Compare commits

...
Author SHA1 Message Date
atlas
c8bd0f7180 test+docs(#2448): multi-agent restart test + doc shapes
- multi_agent_restart_is_one_dag_with_concurrent_per_agent_subgraphs:
  two agents restart in one DAG; both SetWanted heads are roots, each
  acquiring its own agent's lease concurrently. Existing single-agent
  shape/lease tests updated for the templates::restart(&[..], graceful)
  signature.
- coordinator.md + templates.rs module doc: restart takes an agent list;
  a hive-wide restart is one DAG with N per-agent subgraphs.
2026-07-14 23:11:02 +02:00
atlas
1739716fa2 feat(#2448): emit one multi-agent DAG for hivectl restart / restart-all
A hive-wide restart was N separate single-agent DAGs (one submit::restart
per agent). Now that agent is per-node (#2445), make it ONE DAG with a
per-agent restart subgraph each.

- templates::restart takes an agent list: each agent gets an independent
  subgraph (a head SetWanted(Up) root, then its restart chain), so the N
  subgraphs run concurrently on their own leases. One agent = the ordinary
  single-agent restart; unifies the old restart + graceful_restart fns.
- submit::restart / graceful_restart stay as single-agent wrappers over
  the new submit::restart_many(agents, graceful).
- server.rs handle_restart_all + handle_restart_scoped submit one
  restart_many call instead of looping per agent. Infra containers
  unchanged (no lease/DAG, synchronous).

Scope: restart + restart-all only. Broad stop+start is the same pattern
(stop/start templates take agent lists) — a follow-up increment.
2026-07-14 23:11:02 +02:00
5 changed files with 180 additions and 108 deletions

View file

@ -76,12 +76,15 @@ container build:
The power ops write the durable `wanted` intent via a head `SetWanted`
node (not a pre-submit side effect) — it holds the agent lease, so
intent-write + reconcile is atomic per-agent.
intent-write + reconcile is atomic per-agent. `restart` takes an agent
*list*: a hive-wide `hivectl restart` / `restart-all` is ONE DAG with a
per-agent restart subgraph each (independent roots, run concurrently on
their own leases), not N separate DAGs.
```text
rebuild(a): Prebuild(a) → StopForUpdate(a) → Swap(a) →(after-any) Reconcile(a)
graceful-stop(a): SetWanted(a,Off) → Signal(a) → Drain(a) → Reconcile(a)
restart(a): SetWanted(a,Up) → StopForUpdate(a) → Reconcile(a)
restart(a..): per agent: SetWanted(a,Up) → StopForUpdate(a) → Reconcile(a) (N subgraphs, 1 DAG)
graceful-restart(a): SetWanted(a,Up) → Signal(a) → Drain(a) → StopForUpdate(a) → Reconcile(a)
start(a): SetWanted(a,Up) → Reconcile(a) (stale rev ⇒ stale-start below)
stale-start(a): SetWanted(a,Up) → «rebuild subgraph» (prebuild noops — agent is down)

View file

@ -27,12 +27,27 @@ pub fn rebuild(coord: &Arc<Coordinator>, agent: &str, source: Source, reason: St
submit_and_emit(coord, templates::rebuild(agent, source, reason, None, true))
}
/// Restart: mechanical stop + converge to `wanted = Up`. The intent
/// write is the template's head `SetWanted(Up)` node — it matters when
/// `wanted` drifted `Offline` under a running agent (an operator asking
/// for a restart plainly wants it running, not a stop).
/// Restart a single agent: mechanical stop + converge to `wanted = Up`.
/// The intent write is the template's head `SetWanted(Up)` node — it
/// matters when `wanted` drifted `Offline` under a running agent (an
/// operator asking for a restart plainly wants it running, not a stop).
/// Thin wrapper over [`restart_many`] with a one-agent slice.
pub fn restart(coord: &Arc<Coordinator>, agent: &str, source: Source, reason: String) -> u64 {
submit_and_emit(coord, templates::restart(agent, source, reason))
restart_many(coord, &[agent.to_owned()], false, source, reason)
}
/// Restart `agents` (one or many) in a **single** DAG — one per-agent
/// subgraph each, running concurrently on their own leases. `graceful`
/// prepends the signal→drain quiesce per agent. The whole hive-wide
/// `hivectl restart` / `restart-all` is now one DAG instead of N.
pub fn restart_many(
coord: &Arc<Coordinator>,
agents: &[String],
graceful: bool,
source: Source,
reason: String,
) -> u64 {
submit_and_emit(coord, templates::restart(agents, graceful, source, reason))
}
/// Start: `SetWanted(Up)` (a DAG node now) then reconcile. A stale rev
@ -67,17 +82,18 @@ pub fn graceful_stop(coord: &Arc<Coordinator>, agent: &str, source: Source, reas
submit_and_emit(coord, templates::graceful_stop(agent, source, reason))
}
/// Graceful restart: signal → drain → mechanical stop → reconcile (starts
/// it back up) — one atomic DAG, no client-side "await the stop DAG then
/// submit a start DAG" split. The head `SetWanted(Up)` node writes the
/// intent as part of the DAG.
