@angular/build 22.1.5: unit-test builder kills the esbuild service during Building on a 2-CPU/7GB runner (regression from 22.1.4, framework-independent)

[cert-jverdaasdonk]

Command

test

Is this a regression?

• Yes, this behavior used to work in the previous version

The previous version in which this bug was not present was

22.1.4

Description

@angular/build:unit-test (vitest runner) reliably kills the esbuild service during the build phase on a 2-CPU / 7 GB machine — the shape of a private-repo ubuntu-latest GitHub Actions runner. The identical suite builds and runs fine on 22.1.4.

It fails before any test executes, so this is not a test-isolation or teardown problem.

Version bisection (2×2). Our project moved the Angular framework and the builder in one commit, so both were tested independently. Each cell is the identical suite (210 spec files / 3364 tests) in the identical container:

@angular/core	@angular/build + @angular/cli	result
22.1.2	22.1.4	builds, whole suite runs
22.1.3	22.1.4	builds, whole suite runs
22.1.2	22.1.5	build dies — "The service was stopped"
22.1.3	22.1.5	build dies — same

@angular/build 22.1.5 is necessary and sufficient; the framework version is irrelevant.

Already ruled out, so as not to waste triage time:

• Not the Node heap. NODE_OPTIONS=--max-old-space-size=4096 changes nothing. The victim is esbuild's separate service process, not V8.
• Not (visibly) the kernel OOM killer. On GitHub Actions the job dies with SIGTERM (143); there is no exit 137 in any run, and 137/SIGKILL is what a cgroup OOM kill looks like.
• Not test isolation. --isolate=true does not help. It cannot — the failure precedes test execution.
• Not visible on a developer machine. The full suite is green locally on both 22.1.4 and 22.1.5 on a many-core box. Only the constrained machine shows it, which is why it reached us as a CI-only failure.

Minimal Reproduction

No public reproduction repository yet — I could not reduce it to ng new plus a handful of specs, because the failure appears to need a substantial spec set and does not occur on an unconstrained machine. What I can offer instead is a reproduction that needs no CI account and takes ~57 seconds, from any existing Angular workspace using @angular/build:unit-test:

# from the workspace root, with @angular/build + @angular/cli at 22.1.5
docker run --rm --cpus=2 --memory=7g --memory-swap=7g \
  -v "$PWD":/app -w /app node:24 \
  bash -lc 'corepack enable && pnpm install --frozen-lockfile && pnpm exec ng test --watch=false'

Steps:

1. Run the above. The build fails with The service was stopped (see below).
2. Change only @angular/build and @angular/cli to 22.1.4, reinstall, run again. The build succeeds and the suite runs to completion.
3. Change only @angular/core and its siblings between 22.1.2 and 22.1.3 with the builder pinned either way — no effect on the outcome.

--cpus=2 --memory=7g is load-bearing: the failure does not appear without the constraint.

Our workspace passes runnerConfig, but its contents are an empty defineConfig({}), so no project-level vitest tuning is involved.

If a public repro repository is required to action this, I am glad to build one — please say whether you would prefer a generated workspace with N synthetic specs, and roughly what N you would consider acceptable. The bisect loop is under a minute here, so additional data points are cheap; I am happy to run any experiment you would find useful.

Exception or Error

Using Vitest configuration file: /app/vitest-base.config.ts
❯ Building...
✖ Building... [FAILED: The service was stopped]
An exception occurred while creating the test executor:
Error: The service was stopped
    at /app/node_modules/.pnpm/esbuild@0.28.2/node_modules/esbuild/lib/main.js:1259:26
    at responseCallbacks.<computed> (/app/node_modules/.pnpm/esbuild@0.28.2/node_modules/esbuild/lib/main.js:886:9)
    at Socket.afterClose (/app/node_modules/.pnpm/esbuild@0.28.2/node_modules/esbuild/lib/main.js:877:28)
    at Socket.emit (node:events:521:24)
    at endReadableNT (node:internal/streams/readable:1736:12)
    at process.processTicksAndRejections (node:internal/process/task_queues:90:21)

On GitHub Actions the same regression surfaces far less legibly — twice on the same commit:

# run 1 — died 67s in, during the build, before any test
12:09:02  ❯ Building...
12:10:09  undefined
12:10:09  [ERR_PNPM_RECURSIVE_EXEC_FIRST_FAIL] Command was killed with SIGTERM (Termination): ng test '--watch=false'
12:10:09  ##[error]The operation was canceled.

