launcher/src/web/topo.rs

//! Physical topology derivation for the dashboard map.
//!
//! Pure config → map translation: nodes, cables, HA glyphs. The UI draws
//! the fixed scene; the snapshot builder only attaches live HA states.

use crate::config::Config;
use serde::Serialize;

/// One cable on the map: `a` → `b`, where `a` is a switch entity id, the
/// pseudo-switch `"lan"` (the passive 10G), or a node name (DAC), and `b`
/// is a node name.
#[derive(Serialize, Clone, Debug, PartialEq)]
pub struct TopoLink {
    /// "200g" (HA-managed fabric switch), "dac" (direct-attach copper),
    /// "10g" (passive 10G switch).
    pub kind: String,
    pub a: String,
    pub b: String,
}

/// An HA-managed device glyph: the blocking fabric switch ("ib") or an
/// auxiliary entity like a fan ("aux").
#[derive(Serialize, Clone, Debug, PartialEq)]
pub struct TopoSwitch {
    pub entity: String,
    pub kind: String,
}

/// HA state for a map glyph: `on`/`off`, `None` = unknown/unreachable.
#[derive(Serialize, Clone, Debug, PartialEq)]
pub struct TopoState {
    pub entity: String,
    pub on: Option<bool>,
}

/// The fixed physical map, derived from config facts (see [`derive_topo`]).
#[derive(Serialize, Clone, Debug, Default, PartialEq)]
pub struct TopoSnap {
    /// Every configured node, sorted by slot (digits parsed from the name).
    pub nodes: Vec<String>,
    pub links: Vec<TopoLink>,
    pub switches: Vec<TopoSwitch>,
    /// HA states for [`TopoSnap::switches`] (filled by the snapshot builder).
    pub states: Vec<TopoState>,
}

/// Derive the fixed physical topology from the config alone:
///
/// - every configured node occupies its slot (`dgx7.lan` → slot 7) —
///   absent slots render as "planned";
/// - a node with an `ib_ip` hangs on its cluster's *blocking* switch
///   entity that is not a node plug (the 200G fabric, HA-managed) —
///   or, when the cluster has no blocking switch (`2nodes-dac`), the IB
///   link is direct-attach copper between consecutive ranks;
/// - every node hangs on the passive 10G switch (pseudo-entity `"lan"`,
///   no HA state);
/// - extra entities that are neither blocking-fabric switches nor node
///   plugs appear as auxiliary glyphs (a fan).
pub(crate) fn derive_topo(config: &Config) -> TopoSnap {
    let slot = |n: &str| {
        n.chars()
            .filter(|c| c.is_ascii_digit())
            .collect::<String>()
            .parse::<u32>()
            .unwrap_or(u32::MAX)
    };
    let mut nodes: Vec<String> = Vec::new();
    let mut links: Vec<TopoLink> = Vec::new();
    let mut switches: Vec<TopoSwitch> = Vec::new();
    let push_switch = |switches: &mut Vec<TopoSwitch>, entity: &str, kind: &str| {
        if !switches.iter().any(|s| s.entity == entity) {
            switches.push(TopoSwitch {
                entity: entity.to_string(),
                kind: kind.to_string(),
            });
        }
    };
    for c in &config.clusters {
        let node_entities: Vec<String> = c.nodes.iter().filter_map(|n| n.entity.clone()).collect();
        for n in &c.nodes {
            if !nodes.contains(&n.name) {
                nodes.push(n.name.clone());
            }
        }
        let ib: Vec<&crate::config::Node> = c.nodes.iter().filter(|n| n.ib_ip.is_some()).collect();
        let blocking = c
            .extra_entities
            .iter()
            .find(|e| e.blocks() && !node_entities.contains(&e.entity().to_string()));
        if let Some(sw) = blocking {
            push_switch(&mut switches, sw.entity(), "ib");
            for n in ib {
                let link = TopoLink {
                    kind: "200g".into(),
                    a: sw.entity().to_string(),
                    b: n.name.clone(),
                };
                // the same physical cable appears in every cluster that
                // shares the fabric — one line on the map
                if !links.contains(&link) {
                    links.push(link);
                }
            }
        } else {
            // An IB subnet with no switch fabric: direct-attach copper,
            // point-to-point between consecutive ranks.
            let mut ranked = ib;
            ranked.sort_by_key(|n| n.rank);
            for w in ranked.windows(2) {
                links.push(TopoLink {
                    kind: "dac".into(),
                    a: w[0].name.clone(),
                    b: w[1].name.clone(),
                });
            }
        }
        for e in &c.extra_entities {
            if !e.blocks() && !node_entities.contains(&e.entity().to_string()) {
                push_switch(&mut switches, e.entity(), "aux");
            }
        }
    }
    nodes.sort_by_key(|n| (slot(n), n.clone()));
    for n in &nodes {
        links.push(TopoLink {
            kind: "10g".into(),
            a: "lan".into(),
            b: n.clone(),
        });
    }
    TopoSnap {
        nodes,
        links,
        switches,
        states: Vec::new(),
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use std::collections::HashMap;

