fix(scanner): fence unknown tier accounting (#6396)

crates/scanner/src/data_usage_define.rs

@@ -30,7 +30,8 @@ pub use rustfs_data_usage::{
     AllTierStats, BucketTargetUsageInfo, BucketUsageInfo, DATA_USAGE_OBJECT_NAME, DATA_USAGE_OBSERVED_OBJECT_NAME,
     DataUsageEntry, DataUsageHash, DataUsageHashMap, DataUsageInfo, DataUsageSnapshotSetState, LEGACY_DATA_USAGE_OBJECT_NAME,
     PrefixUsageEntry, PrefixUsageQuery, PrefixUsageSummary, ReplTargetSizeSummary, SizeReconciliationEntry,
-    SizeReconciliationScope, SizeSummary, TierStats, hash_path, prefix_usage_in_cache,
+    SizeReconciliationScope, SizeSummary, TierAccountingProof, TierStats, UNKNOWN_TIER, UNKNOWN_TIER_DIAGNOSTIC_BYTE_CAP,
+    UNKNOWN_TIER_DIAGNOSTIC_ENTRY_CAP, UnknownTierStats, hash_path, prefix_usage_in_cache,
 };
 use rustfs_utils::path::{SLASH_SEPARATOR, path_join_buf};
 use tokio::time::{Duration, Instant, sleep, timeout};
@@ -218,25 +219,79 @@ impl ScannerSizeSummaryExt for SizeSummary {
             self.versions = self.versions.saturating_add(1);
         }
 
-        let size = usize::try_from(size.max(0)).unwrap_or(usize::MAX);
+        let logical_size = size.max(0);
+        let size = usize::try_from(logical_size).unwrap_or(usize::MAX);
         self.total_size = self.total_size.saturating_add(size);
+        let logical_bytes = u64::try_from(logical_size).unwrap_or(u64::MAX);
+        let physical_bytes = u64::try_from(oi.size.max(0)).unwrap_or(0);
+        let mut proof = TierAccountingProof {
+            logical_total: logical_bytes,
+            logical_known: 0,
+            physical_total: physical_bytes,
+            physical_known: 0,
+            overflowed: false,
+        };
 
         if oi.transitioned_object.free_version {
+            proof.logical_known = logical_bytes;
+            proof.physical_known = physical_bytes;
+            self.tier_accounting_proof.saturating_add(proof);
             return;
         }
 
-        let mut tier = oi.storage_class.clone().unwrap_or_else(|| storageclass::STANDARD.to_string());
-        if oi.transitioned_object.status == TRANSITION_COMPLETE {
-            tier = oi.transitioned_object.tier.clone();
+        let tier = if oi.transitioned_object.status == TRANSITION_COMPLETE {
+            oi.transitioned_object.tier.as_str()
+        } else {
+            oi.storage_class.as_deref().unwrap_or(storageclass::STANDARD)
+        };
+
+        let builtin_tier = tier == storageclass::STANDARD || tier == storageclass::RRS;
+        let tier_registry_is_empty =
+            self.tier_stats.is_empty() || (self.tier_stats.len() == 1 && self.tier_stats.contains_key(UNKNOWN_TIER));
+        let known_tier = tier != UNKNOWN_TIER && (builtin_tier || self.tier_stats.contains_key(tier));
+
+        // With no configured tier, retain the historical empty-map shape for
+        // ordinary STANDARD/RRS objects. A non-built-in key is still an
+        // observable unknown and must create only the fixed bucket.
+        if tier_registry_is_empty && known_tier {
+            proof.logical_known = logical_bytes;
+            proof.physical_known = physical_bytes;
+            self.tier_accounting_proof.saturating_add(proof);
+            return;
         }
 
