Files
frid/crates/artist-dht/README.md
T
2026-07-16 16:40:48 +03:00

174 lines
6.6 KiB
Markdown

# artist-dht
A proof-of-concept **distributed artist directory** built on top of
[`federation-net`](../federation-net). Every running process is a full DHT
participant — client, router and storage node at once. There are no
bootstrap servers, index nodes or search servers: the "server" is the set of
running peers itself.
## What it demonstrates
```text
/add on Peer B → name normalization → publish index records into the DHT
→ replication to regular peers
/search on A/C → iterative Kademlia-style lookup (no broadcast)
→ record of Peer B is found, with its owner id
```
* Records are published under BLAKE3 keys: one **exact key** for the whole
normalized name and one **token key** per word, so `/search massive` finds
*Massive Attack*.
* Each peer has a stable 256-bit `NodeId` derived from its persistent
`federation-net` endpoint id; records live on the `K = 8` XOR-closest
nodes.
* A peer joins the network knowing **only the network id**: on startup it
finds other peers through a rendezvous record in the public BitTorrent
Mainline DHT (see `federation-net`'s rendezvous). Within the network peers
then discover each other through automatic Hello/PeerExchange gossip;
connections to newly learned peers are opened **on demand** from stored
tickets.
* Deletions propagate as **tombstones** (revision-based, they beat active
records of the same or lower revision). Replicas expire by TTL (30 min
active, 2 h tombstones); owners republish every 10 minutes.
* Local state (identity, own artists, replicas, known peers) lives in
`<data_dir>/state.sqlite3` and `<data_dir>/identity.key`.
Out of scope (by design): fuzzy search, content transfer, CRDTs, consensus,
signatures on DHT records, Sybil protection, accounts, GUI.
## Bootstrapping
There is deliberately **no bootstrap server**. A peer joins a network knowing
only its `--network-id`: peers of the same network find each other through a
shared rendezvous record in the public BitTorrent Mainline DHT (the record's
signing key is derived from the network id). The first peer of a new network
simply publishes itself and waits; every later peer with the same id finds it
within a rendezvous round or two (typically well under a minute). No peer has
a special role.
Because the rendezvous record is world-readable, the network id acts as a
public rendezvous token: anyone who knows it can discover and join the
network. Pick a unique, hard-to-guess name for a private network (e.g.
`myband-artists-prod-7f3a`).
Discovery can be turned off with `--no-bootstrap`; then a peer can only join
through the ticket of *any* already running peer passed via `--connect`
(tickets also work in addition to discovery, e.g. on isolated networks
without internet access).
## Running the demo (three peers)
Start the peers in any order — they only share the network id:
```bash
cargo run -p artist-dht-cli -- \
--data-dir ./peer-a \
--network-id demo-artists \
--name alice
```
```bash
cargo run -p artist-dht-cli -- \
--data-dir ./peer-b \
--network-id demo-artists \
--name bob
```
```bash
cargo run -p artist-dht-cli -- \
--data-dir ./peer-c \
--network-id demo-artists \
--name charlie
```
Within a minute `Peer connected: ...` lines appear on all three. All
processes must use the same `--network-id`; a peer from another network is
rejected during the transport handshake. (With `--no-bootstrap`, pass the
ticket printed by a running peer via `--connect 'fnet...'` instead;
`--connect` may be repeated.)
### Demo scenario
1. On **bob**: `/add Massive Attack` and `/add Portishead` — each prints the
artist id and how many DHT nodes stored the replicas.
2. On **alice**: `/search massive attack` — the DHT results list *Massive
Attack* with bob's endpoint id as the owner.
3. On **charlie**: `/search portishead` (and `/search massive`) — same
records found through an iterative lookup, not a broadcast.
4. Stop **bob** (Ctrl+C). Both alice and charlie still find the records from
replicas until the TTL expires.
5. Restart **bob** with the same `--data-dir`: the endpoint id is unchanged,
the local database is intact and records are republished automatically.
6. On **bob**: `/delete <ARTIST_ID>` (a unique hex prefix is enough) — a
tombstone propagates and the other peers stop returning the record.
### Commands
```text
/help show help
/id show endpoint and node ids
/ticket print the connection ticket
/peers list open connections
/routing list known DHT contacts (connected / last seen)
/list list local artists
/add <ARTIST_NAME> add and publish an artist
/delete <ARTIST_ID> delete a local artist (id or unique hex prefix)
/search-local <QUERY> search the local database only
/search <QUERY> search locally and across the DHT
/republish republish local records now
/quit shut down gracefully
```
## Protocol limits
```text
K = 8, ALPHA = 3, max lookup requests = 32
max contacts per PeerExchange = 32
max records per FindValue response = 100
max artist name = 512 bytes, max tokens = 32
max pending requests = 1024
request timeout = 5 s, lookup timeout = 15 s
on-demand dial timeout = 5 s
dial backoff = 30 s doubling up to 10 min, eviction after 5 failures
```
Value lookups return as soon as the first records arrive; in-flight requests
to slower or dead contacts are cancelled. Contacts that failed to dial are
skipped for an exponentially growing backoff window and evicted after five
consecutive failures (peer exchange re-adds them with a clean slate if they
come back).
## Library usage
The CLI is a thin wrapper around the `artist-dht` library:
```rust
use artist_dht::{ArtistDhtConfig, ArtistDhtService};
use federation_net::NetworkId;
let config = ArtistDhtConfig::builder()
.data_dir("./peer-a")
.network_id(NetworkId::from_name("demo-artists"))
// Optional: discover peers of this network via the mainline DHT.
.rendezvous(artist_dht::RendezvousConfig::default())
.build()?;
let (service, mut events) = ArtistDhtService::start(config).await?;
let (artist, stats) = service.add_artist("Massive Attack".into()).await?;
let outcome = service.search_network("massive").await?;
service.shutdown().await?;
```
## Verification
```bash
cargo fmt --check
cargo clippy --workspace --all-targets --all-features -- -D warnings
cargo test --workspace
```
The integration tests start three real DHT nodes in one process and walk the
whole demo scenario (publish, distributed search, replica survival after the
owner leaves, restart, tombstones), so they need network access and take a
couple of minutes.