//! Passing a writer's changes on along its tree: every member gets them once, over a //! link that works, and as soon as the member before it has them. //! //! The tree is the same on every node, computed from what they share: the graph the //! member list gives (`topology::member_list_graph`: every member's neighbours, as if //! all reached all) and what each member tells its neighbours it reaches //! (`swarm::reach`). Every member's distance from the writer over that graph orders the //! members: by distance, then whether it reaches a member a hop closer, then pair key; //! a member's parent is the neighbour it reaches that comes before it, the closest to //! the writer first, then one that reaches closer, then by pair score //! (`topology::pair_score`). Coming before rules out cycles, and taking only a neighbour //! it reaches routes around links that don't work: where a writer behind NAT reaches a //! single member, that one comes first and the others take it. //! //! Whose a change is, its author says (every node writes as itself: its node ID, by //! pair key here): whoever applies changes passes them on to its own children in each of //! their writers' trees (`forward`), however they came; so does the writer, to all its //! neighbours, with its own writes. A sync message along a tree also says whose //! (`SyncEnvelope::writer`), for one that had them already (`swarm::sync`); like any //! sync message, it needs no answer where the child had nothing the parent lacks //! (`SyncSession::assume_delivered`). //! A member no member it reaches comes before (or whose parent didn't get the change) //! gets it from the announcements (`swarm::relay`), as before. //! //! The graph is computed once per member list (tens of milliseconds for thousands of //! members), the distances once per writer and graph (`MAX_WRITERS` kept). use std::collections::{BTreeMap, BTreeSet, HashMap, VecDeque}; use std::sync::Arc; use tracing::trace; use crate::locks::MutexExt; use crate::state::AppState; use crate::swarm::{reach, topology}; /// The writers whose distances are kept at most. const MAX_WRITERS: usize = 256; type Graph = BTreeMap>; /// The distances from each writer, over the graph they were computed for. #[derive(Default)] pub(crate) struct Trees { graph: Option>, /// The members by pair key, of that graph. by_key: HashMap, distances: HashMap>>, } /// The graph, and every member's distance from `writer` (by pair key) over it; None /// while the member list isn't known, or `writer` isn't in it. fn distances(state: &AppState, writer: u64) -> Option<(Arc, Arc>)> { let graph = topology::member_list_graph(state)?; let mut trees = state.trees.lock_or_recover(); if !trees .graph .as_ref() .is_some_and(|known| Arc::ptr_eq(known, &graph)) { trees.by_key = graph .keys() .map(|member| (topology::pair_key(member), member.clone())) .collect(); trees.distances.clear(); trees.graph = Some(Arc::clone(&graph)); } if let Some(known) = trees.distances.get(&writer) { return Some((graph, Arc::clone(known))); } let from = trees.by_key.get(&writer)?.clone(); let mut distance = HashMap::from([(from.clone(), 0u32)]); let mut queue = VecDeque::from([from]); while let Some(member) = queue.pop_front() { let next = distance[&member] + 1; for neighbour in graph.get(&member).into_iter().flatten() { if !distance.contains_key(neighbour) { distance.insert(neighbour.clone(), next); queue.push_back(neighbour.clone()); } } } if trees.distances.len() >= MAX_WRITERS { trees.distances.clear(); } let distance = Arc::new(distance); trees.distances.insert(writer, Arc::clone(&distance)); Some((graph, distance)) } /// Our children in `writer`'s tree: our neighbours in the graph whose parent we are, /// and whom we reach. /// /// Members are ordered by distance from the writer, then whether they reach a member a /// hop closer (as they say: else, while it hasn't the change from its own side, it is /// as good as farther), then pair key; a member's parent is the neighbour it reaches /// that comes before it, by the same, then pair score. So where the writer reaches few /// (behind NAT), those it reaches come first among the members as far from it, and the /// others take them as parents. What a member reaches we know of ourselves and our /// neighbours, as they said; of others, the graph is taken as right (if we're wrong, /// a change goes twice or by announcement). #[cfg(test)] pub(crate) async fn children(state: &AppState, writer: u64) -> Vec { plan(state, writer) .await .map(|plan| plan.children) .unwrap_or_default() } /// Where `member` comes in the writer's tree (by `distance` from it over `graph`): /// its distance, then whether it reaches a member a hop closer (`reached(by, other)`: /// whether `by` reaches `other`, as known; the writer does), then its pair key. None if /// the graph doesn't