Compare commits

...
Author SHA1 Message Date
atlas
4545dd312e feat(#2450): collapse the startup sweep into one inline DAG
The boot is now ONE DAG, assembled inline in submit_boot_tree — no boot_root
Noop anchor, no per-agent child DAGs, no display-only parent_id grouping: a
sweep MetaLock root (only when something is stale) that grows one rebuild
subgraph per stale agent into the same DAG (via append_subgraph, previous
commit), plus one Reconcile root per drifted agent (independent — a boot
reconcile needs no lock bump).

- submit_boot_tree builds the DagSpec inline; removed the single-use
  templates::{boot_root, startup_sweep} builders (inlined per the operator's
  "don't force single-use shapes into templates.rs" steer).
- fanout_specs simplified to the meta-update cascade path only — the startup
  sweep no longer fans out child DAGs, so its branch was dead.
- test: append_subgraph_roots_on_emitter_and_rebases_local_deps.

Vestigial after this (deliberately left as follow-ups, flagged in the PR):
NodeKind::Noop is now unconstructed (contained to hive-c0re, removable);
Template::StartupSweep is unconstructed but a hive-sh4re wire type
(frontend-coordinated removal).
2026-07-15 18:24:30 +02:00
atlas
b6defdeaaf feat(#2450): append_subgraph primitive + sweep MetaLock grows rebuilds in-DAG
First half of making the startup sweep one DAG. Adds the runtime
subgraph-append machinery and switches the sweep MetaLock from fanning out
child Rebuild DAGs to growing one rebuild subgraph per stale agent into the
same DAG:

- JobQueue::append_subgraph(dag_id, nodes, dep_on) — the multi-node,
  multi-agent generalisation of append_node: rebases a subgraph's local deps
  onto the DAG's id space and roots it on the emitting node.
- NodeOutput.append_subgraph: Vec<Vec<NodeSpec>> — the executor→scheduler
  channel for it; scheduler drains it before completing the emitting node
  (same ordering as append_nodes).
- run_meta_lock sweep branch returns the stale agents' rebuild_nodes
  subgraphs via append_subgraph instead of fanout.

Follow-up commit collapses submit_boot_tree (drop boot_root + per-agent
reconcile child DAGs) so the whole boot is one DAG built inline.
2026-07-15 18:24:30 +02:00
6 changed files with 229 additions and 114 deletions

View file

@ -10,7 +10,7 @@ use std::sync::Arc;
use anyhow::{Context as _, Result};
use super::model::{NodeKind, State, Template};
use super::model::{NodeKind, NodeSpec, State, Template};
use super::{Claim, TerminalDag};
use crate::coordinator::Coordinator;
use crate::power::{ReconcileAction, reconcile_action};
@ -38,6 +38,16 @@ pub struct NodeOutput {
/// these *before* the emitting node's completion so the DAG never
/// rolls terminal with the appended work still pending.
pub append_nodes: Vec<NodeKind>,
/// Whole per-agent *subgraphs* to append into *this same* DAG at
/// runtime — the multi-node generalisation of `append_nodes`. Each
/// inner `Vec<NodeSpec>` is one independent subgraph whose `deps` are
/// local (0-based within that subgraph); the scheduler appends each via
/// [`JobQueue::append_subgraph`], which rebases the deps onto the DAG's
/// node-id space and roots the subgraph on the emitting node. The
/// startup sweep's `MetaLock` uses this to grow one stale-agent rebuild
/// subgraph per agent into the same boot DAG instead of fanning out
/// child DAGs. Same before-completion ordering as `append_nodes`.
pub append_subgraph: Vec<Vec<NodeSpec>>,
}
/// Step-label + build-log sink for one claimed node.
@ -276,8 +286,17 @@ async fn run_meta_lock(
if let Err(e) = crate::meta::lock_update_hyperhive().await {
tracing::warn!(error = ?e, "startup sweep: meta lock_update_hyperhive failed");
}
// Grow one rebuild subgraph per stale agent into *this* boot DAG
// (rooted on this `MetaLock`, so they build against the post-bump
// lock), rather than fanning out child DAGs. `relock = true` — a
// boot sweep relocks per-agent like a manual rebuild.
let append_subgraph = fanout
.unwrap_or_default()
.iter()
.map(|agent| super::templates::rebuild_nodes(agent, true, 0))
.collect();
return Ok(NodeOutput {
fanout: fanout.unwrap_or_default(),
append_subgraph,
..Default::default()
});
}

