DTE workers that fit the work, and a footer that says how full they ran

On vx’s own main branch, nine Linux workers ran 88% busy for one run. The footer prints that for every worker, every run.

xychart-beta
  title "Machine busy, per worker (%)"
  x-axis ["small#1", "#2", "#3", "#4", "#5", "#6", "#7", "#8", "#9", "mac#3"]
  y-axis "%" 0 --> 100
  bar [94, 92, 89, 90, 88, 92, 85, 88, 86, 76]

vx CI on GitHub runners, 160 tasks to run · 4a86e02 · 2026-11-07 · the run

A Mac starts only for a Mac task

dte({
  workers: {
    hosted: {
      machines: {
        small: { max: 9, cpu: 4, memory: '16GB', platform: 'linux-x64' },
        arm: { max: 1, cpu: 4, memory: '16GB', platform: 'linux-arm64' },
        mac: { max: 4, cpu: 3, memory: '7GB', platform: 'darwin-arm64' },
      },
    },
  },
})

A task runs on a machine of its own platform, Linux unless it names another. A Mac starts only when a task needs macOS, and only as many as those tasks fill. exec.worker sends a task to one named type.

A task is booked by what it used last time

flowchart LR
  H[run history] -->|peak memory, average CPU| B[booking]
  B --> F{a core free and the memory fits?}
  F -->|yes| W[sent to the worker]
  F -->|no| Q[waits, longest chain first]

Memory is booked at its peak, since running out kills a task. CPU is booked at its average, so a worker is not left half idle.

A server stops once no task needs it

A dev server or database a test depends on starts on that test’s worker. It stops when no task left in the run uses it, and its CPU is charged to the tasks it served.

The footer shows three numbers per worker

  hosted.small#1  left     12 done · warm-up 91.9s
                  declared 4c/16G · actual 4c/17G
                  online 122.2s · busy 122.1s · idle 0.1s
                  cpu utilization machine 94% · tasks 67% · worker 3%
  best            8 × hosted.small + 1 × hosted.arm + 2 × hosted.mac: 160.9s
  total           cpu utilization machine 88% · tasks 63% · worker 2%

machine counts any process on the box, tasks only your tasks, and worker vx itself. The best line says which workers would have run this run soonest, with the fewest machines.

How DTE places work: vx-dte. More at vznjs.github.io/vx.