活跃时全速干你的活；闲下来就浅睡——内存换出、几乎不占资源，一碰就醒、还是离开时的样子；很久不用就深睡，零占用。它还有一只放猫粮的餐盘（你的磁盘）：浅睡、深睡、换个窝，盘里的数据一口都不少。It runs your work at full speed when awake; dozes when idle — RAM paged out, near-zero footprint, waking in a blink exactly as you left it; sleeps deep when unused, costing nothing. And it keeps a food bowl (your disk): nap it, deep-sleep it, move it to another den — not a bite of data goes missing.
serverless 式的 microVM 来去匆匆——空闲就冻结、负载来了再唤醒、随时可能换到另一台机器上。可你的代码仓库、working tree、git 历史不能跟着一起消失。把存储塞进 guest 内部，VM 一冻结上传就断、一迁移缓存就丢；用 gcsfuse 之类的方案跑 git,又会碰到非原子 rename、弱一致、跑着跑着仓库就坏。dozycat 把数据面从 VM 里搬出来：每个 cat 一个独立 volume，数据持久在对象存储里，VM 怎么折腾，磁盘都还在。Serverless-style microVMs come and go: they freeze when idle, thaw on demand, and can land on a different machine at any moment. But your repo, your working tree, your git history can't vanish with them. Put storage inside the guest and uploads stall the instant you freeze it and the cache is gone the moment it migrates. Try running git on something like gcsfuse and you hit non-atomic renames, weak consistency, and repos that quietly corrupt. dozycat lifts the data plane out of the VM: one volume per cat, data durable in object storage, the disk intact no matter what the VM does.
浅睡 = 暂停 + 把内存换出宿主：几乎不占内存，约 0.1 秒带记忆唤醒，文件随时能读。深睡 = 整机歇火：0 CPU、0 内存，来活儿了重启回来。空闲不烧钱，醒来还是原来那只猫。Light sleep = pause + page its RAM out of the host: almost no memory used, ~0.1s memory-intact wake, files readable the instant it stirs. Deep sleep = powered all the way down: 0 CPU, 0 RAM, reboots back when work arrives. Idle costs nothing; it wakes as the very same cat.
一句 dozycat login，就从云端控制面指挥一整队猫：创建、唤醒、打盹、退役，机队规模下统一调度。基础设施我们运维，休眠的猫不收 compute 费——你只为真正在线的算力付钱。One dozycat login and you command a whole squad of cats from the cloud console: create, wake, nap, retire — all scheduled at fleet scale. We run the infrastructure; sleeping cats cost no compute, so you pay only for what's actually online.
每只猫有自己的餐盘——一个独立 volume，数据落在对象存储里。猫打盹、换窝、被整只换掉，餐盘都在原处，数据一口不少：已写入的，绝不丢。想加餐，还能挂上你自己的 S3 / GCS 桶。Every cat has its own bowl — a dedicated volume with data in object storage. Nap it, move it to another den, swap the whole cat out: the bowl stays put and not a bite is lost — every committed write survives. Bring your own S3 / GCS bucket for seconds.
通过 dozycat 控制面（API、UI 或 CLI，三者同源）建一个 volume,它就是这个租户的一套 dozycat 文件系统:独立元数据命名空间、独立对象存储前缀、独立凭证。建名时同时派生出永久路由 xxxxx.dozyfs。Through the dozycat control plane (API, UI, or CLI, all from one source) create a volume. It's this tenant's own dozycat filesystem: a dedicated metadata namespace, object-store prefix, and credentials. Its permanent route, xxxxx.dozyfs, is derived at the same time.
控制面把执行计划派发给目标节点的 cat-agent,后者驱动 Cloud Hypervisor 启动一个 microVM(cat),并在 host 侧拉起 JuiceFS 客户端和 virtiofsd。The control plane dispatches the plan to the target node's cat-agent, which drives Cloud Hypervisor to boot a microVM (a cat) and brings up the JuiceFS client and virtiofsd on the host side.
cat 通过 virtio-fs 挂上这个 dozycat volume。文件系统流量走 host 本地 vhost-user socket,根本不上网络;凭证留在 host,从不进 guest。The cat mounts the dozycat volume over virtio-fs. Filesystem traffic flows over a host-local vhost-user socket and never touches the network; credentials stay on the host and never enter the guest.