-        if let Some(tier_stats) = self.tier_stats.get_mut(&tier) {
-            *tier_stats = tier_stats.add(&TierStats {
-                total_size: u64::try_from(oi.size).unwrap_or(0),
-                num_versions: 1,
-                num_objects: u64::from(oi.is_latest),
-            });
+        // Configured tiers and the fixed bucket are normally seeded, so the
+        // hot path can mutate them without allocating a key for every object.
+        // The fallback inserts only when a legacy/no-config summary sees its
+        // first unknown key.
+        let tier_stats = if known_tier {
+            if let Some(stats) = self.tier_stats.get_mut(tier) {
+                stats
+            } else {
+                self.tier_stats.entry(tier.to_owned()).or_default()
+            }
+        } else if let Some(stats) = self.tier_stats.get_mut(UNKNOWN_TIER) {
+            stats
+        } else {
+            self.tier_stats.entry(UNKNOWN_TIER.to_string()).or_default()
+        };
+        *tier_stats = tier_stats.add(&TierStats {
+            total_size: physical_bytes,
+            num_versions: 1,
+            num_objects: u64::from(oi.is_latest),
+        });
+        if known_tier {
+            proof.logical_known = logical_bytes;
+            proof.physical_known = physical_bytes;
         }
+        if !known_tier {
+            self.unknown_tier_stats
+                .record_dimensions(tier, logical_bytes, physical_bytes, 1, u64::from(oi.is_latest));
+            if self.unknown_tier_stats.counter_overflowed {
+                proof.overflowed = true;
+            }
+        }
+        self.tier_accounting_proof.saturating_add(proof);
     }
 
     fn actions_accounting_unknown(&mut self, oi: &ObjectInfo) {
@@ -312,6 +367,11 @@ pub struct DataUsageEntryInfo {
     pub name: String,
     pub parent: String,
     pub entry: DataUsageEntry,
+    /// Registry generation used to classify this root entry. Older remote
+    /// workers omit it; callers must reject that result when a frozen cycle
+    /// requires generation fencing.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub tier_registry_generation: Option<u64>,
 }
 
 #[derive(Clone, Copy, Debug, Serialize, Deserialize, PartialEq, Eq, Hash)]
@@ -385,6 +445,10 @@ pub struct DataUsageCacheInfo {
     pub scan_plan_digest: Option<DataUsageScanPlanDigest>,
     #[serde(default)]
     pub cache_key_format: u16,
+    /// Registry generation used for the completed/partial scan. This is
+    /// process-local audit data; older cache writers omit it.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub tier_registry_generation: Option<u64>,
     /// Bounded durable debts for versions whose logical size was not trusted.
     /// The map key is an identity key, never a user-controlled metric label.
     #[serde(default)]
@@ -410,7 +474,8 @@ impl Serialize for DataUsageCacheInfo {
     {
         // Keep this metadata map-encoded so older readers can ignore fields
         // appended by newer scanner versions during rolling upgrades.
-        let field_count = 21 + usize::from(!self.size_reconciliation.is_empty());
+        let field_count =
+            21 + usize::from(self.tier_registry_generation.is_some()) + usize::from(!self.size_reconciliation.is_empty());
         let mut state = serializer.serialize_map(Some(field_count))?;
         state.serialize_entry("name", &self.name)?;
         state.serialize_entry("next_cycle", &self.next_cycle)?;
@@ -428,6 +493,9 @@ impl Serialize for DataUsageCacheInfo {
         state.serialize_entry("snapshot_complete", &self.snapshot_complete)?;
         state.serialize_entry("scan_plan_digest", &self.scan_plan_digest)?;
         state.serialize_entry("cache_key_format", &self.cache_key_format)?;
+        if let Some(generation) = self.tier_registry_generation {
+            state.serialize_entry("tier_registry_generation", &generation)?;
+        }
         if !self.size_reconciliation.is_empty() {
             state.serialize_entry("size_reconciliation", &self.size_reconciliation)?;
         }
@@ -456,6 +524,58 @@ pub(crate) enum DataUsageCachePrepareOutcome {
 }
 