View file

@ -34,7 +34,8 @@ use hive_sh4re::wire_time::now_unix;
use tokio::sync::Notify;
pub use model::{
Dag, DagSpec, DagView, DepWhen, Node, NodeId, NodeKind, PermPayload, Source, State, Template,
Dag, DagSpec, DagView, DepWhen, Node, NodeId, NodeKind, NodeSpec, PermPayload, Source, State,
Template,
};
/// How many terminal DAGs (`Done` / `Failed` / `Cancelled`) to retain
@ -206,6 +207,71 @@ impl JobQueue {
Some(new_id)
}
/// Append a whole *subgraph* into a live (non-terminal) DAG at runtime
/// — the multi-node, multi-agent generalisation of [`Self::append_node`].
/// Each [`NodeSpec`] carries its own `agent` and subgraph-relative `deps`
/// (indices into `nodes`); this rebases those onto the DAG's node-id
/// space (`id == index`, an invariant `append_node` also maintains) and
/// attaches every subgraph *root* — a node with no internal deps — to
/// `dep_on` with an `AfterOk` edge. Used by the startup sweep's
/// `MetaLock` to grow per-agent rebuild subgraphs into the same boot DAG
/// instead of fanning out child DAGs. Same call-*before*-`complete_node`
/// contract as `append_node` (so the DAG can't roll terminal with the
/// appended work still pending). Returns the new node ids; empty if the
/// DAG is gone or `nodes` is empty.
pub fn append_subgraph(
&self,
dag_id: u64,
nodes: Vec<NodeSpec>,
dep_on: NodeId,
) -> Vec<NodeId> {
if nodes.is_empty() {
return Vec::new();
}
let mut inner = self.inner.lock().expect("job_queue mutex poisoned");
let Some(dag) = inner.dags.iter_mut().find(|d| d.id == dag_id) else {
return Vec::new();
};
let base: NodeId = u32::try_from(dag.nodes.len()).unwrap_or(u32::MAX);
let mut new_ids = Vec::with_capacity(nodes.len());
for (i, spec) in nodes.into_iter().enumerate() {
let new_id: NodeId = base + u32::try_from(i).unwrap_or(u32::MAX);
// Subgraph roots (no internal deps) hang off the emitting node;
// internal deps rebase from subgraph-relative onto the DAG id
// space (both start at `base`).
let deps = if spec.deps.is_empty() {
vec![model::Dep {
on: dep_on,
when: DepWhen::AfterOk,
}]
} else {
spec.deps
.into_iter()
.map(|d| model::Dep {
on: base + d.on,
when: d.when,
})
.collect()
};
dag.nodes.push(Node {
id: new_id,
agent: spec.agent,
kind: spec.kind,
deps,
state: State::Queued,
step: None,
build_log_id: None,
started_at: None,
finished_at: None,
error: None,
});
new_ids.push(new_id);
}
drop(inner);
self.notify.notify_one();
new_ids
}
fn push_dag(inner: &mut Inner, spec: DagSpec) -> u64 {
inner.next_id += 1;
let id = inner.next_id;

