Gargantis is a pure peer-to-peer network: there is no central server and no privileged tier. Participants are homogeneous nodes running essentially the same software, and every node is ephemeral — any node may disappear at any time, and the protocol treats this as the normal case, never the exception. A node that stays online continuously can improve discovery latency and data availability at the margin, but it is an ordinary peer the protocol grants no special authority and, crucially, must never depend on: correctness and liveness must hold even if no node is permanent.
Why pure P2P#
- No single point of failure or control. The Gargantisphere survives any individual node leaving.
- Content-addressed data. Scene and stream data is identified by hash and fetched from whoever has it, via Helia / IPFS, not from an owning server.
- Locality of interest. A node subscribes only to the realm partitions it can see, using a Bounding Volume Hierarchy, so traffic scales with attention rather than world size.
Mechanisms#
- Transport & discovery: libp2p (via Helia) — peer discovery, NAT traversal, and pubsub topics per realm/stream.
- Replication: the append-only log model gives a deterministic ordering so peers that consume the same inputs converge to the same state (see Web Stream Message Protocol). This is convergence of state, not a guarantee that any byte is persisted — see Data availability below.
- Data availability under churn: since every node is ephemeral, content is held redundantly across the swarm using erasure coding (an object is split into shards; any sufficient subset reconstructs it) and continuously re-replicated from survivors as nodes leave. Durability is therefore probabilistic, never absolute: if churn outruns repair, an object can fall below its reconstruction threshold and be lost. Some data loss is an accepted property of an all-ephemeral network, designed for and bounded — not a malfunction.
- Always-on peers: optional, non-load-bearing — same code, run as a service. They raise redundancy headroom and lower repair bandwidth, but the protocol must hold with none; nothing may depend on a node persisting.