 impl DataUsageCache {
+    /// Reconcile tier keys loaded from an older cache against the registry
+    /// frozen for this scan. New metadata is already routed through
+    /// `UNKNOWN_TIER`; this pass handles retired keys that predate that rule.
+    /// Legacy `TierStats` carries physical bytes only, so this migration does
+    /// not manufacture a logical unknown-byte value from that physical total.
+    pub(crate) fn fold_retired_tiers(&mut self, tier_names: &[String]) {
+        let known_tiers = tier_names.iter().map(String::as_str).collect::<HashSet<_>>();
+        for entry in self.cache.values_mut() {
+            let Some(tiers) = entry.all_tier_stats.as_mut() else { continue };
+            let existing_unknown = tiers.tiers.get(UNKNOWN_TIER).cloned().unwrap_or_default();
+            let companion_present = entry.unknown_tier_stats.as_ref().is_some_and(|stats| !stats.is_empty());
+            let migrate_existing_unknown = !companion_present;
+            let mut retired = TierStats::default();
+            let mut retired_key_found = false;
+            if migrate_existing_unknown {
+                retired = retired.add(&existing_unknown);
+            }
+            for (tier, stats) in &tiers.tiers {
+                if tier != UNKNOWN_TIER
+                    && tier != storageclass::STANDARD
+                    && tier != storageclass::RRS
+                    && !known_tiers.contains(tier.as_str())
+                {
+                    retired_key_found = true;
+                    retired = retired.add(stats);
+                }
+            }
+            tiers.fold_unknown_tiers(tier_names.iter().map(String::as_str));
+            if !retired.is_empty() && !companion_present {
+                entry.add_unknown_tier_stats(&UnknownTierStats {
+                    // The legacy map stores physical bytes only. Logical
+                    // bytes remain zero until a fresh object scan observes
+                    // them under the current metadata format.
+                    unknown_physical_bytes: retired.total_size,
+                    unknown_objects: retired.num_objects,
+                    unknown_versions: retired.num_versions,
+                    ..Default::default()
+                });
+                // The legacy tier map has no logical-byte dimension, so a
+                // proof that classified this retired key as known cannot be
+                // repaired safely. Mark it unvalidated and require a fresh
+                // scan rather than guessing a logical subtraction.
+                entry.tier_accounting_proof = None;
+            } else if retired_key_found {
+                // A nonempty companion has no provenance tying it to the
+                // retired map keys. Reject the mixed cache until a fresh scan
+                // reconciles the dimensions instead of double-counting them.
+                entry.tier_accounting_proof = None;
+            }
+        }
+    }
+
     /// Prefix-level usage query over this (writer-side) cache; see
     /// [`prefix_usage_in_cache`] for the semantics
     /// (rustfs/backlog#1872).
@@ -945,6 +1065,7 @@ impl DataUsageCache {
             delete_markers_total_count: flat.delete_markers as u64,
             objects_total_size: flat.size as u64,
             tier_stats: flat.all_tier_stats.filter(|tiers| !tiers.is_empty()),
+            unknown_tier_stats: flat.unknown_tier_stats.filter(|stats| !stats.is_empty()),
             buckets_count: u64::try_from(buckets.len()).unwrap_or(u64::MAX),
             buckets_usage,
             ..Default::default()

crates/scanner/src/data_usage_define/tests.rs

@@ -621,7 +621,6 @@ fn size_summary_add_saturates_all_usage_counters() {
             failed_count: usize::MAX,
         },
     );
-
     let mut increment = SizeSummary {
         total_size: 1,
         versions: 1,
@@ -636,6 +635,24 @@ fn size_summary_add_saturates_all_usage_counters() {
         failed_count: 1,
         ..Default::default()
     };
+    summary.tier_stats.insert(
+        UNKNOWN_TIER.to_string(),
+        TierStats {
+            total_size: u64::MAX,
+            num_versions: u64::MAX,
+            num_objects: u64::MAX,
+        },
+    );
+    increment.tier_stats.insert(
+        UNKNOWN_TIER.to_string(),
+        TierStats {
+            total_size: 1,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    );
+    increment.unknown_tier_stats.unknown_bytes = 1;
+    increment.unknown_tier_stats.unknown_physical_bytes = 1;
     increment.repl_target_stats.insert(
         target.clone(),
         ReplTargetSizeSummary {
@@ -672,6 +689,8 @@ fn size_summary_add_saturates_all_usage_counters() {
     assert_eq!(target_summary.failed_size, i64::MAX);
     assert_eq!(target_summary.pending_count, usize::MAX);
     assert_eq!(target_summary.failed_count, usize::MAX);
+    assert_eq!(summary.tier_stats[UNKNOWN_TIER].total_size, u64::MAX);
+    assert_eq!(summary.unknown_tier_stats.unknown_bytes, 1);
 }
 
 #[test]
@@ -721,6 +740,249 @@ fn size_summary_actions_accounting_accumulates_tier_stats() {
     );
 }
 
