diff options
Diffstat (limited to 'src/swarm/topology.rs')
| -rw-r--r-- | src/swarm/topology.rs | 521 |
1 files changed, 521 insertions, 0 deletions
diff --git a/src/swarm/topology.rs b/src/swarm/topology.rs new file mode 100644 index 0000000..57856b6 --- /dev/null +++ b/src/swarm/topology.rs @@ -0,0 +1,521 @@ +//! Which members a node syncs with: its neighbours. +//! +//! In a full mesh every change reaches every node from every other node (Automerge +//! decides per pair, so nobody knows that a third peer already passed it on), so the +//! copies of one write grow with N². Above `k` members, each node keeps only a few +//! neighbours, and changes travel over several hops. +//! +//! Every node computes the graph from the member list and which members it can reach, +//! with no coordination: +//! * a ring ordered by key hash: each member's nearest reachable member in both +//! directions, which keeps the reachable members connected; +//! * plus, for each member, the `k - 2` reachable others with the lowest pair score (a +//! hash of both keys, so both ends agree on it), which keeps paths short. +//! +//! A link exists if either end chose it, so links are symmetric and a node has at least +//! `k` neighbours. With at most `k + 1` reachable members the graph is the full mesh: +//! small swarms behave exactly as before. A member joining, leaving or going offline +//! changes only links near it. + +use std::collections::{BTreeMap, BTreeSet}; + +use sha2::{Digest, Sha256}; +use tokio::time::Instant; +use tracing::info; + +use crate::state::AppState; +use crate::swarm::members; + +/// This node's current neighbours, and the member list they were computed from. +#[derive(Debug, Default)] +pub(crate) struct NeighbourView { + members: Vec<String>, + k: usize, + unreachable: BTreeSet<String>, + stable: BTreeSet<String>, + neighbours: BTreeSet<String>, + /// Every reachable member is a neighbour. + full_mesh: bool, +} + +/// Whether every reachable member is our neighbour (also while the view is unknown). +pub(crate) fn is_full_mesh(state: &AppState) -> bool { + state + .neighbours + .read() + .unwrap_or_else(|poisoned| poisoned.into_inner()) + .as_ref() + .is_none_or(|view| view.full_mesh) +} + +/// The current view in numbers, once known. +#[derive(Debug, Clone, Copy)] +pub(crate) struct Summary { + /// Registered members, us included. + pub(crate) members: usize, + /// Members other than us not currently judged unreachable. + pub(crate) reachable: usize, + pub(crate) stable: usize, + pub(crate) neighbours: usize, + pub(crate) stable_neighbours: usize, + pub(crate) full_mesh: bool, +} + +pub(crate) fn summary(state: &AppState) -> Option<Summary> { + let local = state.local_id(); + state + .neighbours + .read() + .unwrap_or_else(|poisoned| poisoned.into_inner()) + .as_ref() + .map(|view| Summary { + members: view.members.len(), + reachable: view + .members + .iter() + .filter(|m| m.as_str() != local && !view.unreachable.contains(*m)) + .count(), + stable: view.stable.len(), + neighbours: view.neighbours.len(), + stable_neighbours: view + .neighbours + .iter() + .filter(|m| view.stable.contains(*m)) + .count(), + full_mesh: view.full_mesh, + }) +} + +/// Whether we initiate contact with `peer`: HELLOs, keepalives, dormancy checks, sync +/// openings. Everyone is a neighbour until the view is known (while joining, or before +/// any member registered), so joining works as in a full mesh. We always *answer* +/// anyone who contacts us. +pub(crate) fn is_neighbour(state: &AppState, peer: &str) -> bool { + match &*state + .neighbours + .read() + .unwrap_or_else(|poisoned| poisoned.into_inner()) + { + Some(view) => view.neighbours.contains(peer), + None => true, + } +} + +/// How the neighbours changed. +pub(crate) struct NeighbourChange { + pub(crate) added: Vec<String>, + pub(crate) removed: Vec<String>, +} + +/// Recomputes the neighbours once ready, if the member list, the reachable members or +/// `k` changed. +pub(crate) async fn refresh(state: &AppState) -> Option<NeighbourChange> { + if !state.is_ready().await { + return None; + } + let (current, stable) = { + let replica = state.doc.read().await; + let current = members::members(&replica.doc); + let stable: BTreeSet<String> = current + .iter() + .filter(|m| members::tier(&replica.doc, m) == members::Tier::Stable) + .cloned() + .collect(); + (current, stable) + }; + let k = state.neighbours_k; + let now = Instant::now(); + let mut unreachable = BTreeSet::new(); + for member in ¤t { + if let Some(peer_state) = state.get_peer_state(member).await { + if peer_state.lock().await.is_unreachable(&state.timing, now) { + unreachable.insert(member.clone()); + } + } + } + let previous = { + let view = state + .neighbours + .read() + .unwrap_or_else(|poisoned| poisoned.into_inner()); + if view.as_ref().is_some_and(|view| { + view.members == current + && view.k == k + && view.unreachable == unreachable + && view.stable == stable + }) { + return None; + } + view.as_ref().map(|view| view.neighbours.clone()) + }; + + // Before the first view everyone known counted as a neighbour. + let previous = match previous { + Some(previous) => previous, + None => state.peer_states.read().await.keys().cloned().collect(), + }; + let computed = neighbours( + ¤t, + state.local_id(), + k, + &|m| unreachable.contains(m), + &|m| stable.contains(m), + ); + let change = NeighbourChange { + added: computed + .iter() + .filter(|m| !previous.contains(*m)) + .cloned() + .collect(), + removed: previous + .iter() + .filter(|m| !computed.contains(*m)) + .cloned() + .collect(), + }; + info!( + members = current.len(), + unreachable = unreachable.len(), + stable = stable.len(), + k, + neighbours = computed.len(), + added = change.added.len(), + removed = change.removed.len(), + "Neighbours computed from the member list" + ); + let reachable_others = current + .iter() + .filter(|m| m.as_str() != state.local_id() && !unreachable.contains(*m)) + .count(); + *state + .neighbours + .write() + .unwrap_or_else(|poisoned| poisoned.into_inner()) = Some(NeighbourView { + full_mesh: computed.len() >= reachable_others, + members: current, + k, + unreachable, + stable, + neighbours: computed, + }); + Some(change) +} + +const RING_LABEL: &[u8] = b"continuum/topology-ring/1"; +const PAIR_LABEL: &[u8] = b"continuum/topology-pair/1"; + +/// How many members may prefer the same stable member, as a multiple of `k`. Without a +/// bound, 20 stable members among 2000 get about 500 neighbours each: every +/// announcement in the swarm lands on them, and the backbone saturates first. Members +/// that find a stable member full prefer the next one, then go by pair score, so the +/// graph stays connected and symmetric. +const STABLE_FANIN_FACTOR: usize = 4; + +/// The neighbours of `local` among `members` (which may or may not contain `local`), +/// skipping members for which `unreachable` holds. +pub(crate) fn neighbours( + members: &[String], + local: &str, + k: usize, + unreachable: &dyn Fn(&str) -> bool, + stable: &dyn Fn(&str) -> bool, +) -> BTreeSet<String> { + let mut all: Vec<String> = members.to_vec(); + all.push(local.to_owned()); + graph(&all, k, &|m: &str| m != local && unreachable(m), stable) + .remove(local) + .unwrap_or_default() +} + +/// Every reachable member's neighbours. Costs about N² hashes (a few milliseconds for +/// hundreds of members). `stable` members (`swarm::stability`) are preferred for half of +/// each member's scored slots, each by at most `STABLE_FANIN_FACTOR * k` members. +/// +/// Each reachable member walks the ring (by key hash) both ways to the first reachable +/// member, which gives a ring over the reachable members, and adds the `k - 2` +/// reachable members with the lowest pair score. A link exists if either end chose it. +/// With at most `k + 1` reachable members, the reachable members form a full mesh. +/// +/// Each node judges reachability by itself, so nodes may briefly disagree; that only +/// adds links (whoever chose a link initiates, the other answers), it never cuts one. +/// The preference budget is spent in key order, which every node