refactor(#4006): one wanted-state bucket per hive, so a watch can be scoped

A hive reads its own declaration today and that scopes cleanly: DIRECT.GET
carries the key in the subject, so the grant can name it. A *watch* cannot be
scoped that way — a consumer's filter travels in the request payload, so
$JS.API.CONSUMER.CREATE.<stream> grants the whole stream. With every hive in
one bucket, letting a hive watch its own declaration would let it read every
other hive's.

One bucket per hive (hive-wanted-<hive>) makes the stream a hive may hold
exactly as wide as what it is allowed to see, which is what #4006's live-watch
needs. That watch is a separate change; this only moves the boundary.

mara's calls, both on #4006: one stream per hive rather than teaching the auth
responder a hive roster, and a wildcard for the controller — "its okay if
swarm controller can theoretically override hive". A bucket name is a single
subject token with no prefix matching, so no wildcard narrower than * covers N
per-hive buckets; the controller's grant is account-wide by consequence, and
documented as chosen rather than left to look accidental.

The reader arm of #4005's key-layout guard asserted the opposite of that
ruling, so it is replaced rather than deleted: the hive arm survives as
no_hive_may_write_another_role_s_agent_status (with a positive control), and
the_readers_grant_is_deliberately_account_wide pins the decision and names the
ruling, so the width reads as chosen to whoever finds it next.

Two pre-existing negative assertions were silently defanged by the rename --
they matched hive-wanted.beta and $KV.hive-wanted.alpha, strings nothing
produces any more, and kept passing. Both now match current names.

swarm-controller resolves the store per hive per call instead of caching one in
a OnceCell: there is no single handle that serves N buckets, and declarations
change on operator action rather than per tick.
This commit is contained in:
atlas 2026-09-02 21:03:22 +02:00 committed by mara
commit 78637ded0c
6 changed files with 111 additions and 66 deletions

View file

@ -63,7 +63,7 @@ pub async fn pull(coord: &Arc<Coordinator>) -> Result<()> {
// A hung node is worse than a failed one: the dashboard shows it running. // A hung node is worse than a failed one: the dashboard shows it running.
swarm_queue_client::ensure_connected(&client)?; swarm_queue_client::ensure_connected(&client)?;
let Some(store) = swarm_queue_client::wanted::open_read_only(&client).await else { let Some(store) = swarm_queue_client::wanted::open_read_only(&client, &hive).await else {
tracing::info!(%hive, "wanted state: no bucket yet; nothing has been declared"); tracing::info!(%hive, "wanted state: no bucket yet; nothing has been declared");
return Ok(()); return Ok(());
}; };

View file

@ -57,29 +57,31 @@ fn apply(current: Option<&[u8]>, agent: &str, state: AgentState) -> Result<(Hive
/// Writes the wanted-state bucket, and reads it back. /// Writes the wanted-state bucket, and reads it back.
pub struct WantedWriter { pub struct WantedWriter {
client: async_nats::Client, client: async_nats::Client,
store: tokio::sync::OnceCell<async_nats::jetstream::kv::Store>,
} }
impl WantedWriter { impl WantedWriter {
#[must_use] #[must_use]
pub fn new(client: async_nats::Client) -> Self { pub fn new(client: async_nats::Client) -> Self {
Self { Self { client }
client,
store: tokio::sync::OnceCell::new(),
}
} }
/// The bucket handle, created on first use if nothing has made it yet. /// One hive's bucket handle, created on first use if nothing has made it
/// yet.
/// ///
/// Creation lives in [`swarm_queue_client::wanted`] because a bucket is /// Creation lives in [`swarm_queue_client::wanted`] because a bucket is
/// described identically by everyone who may create it. Only the /// described identically by everyone who may create it. Only the
/// controller creates this one; a hive opens it read-only. /// controller creates these; a hive opens its own read-only.
///
/// Resolved per call rather than cached: there is one bucket per hive, so
/// a single cached handle cannot serve them, and the declarations this
/// writes change on operator action rather than on a loop — the extra
/// lookup is per *declaration*, not per tick. A cache here would be a map
/// whose invalidation nobody needs yet.
async fn store( async fn store(
&self, &self,
) -> std::result::Result<&async_nats::jetstream::kv::Store, swarm_queue_client::Error> { hive: &str,
self.store ) -> std::result::Result<async_nats::jetstream::kv::Store, swarm_queue_client::Error> {
.get_or_try_init(|| swarm_queue_client::wanted::open_or_create(&self.client)) swarm_queue_client::wanted::open_or_create(&self.client, hive).await
.await
} }
/// The declaration currently published for `hive`, or `None`. /// The declaration currently published for `hive`, or `None`.
@ -87,7 +89,7 @@ impl WantedWriter {
// An unconnected client does not fail a JetStream request, it hangs // An unconnected client does not fail a JetStream request, it hangs
// on it — see `swarm_queue_client::ensure_connected`. // on it — see `swarm_queue_client::ensure_connected`.
swarm_queue_client::ensure_connected(&self.client)?; swarm_queue_client::ensure_connected(&self.client)?;
let store = self.store().await?; let store = self.store(hive).await?;
let Some(entry) = store let Some(entry) = store
.entry(hive) .entry(hive)
.await .await
@ -110,7 +112,7 @@ impl WantedWriter {
/// fact — the conflict has to be caught here. /// fact — the conflict has to be caught here.
pub async fn set(&self, hive: &str, agent: &str, state: AgentState) -> Result<HiveWanted> { pub async fn set(&self, hive: &str, agent: &str, state: AgentState) -> Result<HiveWanted> {
swarm_queue_client::ensure_connected(&self.client)?; swarm_queue_client::ensure_connected(&self.client)?;
let store = self.store().await?; let store = self.store(hive).await?;
for _ in 0..MAX_ATTEMPTS { for _ in 0..MAX_ATTEMPTS {
let entry = store let entry = store

