@actions/cache: optional S3 storage backend for self-hosted runners

未关闭
#2,572 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 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
28/100
Issue 类型
功能
描述清晰度
基本清楚
活跃度
活跃
技术栈
aws, typescript

调研方向

Start by reading the @actions/cache implementation and the related discussion in actions/cache#679; the issue names no specific source files or tests. The proposed first step is to move the existing v1 and v2 implementations behind CacheBackend without changing behavior. Done means maintainers agree with the direction; implementing S3 and its tests are later, separate steps.

由索引模型根据 Issue 内容生成。

描述

Describe the enhancement

Add an optional S3 storage backend to @actions/cache. The runner administrator turns it on with environment variables on the runner. Without them, nothing changes: the library keeps using the v1 or v2 cache service exactly as today.

On self-hosted runners in AWS, every cache restore and save goes to Azure Blob over the internet. That traffic leaves through a NAT gateway and is billed per GB. A large monorepo restoring 0.3–1.7 GB per job across hundreds of jobs a day turns into terabytes of NAT traffic, for data that could sit in a bucket in the same region and be read over a free S3 gateway endpoint.

Teams solve this today by forking actions/cache (runs-on/cache, tespkg/actions-cache, itchyny/s3-cache-action) or by running a cache server and binary-patching the runner to redirect ACTIONS_RESULTS_URL. The forks only cover workflows that switch uses: to the fork, so setup-* actions and anything else built on @actions/cache stays on the hosted service. None of the forks keep the hosted service's branch scoping, so any job holding the bucket credentials can overwrite the default branch's cache.

Related: #1051, actions/cache#679, #1695.

How the design answers the concerns raised in actions/cache#679

  1. "It should be a repo/org level setting and not a workflow setting." The backend is configured only through environment variables on the runner (ACTIONS_CACHE_S3_BUCKET, ACTIONS_CACHE_S3_REGION, optional ACTIONS_CACHE_S3_PREFIX and ACTIONS_CACHE_S3_ENDPOINT). There are no action inputs, so a workflow author can't redirect caches. It only applies to runners the administrator controls.
  2. "How is the custom cache server endpoint secured?" No custom server is involved, and the runtime token is never sent anywhere new. The backend talks to S3 over HTTPS through the standard AWS SDK credential chain (instance role, IRSA, or credentials from aws-actions/configure-aws-credentials with GitHub OIDC). Access is governed by the bucket's IAM policy.
  3. Cache management experience. Entries in S3 don't appear in the repository's cache UI or API. The bucket owner manages them, with a lifecycle rule for expiry. We'd document that limitation clearly.

Behaviour

  • Object keys are <prefix>/<owner>/<repo>/<ref>/<version>/<key>, where version is the existing getCacheVersion() value, so compression method and cross-OS rules still apply.
  • Saves write only under the job's own ref. Restores look in the job's ref, then GITHUB_BASE_REF, then the default branch, matching the hosted service's scoping. Since the ref is part of the object path, IAM can enforce it. For example, only a role trusted for ref:refs/heads/main may PutObject under …/refs/heads/main/*.
  • Exact keys resolve with HeadObject. Restore keys resolve by prefix with ListObjectsV2, and the newest match wins.
  • ACTIONS_CACHE_MODE, lookupOnly, CacheReadDeniedError and CacheWriteDeniedError keep their current semantics. An S3 AccessDenied maps to the matching denied error.
  • Uploads are multipart and parallel. Downloads use parallel ranged GETs.
  • The AWS SDK is loaded lazily, only when the backend is configured.

Code Snippet

Runner environment, set by the administrator:

ACTIONS_CACHE_S3_BUCKET=my-runner-cache
ACTIONS_CACHE_S3_REGION=us-east-1

Library seam (first PR, no behaviour change):

export interface CacheBackend {
  isAvailable(): boolean
  restoreCache(paths: string[], primaryKey: string, restoreKeys?: string[], options?: DownloadOptions, enableCrossOsArchive?: boolean): Promise<string | undefined>
  saveCache(paths: string[], key: string, options?: UploadOptions, enableCrossOsArchive?: boolean): Promise<number>
}

Additional information

Proposed sequence, each a separate PR:

  1. Refactor only: move the existing v1 and v2 implementations behind CacheBackend, with no behaviour or public API change. A branch is ready to open if you're open to the direction.
  2. Add the S3 backend with unit tests (mocked S3 client), including tests that the scoping matches the hosted service.
  3. In actions/cache: bump @actions/cache, document the backend, and add an end-to-end workflow against a MinIO service container.

Per CONTRIBUTING, we'd like agreement on the direction before opening 2 and 3, and we're happy to write an ADR. Questions for maintainers:

  • Is an administrator-configured S3 backend something you'd accept in @actions/cache, or would you rather keep the library hosted-service-only?
  • If yes, is the extra dependency (@aws-sdk/client-s3, @aws-sdk/lib-storage), loaded lazily, acceptable, or would you prefer the backend as a separate package that registers itself?
主要语言
TypeScript
星标
5.9k
派生
1.8k
平均合并
15 分钟
30 天内合并 PR
1

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

actions/toolkit 的其他 Issue

查看 actions/toolkit 的全部 Issue

相似的 Issue

更多 TypeScript Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。