Skip non-relevant CI stacks based on changed file paths

[Yicong-Huang]

Task Summary

Skip irrelevant CI stacks in the Build workflow based on the PR's changed file paths, so a PR that only touches one stack does not block on every other stack.

Pain points today

The Build workflow runs the full matrix on every PR, regardless of what changed:

• frontend × 3 OS (ubuntu / windows / macos)
• scala (ubuntu-22.04, JDK 11) — sbt build + tests, ~10–15 min
• python × 4 versions (3.10 / 3.11 / 3.12 / 3.13)
• agent-service × 2 OS (ubuntu / macos)
• backport matrix per release/* label, each running the full reusable build

Concrete consequences:

• A docs-only PR (docs/**, *.md) waits for sbt and pip just to be told "nothing changed".
• A pure frontend PR pays the scala / python tax even though no Scala or Python file moved.
• Backport matrix multiplies all of the above per release target — a 1-line frontend tweak with release/v1.1.0-incubating triggers a second full sbt build on the release branch checkout.
• Required checks listed in .asf.yaml for release/* and main are statically named (e.g. frontend (ubuntu-latest, 18), scala (ubuntu-22.04, 11)), so simply gating jobs on if: would leave them as skipped and require coordination with Add a Required Checks aggregator job to consolidate CI status checks #4599 (Required Checks aggregator) so branch protection still passes.

The Build workflow's precheck job already exposes placeholder outputs run_frontend / run_scala / run_python / run_agent_service for exactly this purpose; today they are all hard-coded to true.

Proposal

Implement file-path detection in precheck, with the same rule set as .github/labeler.yml (single source of truth), to drive the existing run_* outputs:

• frontend → frontend/**
• scala → amber/** minus the python subtree, plus **/*.scala, **/build.sbt, **/project.sbt
• python → amber/src/main/python/**, **/*.py, **/requirements.txt
• agent-service → agent-service/**

push events on main / release/* / ci-enable/** keep running everything (no PR file list available, default to all true). Only pull_request events apply the gating.

Once required checks are aggregator-based (#4599), skipped stacks become non-blocking and the optimization is end-to-end.

Prior attempt

PR #3642 "chore(ci): skip non-relevant CIs" tried this in 2025-08 using dorny/paths-filter@v3 for the python subset only, sat open for 9 months, and was closed without merge on 2026-05-01. This task picks up the idea against the current precheck-centric Build workflow and broadens it to all four stacks.

Priority

P2 – Medium

Task Type

• Code Implementation
• DevOps / Deployment