+#[test]
+fn unknown_tier_is_bounded_and_accounted() {
+    let mut summary = SizeSummary::new();
+    summary.tier_stats.insert("WARM".to_string(), TierStats::default());
+    let object = ObjectInfo {
+        storage_class: Some("retired-tier".to_string()),
+        size: 11,
+        is_latest: true,
+        ..Default::default()
+    };
+
+    summary.actions_accounting(&object, 11, 11);
+
+    assert_eq!(summary.tier_stats.len(), 2);
+    assert_eq!(summary.tier_stats.get(UNKNOWN_TIER).map(|stats| stats.total_size), Some(11));
+    assert_eq!(summary.unknown_tier_stats.unknown_bytes, 11);
+    assert_eq!(summary.unknown_tier_stats.unknown_physical_bytes, 11);
+    assert_eq!(summary.unknown_tier_stats.unknown_objects, 1);
+    assert_eq!(summary.tier_accounting_proof.logical_total, 11);
+    assert_eq!(summary.tier_accounting_proof.logical_known, 0);
+    assert_eq!(summary.tier_accounting_proof.physical_total, 11);
+    assert_eq!(summary.tier_accounting_proof.physical_known, 0);
+    assert!(summary.unknown_tier_stats.diagnostics.len() <= UNKNOWN_TIER_DIAGNOSTIC_ENTRY_CAP);
+    assert!(summary.unknown_tier_stats.diagnostics.iter().map(String::len).sum::<usize>() <= UNKNOWN_TIER_DIAGNOSTIC_BYTE_CAP);
+    assert!(
+        summary
+            .unknown_tier_stats
+            .diagnostics
+            .iter()
+            .all(|entry| !entry.contains("retired"))
+    );
+}
+
+#[test]
+fn unknown_tier_is_accounted_when_no_remote_tier_is_configured() {
+    let mut summary = SizeSummary::new();
+    let object = ObjectInfo {
+        storage_class: Some("retired-tier".to_string()),
+        size: 3,
+        is_latest: true,
+        ..Default::default()
+    };
+
+    summary.actions_accounting(&object, 9, 9);
+
+    assert_eq!(summary.tier_stats.len(), 1);
+    assert_eq!(summary.tier_stats[UNKNOWN_TIER].total_size, 3);
+    assert_eq!(summary.unknown_tier_stats.unknown_bytes, 9);
+    assert_eq!(summary.unknown_tier_stats.unknown_physical_bytes, 3);
+
+    let standard = ObjectInfo {
+        storage_class: Some(storageclass::STANDARD.to_string()),
+        size: 4,
+        is_latest: true,
+        ..Default::default()
+    };
+    summary.actions_accounting(&standard, 4, 4);
+    assert_eq!(summary.tier_accounting_proof.logical_total, 13);
+    assert_eq!(summary.tier_accounting_proof.logical_known, 4);
+    assert_eq!(summary.tier_accounting_proof.physical_total, 3 + 4);
+    assert_eq!(summary.tier_accounting_proof.physical_known, 4);
+    assert_eq!(summary.tier_stats.len(), 1, "built-ins preserve the no-tier map shape");
+}
+
+#[test]
+fn million_unique_tier_keys_do_not_grow_stats_map() {
+    let mut summary = SizeSummary::new();
+    summary.tier_stats.insert("WARM".to_string(), TierStats::default());
+    for index in 0..1_000_000_u64 {
+        let object = ObjectInfo {
+            storage_class: Some(format!("untrusted-tier-{index}")),
+            size: 1,
+            ..Default::default()
+        };
+        summary.actions_accounting(&object, 1, 1);
+    }
+
+    assert_eq!(summary.tier_stats.len(), 2);
+    assert_eq!(summary.tier_stats[UNKNOWN_TIER].total_size, 1_000_000);
+    assert_eq!(summary.unknown_tier_stats.unknown_bytes, 1_000_000);
+    assert!(summary.unknown_tier_stats.diagnostics.len() <= UNKNOWN_TIER_DIAGNOSTIC_ENTRY_CAP);
+    assert!(summary.unknown_tier_stats.diagnostics.iter().map(String::len).sum::<usize>() <= UNKNOWN_TIER_DIAGNOSTIC_BYTE_CAP);
+}
+
+#[test]
+fn unknown_tier_never_triggers_transition() {
+    let mut summary = SizeSummary::new();
+    summary.tier_stats.insert("WARM".to_string(), TierStats::default());
+    let mut object = ObjectInfo {
+        storage_class: Some("removed-tier".to_string()),
+        size: 7,
+        ..Default::default()
+    };
+    object.transitioned_object.status = TRANSITION_COMPLETE.to_string();
+    object.transitioned_object.tier = "removed-tier".to_string();
+
+    summary.actions_accounting(&object, 7, 7);
+
+    assert_eq!(summary.tier_stats.get("removed-tier"), None);
+    assert_eq!(summary.tier_stats[UNKNOWN_TIER].total_size, 7);
+}
+
+#[test]
+fn removed_tier_survives_restart_as_unknown() {
+    let mut summary = SizeSummary::new();
+    summary.tier_stats.insert("COLD".to_string(), TierStats::default());
+    summary.tier_stats.insert(
+        "RETIRED".to_string(),
+        TierStats {
+            total_size: 5,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    );
+    let object = ObjectInfo {
+        storage_class: Some("COLD".to_string()),
+        size: 5,
+        ..Default::default()
+    };
+    summary.actions_accounting(&object, 5, 5);
+    let mut entry = DataUsageEntry::default();
+    entry.add_tier_sizes(&summary.tier_stats);
