A distributed operating system built on the Felix log.
Status: design stage. The first image builds and boots in qemu in CI (x86_64 and aarch64); nothing else runs yet. The choices for the first three milestones are made and recorded in docs/decisions. The plan, the questions still open and what each milestone needs from felixd are in docs/design.md, and the work is tracked in milestones.
NineLives treats a cluster as one machine. Every change to the system is a durable, ordered record in a Felix log: what should run, where it runs, what each program received and what it produced. A program's state can be rebuilt from that log on any machine, so when a machine dies its programs carry on somewhere else. That is where the name comes from.
The pieces:
- A bootable image for each machine: a Linux kernel built from upstream source with a hardened, network-tuned config, our own init as PID 1, a minimal read-only Debian root, and atomic upgrades that can roll back (A/B partitions are the current proposal).
- felixd, the node agent that runs workloads from the log. It has its own repo; the image ships its releases, pinned by hash and checked against felixd's signature.
- Felix brokers on some of the machines.
- System programs on the log: the scheduler, the name service and the
fxshell.
The kernel would handle drivers, memory and local processes. NineLives is for what a single machine's kernel can't do across machines: processes that outlive a machine, pipes and files that are durable streams, and a history you can replay.
Existing programs are meant to run unmodified through felixd's native tier, which is not built yet; until then workloads are WebAssembly. Failures are reported, never hidden behind a hang. The reasoning, from what held Plan 9 back, is in Lessons from Plan 9, with a second look at one machine as a failure domain in Lessons from Barrelfish.
| Directory | What it holds |
|---|---|
boot/ |
Bootloader configuration, UEFI and secure-boot signing |
kernel/ |
The pinned Linux version and the config fragments we change |
init/ |
Our init (PID 1), in Rust: mounts, reaping, starting the role's services, shutdown |
rootfs/ |
The image definition: read-only root, data partition, A/B partitions (proposed) |
system/ |
Programs that run on the log: scheduler, name service, fx shell |
tools/ |
The fx CLI, image build wrappers, flashing, the qemu runner |
images/ |
Build targets (qemu x86_64 and aarch64, later clouds and Raspberry Pi) and components.toml, the pinned Felix and felixd releases |
docs/ |
The design, decision records, security and networking requirements, component designs and the felixd contract |
Images are built with mkosi, in CI.
ninelives.getfelix.dev is built from site/. See its README.
The first milestone's image builds and boots in qemu, x86_64 and aarch64, in CI
(tools/image.sh and tools/boot.sh run the same on a Linux host with docker
or podman, aarch64 on an arm64 host). The third milestone boots three of them into a cluster that
survives losing one. Images build on Linux only.
How work moves from decision to design to code, and the standards code is held to, are in CONTRIBUTING.md.
A cat whose head and tail make a 9. The tail is the tail of the log: nine records, one of them lost after a machine died, and the newest glowing amber like the eyes. The files and how to regenerate them are in docs/brand.
Report vulnerabilities privately, as described in SECURITY.md.
MIT for everything in this repository. Built images also contain Linux and other GPL software, and Felix (AGPL-3.0), which keep their own licenses; see docs/design.md.