feat(artifacts): add fenced reconciliation recovery

This commit is contained in:
2026-08-28 00:44:27 +03:00
parent 8784964fb2
commit d2849ea3fe
21 changed files with 3322 additions and 97 deletions
@@ -0,0 +1,214 @@
use std::time::Duration as StdDuration;
use crank_artifacts::ArtifactStore;
use crank_registry::{
ArtifactClaimFinalization, ArtifactClaimFinalizeOutcome, ArtifactClaimOutcome,
ArtifactClaimToken, ClaimArtifactReconciliationRequest,
ClaimExpiredArtifactReconciliationRequest, RegistryError,
};
use super::{
artifact_sources::{TestRoot, timestamp},
common::TestDatabase,
};
#[tokio::test]
async fn locked_database_time_and_expired_recovery_are_fenced() {
let database = TestDatabase::new().await;
let registry = database.registry().await;
let raw_pool = database.raw_pool().await;
let root = TestRoot::new("database-time");
let store = ArtifactStore::open(&root.0).unwrap();
let caller_time = timestamp("2020-01-01T00:00:00Z");
let blocked = store.put_registered(b"row lock expiry\n").unwrap();
assert!(matches!(
registry
.claim_artifact_reconciliation(ClaimArtifactReconciliationRequest {
artifact: &blocked,
token: ArtifactClaimToken::generate(),
claimed_at: caller_time,
lease_expires_at: caller_time + time::Duration::seconds(1),
detached_grace: time::Duration::hours(24),
})
.await
.unwrap(),
ArtifactClaimOutcome::Claimed(_)
));
let mut blocker = raw_pool.begin().await.unwrap();
sqlx::query("select digest from artifact_blobs where digest = $1 for update")
.bind(blocked.artifact_ref().digest_hex())
.fetch_one(&mut *blocker)
.await
.unwrap();
let waiting_registry = registry.clone();
let waiting_artifact = blocked.clone();
let waiting = tokio::spawn(async move {
waiting_registry
.claim_artifact_reconciliation(ClaimArtifactReconciliationRequest {
artifact: &waiting_artifact,
token: ArtifactClaimToken::generate(),
claimed_at: caller_time,
lease_expires_at: caller_time + time::Duration::minutes(5),
detached_grace: time::Duration::hours(24),
})
.await
});
tokio::time::sleep(StdDuration::from_millis(1_100)).await;
blocker.commit().await.unwrap();
assert!(matches!(
waiting.await.unwrap().unwrap(),
ArtifactClaimOutcome::Claimed(_)
));
let recoverable = store.put_registered(b"single winner\n").unwrap();
let prior_token = ArtifactClaimToken::generate();
let prior_claim = match registry
.claim_artifact_reconciliation(ClaimArtifactReconciliationRequest {
artifact: &recoverable,
token: prior_token.clone(),
claimed_at: caller_time,
lease_expires_at: caller_time + time::Duration::minutes(5),
detached_grace: time::Duration::hours(24),
})
.await
.unwrap()
{
ArtifactClaimOutcome::Claimed(claim) => claim,
outcome => panic!("unexpected claim outcome: {outcome:?}"),
};
sqlx::query(
"update artifact_blobs
set claim_expires_at = clock_timestamp() - interval '1 second'
where digest = $1",
)
.bind(recoverable.artifact_ref().digest_hex())
.execute(&raw_pool)
.await
.unwrap();
assert!(matches!(
registry
.claim_artifact_reconciliation(ClaimArtifactReconciliationRequest {
artifact: &recoverable,
token: prior_token.clone(),
claimed_at: caller_time,
