De $32 a $0.90 al mes: un pool de CI que solo existe cuando hace falta

                      pull_request / push a main / cron / dispatch
                                         |
          +------------------------------+-------------------------------+
          |                                                              |
   runner-pool.yml  (ubuntu-latest)                     CI / Deploy * / Security Scan
   despierta: DesiredCapacity=1                          runs-on: self-hosted
          |                                             (los jobs se quedan "queued" hasta
          v                                              que un runner se registra, ~2-4 min)
   ASG lanza 1 caja spot Graviton  -------------------------->  los jobs corren
   el user-data registra 2 runners                                    |
                                                                      v
                                                         workflow_run: completed
                                                                      |
                              +---------------------------------------+------------------+
                              |                                                          |
                runner-pool.yml duerme (ubuntu-latest)              rerun-on-runner-loss.yml (ubuntu-latest)
                DesiredCapacity=0 salvo que otro                    ¿falló + el log dice "shutdown signal"?
                workflow de pool esté queued/en progreso            -> rerun-failed-jobs (máx 3)

const asg = new autoscaling.AutoScalingGroup(this, "RunnerAsg", {
  vpc,
  mixedInstancesPolicy: { /* diez tipos Graviton, solo spot */ },
  minCapacity: 0,
  maxCapacity: 2,
  // desiredCapacity deliberadamente NO declarado.
  capacityRebalance: true,
  groupMetrics: [autoscaling.GroupMetrics.all()],
});

on:
  pull_request:
    branches: [main]
  push:
    branches: [main]
  workflow_dispatch:
  schedule:
    - cron: "50 2 * * 0"     # 10 min antes del security scan del domingo
    - cron: "*/30 * * * *"   # barredor
  workflow_run:
    workflows: [CI, Deploy Backend, Deploy Frontend, Deploy Intelligence, Deploy Marketing, Security Scan]
    types: [completed]

concurrency:
  group: runner-pool
  cancel-in-progress: false

def cmd_sleep(asg, repo, run_id, *, dry_run):
    if any_runner_busy(repo):
        print("A runner is still busy; leaving the pool up.")
        return 0

    others = other_active_runs(repo, run_id)
    if others:
        print("Other runs still need the pool, leaving it up:")
        for name in others:
            print(f"  - {name}")
        return 0

    set_desired(asg, 0, dry_run=dry_run)
    return 0

if ! name=$(aws autoscaling describe-auto-scaling-groups \
  --query "AutoScalingGroups[?Tags[?Key=='Name' && Value=='${RUNNER_ASG_TAG}']].AutoScalingGroupName | [0]" \
  --output text 2>&1); then
  echo "::warning::could not resolve the runner ASG, leaving the pool alone: ${name}"
  name=""
fi

2026-10-10T21:04:51  Successful  Launching a new EC2 instance: i-058255690229cf59c
2026-10-10T00:47:25  Successful  Terminating EC2 instance: i-0327f991f9094b5b1
2026-10-10T00:16:22  Successful  Launching a new EC2 instance: i-0327f991f9094b5b1
2026-10-07T20:32:53  Successful  Terminating EC2 instance: i-0d799f051fa663669
2026-10-07T20:12:34  Successful  Launching a new EC2 instance: i-0d799f051fa663669
2026-10-07T14:19:47  Successful  Terminating EC2 instance: i-0d92076be33dcc045
2026-10-07T14:01:02  Successful  Launching a new EC2 instance: i-0d92076be33dcc045

const timeTravelMs = Number(process.env.TIME_TRAVEL_DAYS || 0) * 86_400_000
if (timeTravelMs) {
  const RealDate = Date
  const realNow = RealDate.now.bind(RealDate)
  class ShiftedDate extends RealDate {
    constructor(...args: ConstructorParameters<typeof Date> | []) {
      if (args.length === 0) super(realNow() + timeTravelMs)
      else super(...(args as ConstructorParameters<typeof Date>))
    }
    static now() {
      return realNow() + timeTravelMs
    }
  }
  globalThis.Date = ShiftedDate as DateConstructor
}

frontend-time-travel:
  needs: changes
  if: needs.changes.outputs.frontend == 'true'
  runs-on: self-hosted
  steps:
    # checkout, pnpm, node, install
    - name: Vitest (clock 400 days ahead)
      env:
        TIME_TRAVEL_DAYS: "400"
      run: cd frontend && pnpm exec vitest run

Wake the pool  2026-10-10T21:14:30Z  gh api /repos/acme/myapp/actions/runners failed (1): gh: Resource not accessible by integration (HTTP 403)
Wake the pool  2026-10-10T21:14:41Z  gh api /repos/acme/myapp/actions/runners failed (1): gh: Resource not accessible by integration (HTTP 403)
Wake the pool  2026-10-10T21:15:11Z  No runner came online within 300s. The self-hosted jobs will queue. If spot has no capacity at all, re-run with the `runner: ubuntu-latest` input.

Sleep the pool if nothing else needs it  gh api /repos/acme/myapp/actions/runners failed (1): gh: Resource not accessible by integration (HTTP 403)
Sleep the pool if nothing else needs it  Other runs still need the pool, leaving it up:
Sleep the pool if nothing else needs it    - Deploy Backend#38086462352 (queued)
Sleep the pool if nothing else needs it    - Deploy Frontend#38086462341 (in_progress)

03:29  10:50  17:38  22:38  02:39  09:47  16:42  21:23  01:02  07:21  13:53  18:06