ci: Build each Linux server target on a runner of its architecture

What does this MR do and why?

Implements the approach agreed in #53 (proposed by @vtak, agreed by @ealcantara): build each Linux server variant natively on a runner of its target architecture, so the packaged native modules actually match the package name.

Problem

Every vscode-reh-web-linux-arm64-*.tar.gz published so far ships x86-64 native binaries. Checked on the latest release (1.94.0-1.0.0-dev-20251209191616):

file in the linux-arm64 tarball ELF machine
./node aarch64 ✅
node_modules/node-pty/build/Release/pty.node x86-64 ❌
node_modules/@vscode/spdlog/build/Release/spdlog.node x86-64 ❌
node_modules/@parcel/watcher/build/Release/watcher.node x86-64 ❌
node_modules/kerberos/build/Release/kerberos.node x86-64 ❌
node_modules/native-watchdog/build/Release/watchdog.node x86-64 ❌
node_modules/@vscode/deviceid/build/Release/windows.node x86-64 ❌
node_modules/@vscode/ripgrep/bin/rg x86-64 ❌

build-server runs one npm ci on an amd64 runner, then loops all five targets through the vscode-reh-web-<os>-<arch>-min-ci gulp tasks. Those are packaging tasks: upstream's build/gulpfile.reh.js copies remote/node_modules verbatim from the host and only downloads the node runtime per target - which is why the arm64 "build" takes 12 seconds and only its node binary is correct. Upstream doesn't support cross-compiling the native modules; Microsoft's own pipelines run one job per architecture with npm_config_arch set (build/azure-pipelines/linux/product-build-linux.yml).

Changes

  • Split build-server into build-server-amd64 (saas-linux-medium-amd64: darwin-x64 darwin-arm64 win32-x64 linux-x64) and build-server-arm64 (saas-linux-medium-arm64: linux-arm64), sharing a .build-server-base template. Both emit into .build/vscode-reh-web, so publish-packages and the released file set are unchanged.
  • scripts/build-and-bundle-server.sh reads the target list from $VSCODE_BUILD_TARGETS (default: the previous hardcoded list, for local builds).
  • The dependency cache key is sharded per architecture (vscode-fork-dependency-cache-{amd64,arm64}): the cache contains compiled native modules, so an arm64 job must never restore amd64 node_modules. push-cache keeps pushing only the amd64 cache for now, so the arm64 job installs cold (~10 extra minutes); an arm64 cache twin can be a follow-up.
  • New scripts/verify-server-package-arch.sh runs after every server build and fails the job if any packaged native binary (*.node, ripgrep, the node runtime) doesn't match the architecture of the host-matching Linux target. This bug has shipped silently since 2022 (#8 (closed), #12 (closed), #53); with the guard it cannot ship silently again. (It reads the ELF e_machine bytes with od, since file isn't installed in the CI image.)

Wall-clock should stay roughly flat: the two jobs run in parallel and each builds fewer targets than before.

Dependencies

  • Depends on gitlab-build-images!1111 (merged): this project's CI image (gitlab-vscode-build) is amd64-only today, so the arm64 job can't start until the image exists for arm64. Draft until that merges.
  • Releases are maintainer-only tag pipelines, so the first fixed linux-arm64 package ships with the next tag cut after this merges.

Open questions for maintainers

  1. The darwin-*/win32-x64 packages have never contained working native modules either (same root cause - they get this host's linux-x64 binaries). This MR deliberately keeps packaging them unchanged on the amd64 job to avoid changing the published artifact set. Is anything consuming them, or should a follow-up drop them?
  2. @ealcantara - the split-web-ide-and-workspaces-forks branch touches the same files; happy to rebase this MR if that lands first.

Follow-ups (tracked in gitlab#550128)

Per the sequencing in gitlab-build-images!905 (merged)#note_2566824178: next comes making the workspaces-tools image fetch per-arch fork tarballs (fetch_assets.sh currently downloads the first asset URL it finds) and building it multi-arch, then the default workspace image, then switching the Rails digest pins back to tags. We are happy to contribute those MRs as well.

Who's asking

Testing attestation (2026-09-04)

Artifacts under test come from the first fully green pipeline, https://gitlab.com/gitlab-org/gitlab-web-ide-vscode-fork/-/pipelines/2820298281 (build-server-arm64 on saas-linux-medium-arm64, 67 min).

1. In-job arch guard (scripts/verify-server-package-arch.sh, job 16307470784):

OK: all 9 ELF binaries in vscode-reh-web-linux-arm64 are linux-arm64 (e_machine 0xb700); 1 non-ELF file(s) skipped

2. Independent ELF audit of the packaged vscode-reh-web-linux-arm64-1.94.0-1.0.0-dev-20260904124619.tar.gz (file(1) on an unrelated machine). Every binary that shipped as x86-64 in previous releases is now aarch64:

binary result
node ELF aarch64
node_modules/node-pty/build/Release/pty.node ELF aarch64
node_modules/@vscode/spdlog/build/Release/spdlog.node ELF aarch64
node_modules/@parcel/watcher/build/Release/watcher.node ELF aarch64
node_modules/kerberos/build/Release/kerberos.node ELF aarch64
node_modules/native-watchdog/build/Release/watchdog.node ELF aarch64
node_modules/@vscode/deviceid/build/Release/windows.node ELF aarch64
node_modules/@vscode/ripgrep/bin/rg ELF aarch64
extensions/ms-vscode.js-debug/src/w32appcontainertokens-*.node PE32+ x86-64 — Windows-only helper that upstream ships on every platform

Highest GLIBC_ symbol referenced by the shipped binaries: 2.28 — the documented workspace container floor is preserved.

3. Runtime test on real hardware, both architectures. Containers running Debian 10 (buster) — i.e. glibc 2.28, the minimum documented workspace platform — on AKS: the linux-arm64 tarball on an arm64 (Ampere) node, and the linux-x64 tarball from the same pipeline on an amd64 node. (libgssapi-krb5-2 installed first; workspace images ship it, bare debian:10 does not.) Results identical on aarch64 and x86_64:

  • ./node -v → v20.16.0
  • require() of node-pty, @vscode/spdlog, kerberos, native-watchdog, @parcel/watcher → all load
  • node_modules/@vscode/ripgrep/bin/rg --version → ripgrep 13.0.0
  • bin/gitlab-webide-server --host 127.0.0.1 --port 8000 --without-connection-token → Extension host agent listening on 8000; GET /version → HTTP 200, body d78a74bcdfad14d5d3b1b782f87255d802b57511
Edited by Miguel Sacristan

Merge request reports

Loading
Loading