Orion

orchestration in a single binary. zero bloat.

i wanted to understand how kubernetes actually works. so i built one from scratch.

Orion is a distributed container orchestrator. it turns a fleet of linux servers into a single, self healing supercomputer.

it's not a wrapper. it's an engine.

---

the problem

you have two choices today:

1. docker-compose: great for localhost, useless if a server catches fire.
2. kubernetes: massive operational overhead. requires etcd, apiservers, and 3 PhDs to maintain.

i needed something in the middle. something that runs as a single binary, uses negligible RAM, and doesn't panic when a node gets overloaded.

the solution: complexity collapse

i stripped the control plane down to the physics of distributed systems:

• no external db: the database is embedded (raft + boltdb).
• no external load balancer: service discovery via gossip.
• symmetric architecture: every node is the same. leader election handles the rest.

---

architecture

• raft: handles consensus. if it's not in the raft log, it didn't happen. guarantees strong consistency (CP).
• gossip: uses memberlist (SWIM protocol) + Lifeguard. it detects "flapping" nodes (high CPU) and prevents false positives.
• docker: direct integration with the docker engine api to spin up containers and dynamic port bindings.

graph TD
subgraph "the cluster"
style L fill:#f9f,stroke:#333,stroke-width:2px
style F1 fill:#e1f5fe,stroke:#333
style F2 fill:#e1f5fe,stroke:#333
L["node 1: leader"] <-->|raft logs TCP| F1["node 2: worker"]
L <-->|raft logs TCP| F2["node 3: worker"]
F1 -.->|gossip UDP| F2
F2 -.->|gossip UDP| L
L -.->|gossip UDP| F1
end
subgraph "external"
CLI[user cli] -->|HTTP POST /tasks| L
Web[web traffic] -->|TCP dynamic port| F2
end

Loading

---

how to run

prerequisites: go 1.24+ and docker.

1. build it

go build -o orion ./cmd/orion/

2. start the leader (bootstrapping)

spin up the first node. it will elect itself leader.

./orion --id node1 --port 8000 --gossip-port 6000 --raft-port 7000 --bootstrap

3. join a worker (auto-discovery)

spin up a second node. just tell it where the leader is.

# make sure to use your LAN IP if running across machines
./orion --id node2 --port 8001 --gossip-port 6001 --raft-port 7001 --join 127.0.0.1:6000

magic moment: watch the leader logs. it detects the join event, automatically adds node 2 to the raft configuration, and starts replicating logs.

4. deploy a payload

schedule an nginx container. the scheduler will bin-pack it to the least loaded node.

curl -X POST localhost:8000/tasks -d '{ "image": "nginx", "memory": 100000000, "portBindings": {"80/tcp": "8080"}}'

---

benchmarks / resilience

metric	result
startup latency	< 50ms (binary start)
scheduling latency	< 10ms (in-memory bin packing)
failover time	~2-4s (raft election + gossip convergence)
network traffic	constant O(1) per node (SWIM gossip)

the "kill" test: start a task on node 2. kill -9 node 2. watch node 1 detect the failure via gossip, mark it dead, and reschedule the task to a healthy node automatically.

---

roadmap

• TUI: the CLI works, but i want a k9s style dashboard.
• overlay networking: replacing port mapping with VXLAN for flat pod IPs.
• persistent storage: CSI driver integration.

---

built with <3 by @bit2swaz.