lease_expires_at: caller_time + time::Duration::minutes(5),
detached_grace: time::Duration::hours(24),
})
.await
.unwrap(),
ArtifactClaimOutcome::HeldByOther
));
let recovery_now = registry.artifact_reconciliation_now().await.unwrap();
let probe = registry
.list_expired_artifact_reconciliation_probes(recovery_now, 1)
.await
.unwrap()
.pop()
.unwrap();
assert!(matches!(
registry
.claim_expired_artifact_reconciliation(
&probe,
ClaimExpiredArtifactReconciliationRequest {
token: prior_token,
claimed_at: caller_time,
lease_expires_at: caller_time + time::Duration::minutes(5),
},
)
.await,
Err(RegistryError::InvalidArtifactSource {
field: "claim_token"
})
));
let first_registry = registry.clone();
let second_registry = registry.clone();
let first_probe = probe.clone();
let (first, second) = tokio::join!(
async move {
first_registry
.claim_expired_artifact_reconciliation(
&first_probe,
ClaimExpiredArtifactReconciliationRequest {
token: ArtifactClaimToken::generate(),
claimed_at: caller_time,
lease_expires_at: caller_time + time::Duration::minutes(5),
},
)
.await
},
async move {
second_registry
.claim_expired_artifact_reconciliation(
&probe,
ClaimExpiredArtifactReconciliationRequest {
token: ArtifactClaimToken::generate(),
claimed_at: caller_time,
lease_expires_at: caller_time + time::Duration::minutes(5),
},
)
.await
}
);
let first = first.unwrap();
let second = second.unwrap();
assert_eq!(
usize::from(first.is_some()) + usize::from(second.is_some()),
1
);
let winner = first.or(second).unwrap();
sqlx::query("update artifact_blobs set claim_expires_at = clock_timestamp() where digest = $1")
.bind(recoverable.artifact_ref().digest_hex())
.execute(&raw_pool)
.await
.unwrap();
for claim in [&winner, &prior_claim] {
assert_eq!(
registry
.finalize_artifact_reconciliation_claim(
claim,
ArtifactClaimFinalization::AlreadyAbsent,
caller_time,
)
.await
.unwrap(),
ArtifactClaimFinalizeOutcome::Stale
);
}
let invalid = store.put_registered(b"invalid bounds\n").unwrap();
assert!(matches!(
registry
.claim_artifact_reconciliation(ClaimArtifactReconciliationRequest {
artifact: &invalid,
token: ArtifactClaimToken::generate(),
claimed_at: caller_time,
lease_expires_at: caller_time + time::Duration::minutes(5),
detached_grace: time::Duration::MAX,
})
.await,
Err(RegistryError::InvalidArtifactSource {
field: "detached_grace"
})
));
assert!(matches!(
registry
.claim_artifact_reconciliation(ClaimArtifactReconciliationRequest {
artifact: &invalid,
token: ArtifactClaimToken::generate(),
claimed_at: caller_time,
lease_expires_at: caller_time
+ time::Duration::minutes(5)
+ time::Duration::seconds(1),
detached_grace: time::Duration::ZERO,
})
.await,
Err(RegistryError::InvalidArtifactSource {
field: "claim_lease"
})
));
database.cleanup().await;
}
@@ -5,24 +5,30 @@ use std::{
os::unix::fs::PermissionsExt,
path::PathBuf,
sync::atomic::{AtomicU64, Ordering},
time::{Duration as StdDuration, SystemTime},
};
use crank_artifacts::{ArtifactRef, ArtifactStore};
use crank_artifacts::{
ArtifactRef, ArtifactStore, ReconciliationMutation, ReconciliationPresence,