shares, so nodes that +/// agree on reachability compute the same links. +pub(crate) fn graph( + members: &[String], + k: usize, + unreachable: &dyn Fn(&str) -> bool, + stable: &dyn Fn(&str) -> bool, +) -> BTreeMap<String, BTreeSet<String>> { + let k = k.max(2); + let all: BTreeSet<&str> = members.iter().map(String::as_str).collect(); + let live: BTreeSet<&str> = all.iter().copied().filter(|m| !unreachable(m)).collect(); + let mut links: BTreeMap<String, BTreeSet<String>> = live + .iter() + .map(|m| ((*m).to_owned(), BTreeSet::new())) + .collect(); + + let mut link = |a: &str, b: &str| { + if a != b { + links.get_mut(a).expect("live member").insert(b.to_owned()); + links.get_mut(b).expect("live member").insert(a.to_owned()); + } + }; + + if live.len() <= k + 1 { + for a in &live { + for b in &live { + link(a, b); + } + } + return links; + } + + // The ring over all members, so its order doesn't change when one goes offline. + let mut ring: Vec<([u8; 32], &str)> = + all.iter().map(|m| (hash(RING_LABEL, &[*m]), *m)).collect(); + ring.sort_unstable(); + let n = ring.len(); + let position: BTreeMap<&str, usize> = + ring.iter().enumerate().map(|(i, (_, m))| (*m, i)).collect(); + let first_live = |from: usize, step: usize| { + (1..n) + .map(|d| ring[(from + d * step) % n].1) + .find(|m| live.contains(m)) + .expect("more than k + 1 live members") + }; + + // How many members have preferred each stable member so far (`live` iterates in key + // order, the same on every node). + let budget = STABLE_FANIN_FACTOR * k; + let mut preferred_by: BTreeMap<&str, usize> = BTreeMap::new(); + + for member in &live { + let at = position[member]; + let next = first_live(at, 1); + let prev = first_live(at, n - 1); + link(member, next); + link(member, prev); + let mut candidates: Vec<([u8; 32], &str)> = live + .iter() + .copied() + .filter(|other| other != member && *other != next && *other != prev) + .map(|other| (pair_score(member, other), other)) + .collect(); + candidates.sort_unstable(); + // Half the slots (rounded up) prefer stable members, so long-running nodes form + // a backbone; the rest go by pair score alone, keeping paths short. A stable + // member enough others prefer already is skipped, so no hub grows with the swarm. + let stable_slots = (k - 2).div_ceil(2); + let preferred: Vec<&str> = candidates + .iter() + .map(|(_, other)| *other) + .filter(|other| stable(other) && preferred_by.get(other).copied().unwrap_or(0) < budget) + .take(stable_slots) + .collect(); + for other in preferred.iter().copied() { + *preferred_by.entry(other).or_default() += 1; + } + let rest = candidates + .iter() + .map(|(_, other)| *other) + .filter(|other| !preferred.contains(other)) + .take(k - 2 - preferred.len()); + for partner in preferred.iter().copied().chain(rest) { + link(member, partner); + } + } + links +} + +/// Symmetric: the same for (a, b) and (b, a). +fn pair_score(a: &str, b: &str) -> [u8; 32] { + let (low, high) = if a <= b { (a, b) } else { (b, a) }; + hash(PAIR_LABEL, &[low, high]) +} + +fn hash(label: &[u8], fields: &[&str]) -> [u8; 32] { + let mut hasher = Sha256::new(); + for field in [label] + .into_iter() + .chain(fields.iter().map(|f| f.as_bytes())) + { + hasher.update((field.len() as u32).to_be_bytes()); + hasher.update(field); + } + hasher.finalize().into() +} + +#[cfg(test)] +mod tests { + use super::*; + use std::collections::VecDeque; + + fn members(n: usize) -> Vec<String> { + (0..n).map(|i| format!("{:064x}", i * 7919 + 13)).collect() + } + + fn all_reachable(_: &str) -> bool { + false + } + + fn none_stable(_: &str) -> bool { + false + } + + fn graph(members: &[String], k: usize) -> BTreeMap<String, BTreeSet<String>> { + super::graph(members, k, &all_reachable, &none_stable) + } + + fn connected(g: &BTreeMap<String, BTreeSet<String>>) -> bool { + let Some(start) = g.keys().next() else { + return true; + }; + let mut seen = BTreeSet::from([start.clone()]); + let mut queue = VecDeque::from([start.clone()]); + while let Some(m) = queue.pop_front() { + for next in &g[&m] { + if seen.insert(next.clone()) { + queue.push_back(next.clone()); + } + } + } + seen.len() == g.len() + } + + fn edges(graph: &BTreeMap<String, BTreeSet<String>>) -> BTreeSet<(String, String)> { + graph + .iter() + .flat_map(|(a, ns)| { + ns.iter().map(move |b| { + if a < b { + (a.clone(), b.clone()) + } else { + (b.clone(), a.clone()) + } + }) + }) + .collect() + } + + #[test] + fn small_swarms_are_a_full_mesh() { + let ms = members(13); + for m in &ms { + assert_eq!