have it. fn rank( graph: &Graph, distance: &HashMap, reached: &dyn Fn(&str, &str) -> bool, member: &str, ) -> Option<(u32, bool, u64)> { let d = *distance.get(member)?; let closer = d == 0 || graph.get(member).into_iter().flatten().any(|other| { distance.get(other.as_str()) == Some(&(d - 1)) && reached(member, other) }); Some((d, !closer, topology::pair_key(member))) } /// `member`'s parent in the writer's tree (`rank`): the neighbour it reaches that comes /// before it, the closest to the writer first, then one that reaches closer, then by /// pair score; None if none does (or it is the writer). fn parent<'g>( graph: &'g Graph, distance: &HashMap, reached: &dyn Fn(&str, &str) -> bool, member: &str, ) -> Option<&'g str> { let its_rank = rank(graph, distance, reached, member)?; graph .get(member) .into_iter() .flatten() .map(String::as_str) .filter(|candidate| { rank(graph, distance, reached, candidate).is_some_and(|r| r < its_rank) && reached(member, candidate) }) .min_by_key(|candidate| { let (d, far, _) = rank(graph, distance, reached, candidate).expect("ranked"); (d, far, topology::pair_score(candidate, member)) }) } /// `writer`'s whole tree, as far as `reached` knows whom each member reaches (the /// dashboard's map, from the links members report): every member's parent, None for /// those without one (they get the writer's changes by announcement). None while the /// member list isn't known, or the writer isn't in it. #[cfg_attr(not(feature = "dashboard"), allow(dead_code))] pub(crate) fn whole( state: &AppState, writer: &str, reached: &dyn Fn(&str, &str) -> bool, ) -> Option>> { let (graph, distance) = distances(state, topology::pair_key(writer))?; Some( graph .keys() .filter(|member| distance.get(member.as_str()).is_some_and(|d| *d > 0)) .map(|member| { let parent = parent(&graph, &distance, reached, member).map(str::to_owned); (member.clone(), parent) }) .collect(), ) } /// Whom we and our neighbours in `graph` reach, by pair key, as far as we know (our /// own, and what our neighbours said). async fn known_reaches<'g>(state: &AppState, graph: &'g Graph) -> HashMap<&'g str, Vec> { let me = state.local_id(); let mut reaches: HashMap<&str, Vec> = HashMap::new(); if let Some((me, _)) = graph.get_key_value(me) { reaches.insert(me.as_str(), reach::published(state)); } for member in graph.get(me).into_iter().flatten() { if let Some(peer_state) = state.get_peer_state(member).await && let Some(keys) = reach::of(peer_state.lock().await.reaches.as_ref()) { reaches.insert(member.as_str(), keys); } } reaches } /// Our own tree, as our writes go along it (`Plan`, as every member computes its part). #[derive(Debug, Clone, Copy, PartialEq, Eq)] pub(crate) struct Shape { /// The other members in it. pub(crate) members: usize, /// The most hops from us to one of them. pub(crate) depth: u32, /// The hops to each, summed (for the mean). pub(crate) hops: u64, /// Members it doesn't reach: without a parent there (no member they reach comes /// before them), below one without, or not connected to us in the member list's /// graph. They get our writes by announcement. pub(crate) off: usize, } /// The shape of our own tree; None while the member list isn't known, or we aren't in /// it. What members other than us and our neighbours reach is taken from the graph, as /// in `plan`. pub(crate) async fn own_shape(state: &AppState) -> Option { let me = state.local_id(); let (graph, distance) = distances(state, topology::pair_key(me))?; let reaches = known_reaches(state, &graph).await; let reached = |by: &str, member: &str| { reaches .get(by) .is_none_or(|keys| keys.contains(&topology::pair_key(member))) }; Some(shape(&graph, &distance, &reached, me)) } /// `own_shape` over `graph`, from `me` at distance 0. fn shape( graph: &Graph, distance: &HashMap, reached: &dyn Fn(&str, &str) -> bool, me: &str, ) -> Shape { let parents: HashMap<&str, Option<&str>> = graph .keys() .filter(|member| distance.get(member.as_str()).is_some_and(|d| *d > 0)) .map(|member| (member.as_str(), parent(graph, distance, reached, member))) .collect(); // Hops from us along parents; None where the chain ends without us. Parents come // before their children, so chains end. let mut hops: HashMap<&str, Option> = HashMap::from([(me, Some(0))]); fn along<'g>( member: &'g str, parents: &HashMap<&'g str, Option<&'g str>>, hops: &mut HashMap<&'g str, Option>, ) -> Option { if let Some(known) = hops.get(member) { return *known; } let found = parents .get(member) .copied() .flatten() .and_then(|parent| along(parent, parents, hops)) .map(|hops| hops + 1); hops.insert(member, found); found } let mut shape = Shape { members: 0, depth: 0, hops: 0, off: graph.len() - 1 - parents.len(), }; for member in parents.keys() { match along(member, &parents, &mut hops) { Some(h) => { shape.members += 1; shape.depth = shape.depth.max(h); shape.hops += u64::from(h); } None => shape.off += 1, } } shape } /// How long our writes take to reach the members along our own tree, by link latency /// alone (the bench's, which knows each link's): the mean over the members it reaches /// (and the others do), against two other trees of ours. In one, each member's parent is chosen as now, but /// among those as close to us and as well placed (`rank`) by latency before pair /// score; in the other, every member takes the fastest path over the links members /// reach (no tree computed alike everywhere gets below it). In milliseconds, one way. #[cfg(test)] #[derive(Debug, Clone, Copy)] pub(crate) struct Latencies { pub(crate) members: usize, pub(crate) tree: f64, pub(crate) by_latency: f64, pub(crate) fastest: f64, } /// `Latencies` of our own tree, by `latency(from, to)` in milliseconds; None while the /// member list isn't known, or we aren't in it, or the tree reaches nobody. #[cfg(test)] pub(crate) async fn own_latencies( state: &AppState, latency: &(dyn Fn(&str, &str) -> f64 + Sync), ) -> Option { let me = state.local_id(); let (graph, distance) = distances(state, topology::pair_key(me))?; let reaches = known_reaches(state, &graph).await; let reached = |by: &str, member: &str| { reaches .get(by) .is_none_or(|keys| keys.contains(&topology::pair_key(member))) }; let members: Vec<&str> = graph .keys() .map(String::as_str) .filter(|member| distance.get(*member).is_some_and(|d| *d > 0)) .collect(); let now: HashMap<&str, Option<&str>> = members .iter() .map(|member| (*member, parent(&graph, &distance, &reached, member))) .collect(); let faster: HashMap<&str, Option<&str>> = members .iter() .map(|member| { let its_rank = rank(&graph, &distance, &reached, member); let candidate = graph .get(*member) .into_iter() .flatten() .map(String::as_str) .filter(|candidate| { its_rank.is_some_and(|its| { rank(&graph, &distance, &reached, candidate).is_some_and(|r| r < its) }) && reached(member, candidate) }) .min_by(|a, b| { let (da, fa, _) = rank(&graph, &distance, &reached, a).expect("ranked"); let (db, fb, _) = rank(&graph, &distance, &reached, b).expect("ranked"); (da, fa) .cmp(&(db, fb)) .then(latency(a, member).total_cmp(&latency(b, member))) .then(topology::pair_score(a, member).cmp(&topology::pair_score(b, member))) }); (*member, candidate) }) .collect(); // Arrival along parents; None where the chain ends without us. Parents come // before their children, so chains end. fn arrival<'g>( member: &'g str, parents: &HashMap<&'g str, Option<&'g str>>, latency: &dyn Fn(&str, &str) -> f64, known: &mut HashMap<&'g str, Option>, ) -> Option { if let Some(at) = known.get(member) { return *at; } let at = parents.get(member).copied().flatten().and_then(|parent| { arrival(parent, parents, latency, known).map(|at| at + latency(parent, member)) }); known.insert(member, at); at } let arrivals = |parents: &HashMap<&str, Option<&str>>| -> HashMap<&str, f64> { let mut known = HashMap::from([(me, Some(0.0))]); members .iter() .filter_map(|member| { arrival(member, parents, latency, &mut known).map(|at| (*member, at)) }) .collect() }; let (now, faster) = (arrivals(&now), arrivals(&faster)); // The fastest paths: Dijkstra over the links members reach (a parent `p` of // `c` where `c` reaches `p`), from us. let mut fastest: HashMap<&str, f64> = HashMap::from([(me, 0.0)]); let mut done = std::collections::HashSet::new(); while let Some((&next, &at)) = fastest .iter() .filter(|(member, _)| !done.contains(*member)) .min_by(|a, b| a.1.total_cmp(b.1)) { done.insert(next); for other in graph.get(next).into_iter().flatten().map(String::as_str) { if !reached(other, next) { continue; } let through = at + latency(next, other); if fastest.get(other).is_none_or(|known| through < *known) { fastest.insert(other, through); } } } // Over the members all three reach (the other tree may leave out others than ours). let common: Vec<&str> = now .keys() .copied() .filter(|member| faster.contains_key(member) && fastest.contains_key(member)) .collect(); if common.is_empty() { return None; } let mean = |times: &HashMap<&str, f64>| { common.iter().map(|member| times[member]).sum::() / common.len() as f64 }; Some(Latencies { members: common.len(), tree: mean(&now), by_latency: mean(&faster), fastest: mean(&fastest), }) } /// What `writer`'s tree means for us: our children in it, and our neighbours it leaves /// out (`uncovered`). pub(crate) struct Plan { pub(crate) children: Vec, /// Those of our children that reach no member but us, as they said: nobody else's /// slow round tells them of a change, should our message to them be lost (it is /// taken as delivered, `SyncSession::assume_delivered`), so ours does. pub(crate) alone: Vec, /// Our neighbours that no member they reach comes before (the tree reaches them