ReconciliationRegistration, RegisteredArtifact,
};
use crank_core::{Workspace, WorkspaceId, WorkspaceStatus};
use crank_registry::{
ArtifactSourceId, ArtifactSourceLifecycle, ArtifactSourceSensitivity,
CreateArtifactSourceRequest, CreateWorkspaceRequest, DetachArtifactSourceRequest,
ListArtifactSourcesQuery, RegistryError,
ArtifactClaimFinalization, ArtifactClaimFinalizeOutcome, ArtifactClaimOutcome,
ArtifactClaimRecheckOutcome, ArtifactClaimToken, ArtifactSourceId, ArtifactSourceLifecycle,
ArtifactSourceSensitivity, ClaimArtifactReconciliationRequest,
ClaimExpiredArtifactReconciliationRequest, CreateArtifactSourceRequest, CreateWorkspaceRequest,
DetachArtifactSourceRequest, ListArtifactSourcesQuery, RegistryError,
};
use serde_json::json;
use time::{OffsetDateTime, format_description::well_known::Rfc3339};
static NEXT_ROOT: AtomicU64 = AtomicU64::new(0);
struct TestRoot(PathBuf);
pub(super) struct TestRoot(pub(super) PathBuf);
impl TestRoot {
fn new(name: &str) -> Self {
pub(super) fn new(name: &str) -> Self {
let path = std::env::temp_dir().join(format!(
"crank-registry-artifacts-{name}-{}-{}",
std::process::id(),
@@ -41,10 +47,26 @@ impl Drop for TestRoot {
}
}
fn timestamp(value: &str) -> OffsetDateTime {
pub(super) fn timestamp(value: &str) -> OffsetDateTime {
OffsetDateTime::parse(value, &Rfc3339).unwrap()
}
fn reconciliation_candidate_for(
store: &ArtifactStore,
artifact: &RegisteredArtifact,
) -> crank_artifacts::ReconciliationCandidate {
let (_, candidates) = store.scan_reconciliation(None, 4096, 32).unwrap();
candidates
.into_iter()
.find(|candidate| {
matches!(
store.register_reconciliation(candidate, StdDuration::ZERO, SystemTime::now()),
Ok(ReconciliationRegistration::Registered(found)) if found == *artifact
)
})
.expect("registered reconciliation candidate")
}
async fn create_workspace(registry: &crank_registry::PostgresRegistry, id: &str) -> WorkspaceId {
let workspace_id = WorkspaceId::new(id);
let created_at = timestamp("2026-08-26T10:00:00Z");
@@ -534,3 +556,300 @@ async fn verified_read_returns_only_digest_and_size_verified_bytes() {
database.cleanup().await;
}
#[tokio::test]
async fn reconciliation_claims_are_fenced_global_and_allow_verified_revival() {
let database = TestDatabase::new().await;
let registry = database.registry().await;
let raw_pool = database.raw_pool().await;
let workspace_a = create_workspace(&registry, "ws_claim_a").await;
let workspace_b = create_workspace(&registry, "ws_claim_b").await;
let root = TestRoot::new("claims");
let store = ArtifactStore::open(&root.0).unwrap();
let unreferenced = store.put_registered(b"unreferenced\n").unwrap();
let started = timestamp("2026-08-27T10:00:00Z");
let lease = timestamp("2026-08-27T10:05:00Z");
let token = ArtifactClaimToken::generate();
assert_eq!(format!("{token:?}"), "ArtifactClaimToken(..)");
let request = ClaimArtifactReconciliationRequest {
artifact: &unreferenced,
token: token.clone(),
claimed_at: started,
lease_expires_at: lease,
detached_grace: time::Duration::hours(24),
};
assert_eq!(
format!("{request:?}"),
"ClaimArtifactReconciliationRequest(..)"
