Why vx is fast: five decisions, not a trick

With nothing changed, vx adds 559 ms to a 9,603-task run. No daemon keeps it warm.

Nothing changed

vx559 ms

Turborepo689 ms

Nx12.07 s

Vite Task10.71 s

Nothing changed, outputs restored

vx1.04 s

Turborepo1.32 s

Nx11.94 s

Vite Task10.57 s

Cold build

vx6.68 s

Turborepo11.18 s

Nx3 min 12 s

Vite Task8.51 s

Cold build: CPU the runner burns

vx42.19 s

Turborepo1 min 19 s

Nx5 min 31 s

Vite Task24.17 s

The table

Time added over the ideal run, 1,601 projects, 9,603 tasks	vx	Turborepo	Nx	Vite Task
Nothing changed	559 ms	689 ms	12.07 s	10.71 s
Nothing changed, outputs restored	1.04 s	1.32 s	11.94 s	10.57 s
Cold build	6.68 s	11.18 s	3 min 12 s	8.51 s
Cold build: CPU the runner burns	42.19 s	1 min 19 s	5 min 31 s	24.17 s

Synthetic monorepo, 1,601 projects, 30 levels · concurrency 10, native configs · benchmarks

Measured: vx 0.0.0 @ ee4c63a57 on 2026-11-12, Turborepo 2.11.10 on 2026-11-12, Nx 24.0.2 on 2026-11-12, Vite Task 1.4.0 on 2026-11-12; linux x64, 4 cores (AMD EPYC 7763 64-Core Processor), 16 GB.

Five design choices make that number. Each one also makes the cache harder to get wrong.

1. Git already hashed your files

xychart-beta horizontal
  title "Hash 2,527 tracked files (ms, min of runs)"
  x-axis ["git ls-files", "git status", "read + hash every file"]
  y-axis "ms" 0 --> 160
  bar [2.3, 5.6, 154.1]

This repo, 48 MB · linux x64, 4 cores, warm page cache · 2026-10-23

Git keeps a hash of every tracked file. vx asks git instead of reading them. Clean-tree key derivation costs zero source-file reads.

A dirty file gets the same hash git would give it, so a commit never changes a key (how).

2. The scheduler counts with bitsets

vx starts the task with the most work below it first. Counting that with a graph walk took 8.5 s of a 10 s run on a 1,090-package graph.

Bitsets do it in milliseconds (how).

3. A hit on an up-to-date tree is a few stat calls

flowchart LR
  H[cache hit] --> S{files on disk match the record?}
  S -->|yes| D[done: no copy, no write]
  S -->|no| C[copy the cached files]

vx wipes outputs before a task runs, so it knows what a hit leaves on disk. If each file still matches (size, mode, millisecond mtime, inode and ctime), there is nothing to restore.

4. A hit is a file copy

xychart-beta horizontal
  title "Restore 1,000 entries of 96 KiB (s, min of 5)"
  x-axis ["from tar.zst", "plain files"]
  y-axis "s" 0 --> 1.4
  bar [1.26, 0.76]

The local cache stores plain files and the kernel copies them.

Metadata and the captured stdout live in the index (the stdout in a side table), so the log comes back with one SQL read (how).

5. Nothing runs that is not needed

A run reads .vx/projects.lock.json and evaluates no config. No telemetry plugin means no listener and no git spawn. A plugin that skips a task costs nothing.

Every change to this path ships with a measured before and after (Optimizations). More at vznjs.github.io/vx.