+    entry.add_unknown_tier_stats(&UnknownTierStats {
+        unknown_bytes: 2,
+        unknown_physical_bytes: 2,
+        unknown_objects: 1,
+        unknown_versions: 1,
+        ..Default::default()
+    });
+    let encoded = rmp_serde::to_vec(&entry).expect("entry should encode");
+    let restored: DataUsageEntry = rmp_serde::from_slice(&encoded).expect("entry should decode");
+    assert_eq!(restored.unknown_tier_stats.as_ref().map(|stats| stats.unknown_bytes), Some(2));
+    assert_eq!(
+        restored.all_tier_stats.as_ref().expect("tier stats persisted").tiers["RETIRED"].total_size,
+        5
+    );
+
+    let mut cache = DataUsageCache::default();
+    cache.replace("bucket", "", restored);
+    cache.fold_retired_tiers(&["COLD".to_string()]);
+    let folded = cache.cache.get(&hash_path("bucket").key()).expect("folded cache entry");
+    assert_eq!(
+        folded.all_tier_stats.as_ref().expect("tier stats persisted").tiers[UNKNOWN_TIER].total_size,
+        5
+    );
+    assert_eq!(folded.unknown_tier_stats.as_ref().map(|stats| stats.unknown_bytes), Some(2));
+}
+
+#[test]
+fn retired_tier_fold_is_idempotent_and_rejects_mixed_companion_provenance() {
+    let mut cache = DataUsageCache::default();
+    let mut entry = DataUsageEntry {
+        all_tier_stats: Some(AllTierStats {
+            tiers: HashMap::from([(
+                "RETIRED".to_string(),
+                TierStats {
+                    total_size: 5,
+                    num_versions: 1,
+                    num_objects: 1,
+                },
+            )]),
+        }),
+        ..Default::default()
+    };
+    entry.unknown_tier_stats = Some(UnknownTierStats {
+        unknown_physical_bytes: 5,
+        ..Default::default()
+    });
+    entry.tier_accounting_proof = Some(TierAccountingProof {
+        physical_total: 5,
+        physical_known: 5,
+        ..Default::default()
+    });
+    cache.replace("bucket", "", entry);
+
+    cache.fold_retired_tiers(&["COLD".to_string()]);
+    let first = cache.cache.get(&hash_path("bucket").key()).expect("entry").clone();
+    assert_eq!(first.all_tier_stats.as_ref().expect("tiers").tiers[UNKNOWN_TIER].total_size, 5);
+    assert_eq!(first.unknown_tier_stats.as_ref().expect("companion").unknown_physical_bytes, 5);
+    assert!(first.tier_accounting_proof.is_none(), "mixed provenance must not publish");
+
+    cache.fold_retired_tiers(&["COLD".to_string()]);
+    let second = cache.cache.get(&hash_path("bucket").key()).expect("entry");
+    assert_eq!(second.all_tier_stats.as_ref().expect("tiers").tiers[UNKNOWN_TIER].total_size, 5);
+    assert_eq!(second.unknown_tier_stats.as_ref().expect("companion").unknown_physical_bytes, 5);
+}
+
+#[test]
+fn tier_registry_refresh_does_not_mix_cycle_generations() {
+    let first = crate::TierRegistrySnapshot {
+        generation: 1,
+        names: Arc::from(["WARM".to_string()]),
+        refresh_failed: false,
+    };
+    let second = crate::TierRegistrySnapshot {
+        generation: 2,
+        names: Arc::from(["COLD".to_string()]),
+        refresh_failed: false,
+    };
+    assert_ne!(first.generation, second.generation);
+    assert_eq!(first.names.as_ref(), ["WARM".to_string()]);
+    assert_eq!(second.names.as_ref(), ["COLD".to_string()]);
+    assert!(first.refreshed(Err(())).refresh_failed);
+    assert!(!second.refreshed(Ok(Arc::from(["HOT".to_string()]))).refresh_failed);
+    assert_eq!(first.refreshed(Err(())).generation, first.generation);
+    assert_eq!(first.refreshed(Err(())).names, first.names);
+}
+
+#[test]
+fn unknown_tier_counter_uses_checked_arithmetic() {
+    let max = TierStats {
+        total_size: u64::MAX,
+        num_versions: u64::MAX,
+        num_objects: u64::MAX,
+    };
+    assert!(max.checked_add(&TierStats::default()).is_some());
+    assert!(
+        max.checked_add(&TierStats {
+            total_size: 1,
+            ..Default::default()
+        })
+        .is_none()
+    );
+
+    let mut unknown = UnknownTierStats {
+        unknown_bytes: u64::MAX,
+        ..Default::default()
+    };
+    unknown.record("overflow", 1, 1, 1);
+    assert_eq!(unknown.unknown_bytes, u64::MAX);
+    assert_eq!(unknown.unknown_objects, 1);
+    assert!(unknown.counter_overflowed);
+    assert!(unknown.checked_add(&UnknownTierStats::default()).is_none());
+    assert!(
+        unknown
+            .checked_add(&UnknownTierStats {
+                unknown_bytes: 1,
+                ..Default::default()
+            })
+            .is_none()
+    );
+}
+
 #[test]
 fn size_summary_unknown_accounting_keeps_physical_tier_and_version_only() {
     let mut summary = SizeSummary::default();

