Add phase-1 native arm64 nightly CI (regress + isolation on ubuntu-24.04-arm)

[ibrahim halatci (ihalatci)]

What

Adds a phase-1, bounded nightly GitHub Actions workflow that natively compiles Citus and runs its core regression suites on arm64 (ubuntu-24.04-arm). This is the upstream half of the multi-arch effort tracked by citus#8612.

Docker (multi-arch alpine image) and packaging (gated arm64 .deb legs) have already shipped their arm64 support; the remaining gap was that no CI validated that Citus itself builds and its tests pass on arm64 — a search of .github/ finds zero arm64 / aarch64 / ubuntu-24.04-arm references before this change. This PR closes that gap.

Files changed

• NEW .github/workflows/nightly_arm64.yml
• MODIFY .github/actions/setup_cassert_pg/action.yml (optional cassert input; cache-key hardening)

Design

Reuses the arch-portable from-source build seam

The workflow reuses .github/actions/setup_cassert_pg, which builds PostgreSQL from source via pgenv and then ./configure && make all && make install-all for Citus on a bare runner — no prebuilt, amd64-locked container. Getting native arm64 coverage is therefore mostly a matter of runs-on: ubuntu-24.04-arm.

Bounded phase-1 matrix (intentional, to cap arm-runner cost)

• PostgreSQL: 17 (17.9) and 18 (18.3) only.
• Suite groups: regress and isolation only (≈4 test jobs), reusing the exact make-target lists from nightly_cassert.yml.
• Deliberately dropped for phase 1 (with an in-file comment to expand once green): arbitrary-configs, pg-upgrade, citus-upgrade, columnar, failure, tap, generator.
• Triggers: schedule at 0 4 * * * (04:00 UTC — offset from the 03:00 cassert nightly cron) + workflow_dispatch.

Clean pass/fail signal → assert-off build

The purpose of this nightly is arm validation (green = "Citus works on arm"). setup_cassert_pg normally builds an --enable-cassert PG whose nightly is "expected red" as pre-existing asserts surface — that would muddy the arm signal. So this workflow builds assert-off:

• A new optional input cassert (default "true") was added to setup_cassert_pg. The default preserves the existing cassert nightly's behavior byte-for-byte.
• When cassert: "false" (which the arm workflow passes), the action strips --enable-cassert from the copied pgenv default.conf (the source config is untouched) and skips the guard that otherwise hard-fails a non-cassert build.

Tradeoff: an assert-off build won't catch arm-specific assertion violations. That's acceptable for a phase-1 pass/fail gate; a cassert arm leg can be layered on later once the basic gate is reliably green.

Cache-key correctness (arch + cassert segments)

The pgenv cache key gained ${{ runner.arch }} and a cassert${{ inputs.cassert }} segment. runner.os is Linux on both amd64 and arm64 runners, so without an arch segment an amd64 ~/.pgenv (prebuilt binaries) could be restored onto an arm64 runner. The cassert segment keeps assert-on and assert-off builds in separate cache entries (the assert-off edit is applied to the copied config, so it does not change hashFiles('.devcontainer/pgenv/config/default.conf')). Net effect on the existing cassert nightly: a one-time cache rebuild, then identical behavior.

Failure notifications

The notify job mirrors the cassert nightly's auto-file-issue-on-failure logic but uses an arm64-specific label (nightly-arm64), marker, and title, so its failure issues never collide with the cassert nightly's.

Explicitly out of scope (follow-up)

Gating pull requests on arm64. build_and_test.yml / run_tests.yml run inside prebuilt, amd64-only containers (ghcr.io/citusdata/extbuilder:*, ghcr.io/citusdata/exttester:*). arm64 PR-gating would first require rebuilding that whole builder/tester image toolchain multi-arch (in citusdata/the-process). That is the natural next step if/when arm64 PR-gating is desired — not part of this PR.

Guarantees / blast radius

• The existing amd64 pipelines (build_and_test.yml, run_tests.yml, nightly_cassert.yml) are not modified. The only shared change is the optional, default-preserving cassert input on setup_cassert_pg.
• Validated with actionlint (clean on the new workflow) and YAML parse.

Notes

• This PR is opened as a DRAFT. First scheduled/dispatched runs may reveal arm-specific compile or test issues; the intent is to surface those for triage rather than to deep-fix arm bugs in this PR.

Refs citus#8612.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 88.72%. Comparing base (92d0b92) to head (b79d864).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #8816      +/-   ##
==========================================
- Coverage   88.73%   88.72%   -0.01%     
==========================================
  Files         289      289              
  Lines       65013    65013              
  Branches     8203     8203              
==========================================
- Hits        57691    57686       -5     
- Misses       4954     4956       +2     
- Partials     2368     2371       +3     

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.