Skip to content

Releases: acmestack/gpi

Release list

v0.0.1

v0.0.1 Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 16 Aug 13:47
v0.0.1
d6a483f
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

gpi v0.0.1

gpi v0.0.1

gpi 是一个轻量、自托管、多云算力调度器（multi-cloud compute scheduler），用 Go 编写、对标 SkyPilot。用一个任务 YAML 声明资源与运行命令，自动完成比价选型、跨云置备、环境准备与任务执行。

gpi is a lightweight, self-hosted, multi-cloud compute scheduler written in Go, positioned against SkyPilot. Declare resources and run commands in a single task YAML; gpi handles price-based selection, cross-cloud provisioning, environment setup, and task execution automatically.

核心能力

Core features

  • 任务即代码：一份 YAML 声明 resources / setup / run，可在任何支持的云上运行。

  • Task-as-code: one YAML declares resources / setup / run, runnable on any supported cloud.

  • 价格驱动调度：跨全部已注册云做资源匹配与实时比价，输出 failover placement plan。

  • Price-driven scheduling: resource matching and real-time price comparison across all registered clouds, producing a failover placement plan.

  • 可插拔 Optimizer：内置 cost（默认）与 time 两种指标，支持 --optimizer cost,time 等字典序多指标策略；通过 Metric 接口可扩展自定义指标（时延/碳排/预算等）。

  • Pluggable Optimizer: built-in cost (default) and time metrics, plus lexicographic multi-metric strategies like --optimizer cost,time; extend via the Metric interface with custom metrics (latency / carbon / budget, etc.).

  • 多云置备：已支持 aliyun、aws、gcp、azure、kubernetes（云厂商零官方 SDK，标准库实现签名；Kubernetes 原生 API），完整实例生命周期（VPC/SG/KeyPair/镜像/弹性 IP）。

  • Multi-cloud provisioning: aliyun, aws, gcp, azure, kubernetes supported (zero official SDKs for cloud providers, signature via standard library; Kubernetes native API), full instance lifecycle (VPC/SG/KeyPair/image/elastic IP).

  • 节点 Agent（gpilet）：跑在每个节点上，采集 CPU/内存/磁盘/GPU/Ray 状态（对标 skylet）。

  • Node Agent (gpilet): runs on every node, collecting CPU/memory/disk/GPU/Ray status (SkyPilot's skylet equivalent).

  • 多节点 Ray 集群：num_nodes>1 自动组成 head/worker Ray 集群。

  • Multi-node Ray clusters: num_nodes>1 automatically forms head/worker Ray clusters.

  • 服务化部署：多副本服务、健康检查（对标 SkyServe）。

  • Serving: multi-replica services with health checks (SkyServe equivalent).

  • 定时任务：cron 调度 + 失败自动重试（对标 Sky Jobs）。

  • Scheduled jobs: cron scheduling + automatic retry on failure (Sky Jobs equivalent).

  • 多执行方式：云 VM / 挂已有主机 SSH / 本地 Docker / 本机直跑。

  • Multiple execution modes: cloud VM / SSH to existing hosts / local Docker / run directly on the local machine.

  • 可插拔持久化：file（默认）/ sqlite / mysql / redis。

  • Pluggable persistence: file (default) / sqlite / mysql / redis.

  • REST API + 调度器：gpi server，含响应格式定制、Request ID、key 风格、Bearer 认证、中间件、OpenAPI/Swagger（GitLab 内建渲染）。

  • REST API + scheduler: gpi server, with response format customization, Request ID, key style, Bearer auth, middleware, OpenAPI/Swagger (rendered natively by GitLab).

元数据设计

Metadata design

  • 全动态、无静态数据：规格与价格都从云 API 实时拉取，按每云 TTL 缓存（规格 24h、价格 10min，stale-while-error）。

  • Fully dynamic, no static data: specs and prices are pulled live from cloud APIs, cached per-cloud with TTL (specs 24h, prices 10min, stale-while-error).

  • 契约层 internal/cloud/catalog（对应 SkyPilot sky/catalogs），TTL 缓存运行时 internal/metacache。

  • Contract layer internal/cloud/catalog (SkyPilot's sky/catalogs equivalent), TTL cache runtime internal/metacache.

  • 新增云只需一个包 + 一个 struct（Provider 同时实现生命周期与元数据）。

  • Adding a cloud requires only one package + one struct (Provider implements both lifecycle and metadata).

本次发布内容

In this release

  • 首个可运行版本，覆盖上述全部核心能力。

  • First runnable release covering all core features above.

  • 新增 Kubernetes cloud backend：kubeconfig context = Region，Pod = Instance，GPU 自动检测 + nodeAffinity。

  • Added Kubernetes cloud backend: kubeconfig context = Region, Pod = Instance, automatic GPU detection + nodeAffinity.

  • 新增 GCP Compute Engine provider（OAuth2 + REST API）和 Azure Resource Manager provider（Client Credentials + REST API）。

  • Added GCP Compute Engine provider (OAuth2 + REST API) and Azure Resource Manager provider (Client Credentials + REST API).

  • CLI：gpi launch / optimize / status / cluster / serve / jobs / server 等。

  • CLI: gpi launch / optimize / status / cluster / serve / jobs / server, etc.

  • 构建：make build；Docker 镜像多阶段构建；K8s 部署清单 deploy/k8s/。

  • Build: make build; multi-stage Docker image build; K8s deployment manifests in deploy/k8s/.

  • 文档：架构设计、接入新云指南、Optimizer 扩展指南、增强对比清单。

  • Docs: architecture design, adding-a-new-cloud guide, Optimizer extension guide, enhancement comparison list.

快速开始

Quick start

make build
./gpi optimize my_task.yaml     # 只看 placement plan
./gpi launch my_task.yaml -y    # 真正部署
make build
./gpi optimize my_task.yaml     # see the placement plan only
./gpi launch my_task.yaml -y    # actually deploy

获取

Getting it

已知限制

Known limitations

  • 平台仅支持 Linux / macOS（无 Windows）。

  • Linux / macOS only (no Windows).

  • 阿里云开实例受账号余额限制（价格/元数据 API 不受影响）。

  • Creating Alibaba Cloud instances is subject to account balance (price/metadata APIs are unaffected).

  • 安全组默认放行 0.0.0.0/0:22 与 1024-65535，生产需收敛。

  • Security groups default to opening 0.0.0.0/0:22 and 1024-65535; tighten for production.

Assets 6
Loading