protoc binary #248

However we get more benefit from Aspect's product:

• dogfood Buildbarn building on a Buildbarn cluster
• keeping warm, persistent CI runners which have the Bazel JVM already running and its in-memory caches populated
• improves cold-start time for scale-out of CI runner machines, by restoring caches using things like parallel downloads, https://zlib.net/pigz/ for extraction

[EdSchouten]

Awesome! Then let’s do it!