crates/scanner/src/lib.rs

@@ -24,8 +24,11 @@
 use bytes::Bytes;
 use http::HeaderMap;
 use rustfs_config::server_config::{Config as ServerConfig, get_global_server_config as config_get_global_server_config};
+use sha2::{Digest as _, Sha256};
+use std::collections::HashMap;
 use std::path::PathBuf;
 use std::sync::Arc;
+use std::sync::LazyLock;
 use std::sync::RwLock;
 use std::time::{Duration, Instant};
 use storage_api::owner::{
@@ -96,6 +99,45 @@ static SCANNER_RUNTIME_INSTANCES: AtomicU64 = AtomicU64::new(0);
 static SCANNER_FOREGROUND_READ_ACTIVITY: AtomicU64 = AtomicU64::new(0);
 static SCANNER_FOREGROUND_STREAM_READS: AtomicU64 = AtomicU64::new(0);
 
+/// Immutable tier registry captured at the beginning of a folder scan.
+/// Generation makes it possible to prove that a result was classified against
+/// one registry even when the process-wide TTL cache refreshes concurrently.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub(crate) struct TierRegistrySnapshot {
+    pub(crate) generation: u64,
+    pub(crate) names: Arc<[String]>,
+    /// True when the last refresh attempt failed and `names` is therefore a
+    /// retained last-good snapshot rather than a newly read registry.
+    pub(crate) refresh_failed: bool,
+}
+
+impl TierRegistrySnapshot {
+    /// Apply a refresh only when the registry read succeeds. A failed refresh
+    /// retains the prior generation, preventing a transient config failure
+    /// from classifying the remainder of a scan against an empty registry.
+    pub(crate) fn refreshed(&self, names: Result<Arc<[String]>, ()>) -> Self {
+        match names {
+            Ok(names) => Self {
+                generation: self.generation.saturating_add(1),
+                names,
+                refresh_failed: false,
+            },
+            Err(()) => Self {
+                refresh_failed: true,
+                ..self.clone()
+            },
+        }
+    }
+
+    pub(crate) fn initial(names: Arc<[String]>) -> Self {
+        Self {
+            generation: 1,
+            names,
+            refresh_failed: false,
+        }
+    }
+}
+
 pub fn current_scanner_activity() -> u64 {
     SCANNER_ACTIVE_WORK_UNITS.load(Ordering::Relaxed)
 }
@@ -373,6 +415,7 @@ pub(crate) fn resolve_scanner_server_config() -> Option<ServerConfig> {
 /// How long the scanner caches the runtime tier-name list before re-reading
 /// the tier configuration manager.
 const TIER_NAME_CACHE_TTL: Duration = Duration::from_secs(30);
+const MAX_TIER_REGISTRY_NAME_BYTES: usize = 256;
 