View file

@ -17,7 +17,7 @@ use std::collections::HashMap;
use std::sync::Arc;
use super::exec::{self, NodeOutput};
use super::{Claim, Source, Template, templates};
use super::{Claim, Source, templates};
use crate::coordinator::Coordinator;
struct NodeDone {
@ -116,6 +116,15 @@ async fn handle_completion(
.job_queue
.append_node(claim.dag_id, kind, claim.node_id);
}
// Same before-completion ordering as `append_nodes`, but for
// whole per-agent subgraphs (the startup sweep's rebuild
// subgraphs growing into the boot DAG) — each an independent
// subgraph rooted on this node.
for subgraph in output.append_subgraph {
coord
.job_queue
.append_subgraph(claim.dag_id, subgraph, claim.node_id);
}
coord
.job_queue
.complete_node(claim.dag_id, claim.node_id, Ok(()));
@ -157,26 +166,29 @@ async fn process_terminals(
}
}
/// Child `Rebuild` specs for a completed `MetaLock` fan-out, grouped
/// under the parent via `parent_id`. Meta-update children skip the
/// per-agent relock (it would revert the bump the parent just
/// committed); sweep children relock like a manual rebuild.
/// Child `Rebuild` specs for a completed meta-update `MetaLock` fan-out,
/// grouped under the parent via `parent_id`. Meta-update children skip the
/// per-agent relock (`relock = false`) — it would revert the bump the parent
/// just committed. This is now the meta-update cascade path only: the startup
/// sweep no longer fans out child DAGs — it grows one rebuild subgraph per
/// stale agent into its own DAG via `append_subgraph` (see
/// `exec::run_meta_lock`).
fn fanout_specs(claim: &Claim, agents: Vec<String>) -> Vec<super::DagSpec> {
let sweep = claim.template == Template::StartupSweep;
let (source, relock) = if sweep {
(Source::StartupSweep, true)
} else {
(Source::MetaUpdate, false)
};
let reason = if sweep {
"startup sweep".to_owned()
} else if let Some(approval_id) = claim.approval_id {
let reason = if let Some(approval_id) = claim.approval_id {
format!("approval #{approval_id} meta input cascade")
} else {
"meta-update cascade".to_owned()
};
agents
.into_iter()
.map(|agent| templates::rebuild(&agent, source, reason.clone(), Some(claim.dag_id), relock))
.map(|agent| {
templates::rebuild(
&agent,
Source::MetaUpdate,
reason.clone(),
Some(claim.dag_id),
false,
)
})
.collect()
}

View file

@ -216,49 +216,11 @@ pub fn meta_update(
}
}
/// Boot-time root anchor DAG: a single [`NodeKind::Noop`] node that groups
/// this boot's `StartupSweep` + per-agent `Reconcile` child DAGs (linked via
/// `parent_id`) into one tree so the dashboard renders the boot as one entry.
/// Holds no lease and does no work — the children it anchors still run
/// concurrently. `auto_update::run` submits it first (when there's any boot
/// work), then parents the sweep + reconciles onto its id.
pub fn boot_root(reason: String) -> DagSpec {
DagSpec {
template: Template::Boot,
source: Source::AutoUpdate,
reason,
parent_id: None,
approval_id: None,
inputs: Vec::new(),
perm_payload: None,
transient: None,
nodes: vec![node("hyperhive", NodeKind::Noop, Vec::new())],
}
}
/// Boot-time sweep parent: bump meta's hyperhive input (non-fatal),
/// then fan out `Rebuild` children for the precomputed stale agent
/// list (topology-sorted by the caller).
pub fn startup_sweep(reason: String, stale_agents: Vec<String>) -> DagSpec {
DagSpec {
template: Template::StartupSweep,
source: Source::AutoUpdate,
reason,
parent_id: None,
approval_id: None,
inputs: Vec::new(),
perm_payload: None,
transient: None,
nodes: vec![node(
"hyperhive",
NodeKind::MetaLock {
sweep: true,
fanout: Some(stale_agents),
},
Vec::new(),
)],
}
}
// The boot is now assembled inline in `workers/auto_update.rs::submit_boot_tree`
// as ONE DAG (a sweep `MetaLock` root that grows rebuild subgraphs in-DAG, plus
// a `Reconcile` root per drifted agent) — no `boot_root` Noop anchor, no
// `startup_sweep` parent template, no per-agent child DAGs. The old single-use
// `boot_root` / `startup_sweep` builders were inlined there and removed.
/// Validate a spec before it enters the queue: node ids are dense
/// (index = id), deps reference existing nodes, and the dep graph is