( + neighbours(&ms, m, 12, &all_reachable, &none_stable).len(), + 12 + ); + } + } + + #[test] + fn links_are_symmetric_and_degree_at_least_k() { + let ms = members(250); + let g = graph(&ms, 8); + for (a, ns) in &g { + assert!(ns.len() >= 8, "{a} has {} neighbours", ns.len()); + assert!(!ns.contains(a)); + for b in ns { + assert!(g[b].contains(a), "{a} -> {b} not symmetric"); + } + } + } + + #[test] + fn graph_is_connected() { + for (n, k) in [(250, 8), (250, 2), (60, 3)] { + assert!(connected(&graph(&members(n), k)), "n={n} k={k}"); + } + } + + #[test] + fn one_member_joining_changes_few_links() { + let ms = members(200); + let before = edges(&graph(&ms, 8)); + let mut more = ms.clone(); + more.push(format!("{:064x}", 999_999)); + let after = edges(&graph(&more, 8)); + let changed = before.symmetric_difference(&after).count(); + assert!(changed <= 4 * 8, "{changed} links changed"); + } + + #[test] + fn reachable_members_stay_connected_when_most_are_offline() { + let ms = members(100); + // Every tenth member is online. + let online: BTreeSet<String> = ms.iter().step_by(10).cloned().collect(); + let offline = |m: &str| !online.contains(m); + let g = super::graph(&ms, 12, &offline, &none_stable); + assert_eq!(g.len(), 10); + // Ten reachable members and k = 12: a full mesh among them. + assert!(g.values().all(|ns| ns.len() == 9)); + + let online: BTreeSet<String> = ms.iter().step_by(4).cloned().collect(); + let offline = |m: &str| !online.contains(m); + let g = super::graph(&ms, 8, &offline, &none_stable); + assert_eq!(g.len(), 25); + assert!(connected(&g)); + assert!( + g.values() + .all(|ns| ns.len() >= 8 && ns.iter().all(|n| online.contains(n))) + ); + } + + #[test] + fn stable_members_become_hubs_with_bounded_degree() { + let ms = members(250); + let stable: BTreeSet<String> = ms.iter().step_by(12).take(20).cloned().collect(); + let is_stable = |m: &str| stable.contains(m); + let g = super::graph(&ms, 12, &all_reachable, &is_stable); + assert!(connected(&g)); + + let degree = |pred: &dyn Fn(&String) -> bool| -> (usize, f64) { + let degrees: Vec<usize> = g + .iter() + .filter(|(m, _)| pred(m)) + .map(|(_, ns)| ns.len()) + .collect(); + let max = *degrees.iter().max().unwrap(); + ( + max, + degrees.iter().sum::<usize>() as f64 / degrees.len() as f64, + ) + }; + let (stable_max, stable_mean) = degree(&|m| stable.contains(m)); + let (_, transient_mean) = degree(&|m| !stable.contains(m)); + // 230 transient members put about 5 of their 10 scored slots on 20 stable ones. + assert!( + stable_mean > 3.0 * transient_mean, + "{stable_mean} vs {transient_mean}" + ); + assert!( + stable_max <= 100, + "a stable member has {stable_max} neighbours" + ); + } + + #[test] + fn hub_degree_is_bounded_however_few_stable_members() { + // Few stable members among many: without the fan-in budget each would be chosen + // by most of the swarm (about 130 neighbours with 10 of 250). With it, a hub has + // at most the budget plus the links others choose by pair score and the ring. + for (n, stable_count) in [(250, 10), (1000, 10), (1000, 20)] { + let k = 12; + let ms = members(n); + let stable: BTreeSet<String> = ms.iter().step_by(n / stable_count).cloned().collect(); + let is_stable = |m: &str| stable.contains(m); + let g = super::graph(&ms, k, &all_reachable, &is_stable); + assert!(connected(&g), "n={n} stable={stable_count}"); + let hub_max = g + .iter() + .filter(|(m, _)| stable.contains(*m)) + .map(|(_, ns)| ns.len()) + .max() + .unwrap(); + assert!( + hub_max <= 8 * k, + "n={n} stable={stable_count}: a stable member has {hub_max} neighbours" + ); + } + } +} |
