Refresh README and config example

Document the event-sourced store, gossip replication, push-on-commit,
compaction and the TLS peer link; drop the stale persistence status.
This commit is contained in:
Jean Chevronnet 2026-07-12 07:31:16 +00:00
parent 0af7f10c3d
commit d30f5c05a6
No known key found for this signature in database
2 changed files with 64 additions and 10 deletions

View file

@ -2,18 +2,56 @@
A federated IRC services daemon in Rust. Protocol-agnostic core, InspIRCd link first.
Clean-room inspired by the ideas behind modern event-log/gossip services, not
Clean-room, inspired by the ideas behind modern event-log/gossip services, not
derived from any project's source.
## Design
- **`proto/`** — the ircd link layer. A `Protocol` trait maps raw server-to-server
lines to/from a normalized `NetEvent`/`NetAction` model. The engine never touches
a raw line, so a new ircd = one new module. InspIRCd is the first.
- **`engine/`** — the services engine: live network `state`, an `EventLog` (every
persistent change is an `Event`; state is a fold over the log — this is what makes
replicating across nodes a later addition, not a rewrite), and the `Service` trait.
- **`services/`** — the pseudo-clients: NickServ first, then ChanServ / OperServ / …
- **`proto/`** the ircd link layer. A `Protocol` trait maps raw server-to-server
lines to and from a normalized `NetEvent`/`NetAction` model, so the engine never
touches a raw line and a new ircd is one new module. InspIRCd is the first.
- **`engine/`** the services engine: live network `state`, the account store, and
the `Service` trait. The store is event-sourced: every change is an `Event`
appended to a per-node log, and state is a fold over that log.
- **`gossip/`** node-to-node replication over the logs.
- **`services/`** the pseudo-clients: NickServ first, then ChanServ / OperServ.
## Replication
Each log entry carries an `origin`, a per-origin `seq`, and a Lamport clock. Peers
connect, exchange version vectors, and send each other the entries the other
lacks; a freshly committed entry is also pushed immediately, so a registration
propagates in milliseconds. Ingest is idempotent, so re-delivery and reconnect
after a split both converge. The log is compacted to one entry per account as it
grows. An account registered on any node works on all of them.
## Config
```toml
[uplink]
host = "127.0.0.1"
port = 7000
password = "linkpw"
[server]
name = "services.example"
sid = "42S"
[gossip] # omit to run a single, non-replicating node
bind = "0.0.0.0:16700"
secret = "shared-secret"
[gossip.tls] # omit for a plaintext link
cert = "certs/nodeA.crt"
key = "certs/nodeA.key"
ca = "certs/ca.crt" # a peer must present a certificate signed by this CA
[[peer]]
addr = "nodeB.example:16700"
name = "nodeB" # TLS name to expect; must match the peer certificate
```
Generate certificates with `scripts/gen-certs.sh`.
## Run
@ -24,5 +62,7 @@ cargo run -- config.toml
## Status
Early scaffold: links to an InspIRCd uplink and introduces NickServ. Burst/mode
handling and account persistence are next.
Links to an InspIRCd uplink and runs NickServ (REGISTER, IDENTIFY, LOGOUT, CERT,
and SASL PLAIN / SCRAM-SHA-256/512 / EXTERNAL) over an event-sourced account store
replicated between nodes by gossip, with an optional mutually authenticated TLS
peer link. Next: ChanServ.