View file

@ -450,6 +450,58 @@ fn offline_agents_skip_mechanical_nodes_but_keep_reconcile() {
);
}
#[test]
fn append_subgraph_roots_on_emitter_and_rebases_local_deps() {
// The startup-sweep mechanism: a `MetaLock` emitter grows one rebuild
// subgraph per stale agent into its OWN DAG. Each subgraph is rooted on
// the emitter and its LOCAL 0-based deps are rebased onto the DAG.
let q = JobQueue::new(4);
let spec = DagSpec {
template: Template::Boot,
source: Source::AutoUpdate,
reason: "sweep".to_owned(),
parent_id: None,
approval_id: None,
inputs: Vec::new(),
perm_payload: None,
transient: None,
nodes: vec![NodeSpec {
agent: "hyperhive".to_owned(),
kind: NodeKind::MetaLock {
sweep: true,
fanout: None,
},
deps: Vec::new(),
}],
};
let id = submit(&q, spec);
let emitter = claim_one(&q);
assert_eq!(emitter.kind.as_str(), "meta_lock");
// Two independent per-agent subgraphs — the REAL production shape the
// sweep MetaLock grows (`rebuild_nodes(_, true, 0)`: root Prebuild →
// StopForUpdate → Swap → Reconcile, local 0-based deps), so this test
// tracks any drift in that builder's root-first (`base = 0`) shape.
let subgraph = |agent: &str| templates::rebuild_nodes(agent, true, 0);
// Must append BEFORE completing the emitter (the documented contract).
q.append_subgraph(id, subgraph("a"), emitter.node_id);
q.append_subgraph(id, subgraph("b"), emitter.node_id);
q.complete_node(id, emitter.node_id, Ok(()));
// Still ONE DAG; both subgraph roots become ready once the emitter is
// Done (rooted on it), each on its own agent lease.
assert_eq!(q.snapshot().len(), 1);
let next = q.claim_ready();
let mut kinds: Vec<(&str, &str)> = next
.iter()
.map(|c| (c.agent.as_str(), c.kind.as_str()))
.collect();
kinds.sort_unstable();
assert_eq!(
kinds,
vec![("a", "prebuild"), ("b", "prebuild")],
"both rebuild subgraphs root on the emitter and run concurrently in one DAG"
);
}
// ---- failure: cancel-downstream + AfterAny ----
#[test]