View file

@ -138,14 +138,14 @@ impl Policy {
format!("KV_{}", self.bucket) format!("KV_{}", self.bucket)
} }
/// The stream backing the wanted-state bucket. /// The stream backing one hive's wanted-state bucket.
/// ///
/// Derived from the crate constant rather than from a second configurable /// Derived from the crate's own name function rather than from a
/// bucket name: `wanted::BUCKET` is a `const` precisely so writer and /// configurable bucket name: `wanted::bucket` is what the writer and the
/// reader cannot disagree about it, and a flag here would reintroduce the /// reader both call, so a flag here would let two deployments disagree
/// disagreement one layer out. /// about a name they must share.
fn wanted_stream() -> String { fn wanted_stream(hive: &str) -> String {
format!("KV_{}", swarm_queue_client::wanted::BUCKET) format!("KV_{}", swarm_queue_client::wanted::bucket(hive))
} }
/// The stream backing the per-agent status bucket. /// The stream backing the per-agent status bucket.
@ -256,19 +256,22 @@ impl Policy {
// pointing at a subject nobody publishes to. Not `extra_hive_subjects`: // pointing at a subject nobody publishes to. Not `extra_hive_subjects`:
// that is for streams this crate does not know about. // that is for streams this crate does not know about.
subjects.push(swarm_queue_client::notices::subject(hive)); subjects.push(swarm_queue_client::notices::subject(hive));
// Reading the wanted-state bucket, and **only this hive's key**. // Reading the wanted-state bucket, which is **this hive's own**.
// //
// A KV read is a publish: `store.get` is a request, and the direct-get // One bucket per hive rather than one keyed by hive, so that a *watch*
// subject carries the key, so the grant scopes to one hive rather than // can be granted without widening the read: a consumer's filter travels
// to the bucket. `$KV.<wanted>.<hive>` is deliberately absent — the // in the request payload, so `CONSUMER.CREATE` grants a whole stream and
// controller declares this and a hive converges to it, so a hive that // cannot be scoped to a key the way `DIRECT.GET` can.
// could write its own key could declare its own desired state. //
// `$KV.<bucket>.<hive>` stays deliberately absent — the controller
// declares this and a hive converges to it, so a hive that could write
// its own key could declare its own desired state.
let wanted = Self::wanted_stream(hive);
subjects.extend([ subjects.extend([
format!("$JS.API.STREAM.INFO.{}", Self::wanted_stream()), format!("$JS.API.STREAM.INFO.{wanted}"),
format!( format!(
"$JS.API.DIRECT.GET.{}.$KV.{}.{hive}", "$JS.API.DIRECT.GET.{wanted}.$KV.{}.{hive}",
Self::wanted_stream(), swarm_queue_client::wanted::bucket(hive)
swarm_queue_client::wanted::BUCKET
), ),
]); ]);
// Publishing this hive's agents into the per-agent status bucket, and // Publishing this hive's agents into the per-agent status bucket, and
@ -349,13 +352,25 @@ impl Policy {
// admission). Same failure mode as the knowledge event above: a // admission). Same failure mode as the knowledge event above: a
// refused publish reaches the client as a timeout. // refused publish reaches the client as a timeout.
swarm_queue_client::DEPLOY_SUBJECT_WILDCARD.to_owned(), swarm_queue_client::DEPLOY_SUBJECT_WILDCARD.to_owned(),
// The wanted-state bucket, which the controller creates and writes // The wanted-state buckets — one per hive, all created and written
// for every hive. `.>` here against one key per hive above: this is // by this single client.
// the single writer, so scoping it per hive would be a list of //
// hives to keep, which `hive_name`'s doc argues against. // Wildcards rather than a name per hive: a bucket name is one
format!("$JS.API.STREAM.INFO.{}", Self::wanted_stream()), // subject token and subjects have no prefix matching, so
format!("$JS.API.STREAM.CREATE.{}", Self::wanted_stream()), // `KV_hive-wanted-*` is not expressible. The choice is an exact
format!("$KV.{}.>", swarm_queue_client::wanted::BUCKET), // name per hive — which needs a hive roster in this responder,
// which `hive_name`'s doc argues against — or `*`.
//
// ⚠️ `*` therefore reaches every stream in the account, including
// the buckets hives publish themselves, so this client *can*
// overwrite a hive's self-reported status. That is an accepted
// operator decision, not an oversight: the controller is the
// swarm's writer of record, and a swarm that cannot trust it has a
// larger problem than this grant.
"$JS.API.STREAM.INFO.*".to_owned(),
"$JS.API.STREAM.CREATE.*".to_owned(),
"$JS.API.DIRECT.GET.*.>".to_owned(),
"$KV.*.>".to_owned(),
]); ]);
// The per-agent status bucket, read-side only. Same five subjects as // The per-agent status bucket, read-side only. Same five subjects as
// the hive-status block above and for the same measured reasons — // the hive-status block above and for the same measured reasons —