);
let claim = match registry
.claim_artifact_reconciliation(request)
.await
.unwrap()
{
ArtifactClaimOutcome::Claimed(claim) => claim,
outcome => panic!("unexpected claim outcome: {outcome:?}"),
};
assert_eq!(format!("{claim:?}"), "ArtifactReconciliationClaim(..)");
assert!(matches!(
registry
.claim_artifact_reconciliation(ClaimArtifactReconciliationRequest {
artifact: &unreferenced,
token,
claimed_at: timestamp("2026-08-27T10:01:00Z"),
lease_expires_at: lease,
detached_grace: time::Duration::hours(24),
})
.await
.unwrap(),
ArtifactClaimOutcome::Claimed(_)
));
assert_eq!(
registry
.recheck_artifact_reconciliation_claim(&claim, started, time::Duration::hours(24))
.await
.unwrap(),
ArtifactClaimRecheckOutcome::Mutate
);
assert!(matches!(
registry
.create_artifact_source(CreateArtifactSourceRequest {
workspace_id: &workspace_a,
source_id: &ArtifactSourceId::new("src_blocked"),
artifact: &unreferenced,
mime_type: "text/plain",
sensitivity: ArtifactSourceSensitivity::Internal,
created_at: started,
})
.await,
Err(RegistryError::ArtifactClaimInProgress)
));
assert_eq!(
registry
.finalize_artifact_reconciliation_claim(
&claim,
ArtifactClaimFinalization::Quarantined,
timestamp("2026-08-27T10:01:00Z"),
)
.await
.unwrap(),
ArtifactClaimFinalizeOutcome::Retained
);
let final_candidate = reconciliation_candidate_for(&store, &unreferenced);
assert!(matches!(
store.quarantine_reconciliation(final_candidate),
Ok(ReconciliationMutation::Quarantined | ReconciliationMutation::AlreadyQuarantined)
));
let quarantined_candidate = reconciliation_candidate_for(&store, &unreferenced);
assert!(matches!(
store.delete_quarantined_reconciliation(quarantined_candidate),
Ok(ReconciliationMutation::Deleted | ReconciliationMutation::AlreadyAbsent)
));
assert_eq!(
store
.reconciliation_presence(unreferenced.artifact_ref())
.unwrap(),
ReconciliationPresence::Absent
);
sqlx::query(
"update artifact_blobs
set claim_expires_at = clock_timestamp() - interval '1 second'
where digest = $1",
)
.bind(unreferenced.artifact_ref().digest_hex())
.execute(&raw_pool)
.await
.unwrap();
let recovery_now = registry.artifact_reconciliation_now().await.unwrap();
let probes = registry
.list_expired_artifact_reconciliation_probes(recovery_now, 1)
.await
.unwrap();
assert_eq!(
format!("{:?}", probes.first().expect("expired claim probe")),
"ArtifactExpiredClaimProbe(..)"
);
let winner = registry
.claim_expired_artifact_reconciliation(
probes.first().expect("expired claim probe"),
ClaimExpiredArtifactReconciliationRequest {
token: ArtifactClaimToken::generate(),
claimed_at: recovery_now,
lease_expires_at: recovery_now + time::Duration::minutes(5),
},
)
.await
.unwrap()
.expect("expired claim was not recovered");
assert_eq!(
registry
.finalize_artifact_reconciliation_claim(
&claim,
ArtifactClaimFinalization::Deleted,
timestamp("2026-08-27T10:05:01Z"),
)
.await
.unwrap(),
ArtifactClaimFinalizeOutcome::Stale
);
assert_eq!(
registry
.recover_artifact_reconciliation_claim(
&winner,
ArtifactClaimFinalization::AlreadyAbsent,
timestamp("2026-08-27T10:05:02Z"),
)
.await
.unwrap(),
ArtifactClaimFinalizeOutcome::Unavailable
);
let republished = store.put_registered(b"unreferenced\n").unwrap();
let revived = registry
.create_artifact_source(CreateArtifactSourceRequest {
workspace_id: &workspace_a,
source_id: &ArtifactSourceId::new("src_revived"),
artifact: &republished,
mime_type: "text/plain",
sensitivity: ArtifactSourceSensitivity::Internal,
created_at: timestamp("2026-08-27T10:06:00Z"),
})
.await
.unwrap();
assert_eq!(
revived.blob.lifecycle,
crank_registry::ArtifactBlobLifecycle::Available
);
let referenced = store.put_registered(b"referenced\n").unwrap();
let active_id = ArtifactSourceId::new("src_active");