工作负载从永久名 xxxxx.dozyfs 进入。即便 cat 当时是深冷的,网关也会先唤醒它、等到 ready、再代理过去——「来一个包就激活」。换了节点?名字纹丝不动。Workloads come in via the permanent name xxxxx.dozycat. Even if the cat is deep-cold, the gateway wakes it, waits for ready, then proxies the request through. Activate-on-traffic. Land on a new node? The name doesn't budge.
自己托管、完全免费,或交给我们全托管。随时可升级。Self-host for free, or let us run it for you. Upgrade whenever.
是。开源版包含完整的控制面(dozycat 的 API、UI、CLI)、每节点的 cat-agent,以及 dozycat 文件系统层,采用 Apache-2.0 风格许可,可永久免费自托管。想省去运维就选全托管的 Cloud 版本。Yes. The open source edition includes the full control plane (dozycat's API, UI, and CLI), the per-node cat-agent, and the dozycat filesystem layer, under an Apache-2.0-style license, free to self-host forever. Prefer not to run it yourself? Use the managed Cloud.
gcsfuse 把一个文件映射成一个 object,POSIX 语义弱、rename 非原子、没有真正的多写一致性——跑 git 容易坏。dozycat 基于 JuiceFS,提供完整 POSIX、原子 rename、跨挂载锁和 close-to-open 一致性,而且数据面跑在 host 侧、跟 VM 生命周期解耦——这是网络文件系统给不了的:VM 冻结/迁移时上传照样进行。gcsfuse maps one file to one object, with weak POSIX semantics, non-atomic renames, and no real multi-writer consistency, so git tends to corrupt. dozycat is built on JuiceFS with full POSIX, atomic renames, cross-mount locks, and close-to-open consistency, plus a host-side data plane decoupled from the VM lifecycle, something a network filesystem can't give you: uploads continue while the VM is frozen or migrating.
cat 就是一个 Cloud Hypervisor microVM——dozycat 里运行工作负载的单位。这名字配得上我们的像素猫吉祥物,也呼应了它的脾气:大部分时间在打盹(冷),需要时一下就醒(热)。一个 cat 跑在一个节点上,挂一个 volume。A cat is a Cloud Hypervisor microVM, the unit that runs your workload in dozycat. The name suits our pixel-cat mascot and its temperament: mostly napping (cold), instantly awake when needed (hot). One cat runs on one node and mounts one volume.
数据落在 S3 或 GCS 这类对象存储上(权威、持久、容量不封顶),热数据缓存在节点本地盘。元数据走一个 HA 引擎——默认推荐托管 PostgreSQL;在超高元数据 OPS 或十万级 volume 时可换 Redis 或 TiKV。Data lives in object storage like S3 or GCS (authoritative, durable, effectively uncapped), with hot data cached on local node disk. Metadata runs on an HA engine; managed PostgreSQL is the recommended default, with Redis or TiKV for very high metadata OPS or 100k-volume scale.
两个持久性锚点都必须 hold:数据 block 必须成功上传到对象存储才算持久,writeback 暂存盘是持久块设备、崩溃后可续传;元数据落在 HA、同步复制的引擎上(真机验证过:杀主库切换零丢写)。每次停机/降级前强制 drain——已 fsync 的写,一个都不丢。Two durability anchors must both hold: data blocks count as durable only after a successful upload to object storage, and the writeback staging disk is a persistent block device that resumes after a crash; metadata lands on an HA, synchronously replicated engine (validated on real hardware: killing the primary failed over with zero write loss). A forced drain runs before every stop or downgrade, so every fsynced write survives.
三部分:活跃 cat 小时 + 存储 GB-月 + egress。因为 cat 会轮换到冷态,休眠的 cat 不产生 compute 费用——你只为真正在线的算力付费。先用免费额度跑通,再按需扩。Three parts: active cat-hours + storage GB-month + egress. Because cats rotate into cold states, sleeping cats incur no compute charge, so you pay only for compute that's actually online. Start on the free tier and scale as you go.
几分钟建好第一个 volume,启动一只 cat,从 xxxxx.dozyfs 访问它。无论它休眠、迁移还是被换掉,数据都还在原处。Spin up your first volume in minutes, launch a cat, and reach it at xxxxx.dozycat. Whether it sleeps, migrates, or gets replaced, the data stays right where you left it.