View file

@ -288,12 +288,14 @@ pub async fn run(coord: Arc<Coordinator>) -> Result<()> {
Ok(())
}
/// Submit this boot's DAGs under one `Boot` root: a `Noop` anchor with the
/// startup sweep + per-agent reconciles hung off it via `parent_id`, so the
/// dashboard renders the boot as a single tree instead of N+1 rows. No-op when
/// there's nothing to do. The `parent_id` link is a display grouping, not a
/// dependency edge — the children run concurrently, so the reconciles never
/// wait behind the lock bump.
/// Submit this boot's work as **one DAG** (no `boot_root` Noop anchor, no
/// per-agent child DAGs). Node 0 is the sweep `MetaLock` (only when
/// something is stale) — its executor bumps the hyperhive lock, then grows
/// one rebuild subgraph per stale agent into *this same* DAG (rooted on the
/// `MetaLock`, so they build against the post-bump lock; see
/// `exec::run_meta_lock`). Every drifted agent gets a boot `Reconcile` as an
/// independent root — a boot reconcile needs no lock bump, so it converges
/// concurrently with the sweep. No-op when there's nothing to do.
fn submit_boot_tree(
coord: &Arc<Coordinator>,
any_stale: bool,
@ -302,59 +304,61 @@ fn submit_boot_tree(
n_deferred: usize,
n_skipped: usize,
) {
// Only emit a boot root when there's actually boot work — a fully-quiet
// boot (nothing stale, nothing drifted) submits nothing, exactly as before.
let boot_root_id = if any_stale || !drifted.is_empty() {
let reason = format!(
"boot: {} rebuild(s), {} reconcile(s), {} deferred (offline), {} up-to-date",
fanout.len(),
drifted.len(),
n_deferred,
n_skipped,
);
match coord
.job_queue
.submit(crate::job_queue::templates::boot_root(reason))
{
Ok(id) => Some(id),
Err(e) => {
tracing::warn!(error = ?e, "boot reconcile: boot-root submit failed");
None
}
}
} else {
None
};
use crate::job_queue::{DagSpec, NodeKind, NodeSpec, Source, Template};
// Fully-quiet boot (nothing stale, nothing drifted) submits nothing.
if !any_stale && drifted.is_empty() {
return;
}
let reason = format!(
"boot: {} rebuild(s), {} reconcile(s), {} deferred (offline), {} up-to-date",
fanout.len(),
drifted.len(),
n_deferred,
n_skipped,
);
let mut nodes: Vec<NodeSpec> = Vec::new();
// Sweep whenever ANY marker is stale — even when every stale agent is
// wanted-offline: the hyperhive lock bump must land now so their later
// start-upgrade rebuilds build against it. No stale agents ⇒ no sweep ⇒ no
// meta commit on a no-change boot.
// start-upgrade rebuilds build against it. No stale agents ⇒ no MetaLock
// ⇒ no meta commit on a no-change boot. The `fanout` list rides the
// MetaLock into `run_meta_lock`, which appends the rebuild subgraphs.
if any_stale {
let reason = format!(
"startup sweep: {} rebuild(s), {} deferred (offline), {} up-to-date",
fanout.len(),
n_deferred,
n_skipped,
);
let mut spec = crate::job_queue::templates::startup_sweep(reason, fanout);
spec.parent_id = boot_root_id;
if let Err(e) = coord.job_queue.submit(spec) {
tracing::warn!(error = ?e, "boot reconcile: sweep submit failed");
}
nodes.push(NodeSpec {
agent: "hyperhive".to_owned(),
kind: NodeKind::MetaLock {
sweep: true,
fanout: Some(fanout),
},
deps: Vec::new(),
});
}
// One boot Reconcile per drifted agent — independent roots.
for name in drifted {
let mut spec = crate::job_queue::templates::reconcile_only(
crate::job_queue::Template::Reconcile,
&name,
crate::job_queue::Source::AutoUpdate,
"boot reconcile".to_owned(),
None,
);
spec.parent_id = boot_root_id;
if let Err(e) = coord.job_queue.submit(spec) {
tracing::warn!(%name, error = ?e, "boot reconcile: submit failed");
}
nodes.push(NodeSpec {
agent: name,
kind: NodeKind::Reconcile,
deps: Vec::new(),
});
}
let spec = DagSpec {
template: Template::Boot,
source: Source::AutoUpdate,
reason,
parent_id: None,
approval_id: None,
inputs: Vec::new(),
perm_payload: None,
// Rebuilding when the sweep will grow rebuild subgraphs (per-agent
// crash-watch suppression during their Swap, applied at claim time);
// a reconcile-only boot needs no transient.
transient: any_stale.then_some(crate::coordinator::TransientKind::Rebuilding),
nodes,
};
if let Err(e) = coord.job_queue.submit(spec) {
tracing::warn!(error = ?e, "boot: sweep DAG submit failed");
}
coord.emit_rebuild_queue_snapshot();
}