let active = registry
.create_artifact_source(CreateArtifactSourceRequest {
workspace_id: &workspace_b,
source_id: &active_id,
artifact: &referenced,
mime_type: "text/plain",
sensitivity: ArtifactSourceSensitivity::Internal,
created_at: started,
})
.await
.unwrap();
assert!(matches!(
registry
.claim_artifact_reconciliation(ClaimArtifactReconciliationRequest {
artifact: &referenced,
token: ArtifactClaimToken::generate(),
claimed_at: timestamp("2026-08-27T10:01:00Z"),
lease_expires_at: lease,
detached_grace: time::Duration::hours(24),
})
.await
.unwrap(),
ArtifactClaimOutcome::ActiveReference
));
registry
.detach_artifact_source(DetachArtifactSourceRequest {
workspace_id: &workspace_b,
source_id: &active_id,
expected_updated_at: Some(active.updated_at),
detached_at: timestamp("2026-08-27T10:02:00Z"),
})
.await
.unwrap();
assert!(matches!(
registry
.claim_artifact_reconciliation(ClaimArtifactReconciliationRequest {
artifact: &referenced,
token: ArtifactClaimToken::generate(),
claimed_at: timestamp("2026-08-27T10:03:00Z"),
lease_expires_at: lease,
detached_grace: time::Duration::hours(24),
})
.await
.unwrap(),
ArtifactClaimOutcome::DetachedReferenceInGrace
));
let concurrent = store.put_registered(b"concurrent claim\n").unwrap();
let first_registry = registry.clone();
let second_registry = registry.clone();
let first_artifact = concurrent.clone();
let second_artifact = concurrent.clone();
let (first, second) = tokio::join!(
async move {
first_registry
.claim_artifact_reconciliation(ClaimArtifactReconciliationRequest {
artifact: &first_artifact,
token: ArtifactClaimToken::generate(),
claimed_at: timestamp("2026-08-27T11:00:00Z"),
lease_expires_at: timestamp("2026-08-27T11:05:00Z"),
detached_grace: time::Duration::hours(24),
})
.await
},
async move {
second_registry
.claim_artifact_reconciliation(ClaimArtifactReconciliationRequest {
artifact: &second_artifact,
token: ArtifactClaimToken::generate(),
claimed_at: timestamp("2026-08-27T11:00:00Z"),
lease_expires_at: timestamp("2026-08-27T11:05:00Z"),
detached_grace: time::Duration::hours(24),
})
.await
}
);
let first = first.unwrap();
let second = second.unwrap();
assert_eq!(
usize::from(matches!(first, ArtifactClaimOutcome::Claimed(_)))
+ usize::from(matches!(second, ArtifactClaimOutcome::Claimed(_))),
1
);
assert_eq!(
usize::from(matches!(first, ArtifactClaimOutcome::HeldByOther))
+ usize::from(matches!(second, ArtifactClaimOutcome::HeldByOther)),
1
);
let attach_race = store.put_registered(b"attach race\n").unwrap();
let claim_registry = registry.clone();
let attach_registry = registry.clone();
let claim_artifact = attach_race.clone();
let attach_artifact = attach_race.clone();
let attach_workspace = workspace_a.clone();
let (claim_race, attach_race) = tokio::join!(
async move {
claim_registry
.claim_artifact_reconciliation(ClaimArtifactReconciliationRequest {
artifact: &claim_artifact,
token: ArtifactClaimToken::generate(),
claimed_at: timestamp("2026-08-27T12:00:00Z"),
lease_expires_at: timestamp("2026-08-27T12:05:00Z"),
detached_grace: time::Duration::hours(24),
})
.await
},
async move {
attach_registry
.create_artifact_source(CreateArtifactSourceRequest {
workspace_id: &attach_workspace,
source_id: &ArtifactSourceId::new("src_attach_race"),
artifact: &attach_artifact,
mime_type: "text/plain",
sensitivity: ArtifactSourceSensitivity::Internal,
created_at: timestamp("2026-08-27T12:00:00Z"),
})
.await
}
);
match (claim_race.unwrap(), attach_race) {
(ArtifactClaimOutcome::ActiveReference, Ok(_)) => {}
(ArtifactClaimOutcome::Claimed(_), Err(RegistryError::ArtifactClaimInProgress)) => {}
(claim, attach) => panic!("unsafe claim/attach race outcome: {claim:?}, {attach:?}"),
}
database.cleanup().await;
}