K7d – Fork live Kubernetes clusters in GRPO-train AI on infra
Show HN: K7d – Fork live Kubernetes clusters in GRPO-train AI on infra
我们发现了什么
Show HN: K7d – Fork live Kubernetes clusters in GRPO-train AI on infra. I have two goals here: 1) enable large scale GRPO/RL training of AI on Kubernetes infra, which IMO is a great playgrou
- 来源：Hacker News（发现于 2026-08-19）
- 证据等级：D · 发现产品或需求信号，暂未获得可核验的商业证据。
- 商业模式：API / Usage-based
- 主题：独立产品
- 初筛评分：18.4/100 · 收录 1 次
证据，比故事更重要。
规则清洗与初筛，未经人工商业核验。原文语境、实际客户和付费情况仍需自行验证。
引用与数字披露
来源类型（原作者自述／第三方测算／媒体转引）需采集端标注，本版尚未落字段。
- 作者
- 未标注
- 抓取日期
- 来源类型
- 未标注
- 币种
- 未标注
- 口径
- 未标注
- 披露主体
- 未标注
- 披露日期
- 未标注
中文辅助译文（全文）
大家好，HN，我是 Gary。今天我想向大家介绍 k7d，这是一款采用 Apache 2.0 协议的紧凑 Rust VMM + shim，实现了此前不可能做到的事情：对正在运行的虚拟化多节点 k8s 集群进行快速分叉（fork），同时保持进行中的连接不中断。一个 3 节点 VM 的 K8s 集群可在 105ms 内完成分叉，在一台 64GB 内存的机器上，一个 3 节点 VM 集群的 50 倍分叉耗时 4.1 秒。我有两个目标：1) 在 Kubernetes 基础设施上支持大规模 GRPO/RL 训练，在我看来这是推理训练的优秀试验场，除此之外还训练一种真正有用的能力。这不仅要求快速重置 episode（因为在 RL 后训练中你需要数万次多轮运行），还极大受益于快速分叉，从而可以在 RL 训练中进行并行分支探索、回滚和剪枝。忠实的分叉还能为 GRPO 中的 G 提供字节级一致的起点，从而降低组内的方差。2) 在 3 秒以内实现带 Docker-in-VM 的沙箱的 VM 快照暂停/恢复/分叉，用于我的另一个项目 K7，该项目以用户友好的 CLI / API / Python SDK 以及 Kubernetes 原生方式大规模提供自托管的 VM 沙箱基础设施。除此之外，k7d 还具备以下特性：- 用于资源管理的树形 API：正如你所猜测的，分叉时即使采用了优化的写时复制（CoW）页面共享，你仍然需要妥善管理资源（内存 + 磁盘），因此你需要知道如何在运行时进行淘汰。
所以我设计了树形逻辑来跟踪子节点如何与父节点共享页面，并让你的 AI 智能体保护有前景的树分支、淘汰无前景的分支，或在资源紧张时由类 LRU 逻辑自动淘汰。这种基于树的逻辑既适用于单 VM 沙箱，也适用于它们自建 Linux 网桥上的多 VM 集群。- 形式化验证：当然不是全部，但 k7d 中选定的关键子部分经过了形式化验证：我在 unsafe 路径的内存运算上使用 Kani，并使用 Aeneas（后端为 Lean）形式化证明上述基于树的逻辑，以确保淘汰永远不会释放被存活后代引用的页面。- 将延迟作为 CI 指标：我严格跟踪最重要操作的延迟，并通过一套集成测试进行检查和强制约束。我曾非常努力地尝试不自己构建 VMM，最终先为 K7 构建了另一个后端，取代了我最初的 "kfd"（Kata + Firecracker + 基于 LVM 精简池的 Devmapper-snapshotter），新后端我称之为 "kql"，即 Kata + Qemu + Longhorn。如果你了解这套技术栈，你会立刻猜到：Longhorn 非常适合跨节点复制，所以我利用它让快照在节点间复制，从而 "snapshot resume" 始终可用/具备高可用。这里的 Qemu 是因为 Longhorn 的块存储要求与 Firecracker 不兼容——Firecracker 需要 Devmapper-snapshotter，而我不想自己为该后端构建跨节点复制逻辑。但是 "kql" 后端由于 Longhorn 的实现方式，分叉需要 45 秒，对那些要求 K7 实现快速分叉的用户来说太慢了。于是……
译文由上游机器翻译生成，可能有误；判断请以英文原文为准。
英文原文（来源本站未改写）
Hey HN, Gary here.Today I want to present k7d which is an Apache 2.0, tight Rust VMM + shim enabling something not possible before: fast forking of live running virtualized multi-node k8s clusters with surviving of in-flight connections.A 3-VM nodes K8s cluster gets forked in 105ms, and a 50x fork of a 3-VM cluster in 4.1s on a 64GB RAM box.I have two goals here: 1) enable large scale GRPO/RL training of AI on Kubernetes infra, which IMO is a great playground for reasoning training, besides training a capability that's actually useful.
And this requires not only fast episode reset (as you need tens of thousand of multi-turn runs during RL post-training) but also greatly benefits from fast forking so you can do parallel branch exploration, rollback, pruning during RL training.Faithful forks also give you byte-identical starts for the G of GRPO, which gives variance reduction across the group. 2) enable <3s VM snapshot pause/resume/fork of sandboxes with Docker-in-VM, for my other project K7 which provides self-hosted infra for VM sandboxes at scale, with a user-friendly CLI / API / Python SDK, and Kubernetes native.
Besides that, k7d is equipped with: - A Tree-shaped API for resource management: as you can guess when you fork, even with optimized CoW-page-sharing, you want to properly manage your resources (memory + disk) and hence you need to know how to evict while things run.So I have tree-shaped logic to keep track of how children share pages with parents, and let your AI agent protect a promising tree branch, evict an unpromising one, or let LRU-ish logic auto-evict when resources get tight.
This tree-based logic applies both to single VM sandboxes, and to multi-VM clusters on their own Linux bridge. - Formal verification: of course not all of it, but selected critical subparts of k7d are formally verified: I use Kani for memory arithmetics in the unsafe paths, and Aeneas (with Lean backend) to formally prove the tree-based logic explained above so that eviction never frees a page referenced by a live descendent. - Latencies as CI: I rigorously keep track of latency for most important operations which remain checked/enforced via a suite of integration tests.
I really tried hard not building my own VMM and first ended up building another backend for K7 than my initial "kfd" (Kata + Firecracker + Devmapper-snapshotter over LVM thin-pool), which I called "kql" for Kata + Qemu + Longhorn.If you know this stack you'll guess it right away: Longhorn is great for cross-node replication so I used it to have my snapshots replicated across nodes, so "snapshot resume" always works / HA.Qemu here is because Longhorn's block storage requirements was incompatible with Firecracker who wants Devmapper-snapshotter, a backend for which I would not want to build myself the cross-node replication logic.
But this "kql" backend yielded forks in 45s due to how Longhorn is built, which was too slow for the users who asked me to enable fast forking for K7.So t
这条还缺什么证据？
下面每条都由本条已有字段推出（等级、理由、商业模式、来源次数、是否演示）， 本站不生成推测性结论；通用验证方法放在方法论页。
- 可核验的收入或付费证据查官网定价页与付费口径；第三方数据源（如 GetLatka）只作旁证，需标注来源与时点。
- 只有单一来源找一手站点或其他渠道是否重复出现同一产品；社区热帖数量不等于商业进展。
通用验证清单（谁有这个问题／谁愿意付费／一个人能交付哪一小步）见我们的筛选方法。