CI workflows

Apache Arrow's CI is implemented as GitHub Actions. Every workflow lives in .github/workflows/. This page summarizes what each one does.

Build / test workflows (one per language)

Workflow	Triggers	What it builds	Where it runs
cpp.yml	Push, PR	C++ core + Parquet + Acero + Dataset, plus c_glib in many jobs	Ubuntu, Fedora, AlmaLinux, macOS, also static-only and bundled-only variants
cpp_extra.yml	Manual / scheduled	Extended C++ matrix (older toolchains, sanitizers, alternative deps)	Various
cpp_windows.yml	Push, PR	C++ + Parquet on Windows	windows-latest
cuda_extra.yml	Manual	CUDA-enabled C++ build	Self-hosted GPU runners
python.yml	Push, PR	PyArrow with various component combinations	Ubuntu, macOS, Windows
r.yml	Push, PR	R package on Linux + macOS + Windows	Multiple R versions
r_extra.yml	Manual / scheduled	Extended R matrix (cran-likeness checks, dev R, valgrind)	Linux
r_nightly.yml	Scheduled	R-universe nightly builds	Ubuntu
ruby.yml	Push, PR	Ruby gems on Ubuntu, Fedora, AlmaLinux, Debian, macOS	Various
matlab.yml	Push, PR	MATLAB bindings	macOS, Linux, Windows with MATLAB runners
archery.yml	Push, PR	The archery CLI itself	Ubuntu

Cross-cutting workflows

Workflow	Purpose
integration.yml	Runs the cross-language integration test suite (C++ ↔ Java ↔ Go ↔ JS ↔ Rust ↔ C# ↔ nano-arrow).
dev.yml	Repository hygiene: pre-commit, clang-format, RAT (Release Audit Tool), generated-file drift checks.
docs.yml	Builds the Sphinx documentation.
docs_light.yml	Lightweight docs build for PRs that only touch documentation.

Bots and automation

Workflow	Purpose
dev_pr.yml	Validates PR titles against the GH-NNN: [Component] Title convention.
dev_pr/	Sub-workflows triggered by PR-related events.
pr_bot.yml	The bot that handles @github-actions crossbow submit ... comments.
pr_review_trigger.yml	Triggers downstream checks when reviews complete.
comment_bot.yml	Handles the various comment commands (rebase, label, ...).
check_labels.yml	Validates issue/PR labels.
issue_bot.yml	Handles the take comment to assign issues.
stale.yml	Closes stale issues and PRs.
report_ci.yml	Aggregates CI results into a single status.

Release workflows

Workflow	Purpose
release.yml	Builds release artifacts: source tarball, binary wheels, deb/rpm packages. Triggered by tag pushes.
release_candidate.yml	Builds release candidates and posts them for verification.
verify_rc.yml	Verifies an RC: signature checks, integration tests, source builds across platforms.
package_linux.yml	Builds Linux distro packages (Apt, Yum).

What "required" looks like for a PR

Every PR triggers the per-language workflow for the directories it touched (cpp.yml if cpp/ changed, etc.). The integration workflow runs whenever format files or core C++ change. The lint and PR-validation workflows run on every PR. Crossbow runs are triggered manually by maintainers via pr_bot.yml.

Reusable scripts

Workflows mostly delegate the actual work to scripts in ci/scripts/:

• cpp_build.sh, cpp_test.sh — C++.
• python_build.sh, python_test.sh — Python.
• r_install.sh, r_test.sh, r_pkgdown.sh — R.
• ruby_test.sh — Ruby + c_glib.
• c_glib_build.sh, c_glib_test.sh — c_glib.
• integration_arrow.sh, integration_dask.sh, integration_substrait.sh — Cross-implementation tests.
• install_minio.sh, install_azurite.sh, install_gcs_testbench.sh — Local cloud-storage emulators.
• install_thrift.sh, install_protobuf.sh, etc. — Toolchain installers used when building from source.

All of these are also runnable locally inside the corresponding compose.yaml services.

Self-hosted runners

A few workflows use self-hosted runners — notably the CUDA and ARM nightly jobs. The configuration is declared in the workflow files via runs-on: [self-hosted, ...] labels and is provisioned by runs-on.com. The README acknowledges runs-on as a CI sponsor.

Crossbow dispatcher

Crossbow tasks (dev/tasks/) are dispatched via the GitHub Actions in the apache/arrow-crossbow sibling repo. The PR bot in pr_bot.yml triggers a Crossbow run by posting to the dispatcher with a list of tasks. Each task's success/failure is reported back to the originating PR via comment.