    fn node(name: &str, ib: Option<&str>, entity: Option<&str>, rank: u32) -> crate::config::Node {
        crate::config::Node {
            name: name.into(),
            ssh_ip: "10.0.0.1".into(),
            ib_ip: ib.map(String::from),
            rank,
            host_key: None,
            entity: entity.map(String::from),
        }
    }

    fn cluster(
        name: &str,
        nodes: Vec<crate::config::Node>,
        extra: Vec<crate::config::ExtraEntity>,
    ) -> crate::config::Cluster {
        crate::config::Cluster {
            name: name.into(),
            master_ib_ip: None,
            nodes,
            extra_entities: extra,
        }
    }

    #[test]
    fn derive_topo_fabric_dac_lan() {
        let config = Config {
            ssh_user: "u".into(),
            ssh_key_path: "/k".into(),
            nccl: HashMap::new(),
            hf_cache_dir: None,
            init_port: 25000,
            before_serve: HashMap::new(),
            exec_aliases: HashMap::new(),
            home_assistant: None,
            mikrotik: None,
            global_mods_dir: None,
            drop_caches: true,
            mount_aliases: HashMap::new(),
            recipe_sources: Vec::new(),
            docker_image_sources: Vec::new(),
            cluster_sources: Vec::new(),
            clusters: vec![
                // fabric: blocking switch + ib nodes → 200g links
                cluster(
                    "fabric",
                    vec![
                        node("dgx2.lan", Some("10.0.200.2"), Some("switch.dgx2"), 1),
                        node("dgx1.lan", Some("10.0.200.1"), Some("switch.dgx1"), 0),
                    ],
                    vec![
                        crate::config::ExtraEntity::Simple("switch.sw6".into()),
                        crate::config::ExtraEntity::Detailed {
                            entity: "switch.dgx1".into(),
                            blocks: true,
                        },
                    ],
                ),
                // dac: ib nodes, no blocking switch → direct chain by rank
                cluster(
                    "dac",
                    vec![
                        node("dgx1.lan", Some("10.0.100.1"), Some("switch.dgx1"), 0),
                        node("dgx2.lan", Some("10.0.100.2"), Some("switch.dgx2"), 1),
                    ],
                    vec![],
                ),
                // lan-only nodes (no ib_ip), plus a fan extra
                cluster(
                    "solo",
                    vec![node("dgx10.lan", None, Some("switch.dgx10"), 0)],
                    vec![crate::config::ExtraEntity::Detailed {
                        entity: "switch.fan".into(),
                        blocks: false,
                    }],
                ),
            ],
        };
        let t = derive_topo(&config);
        // slots: dgx1 < dgx2 < dgx10 (numeric, not lexicographic)
        assert_eq!(t.nodes, vec!["dgx1.lan", "dgx2.lan", "dgx10.lan"]);
        // 200g: every ib node of the fabric cluster → the blocking switch
        assert_eq!(t.links.iter().filter(|l| l.kind == "200g").count(), 2);
        assert!(
            t.links
                .iter()
                .any(|l| l.kind == "200g" && l.a == "switch.sw6" && l.b == "dgx1.lan")
        );
        // dac chain between the ranks (dgx1 → dgx2)
        assert!(
            t.links
                .iter()
                .any(|l| l.kind == "dac" && l.a == "dgx1.lan" && l.b == "dgx2.lan")
        );
        // every node hangs on the passive lan
        assert_eq!(t.links.iter().filter(|l| l.kind == "10g").count(), 3);
        // switch glyphs: the fabric switch (ib, node-plug entities excluded)
        // and the fan (aux) — deduped across clusters
        assert_eq!(
            t.switches,
            vec![
                TopoSwitch {
                    entity: "switch.sw6".into(),
                    kind: "ib".into()
                },
                TopoSwitch {
                    entity: "switch.fan".into(),
                    kind: "aux".into()
                }
            ]
        );
    }
}