megascaler

WIP Autoscaler for self-hosted runners.

This program will communticate with GitHub to receive job requests and will delegate running the job to a worker program, while the worker program has no access to the runner / PAT tokens if it is spawned in an isolated environment.

Works with both the official https://github.com/actions/runner and the unofficial https://github.com/ChristopherHX/github-act-runner.

Known problems

• you should avoid using multiple labels
  • if you really want to use them always provide an entry for any possible configuration
  • your jobs with more labels may wait longer for execution, due to other jobs taking away your runner
  • might need to create runners with more labels than requested to minimize queue time
• webhooks cannot always be delivered, need to add github api polling for reliability if waiting long

Goals

• listen for workflow_job webhooks
• create an inmemory runner with a PAT token
• track job assignment, so you can control in which environment the job executes
• cache runner registration token, valid for ca. 1 hour
• cache runner registration jwt, very short lived you can still register ca. > 1000 runners without extra rate limit cost
• use both actions/runner and nektos/act worker
• provide a script to allocate a target enviroment for a job
• provide a script to execute the worker process on the target enviroment, by forwarding stdin

Extras for actions/runner

• ${{ github.megascaler.jobid }} is assigned to the rest api job id
• ${{ github.megascaler.job_url }} is assigned to the job html url
• ${{ github.megascaler.run_url }} is assigned to the workflow run html url
• ${{ github.megascaler.labels }} is assigned to an array of string with all labels from runs-on of the current instance
• ${{ github.megascaler.variables }} is assigned to the otherwise hidden system variables of Actions
• ${{ github.megascaler.jobDisplayName }} is assigned to the otherwise hidden job display name

Non representative Statistics

• 68 of 81 jobs have received the wrong runner
• average time between job start and webhook in_progress is between 1s and 60s

Getting started

create a config.yml

worker:                                             # list of worker to spawn if the label list has an exact match
- labels:
  - github-act-runner                               # labels to spawn the specfic worker, if this matches what you added to the `runs-on` key this program will spawn a worker
  args:
  - /workspaces/megascaler/runner/github-act-runner # Path to https://github.com/ChristopherHX/github-act-runner binary
  - worker                                          # Execute in worker mode
- labels:
  - actions-runner                                  # labels to spawn the specfic worker, if this matches what you added to the `runs-on` key this program will spawn a worker
  args:                                             # create a `/workspaces/megascaler/runner/.runner` textfile with content `{"workFolder": "_work"}`
  - pwsh                                            # Path to https://github.com/powershell/powershell binary or add it to your `PATH` env variable
  - actions-runner-worker.ps1                       # Get the file from https://github.com/ChristopherHX/github-act-runner/blob/main/compat/actions-runner-worker.ps1
  - /workspaces/megascaler/runner/bin/Runner.Worker # Path to https://github.com/actions/runner binary
address: 0.0.0.0:9403                               # webhook listener to receive `workflow_job` events from GitHub
# secret: mysecret                                  # validate the webhook signature to harden the webhook endpoint
pat: ghp_XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX       # GitHub Classic PAT with full `repo` scope

You can also pipe the worker via the args list via ssh or other programs, it will connect to GitHub via http and communicates over stdin/stdout.

run

go build
./megascaler

or debug main.go