/// Graceful restart of a single agent: signal → drain → mechanical stop →
/// reconcile (starts it back up) — one atomic DAG, no client-side "await
/// the stop DAG then submit a start DAG" split. The head `SetWanted(Up)`
/// node writes the intent as part of the DAG. Thin wrapper over
/// [`restart_many`] with `graceful = true`.
pub fn graceful_restart(
coord: &Arc<Coordinator>,
agent: &str,
source: Source,
reason: String,
) -> u64 {
submit_and_emit(coord, templates::graceful_restart(agent, source, reason))
restart_many(coord, &[agent.to_owned()], true, source, reason)
}
/// Perm change: commit the JSON file(s) then rebuild.

View file

@ -4,9 +4,11 @@
//! `Vec<Node>` + `deps`).
//!
//! Every node carries its own `agent` (there is no DAG-level agent) — the
//! `node` helper stamps the template's agent onto each. Today's templates
//! are single-agent (every node shares one agent); a future multi-agent
//! template would stamp different agents per subgraph.
//! `node` helper stamps the template's agent onto each. `restart` takes an
//! agent *list* and stamps each agent onto its own subgraph, so a hive-wide
//! `hivectl restart` is ONE DAG with N independent per-agent subgraphs
//! (each a root chain, run concurrently on its own lease) rather than N
//! separate DAGs. The other templates are still single-agent.
//!
//! The power ops write the durable `wanted` intent via a head
//! `SetWanted(w)` node (not a pre-submit side effect); it holds the agent
@ -146,37 +148,36 @@ pub fn graceful_stop(agent: &str, source: Source, reason: String) -> DagSpec {
}
}
/// Restart: write `wanted = Up` (head `SetWanted` node), mechanical stop,
/// then converge — a stop + start like the old `lifecycle::restart`
/// regardless of prior intent drift.
pub fn restart(agent: &str, source: Source, reason: String) -> DagSpec {
DagSpec {
template: Template::Restart,
source,
reason,
parent_id: None,
approval_id: None,
inputs: Vec::new(),
perm_payload: None,
transient: Some(TransientKind::Restarting),
nodes: vec![
node(agent, NodeKind::SetWanted { up: true }, Vec::new()),
node(agent, NodeKind::StopForUpdate, after_ok(0)),
node(agent, NodeKind::Reconcile, after_ok(1)),
],
/// Restart one or more agents in a **single** DAG. Each agent gets an
/// independent subgraph — a head `SetWanted(Up)` (a root: no cross-agent
/// dep) then its restart chain — so all agents' restarts run concurrently,
/// each taking its own agent lease. `graceful` inserts `Signal → Drain`
/// before the mechanical `StopForUpdate` (each agent's subgraph is still one
/// atomic restart). One agent = the ordinary single-agent restart; many =
/// a hive-wide `hivectl restart` as one DAG instead of N separate ones.
pub fn restart(agents: &[String], graceful: bool, source: Source, reason: String) -> DagSpec {
let mut nodes = Vec::new();
for agent in agents {
let base = u32::try_from(nodes.len()).unwrap_or(u32::MAX);
// Head of this agent's subgraph — a root (empty deps), so the N
// per-agent subgraphs are independent and run concurrently.
nodes.push(node(agent, NodeKind::SetWanted { up: true }, Vec::new()));
if graceful {
nodes.push(node(agent, NodeKind::Signal, after_ok(base)));
nodes.push(node(agent, NodeKind::Drain, after_ok(base + 1)));
nodes.push(node(agent, NodeKind::StopForUpdate, after_ok(base + 2)));
nodes.push(node(agent, NodeKind::Reconcile, after_ok(base + 3)));
} else {
nodes.push(node(agent, NodeKind::StopForUpdate, after_ok(base)));
nodes.push(node(agent, NodeKind::Reconcile, after_ok(base + 1)));
}
}
}
/// Graceful restart: write `wanted = Up` (head `SetWanted`), signal →
/// drain → mechanical stop → converge. One atomic DAG start to finish
/// (no client- or server-side "submit one DAG, await it, submit the next"
/// composition): the `Drain` node is the same bounded harness-checkpoint
/// wait `graceful_stop` uses, then `StopForUpdate` (mechanical, ignores
/// `wanted`) and the tail `Reconcile` (converges to `wanted = Up`, i.e.
/// starts it back up) chain exactly like `restart`'s tail.
pub fn graceful_restart(agent: &str, source: Source, reason: String) -> DagSpec {
DagSpec {
template: Template::GracefulRestart,
template: if graceful {
Template::GracefulRestart
} else {
Template::Restart
},
source,
reason,
parent_id: None,
@ -184,13 +185,7 @@ pub fn graceful_restart(agent: &str, source: Source, reason: String) -> DagSpec
inputs: Vec::new(),
perm_payload: None,
transient: Some(TransientKind::Restarting),
nodes: vec![
node(agent, NodeKind::SetWanted { up: true }, Vec::new()),
node(agent, NodeKind::Signal, after_ok(0)),
node(agent, NodeKind::Drain, after_ok(1)),
node(agent, NodeKind::StopForUpdate, after_ok(2)),
node(agent, NodeKind::Reconcile, after_ok(3)),
],
nodes,
}
}

View file

@ -67,7 +67,12 @@ fn distinct_submits_never_collapse() {
let b = submit(&q, rebuild("agent-b", "r"));
let c = submit(
&q,
templates::restart("agent-a", Source::Manual, "r".to_owned()),
templates::restart(
&["agent-a".to_owned()],
false,
Source::Manual,
"r".to_owned(),
),
);
assert_ne!(a, b);
assert_ne!(a, c);
@ -200,7 +205,12 @@ fn lease_serializes_two_lifecycle_dags_for_same_agent() {
let q = JobQueue::new(4);
let restart = submit(
&q,
templates::restart("agent-a", Source::Manual, "restart".to_owned()),
templates::restart(
&["agent-a".to_owned()],
false,
Source::Manual,
"restart".to_owned(),
),
);
let stop = submit(
&q,
@ -284,16 +294,60 @@ fn agents_do_not_contend_on_each_others_leases() {
let q = JobQueue::new(4);
submit(
&q,
templates::restart("agent-a", Source::Manual, "r".to_owned()),
templates::restart(
&["agent-a".to_owned()],
false,
Source::Manual,
"r".to_owned(),
),
);
submit(
&q,
templates::restart("agent-b", Source::Manual, "r".to_owned()),
templates::restart(
&["agent-b".to_owned()],
false,
Source::Manual,
"r".to_owned(),
),
);
let claims = q.claim_ready();
assert_eq!(claims.len(), 2, "different agents run concurrently");
}
#[test]
fn multi_agent_restart_is_one_dag_with_concurrent_per_agent_subgraphs() {
let q = JobQueue::new(4);
let id = submit(
&q,
templates::restart(
&["agent-a".to_owned(), "agent-b".to_owned()],
false,
Source::Manual,
"hive-wide".to_owned(),
),
);
// A hive-wide restart is ONE DAG, not one-per-agent.
assert_eq!(q.snapshot().len(), 1);
// Each agent's subgraph head (SetWanted) is a root, so both are
// claimable at once — each takes its OWN agent's lease (no contention
// across distinct agents), all inside the single DAG.
let claims = q.claim_ready();
assert!(claims.iter().all(|c| c.dag_id == id));
let mut heads: Vec<(&str, &str, bool)> = claims
.iter()
.map(|c| (c.agent.as_str(), c.kind.as_str(), c.lease_acquired))
.collect();
heads.sort_unstable();
assert_eq!(
heads,
vec![
("agent-a", "set_wanted", true),
("agent-b", "set_wanted", true),
],
"both per-agent subgraphs start concurrently, each acquiring its own lease"
);
}
// ---- failure: cancel-downstream + AfterAny ----
#[test]
@ -514,7 +568,12 @@ fn terminal_dag_reported_exactly_once_and_lease_released() {
let q = JobQueue::new(1);
let id = submit(
&q,
templates::restart("agent-a", Source::Manual, "r".to_owned()),
templates::restart(
&["agent-a".to_owned()],
false,
Source::Manual,
"r".to_owned(),
),
);
// restart = SetWanted → StopForUpdate → Reconcile; not terminal until
// the last node completes.
@ -622,7 +681,12 @@ fn trim_keeps_terminal_parent_with_live_children() {
);
let lock = claim_one(&q);
q.complete_node(meta, lock.node_id, Ok(()));
let mut child_spec = templates::restart("agent-x", Source::MetaUpdate, "cascade".to_owned());
let mut child_spec = templates::restart(
&["agent-x".to_owned()],
false,
Source::MetaUpdate,
"cascade".to_owned(),
);
child_spec.parent_id = Some(meta);
let child = submit(&q, child_spec);
// Churn > MAX_HISTORY_PER_TEMPLATE terminal meta_update DAGs.

View file

@ -552,28 +552,29 @@ fn submit_single(coord: &Arc<Coordinator>, name: &str, verb: Verb) -> HostRespon
HostResponse::queued(vec![id])
}
/// Restart every container by submitting one restart DAG per agent —
/// each serializes on its own lease, so unrelated agents' restarts
/// overlap while nothing races an in-flight rebuild. Returns once all
/// are queued; per-agent results surface on the queue.
/// Restart every container in **one** DAG — a per-agent restart subgraph
/// each, running concurrently on their own leases (so unrelated agents'
/// restarts overlap while nothing races an in-flight rebuild). Returns
/// once queued; per-node progress surfaces on the single DAG.
async fn handle_restart_all(coord: &Arc<Coordinator>) -> Result<HostResponse> {
tracing::info!("restart-all");
let agents = lifecycle::list().await?;
let mut ok_agents: Vec<String> = Vec::new();
let mut queued: Vec<u64> = Vec::new();
for agent in &agents {
let Some(logical) = agent.strip_prefix(lifecycle::AGENT_PREFIX) else {
continue;
};
queued.push(crate::job_queue::submit::restart(
let containers = lifecycle::list().await?;
let agents: Vec<String> = containers
.iter()
.filter_map(|a| a.strip_prefix(lifecycle::AGENT_PREFIX).map(str::to_owned))
.collect();
let queued = if agents.is_empty() {
Vec::new()
} else {
vec![crate::job_queue::submit::restart_many(
coord,
logical,
&agents,
false,
crate::job_queue::Source::Manual,
"manual restart via hivectl restart-all".to_owned(),
));
ok_agents.push(logical.to_owned());
}
let mut resp = HostResponse::list(ok_agents);
)]
};
let mut resp = HostResponse::list(agents);
resp.queued_dags = Some(queued);
Ok(resp)
}
@ -718,16 +719,14 @@ async fn handle_start(
}
/// Restart containers hive-wide (`hivectl restart`) — the DAG-based
/// sibling of [`handle_stop`]/[`handle_start`], replacing the old
/// client-side stop-then-start composition (issue tracker "dagify hivectl
/// commands"). Each targeted agent gets exactly one atomic DAG submitted
/// up front: the `Restart` template (mechanical stop + reconcile) in the
/// common case, or `GracefulRestart` (signal → drain → mechanical stop →
/// reconcile) with `graceful` set — no "submit a DAG, wait for it, submit
/// another" composition on either path, so a dropped `hivectl` connection
/// never strands an agent, the same way `handle_restart_all` already
/// avoids it. Infra containers have no lease/DAG and restart
/// synchronously (stop then start).
/// sibling of [`handle_stop`]/[`handle_start`]. All targeted agents ride
/// **one** DAG (a per-agent restart subgraph each: `SetWanted → [Signal →
/// Drain →] StopForUpdate → Reconcile`, independent roots that run
/// concurrently on their own leases), not N separate DAGs — a hive-wide
/// restart is one job. `graceful` prepends signal→drain per agent. No
/// client-side stop-then-start composition, so a dropped `hivectl`
/// connection never strands an agent. Infra containers have no lease/DAG
/// and restart synchronously (stop then start).
async fn handle_restart_scoped(
coord: &Arc<Coordinator>,
scope: &LifecycleScope,
@ -740,30 +739,25 @@ async fn handle_restart_scoped(
let mut errors: Vec<String> = Vec::new();
let mut queued: Vec<u64> = Vec::new();
// One atomic DAG per agent, submitted up front — `Restart`
// (mechanical stop + reconcile) or, with `--graceful`,
// `GracefulRestart` (signal → drain → mechanical stop → reconcile).
// No client- or server-side "submit a stop DAG, await it, then
// submit a start DAG" composition: that window is exactly the
// dropped-connection gap this DAG-based path exists to close.
for agent in &agents {
let id = if graceful {
crate::job_queue::submit::graceful_restart(
coord,
agent,
crate::job_queue::Source::Manual,
"manual via hivectl restart --graceful".to_owned(),
)
} else {
crate::job_queue::submit::restart(
coord,
agent,
crate::job_queue::Source::Manual,
"manual restart via hivectl restart".to_owned(),
)
};
queued.push(id);
ok_items.push(agent.clone());
// One DAG for all targeted agents — a per-agent restart subgraph each
// (`SetWanted → [Signal → Drain →] StopForUpdate → Reconcile`),
// independent roots that run concurrently on their own leases. A
// hive-wide `hivectl restart` is now a single DAG, not N. No
// client-side stop-then-start composition — the whole restart survives
// a dropped connection because the DAG owns it.
if !agents.is_empty() {
queued.push(crate::job_queue::submit::restart_many(
coord,
&agents,
graceful,
crate::job_queue::Source::Manual,
if graceful {
"manual via hivectl restart --graceful".to_owned()
} else {
"manual restart via hivectl restart".to_owned()
},
));
ok_items.extend(agents.iter().cloned());
}
for &container in &infra {