# run 2 — same commit, re-run: tests ran and PASSED, then 63s of silence
12:26:04  ✓ src/app/shared/components/cfm-form/cfm-form-validators.spec.ts (89 tests)
12:27:07  ##[error]The operation was canceled.

The bare undefined in run 1 appears to be the parent trying to log an error object from a child that was already gone.

Your Environment

Angular CLI       : 22.1.4      <-- works;  22.1.5 fails
Angular           : 22.1.3      <-- not the variable; 22.1.2 behaves identically
Node.js           : 24.19.0
Package Manager   : pnpm 11.22.0
Operating System  : win32 x64 (local, passes) / linux x64 container, 2 CPU / 7 GB (fails)

┌───────────────────────────┬───────────────────┬───────────────────┐
│ Package                   │ Installed Version │ Requested Version │
├───────────────────────────┼───────────────────┼───────────────────┤
│ @angular/build            │ 22.1.4            │ 22.1.4            │
│ @angular/cdk              │ 22.1.3            │ 22.1.3            │
│ @angular/cli              │ 22.1.4            │ 22.1.4            │
│ @angular/common           │ 22.1.3            │ 22.1.3            │
│ @angular/compiler         │ 22.1.3            │ 22.1.3            │
│ @angular/compiler-cli     │ 22.1.3            │ 22.1.3            │
│ @angular/core             │ 22.1.3            │ 22.1.3            │
│ @angular/forms            │ 22.1.3            │ 22.1.3            │
│ @angular/language-service │ 22.1.3            │ 22.1.3            │
│ @angular/material         │ 22.1.3            │ 22.1.3            │
│ @angular/platform-browser │ 22.1.3            │ 22.1.3            │
│ @angular/router           │ 22.1.3            │ 22.1.3            │
│ @angular/service-worker   │ 22.1.3            │ 22.1.3            │
│ rxjs                      │ 7.8.2             │ ^7.8.2            │
│ typescript                │ 6.0.3             │ ~6.0.3            │
│ vitest                    │ 4.1.11            │ ^4.1.11           │
│ zone.js                   │ 0.16.2            │ ^0.16.2           │
└───────────────────────────┴───────────────────┴───────────────────┘

esbuild 0.28.2 (transitive)
Suite size: 210 spec files, 3364 tests

Anything else relevant?

I searched open and closed issues first. Closest matches, none of which report this as a 22.1.4 → 22.1.5 regression:

• The service was stopped #28223 and The service was stopped #28227 — same The service was stopped error, but on much older CLI versions and centred on Windows / Node 22.
• [Bug]: service was stopped error with esbuild-wasm incorrectly loading (CI Only) thymikee/jest-preset-angular#1709 — Jest rather than vitest, but notes the same "4 vCPU / 7 GB CI agent" class of environment, and that it does not occur with native esbuild.
• @angular/build:unit-test virtual init-testbed.js guards initTestEnvironment() behind a once-per-worker symbol → stale DomAdapter under vitest ≥4.0.5 + isolate: false #33047 and unit-test (vitest): isolate:false default silently breaks cross-file vi.mock() of shared modules on low-CPU workers #33570 concern isolate: false and low-CPU workers. They look like a different bug in the same neighbourhood — --isolate=true has no effect on the failure reported here, which occurs before tests run.

Package manager is pnpm, which puts esbuild's binary under node_modules/.pnpm/…; I have not tested whether npm or yarn behave differently.