only /// by announcement), and those not in the member list's graph (whom no tree has). pub(crate) uncovered: Vec, /// Our neighbours that get the writer's changes from another parent, and the writer /// itself: they have them, or will, without us (`forward`). pub(crate) covered: Vec, } /// Our part in `writer`'s tree (`children`, `Plan`); None while the member list isn't /// known, or the writer or we aren't in it. pub(crate) async fn plan(state: &AppState, writer: u64) -> Option { let (graph, distance) = distances(state, writer)?; let me = state.local_id(); if !distance.contains_key(me) { return None; } let reaches = known_reaches(state, &graph).await; let reached = |by: &str, member: &str| { reaches .get(by) .is_none_or(|keys| keys.contains(&topology::pair_key(member))) }; let my_rank = rank(&graph, &distance, &reached, me)?; let my_key = topology::pair_key(me); let mut children = Vec::new(); let mut alone = Vec::new(); let mut uncovered = Vec::new(); let mut covered = Vec::new(); for member in graph.get(me).into_iter().flatten() { let Some(its_rank) = rank(&graph, &distance, &reached, member) else { continue; }; if distance[member.as_str()] == 0 { covered.push(member.clone()); continue; } match parent(&graph, &distance, &reached, member) { None => uncovered.push(member.clone()), Some(parent) if parent == me && its_rank > my_rank && topology::reaches(state, member).await => { if reaches .get(member.as_str()) .is_some_and(|keys| keys.iter().all(|key| *key == my_key)) { alone.push(member.clone()); } children.push(member.clone()); } Some(parent) if parent != me => covered.push(member.clone()), // Ours by the graph, but we don't reach it, or it comes before us. Some(_) => {} } } // Neighbours the member list's graph doesn't give us (not listed yet, or only // connecting out): no tree has them. for (peer, _) in state.peers_snapshot().await { if topology::is_neighbour(state, &peer) && !graph.get(me).is_some_and(|ours| ours.contains(&peer)) { uncovered.push(peer); } } Some(Plan { children, alone, uncovered, covered, }) } /// We applied `changes` of `writer`'s that `from` passed on (none if we had them /// already): to our children in its tree (but `from`), a sync message each, saying /// whose. Returns our part in the tree (`Plan`), None if we have none (announcements /// then do it all). /// /// Our neighbours that get them from another parent (`Plan::covered`) count as having /// them (`SyncSession::known_to_peer`): else, with writers in turn, each change coming /// along another tree, the next message we pass on to one would carry all we got since /// our last to it, mostly what its other parents brought it. Should one lack them after /// all, it asks, and gets them (`resend_if_needed`). But to our children, those of /// them that `changes` build on go along, if they were counted so lately /// (`SyncSession::with_predecessors`). pub(crate) async fn forward( state: &AppState, writer: u64, from: &str, changes: &[crate::document::ChangeHash], ) -> Option { let mut plan = plan(state, writer).await?; if !changes.is_empty() { for neighbour in &plan.covered { if let Some(peer_state) = state.get_peer_state(neighbour).await { peer_state .lock() .await .sync .known_to_peer(changes, tokio::time::Instant::now()); } } } plan.children.retain(|child| child != from); let children = &plan.children; let replica = if changes.is_empty() { None } else { Some(state.doc.read().await) }; for child in children { if let Some(peer_state) = state.get_peer_state(child).await { let mut peer_state = peer_state.lock().await; peer_state.tree_writer = Some(writer); if let Some(replica) = &replica { peer_state.sync.with_predecessors( &replica.doc, changes, tokio::time::Instant::now(), ); } } state.stats.tree_forward(); state.request_sync(child); } if !children.is_empty() { trace!( writer, children = children.len(), "Passing a change on along the writer's tree" ); } Some(plan) } #[cfg(test)] mod tests { use super::*; /// Hops along parents from us; members without a parent, below one without, or /// not connected count as off the tree. #[test] fn our_tree_in_numbers() { let link = |graph: &mut Graph, a: &str, b: &str| { graph.entry(a.into()).or_default().insert(b.into()); graph.entry(b.into()).or_default().insert(a.into()); }; let mut graph = Graph::new(); link(&mut graph, "me", "a"); link(&mut graph, "a", "b"); link(&mut graph, "me", "c"); link(&mut graph, "c", "e"); graph.insert("d".into(), BTreeSet::new()); let distance: HashMap = [("me", 0), ("a", 1), ("c", 1), ("b", 2), ("e", 2)] .into_iter() .map(|(member, d)| (member.to_owned(), d)) .collect(); // c doesn't reach us: no parent, nor e below it. let reached = |by: &str, member: &str| !(by == "c" && member == "me"); assert_eq!( shape(&graph, &distance, &reached, "me"), Shape { members: 2, depth: 2, hops: 3, off: 3, } ); } }