vx on real repos: wall times

The synthetic benchmark runs one generated
graph. This post times repos someone else wrote: open-source monorepos
that already use Turborepo or Nx, each moved to native vx config, then
run under both tools on the same machine.

Both tools build the same outputs; that is the assumption behind every
row.

Method

• One cloud container: 4 vCPU (Intel Xeon @ 2.10GHz), 15 GB RAM, Linux, nothing else running. Bun 1.4.2, Node 22.22.0.
• vx built from main with bun build --compile: 522c23b for create-t3-turbo, react-email, trpc and typescript-eslint; 731cf92 for vercel/ai, medusa and TanStack query; 1ee058c for the react-email clone row.
• Configs written by bunx @vzn/vx-migrate --native --no-install, then vx lock.
• Concurrency 10 for every tool: Turborepo --concurrency=10, Nx --parallel=10, vx --concurrency 10.
• Local cache only: no remote cache, no daemon, no telemetry.
• Four arms, each tool wiping only its own cache:
  • cold: cache and every build output wiped.
  • restore: outputs wiped, cache kept.
  • edit: one source file gets a new line (new content each rep), outputs kept.
  • clone: a fresh git clone at another path, node_modules copied, sharing the first checkout's cache, with no outputs.
• Three reps, the order alternating between tools. Each cell is the minimum of the three, in seconds. Every run exited 0.

Results

Seconds, min of 3. Shorter bars are less time; each chart keeps its table one click away.

create-t3-turbo vs Turborepo 2.5.8

build typecheck, 11 tasks, commit 8f945b7.

Arm	Turborepo 2.5.8	vx
cold	13.63 s	10.97 s
restore	0.28 s	0.09 s
edit	4.69 s	4.89 s
edit, clean: false	4.61 s	3.83 s
clone	0.28 s	0.11 s

Both edit rows are a rerun on vx 5b2b0a1 (same machine shape, min of
3). The second sets cache.outputs.clean: false on the three tsc build
tasks: by default vx deletes a task's declared outputs before it runs on
a miss, tsc's .cache/tsbuildinfo.json included, so tsc rebuilds api
from scratch.

react-email vs Turborepo 2.9.14

build, 4 tasks, commit 1531a39.

Arm	Turborepo 2.9.14	vx
cold	28.92 s	25.11 s
restore	0.32 s	0.22 s
edit	28.23 s	26.29 s
clone	0.34 s	0.26 s

trpc vs Turborepo 2.10.12

build, 7 tasks, commit d756e59.

Arm	Turborepo 2.10.12	vx
cold	9.29 s	7.48 s
restore	0.17 s	0.15 s
edit	9.40 s	7.04 s
clone	0.16 s	0.13 s

vercel/ai vs Turborepo 2.9.14

build, 82 tasks, commit e8c8e32.

Arm	Turborepo 2.9.14	vx
cold	144.96 s	135.33 s
restore	0.92 s	0.45 s
edit	156.38 s	145.61 s
clone	0.97 s	0.65 s

medusa vs Turborepo 1.13.4

build build:plugin, 83 tasks, commit 5018ab67.

Arm	Turborepo 1.13.4	vx
cold	316.57 s	307.81 s
restore	3.56 s	3.45 s
edit	268.92 s	259.96 s
clone	2.59 s	3.54 s

Clone row

vx's clone takes what its restore takes (3.54 s and 3.45 s, all 83 hits).
Turborepo's clone takes less time than its own restore.

TanStack query vs Nx 23.2.1

build, 25 tasks, commit 817bd02.

Arm	Nx 23.2.1	vx
cold	35.01 s	29.37 s
restore	0.94 s	0.32 s
edit	34.10 s	29.03 s
clone	0.96 s	0.58 s

Clone row

Nx's first clone rep missed its cache (36.42 s): Nx 23 indexes its local
cache in the checkout's own .nx/workspace-data. The chart shows the
minimum.

typescript-eslint vs Nx 23.2.0

build, 18 tasks, commit 57e7cccce.

Arm	Nx 23.2.0	vx
cold	52.50 s	41.77 s
restore	4.45 s	5.32 s
edit	40.69 s	38.97 s
clone	4.15 s	5.37 s

Restore and clone rows

vx hits 16 of 18, Nx 17. ast-spec#build stays a miss by design: its
inputs include committed snapshots its own test may rewrite, and vx never
keys a task on files another task in the run can write.

Nx's first clone rep missed its cache (56.82 s), as with TanStack query.

Run it on yours

npx @vzn/vx init
npx @vzn/vx run build --all

Then time your runner's command and vx run on the same machine, each
on its own 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.}