Repository navigation
CI: web, Linux and Android in one x86_64 job; BuildBuddy's remote cache; a Waydroid probe - #229
Merged
Merged
Conversation
…Buddy's remote cache, a Waydroid probe (scripts; the workflow's change is in the pull request's body)
…ther, from cold (reverted before merging)
…est's); a place for a read-only key
…oid together, from cold (reverted before merging)" This reverts commit 4b0e906.
… key for main and releases alone; a Waydroid probe by hand
AttilaTheFun
approved these changes
Nov 16, 2026
AttilaTheFun
approved these changes
Nov 16, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Three changes to CI. The workflow part (
.github/workflows/build.yml) can't be pushed from this branch: the account that opened it has noworkflowspermission. It's the diff at the end of this body. Someone who can push workflows needs to commit it to this branch before merging. Everything else is on the branch.1. The web, Linux and Android in one x86_64 job
Until now, the web and Linux targets built on
ubuntu-24.04-armin the project's image, and Android built separately onubuntu-24.04(x86_64) in Swift's image, because Android's SDK tools are x86_64 programs. With this change, each stage runs one job onubuntu-24.04with one Bazel output base on the runner's second disk (/mnt/bazel) and one cache group,ubuntu:container(the required check keeps its name) replacescontainerandcontainer-android.populatereplacespopulateandpopulate-android. It builds everything, then the published Android app and the site.verifyreplacesverifyandverify-android.androidnow needsverify, andverify-androidis gone fromalpha-apple's needs and from tools/ci/verified.sh.bazel8is one x86_64 job instead of a two-runner matrix.container-verify(the headful web and Linux tests under Xvfb and Chrome) andproof's web build moved to x86_64 too, so they read the same cache group. Onlyproof-comparestays on arm64, and it uses no Bazel cache.The image.
ghcr.io/attilathefun/visor-agentis arm64 only: every tag's manifest has one platform, linux/arm64. So the combined job runs inswift:6.4.2-noble(by digest, as the Android jobs already did). When the job builds Linux,tools/ci/job_tools.shadds what the Linux apps link, matching the agent image's Dockerfile: SDL2, SDL2_ttf, GL, EGL, GTK 3, WebKitGTK, lld, libcurl, SQLite and NSS. It checks for them with pkg-config first, so in the agent image it does nothing..bazelrc'sbuild:linuxalready sets CC=clang, and MODULE.bazel already registers swift.org's ubuntu24.04 toolchains for x86_64 (exec, wasm and Android) beside aarch64. Chrome for Testing is pinned for both architectures.The cache. The
linux(3 GB) andandroid(2 GB) groups merge intoubuntu, trimmed to 5 GB, so the total in GitHub's 10 GB stays as it was.apple(2) andwindows(1) are unchanged. The cache actions can also put the caches where a job'sISOMER_BAZEL_CACHEsays. No job sets it yet: the disk cache stays in the home directory on the root disk, and Bazel's tree stays on/mnt.container-appleand the Windows jobs are unchanged.2. BuildBuddy's remote cache
.bazelrcandexamples/.bazelrcget abuildbuddyconfig, with no key: the BES and cache endpoints,--remote_timeout=10m,--remote_cache_compressionand--nolegacy_important_outputs.tools/ci/ci_bazel.sh setupreadsBUILDBUDDY_API_KEY. When it's set, setup writes--config=buildbuddyand the key's two headers into a file outside the checkout (~/.config/isomer-ci/buildbuddy.bazelrc, mode 600)..bazelrc.userimports that file by its path, so the key is never on a command line or in the log. Then setup unsets the variable.buildbuddy-api-keyinput to the setup step alone, only whengithub.event_name == 'release' || (github.event_name != 'pull_request' && github.ref == 'refs/heads/main'). Pull requests and branch runs build on GitHub's cache alone.populate,bazel8), a published release's jobs, andalpha/alpha-appleon main. They also download every output (--remote_download_all), so the disk cache that GitHub's cache keeps stays whole. Main's other runs by hand (verify, proof) only read (--noremote_upload_local_results).BUILDBUDDY_READ_KEY. ci_bazel.sh has a commented place for it. It isn't wired yet.3. A Waydroid probe (
run: waydroid-probe)This tests whether Android GUI tests could run in the same job as everything else on GitHub's hosted Ubuntu. It runs by hand only, on
ubuntu-24.04, directly on the runner (not in a container), and stops after 20 minutes. It usescontinue-on-errorand nothing depends on it.tools/ci/waydroid_probe.shdoes the following, each step time-limited:uname -aand the kernel config's binder and ashmem options, lists/dev/binder*and/dev/binderfs, and checks/proc/filesystemsfor binder;sudo modprobe binder_linux devices=binder,hwbinder,vndbinder. If that fails, it installs the kernel'slinux-modules-extrapackage and tries again. Then it mounts binderfs and checks for ashmem;waydroid init;waydroid-container), then starts a session under a headless Weston;sys.boot_completed, then runswaydroid status,adb connect <waydroid IP>:5555andadb devices.It writes a table of each step's result and a verdict to the job summary, and uploads its logs as the
waydroid-probeartifact.Testing
The workflow change can't run from this branch until it's committed, so nothing here has run the combined job, BuildBuddy or the probe under the new workflow. To test the combined build anyway, a temporary commit made the current workflow's x86_64 job (
container-android, Swift's image,/mnt) build what the combinedcontainerbuilds (ISOMER_PLATFORMS="web linux android",container_build.sh pr), from cold (no cache). That commit is reverted before this is ready.The result (run 43780835118, job
container-androidat the temporary commit). Every build and test passed, cold, in 72.5 minutes, in one output base:job_tools.sh: 35 s, Linux libraries included.Before, cold, the two jobs took 38 minutes (Linux, arm64) and 50 minutes (Android, x86_64), so 88 runner-minutes, but running side by side the check finished in 50 minutes. Combined, it's 72.5 runner-minutes but about 22 minutes longer to wait for, cold. Warm runs, from the merged cache, should be well under that.
Disk. On this runner
/mntis the root disk (145 GB, 81 GB free at the start). Bazel's tree used 39 GB, leaving 43 GB free with no disk cache in play. A coldpopulatewrites a disk cache of about the same size beside it, which should still fit.The arm64
containerjob (old workflow, branch scripts) also passed, in 35 minutes. In the project's imagejob_tools.shtook 7 s, the same as on main, so it added none of the Linux libraries.container-windowsfails here as it does on main (populate-windows, MAX_PATH), which is unrelated.BuildBuddy has no link from this pull request's runs, by design (pull requests don't get the key). Main's first
populateafter merging is the first run that connects, fills the cache and prints theStreaming build results to: https://app.buildbuddy.io/invocation/…links. Main's next build shows the hit rate and the warm time. Pull requests stay on GitHub's cache until there is a read-only key.The workflow's change
git applyaccepts it on this branch's head: