swarm-controller: let creating an agent reuse a destroyed name

Review caught a path the new declaration node breaks: an operator may
already declare an agent `Destroyed` over the per-agent state endpoint,
and `apply` then refuses any transition off that state. Before the
wanted-state node existed the refusal was inert at creation time, but
now creating an agent under a previously-destroyed name builds its
identity, repo and config, fails the declaration, and silently cancels
the deploy — while the caller sees a 200 and a job id.

`AgentState::Destroyed`'s own doc comment already sanctions this case
("no state that brings a destroyed agent back short of a fresh deploy");
nothing implemented it. Give `apply` an `Intent`, keep the refusal for
redeclares, and add `WantedWriter::create` for the one caller that is a
fresh deploy. A separate method rather than a parameter on `set`, so no
other caller can reach the override by passing an argument wrong.
This commit is contained in:
iris 2026-09-18 18:09:47 +02:00 committed by mara
commit 48aa7a1e79
2 changed files with 152 additions and 25 deletions

View file

@ -340,7 +340,12 @@ async fn declare_new_agent(
) -> hive_jobq::scheduler::Outcome {
use hive_jobq::scheduler::Outcome;
match writer.set(hive, agent, NEW_AGENT_WANTED_STATE).await {
// `create`, not `set`: a name that was destroyed earlier still carries a
// terminal entry, and recreating the agent is the fresh deploy that
// entry's own doc comment names as the way back. `set` would refuse it,
// cancelling the deploy of an agent whose identity, repo and config the
// swarm has just built.
match writer.create(hive, agent, NEW_AGENT_WANTED_STATE).await {
Ok(_) => Outcome::Done,
Err(e) => Outcome::Failed(format!("{e:#}")),
}
@ -1387,11 +1392,11 @@ fn declare_agent_job(
// unconfigured option into an agent nobody runs.
//
// `after_ok` on the declaration, though: an agent deployed without its
// pause landing first is the race this node exists to close, and the
// only way the declaration fails without a live queue is a host that has
// no queue at all — on which `TriggerDeploy` has nothing to publish to
// either, so the strict edge cancels a node that could not have
// succeeded.
// pause landing first is the race this node exists to close, so a
// declaration that did not land must not be deployed past. A host with no
// queue at all fails this node — but `TriggerDeploy` has nothing to
// publish to on such a host either, so the strict edge cancels a node
// that could not have succeeded anyway.
let _trigger_deploy = b
.node(SwarmNodeKind::TriggerDeploy {
hive: hive.to_owned(),

View file

@ -57,6 +57,28 @@ enum Wrote {
Failed(anyhow::Error),
}
/// Why a caller is declaring a state, which is what decides whether a
/// `Destroyed` entry stands in its way.
///
/// The distinction is the one [`AgentState::Destroyed`]'s own doc comment
/// already promises — "no state that brings a destroyed agent back **short of
/// a fresh deploy**". Redeclaring is not a fresh deploy and must keep being
/// refused; creating the agent again *is* one, and the whole of its identity,
/// repo and config has just been built anew alongside this declaration.
///
/// An enum rather than a bool because the two call sites read as opposites
/// only if the argument says which one it is — `apply(.., true)` at the create
/// path would be a flag nobody can check at a glance.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum Intent {
/// An operator (or anything else) changing what an existing agent should
/// be doing. Cannot move off `Destroyed`.
Redeclare,
/// The agent is being created from nothing — a fresh deploy. Replaces a
/// `Destroyed` entry left behind by an agent that had this name before.
Create,
}
/// How many times a losing writer re-reads and re-applies before giving up.
///
/// A conflict means another writer changed a *different* agent between this
@ -75,7 +97,15 @@ const MAX_ATTEMPTS: usize = 5;
/// value is an **error**, never treated as absent: overwriting a document
/// nobody can read discards the declarations of every other agent on that
/// hive, which is exactly what a fresh-start fallback would do quietly.
fn apply(current: Option<&[u8]>, agent: &str, state: AgentState) -> Result<(HiveWanted, Vec<u8>)> {
///
/// `intent` decides whether an existing `Destroyed` entry for this agent is a
/// wall or something to write over — see [`Intent`].
fn apply(
current: Option<&[u8]>,
agent: &str,
state: AgentState,
intent: Intent,
) -> Result<(HiveWanted, Vec<u8>)> {
let mut declaration = match current {
Some(raw) => serde_json::from_slice::<HiveWanted>(raw)
.context("the hive's current declaration is not decodable")?,
@ -84,8 +114,11 @@ fn apply(current: Option<&[u8]>, agent: &str, state: AgentState) -> Result<(Hive
// Redeclaring the same terminal state is a no-op, not a transition — a
// client that already sees the agent as destroyed and asks again should
// not be punished for it. Anything else moving off `Destroyed` is the
// one transition this module exists to refuse.
if let Some(existing) = declaration.agents.get(agent)
// one transition this module exists to refuse — unless the caller is
// creating the agent from nothing, which is the "fresh deploy" carve-out
// `Intent` documents.
if intent == Intent::Redeclare
&& let Some(existing) = declaration.agents.get(agent)
&& existing.state == AgentState::Destroyed
&& state != AgentState::Destroyed
{
@ -149,15 +182,47 @@ impl WantedWriter {
.with_context(|| format!("the declaration for {hive} is not decodable"))
}
/// Declare `agent` on `hive` to be in `state`, and return the whole
/// Redeclare `agent` on `hive` to be in `state`, and return the whole
/// declaration as published.
///
/// Refuses to move an agent off `Destroyed` — see
/// [`create`](Self::create) for the one caller that may.
pub async fn set(&self, hive: &str, agent: &str, state: AgentState) -> Result<HiveWanted> {
self.write(hive, agent, state, Intent::Redeclare).await
}
/// Declare a *newly created* `agent` on `hive`, and return the whole
/// declaration as published.
///
/// Same write as [`set`](Self::set) but with [`Intent::Create`], so a
/// `Destroyed` entry left by an earlier agent of this name does not
/// refuse it. Recreating a destroyed name is precisely the "fresh deploy"
/// [`AgentState::Destroyed`]'s doc comment names as the way back, and by
/// the time this runs the swarm has already built that agent's identity,
/// repo and config from nothing.
///
/// A separate method rather than an `Intent` parameter on `set`: the
/// caller that may overwrite a terminal state is the agent-creation job
/// node and nothing else, and an argument every other caller has to pass
/// correctly is an argument one of them eventually passes wrong.
pub async fn create(&self, hive: &str, agent: &str, state: AgentState) -> Result<HiveWanted> {
self.write(hive, agent, state, Intent::Create).await
}
/// The write both of the above are.
///
/// Read-modify-write against the entry's revision rather than a plain
/// `put`: the value is the hive's whole agent map, so a blind write
/// would drop a concurrent change to a different agent. The bucket keeps
/// one revision of history, so a lost write is not recoverable after the
/// fact — the conflict has to be caught here.
pub async fn set(&self, hive: &str, agent: &str, state: AgentState) -> Result<HiveWanted> {
async fn write(
&self,
hive: &str,
agent: &str,
state: AgentState,
intent: Intent,
) -> Result<HiveWanted> {
swarm_queue_client::ensure_connected(&self.client)?;
let store = self.store(hive).await?;
@ -167,8 +232,12 @@ impl WantedWriter {
.await
.with_context(|| format!("reading the declaration for {hive}"))?;
let revision = entry.as_ref().map(|e| e.revision);
let (declaration, encoded) =
apply(entry.as_ref().map(|e| e.value.as_ref()), agent, state)?;
let (declaration, encoded) = apply(
entry.as_ref().map(|e| e.value.as_ref()),
agent,
state,
intent,
)?;
// `update` and `create` have separate error types, and only one
// variant of each means "someone else got there first". Every
@ -209,13 +278,24 @@ impl WantedWriter {
#[cfg(test)]
mod tests {
use super::apply;
use swarm_queue_client::wanted::AgentState;
use super::{Intent, apply};
use swarm_queue_client::wanted::{AgentState, HiveWanted};
/// Every case below that predates the create carve-out is a redeclare;
/// spelling the intent out in each of them would say nothing the test
/// name doesn't. The carve-out's own tests call `apply` directly.
fn redeclare(
current: Option<&[u8]>,
agent: &str,
state: AgentState,
) -> anyhow::Result<(HiveWanted, Vec<u8>)> {
apply(current, agent, state, Intent::Redeclare)
}
#[test]
fn declaring_one_agent_preserves_every_other() {
let current = br#"{"agents":{"iris":{"state":"up"},"argus":{"state":"offline"}}}"#;
let (declaration, _) = apply(Some(current), "atlas", AgentState::Up).unwrap();
let (declaration, _) = redeclare(Some(current), "atlas", AgentState::Up).unwrap();
assert_eq!(declaration.agents.len(), 3);
assert_eq!(declaration.agents["iris"].state, AgentState::Up);
assert_eq!(declaration.agents["argus"].state, AgentState::Offline);
@ -225,7 +305,7 @@ mod tests {
#[test]
fn redeclaring_an_agent_replaces_only_its_own_state() {
let current = br#"{"agents":{"iris":{"state":"up"},"atlas":{"state":"up"}}}"#;
let (declaration, _) = apply(Some(current), "atlas", AgentState::Offline).unwrap();
let (declaration, _) = redeclare(Some(current), "atlas", AgentState::Offline).unwrap();
assert_eq!(declaration.agents.len(), 2);
assert_eq!(declaration.agents["iris"].state, AgentState::Up);
assert_eq!(declaration.agents["atlas"].state, AgentState::Offline);
@ -233,7 +313,7 @@ mod tests {
#[test]
fn a_hive_with_no_declaration_yet_gets_a_one_agent_one() {
let (declaration, _) = apply(None, "atlas", AgentState::Up).unwrap();
let (declaration, _) = redeclare(None, "atlas", AgentState::Up).unwrap();
assert_eq!(declaration.agents.len(), 1);
assert_eq!(declaration.agents["atlas"].state, AgentState::Up);
}
@ -242,7 +322,7 @@ mod tests {
// here would publish a one-agent document over a hive's whole set.
#[test]
fn an_undecodable_declaration_is_an_error_not_a_fresh_start() {
let err = apply(Some(b"{not json"), "atlas", AgentState::Up).unwrap_err();
let err = redeclare(Some(b"{not json"), "atlas", AgentState::Up).unwrap_err();
assert!(
err.to_string().contains("not decodable"),
"unexpected error: {err}"
@ -252,13 +332,13 @@ mod tests {
#[test]
fn an_unknown_state_in_the_current_value_is_also_an_error() {
let current = br#"{"agents":{"iris":{"state":"sideways"}}}"#;
assert!(apply(Some(current), "atlas", AgentState::Up).is_err());
assert!(redeclare(Some(current), "atlas", AgentState::Up).is_err());
}
#[test]
fn the_encoded_form_round_trips() {
let (_, encoded) = apply(None, "atlas", AgentState::Offline).unwrap();
let (again, _) = apply(Some(&encoded), "iris", AgentState::Up).unwrap();
let (_, encoded) = redeclare(None, "atlas", AgentState::Offline).unwrap();
let (again, _) = redeclare(Some(&encoded), "iris", AgentState::Up).unwrap();
assert_eq!(again.agents["atlas"].state, AgentState::Offline);
assert_eq!(again.agents["iris"].state, AgentState::Up);
}
@ -270,7 +350,7 @@ mod tests {
fn moving_a_destroyed_agent_to_any_other_state_is_refused() {
let current = br#"{"agents":{"atlas":{"state":"destroyed"}}}"#;
for state in [AgentState::Up, AgentState::Offline, AgentState::Paused] {
let err = apply(Some(current), "atlas", state).unwrap_err();
let err = redeclare(Some(current), "atlas", state).unwrap_err();
assert!(
err.downcast_ref::<super::TerminalStateError>().is_some(),
"expected a TerminalStateError for {state:?}, got: {err}"
@ -284,7 +364,7 @@ mod tests {
#[test]
fn redeclaring_destroyed_as_destroyed_is_not_a_transition() {
let current = br#"{"agents":{"atlas":{"state":"destroyed"}}}"#;
let (declaration, _) = apply(Some(current), "atlas", AgentState::Destroyed).unwrap();
let (declaration, _) = redeclare(Some(current), "atlas", AgentState::Destroyed).unwrap();
assert_eq!(declaration.agents["atlas"].state, AgentState::Destroyed);
}
@ -293,8 +373,50 @@ mod tests {
#[test]
fn a_destroyed_agent_does_not_block_declaring_a_different_one() {
let current = br#"{"agents":{"atlas":{"state":"destroyed"},"iris":{"state":"up"}}}"#;
let (declaration, _) = apply(Some(current), "iris", AgentState::Offline).unwrap();
let (declaration, _) = redeclare(Some(current), "iris", AgentState::Offline).unwrap();
assert_eq!(declaration.agents["atlas"].state, AgentState::Destroyed);
assert_eq!(declaration.agents["iris"].state, AgentState::Offline);
}
/// The carve-out `AgentState::Destroyed`'s doc comment already promised
/// and nothing implemented: a *fresh deploy* does bring a destroyed name
/// back. Without this, creating an agent under a name that was destroyed
/// earlier builds its identity, repo and config and then fails the
/// declaration — cancelling the deploy while the operator sees a 200.
#[test]
fn creating_an_agent_may_reuse_a_destroyed_name() {
let current = br#"{"agents":{"atlas":{"state":"destroyed"}}}"#;
let (declaration, _) =
apply(Some(current), "atlas", AgentState::Paused, Intent::Create).unwrap();
assert_eq!(declaration.agents["atlas"].state, AgentState::Paused);
}
/// The carve-out is scoped to the agent being created — recreating one
/// name must not disturb another destroyed agent on the same hive.
#[test]
fn creating_an_agent_leaves_other_destroyed_agents_destroyed() {
let current = br#"{"agents":{"atlas":{"state":"destroyed"},"iris":{"state":"destroyed"}}}"#;
let (declaration, _) =
apply(Some(current), "atlas", AgentState::Paused, Intent::Create).unwrap();
assert_eq!(declaration.agents["atlas"].state, AgentState::Paused);
assert_eq!(declaration.agents["iris"].state, AgentState::Destroyed);
}
/// `Intent::Create` relaxes exactly one rule. An undecodable current
/// value is still an error, for the same reason it is on a redeclare:
/// writing over it discards every other agent on the hive.
#[test]
fn creating_an_agent_still_refuses_an_undecodable_declaration() {
let err = apply(
Some(b"{not json"),
"atlas",
AgentState::Paused,
Intent::Create,
)
.unwrap_err();
assert!(
err.to_string().contains("not decodable"),
"unexpected error: {err}"
);
}
}