juliagpu / buildkite Goto Github PK

For posterity, here is what I did. I went to https://buildkite.com/julialang/dagger-dot-jl, and I clicked on the "Edit Steps" button. The contents looked like this originally:

steps:
  - command: buildkite-agent pipeline upload

I instead changed the contents to look like this (taken from the steps for the CUDA.jl pipeline):

steps:
  - label: ":buildkite: Pipeline upload"
    command: buildkite-agent pipeline upload
    branches: "!gh-pages"
    agents:
      queue: "juliagpu"

It's kind of annoying that once Julia 1.6 is released, we have to make a bunch of pull requests to change Buildkite from '1.5' to '1.6'. Would be nice to support something like '1' or '1-stable'. I don't feel strongly about the syntax.

As far as implementing it, probably its easiest to just do the following:

1. Download the https://julialang-s3.julialang.org/bin/versions.json file and parse it.
2. If we have version = '1-stable', then collect all versions that begin with 1. and have stable=true. For each version, extract the minor version number. Take the largest minor version number.

An alternate implementation would be to use something like https://github.com/npm/node-semver, but that might be fancier than we need.

All we really need is something that expands to the latest 1.X.

https://buildkite.com/organizations/julialang/repository-providers/new seems broken?

Setting-up a repo currently doesn't give Buildkite proper permissions to post a status update. The suggested app integration in the README only works for Buildkite org <-> GitHub org integration, which requires admin permissions at both sides.

Alternatively, by creating a Buildkite account for the user, we could maybe use that to post a status update: https://buildkite.com/docs/integrations/github#to-connect-your-buildkite-account-to-github-using-oauth

Furthermore, that user account could also be user to cancel/retry builds, which currently isn't possible.

• Buildkite admin needs GH org admin priviliges, because adding the add is insufficient, it also needs to be set-up as a provider https://buildkite.com/organizations/julialang/repository-providers/new (this seems to fail if you're not a GH org admin)
• better split the instructions in a part for the BK admin and the user, it turns out to be confusing as it is right now

That way we can just stop agents and expect the latest configuration to be used when restarting.

... and override it by setting JULIA_CPU_THREADS in the configuration.

This aligns with what Elliot's infrastructure does, and also avoids defaulting to potentially unwanted multithreaded behavior (we use the number to do multi-process testing in CUDA.jl/oneAPI.jl).

Some applications complain about that, like mpiexec. Maybe we should use a regular user with sudo permissions, like PkgEval does: https://github.com/JuliaComputing/NewPkgEval.jl/blob/master/runner/Dockerfile

For NVML.