[INFRA] Run the Velox x86, ARM and Kafka PR pipelines in contributors' forks
维护者通常 3 天内回复
还没有人认领这个 Issue。
评估
- 难度
- 5/5
- 预计耗时
- 一周以上
- 新手友好度
- 25/100
- Issue 类型
- 功能
- 描述清晰度
- 基本清楚
- 活跃度
- 活跃
- 技术栈
- github-actions, scala, yaml
- 领域
- ci-cd, infrastructure
调研方向
Read #13535 and inspect .github/workflows/velox_backend_x86.yml to understand the draft fork workflow and its pilot runs. The issue leaves open how to support contributors who cannot enable Actions and says the cache fix still needs CI validation. Done means the fork run and PR result reporting work reliably within the stated ASF constraints, with those open cases addressed.
由索引模型根据 Issue 内容生成。
描述
Problem
In the seven days from 2026-10-27 to 2026-11-02, apache/gluten used about 292,000 GitHub Actions runner-minutes (job durations summed from the Actions API, re-runs included). The ASF policy limits a project's average weekly usage to 250,000 minutes, and projects that keep exceeding it lose Actions access until they fix their builds.
| Workflow (event) | Runs | Runner-minutes | Share |
|---|---|---|---|
Velox Backend (x86), pull_request |
165 | 209,502 | 71.8% |
Delta Spark UT (Gluten), pull_request |
57 | 36,876 | 12.6% |
| All other workflows | 45,590 | 15.6% |
All 165 x86 runs and all 57 Delta runs came from fork PRs.
Proposal
Run velox_backend_x86.yml, velox_backend_arm.yml and velox_kafka.yml for fork PRs in the contributor's fork on push, and report the results on the PR, as Spark has done since SPARK-35048. Fork PR runs of the ARM and Kafka pipelines used another 7,560 and 5,928 runner-minutes that week. Runs in a fork count against the fork owner's account, standard runners are free on public repositories, and a public fork gets the same 4-CPU / 16 GB runners as apache/gluten.
#13535 implements this, and its description has the details and the pilot runs. In short, the fork run tests the PR squash-merged onto the current apache/gluten main and restores the native-build ccache that apache/gluten saves on main. A scheduled workflow in apache/gluten then copies the fork run's result into a check on the PR, without pull_request_target, which the ASF disables from November 2026.
In the pilot, the native build took 4 to 5 minutes with a restored ccache and 113 minutes without one. A full run used about 1,200 runner-minutes of the fork's own quota.
What changes for contributors and reviewers
- Authors have to enable Actions in their fork. Of the 47 forks behind the open PRs on 2026-11-04, 39 had Actions runs since 2026-10-01.
- The check shows up on the PR at the next pass of the scheduled workflow, every 5 minutes, and links to the run in the fork. It fails as soon as a job in that run fails. GitHub can delay scheduled runs.
- Reviewers can't rerun a flaky job in someone else's fork. The author has to, or a committer can push the branch to their own fork and link that run on the PR.
- A free account runs at most 20 jobs at a time, which is also the ASF limit for one workflow. An author with several PRs running at once splits those 20 between them.
- Open PRs get no x86, ARM or Kafka run until their branch is rebased onto
mainafter #13535 and pushed again. - A fork run uses the workflow files of the pushed commit, so a CI change on
mainreaches a PR once its branch is merged or rebased onto thatmain. - The trust model doesn't change:
pull_requestalready runs the PR's own copy of the workflow files.
Follow-ups and open questions
- I don't have a good answer yet for authors who can't enable Actions in their fork. One option is a committer-only comment trigger like
/ansi-test. - Fork PRs don't get the "Report test results" check, because runs in a fork don't trigger
test_report.yml'sworkflow_run.update-fork-build-status.jscould publish the fork runs'*-reportartifacts as that check. - Delta Spark UT can move the same way.
- Switch the status script to
apache/infrastructure-actions/fork-cionce apache/infrastructure-actions#1512 and apache/infrastructure-actions#1513 are fixed. - A fork with a name other than
gluten, such asincubator-gluten, builds under a different workspace path. In a probe it got 0 of 272 ccache hits from the upstream stash. #13535 setsCCACHE_BASEDIRin the CI build scripts, and with it a second probe got 259 hits against the 269 entries it had cached. Existing stashes miss once after the merge, until the cache workflow saves new ones. pr_bot.ymlandclickhouse_be_trigger.ymlstill usepull_request_target. Moving them to theworkflow_runpattern on the ASF policy page is a separate change.- This doesn't replace the ANSI test tags discussed in #13263.
I'd like to hear what you think. cc @philo-he @zhztheplayer
Generated-by: Claude Opus 5.5
- 主要语言
- Scala
- 星标
- 1.6k
- 派生
- 685
- 平均合并
- 4 天 16 小时
- 30 天内合并 PR
- 84
环境准备
在浏览器里用你自己的 GitHub 账号启动这个项目的开发容器。
- 没有 Dockerfile 或 Docker Compose 文件
- 有 Pull Request 模板
- 阅读贡献指南
从这里开始
- 先读完整个 Issue,再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库,在一个分支上完成修改。
- 提交 Pull Request,并在描述里引用这个 Issue 编号。
apache/gluten 的其他 Issue
-
[CORE] Selecting a cost model the backend has no coster for fails with ClassNotFoundException可能已有人在做 @LuciferYang 于 1 天前认领。 未关闭
难度 2/5 1-3 小时 新手友好度 66/100
维护者通常 3 天内回复
-
bug unit testing velox backend
难度 2/5 1-3 小时 新手友好度 78/100
维护者通常 3 天内回复
-
难度 1/5 1 小时以内 新手友好度 82/100
维护者通常 3 天内回复
-
[VL] madvise(WILLNEED) call fails in MmapFileStream because of wrong calculation of fetching length可能重新可做 关联的 PR 已关闭且未合并。 未关闭bug triage
难度 2/5 1-3 小时 新手友好度 82/100
维护者通常 3 天内回复
-
难度 2/5 1-3 小时 新手友好度 78/100
维护者通常 3 天内回复
相似的 Issue
-
难度 2/5 1-3 小时 新手友好度 78/100
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 72/100
rchillyard/TableParser#206 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 72/100
维护者通常 2 天内回复
-
? - Needs Triage bug
难度 2/5 1-3 小时 新手友好度 66/100
NVIDIA/cudf-spark#17063 ·
维护者通常 2 天内回复
-
bug 🐞 User issue 🚨
难度 2/5 1-3 小时 新手友好度 76/100
valkey-io/valkey-glide#7813 · 2 条评论 ·
维护者通常 1 天内回复