@ -414,23 +429,38 @@ mod tests {
} }
#[test] #[test]
fn the_reader_gets_no_write_subject_for_the_agent_status_bucket() { fn no_hive_may_write_another_role_s_agent_status() {
// The controller reads this bucket and never writes it. A `$KV.` // A hive publishes its *own* agents' status under
// subject here would let a reader forge any agent's status on any // `$KV.agent-status.<hive>.*` and must reach no further. This is the
// hive — the containment `hive_subjects` establishes is only worth // arm of the old reader-side guard that still holds.
// anything if the other role cannot bypass it.
let bucket = swarm_queue_client::agent_status::BUCKET; let bucket = swarm_queue_client::agent_status::BUCKET;
let prefix = format!("$KV.{bucket}."); let subjects = policy().hive_subjects("alpha");
for forbidden in [
format!("$KV.{bucket}.>"),
format!("$KV.{bucket}.*"),
format!("$KV.{bucket}.beta.*"),
] {
assert!(!subjects.contains(&forbidden), "hive alpha got {forbidden}");
}
// Control: the containment being asserted is not vacuous — the hive
// does hold its own slice.
assert!(subjects.contains(&format!("$KV.{bucket}.alpha.*")));
}
#[test]
fn the_readers_grant_is_deliberately_account_wide() {
// ⚠️ This pins a DECISION, not a safety property, and it replaced a
// guard that asserted the opposite — the operator's call was that the
// controller being able to override a hive is acceptable.
//
// A bucket name is one subject token and subjects have no prefix
// matching, so per-hive wanted-state buckets cannot be covered by any
// wildcard narrower than `*`. Scoping the controller would mean a hive
// roster inside this responder. If that trade is ever revisited, this
// test is the thing to change — it exists so the width reads as chosen.
let subjects = policy().reader_subjects(); let subjects = policy().reader_subjects();
assert!( assert!(subjects.contains(&"$KV.*.>".to_owned()));
!subjects.iter().any(|s| s.starts_with(&prefix)), assert!(subjects.contains(&"$JS.API.STREAM.INFO.*".to_owned()));
"the reader was granted a {prefix}* subject"
);
// Control: the prefix test does fire on a bucket the reader
// legitimately writes, so the pass above is about agent-status rather
// than about a matcher that never matches anything.
let wanted = format!("$KV.{}.", swarm_queue_client::wanted::BUCKET);
assert!(subjects.iter().any(|s| s.starts_with(&wanted)));
} }
#[test] #[test]
@ -736,8 +766,9 @@ mod tests {
.permissions("hive-alpha") .permissions("hive-alpha")
.expect("a hive is admitted"); .expect("a hive is admitted");
assert!( assert!(
p.publish p.publish.contains(
.contains(&"$JS.API.DIRECT.GET.KV_hive-wanted.$KV.hive-wanted.alpha".to_owned()), &"$JS.API.DIRECT.GET.KV_hive-wanted-alpha.$KV.hive-wanted-alpha.alpha".to_owned()
),
"got: {:?}", "got: {:?}",
p.publish p.publish
); );
@ -749,7 +780,12 @@ mod tests {
let p = policy() let p = policy()
.permissions("hive-alpha") .permissions("hive-alpha")
.expect("a hive is admitted"); .expect("a hive is admitted");
assert!(!p.publish.iter().any(|s| s.contains("hive-wanted.beta"))); // ⚠️ Matched against the CURRENT bucket name. The per-hive split moved
// this from `hive-wanted.beta` to `hive-wanted-beta`, and the old
// spelling kept passing against a subject set that no longer contains
// it — a pass that had stopped meaning anything.
assert!(!p.publish.iter().any(|s| s.contains("hive-wanted-beta")));
assert!(!p.publish.iter().any(|s| s.contains("beta")));
assert!( assert!(
!p.publish !p.publish
.iter() .iter()
@ -764,8 +800,12 @@ mod tests {
let p = policy() let p = policy()
.permissions("hive-alpha") .permissions("hive-alpha")
.expect("a hive is admitted"); .expect("a hive is admitted");
assert!(!p.publish.iter().any(|s| s == "$KV.hive-wanted.alpha")); // Current names, for the reason the sibling test above records: the
assert!(!p.publish.iter().any(|s| s == "$KV.hive-wanted.>")); // pre-split strings `$KV.hive-wanted.{alpha,>}` still "passed" after
// the rename because nothing produces them any more.
assert!(!p.publish.iter().any(|s| s == "$KV.hive-wanted-alpha.alpha"));
assert!(!p.publish.iter().any(|s| s == "$KV.hive-wanted-alpha.>"));
assert!(!p.publish.iter().any(|s| s.starts_with("$KV.hive-wanted")));
} }
#[test] #[test]
@ -775,11 +815,11 @@ mod tests {
let p = policy() let p = policy()
.permissions("swarm-controller") .permissions("swarm-controller")
.expect("a reader is admitted"); .expect("a reader is admitted");
assert!(p.publish.contains(&"$KV.hive-wanted.>".to_owned())); // Account-wide since the per-hive split: a bucket name is one subject
assert!( // token, so no wildcard narrower than `*` covers N per-hive buckets.
p.publish // See `the_readers_grant_is_deliberately_account_wide`.
.contains(&"$JS.API.STREAM.CREATE.KV_hive-wanted".to_owned()) assert!(p.publish.contains(&"$KV.*.>".to_owned()));
); assert!(p.publish.contains(&"$JS.API.STREAM.CREATE.*".to_owned()));
} }
#[test] #[test]

