When running beats downloading

If a download would take longer than running the task, vx runs the task.
Here 500 tasks are remote hits, and 20 of them are 8 MB on a 4 MB/s
link:

$ vx run build --all --dry
500 task(s) planned, 500 cache hits (500 remote, 153 MB).

$ vx run build --all
  cache     local 20↑ · remote 480↓ 129 KB
  time      1.02s

Those 20 would take 2 s each to download. Their tasks run in
milliseconds, so vx ran them here (local 20↑) and downloaded the
other 480.

The remote says	Full run	Downloaded
size and task time	1.15 s	129 KB
size only	3.12 s	153 MB

Min of 3 interleaved runs, empty local cache · private/vx-bench/generate.ts, 500 projects, 20 writing 8 MB · Turborepo-protocol stub 30 ms away at 4 MB/s, turboCache() · Linux x64, 4 cores · vx 22f6eb7 · 2026-10-23

Weigh the download against the task

The remote has to tell vx the entry's size and its task's time. Turbo's
cache API sends both (Content-Length, x-artifact-duration), and
turboCache() passes them on. Twice the task time, because a run here
also saves its outputs. The task counts as a miss and uploads nothing.

A remote that doesn't send the time is always waited for, as in the
second row. vx won't guess.

Know what a run will pull

The footer shows the bytes each way. --dry shows what the remote hits
would download, before anything runs.

For plugin authors

A remote's has returns { size, durationMs }, either one, {}, or
null for a missing key. hasMany returns a Map of the same. Send a
size and you get the --dry figure. Send a duration too and you get
the slow-download check. Your own cache
shows a whole remote.

Setup: Remote cache. More at
vznjs.github.io/vx.

---

Originally published on the vx blog. vx is open-source (MIT) task orchestration with a build cache: vznjs.github.io/vx · GitHub.

_{Written with AI assistance.}