View file

@ -125,7 +125,7 @@ pub async fn open_or_create(
}) })
.await .await
.map_err(|source| Error::CreateBucket { .map_err(|source| Error::CreateBucket {
bucket: BUCKET, bucket: BUCKET.to_owned(),
source, source,
}) })
} }

View file

@ -67,7 +67,7 @@ pub async fn open_or_create(
}) })
.await .await
.map_err(|source| Error::CreateBucket { .map_err(|source| Error::CreateBucket {
bucket: BUCKET, bucket: BUCKET.to_owned(),
source, source,
}) })
} }

View file

@ -170,7 +170,10 @@ mod tests {
// the grant scoping in `swarm-nats-auth` names this string, so a bucket // the grant scoping in `swarm-nats-auth` names this string, so a bucket
// the client refuses to open would surface as a permissions problem // the client refuses to open would surface as a permissions problem
// rather than as the naming problem it is. // rather than as the naming problem it is.
let legal = |s: &str| s.chars().all(|c| c.is_ascii_alphanumeric() || c == '_' || c == '-'); let legal = |s: &str| {
s.chars()
.all(|c| c.is_ascii_alphanumeric() || c == '_' || c == '-')
};
assert!(legal(&bucket("alpha"))); assert!(legal(&bucket("alpha")));
assert!(legal(&bucket("a-hive-with-hyphens"))); assert!(legal(&bucket("a-hive-with-hyphens")));
assert!(legal(&bucket("h9"))); assert!(legal(&bucket("h9")));