 /// Process-wide TTL cache of runtime tier names.
 ///
@@ -385,24 +428,192 @@ const TIER_NAME_CACHE_TTL: Duration = Duration::from_secs(30);
 /// `TIER_NAME_CACHE_TTL` later; a removed tier can leave an all-zero
 /// `TierStats` seed behind for one cache generation, which merges harmlessly
 /// by key in per-object accounting and disappears on the next refresh.
-static TIER_NAME_CACHE: RwLock<Option<(Instant, Arc<[String]>)>> = RwLock::new(None);
+static TIER_NAME_CACHE: RwLock<Option<(Instant, TierRegistrySnapshot)>> = RwLock::new(None);
+static TIER_REGISTRY_GENERATION: AtomicU64 = AtomicU64::new(0);
+static TIER_CYCLE_SNAPSHOTS: LazyLock<RwLock<HashMap<(u64, u64), TierRegistrySnapshot>>> =
+    LazyLock::new(|| RwLock::new(HashMap::new()));
+static TIER_ACTIVE_CYCLES: LazyLock<RwLock<HashMap<(u64, u64), usize>>> = LazyLock::new(|| RwLock::new(HashMap::new()));
+static TIER_NAME_REFRESH_LOCK: LazyLock<tokio::sync::Mutex<()>> = LazyLock::new(|| tokio::sync::Mutex::new(()));
+
+/// Return one immutable registry snapshot for a scanner unit of work.
+pub(crate) async fn runtime_tier_registry() -> TierRegistrySnapshot {
+    {
+        let cached = TIER_NAME_CACHE.read().unwrap_or_else(|err| err.into_inner()).clone();
+        if let Some((refreshed_at, snapshot)) = cached
+            && refreshed_at.elapsed() < TIER_NAME_CACHE_TTL
+        {
+            return snapshot;
+        }
+    }
+
+    // Serialize refreshes so a slower read of the old config cannot overwrite
+    // a newer snapshot published by a concurrent caller.
+    let _refresh_guard = TIER_NAME_REFRESH_LOCK.lock().await;
+    {
+        let cached = TIER_NAME_CACHE.read().unwrap_or_else(|err| err.into_inner()).clone();
+        if let Some((refreshed_at, snapshot)) = cached
+            && refreshed_at.elapsed() < TIER_NAME_CACHE_TTL
+        {
+            return snapshot;
+        }
+    }
+
+    let previous = TIER_NAME_CACHE
+        .read()
+        .unwrap_or_else(|err| err.into_inner())
+        .as_ref()
+        .map(|(_, snapshot)| snapshot.clone());
+    let names = ecstore_get_global_tier_config_mgr()
+        .read()
+        .await
+        .list_tiers()
+        .into_iter()
+        .map(|tier| tier.name)
+        .collect::<Vec<_>>();
+    let snapshot = match validate_tier_registry_names(names) {
+        Ok(names) => {
+            let generation = next_tier_registry_generation();
+            match previous {
+                Some(previous) => TierRegistrySnapshot {
+                    generation,
+                    ..previous.refreshed(Ok(names))
+                },
+                None => TierRegistrySnapshot {
+                    generation,
+                    ..TierRegistrySnapshot::initial(names)
+                },
+            }
+        }
+        Err(()) => match previous {
+            Some(previous) => previous.refreshed(Err(())),
+            None => TierRegistrySnapshot {
+                generation: next_tier_registry_generation(),
+                names: Arc::new([]),
+                refresh_failed: true,
+            },
+        },
+    };
+    *TIER_NAME_CACHE.write().unwrap_or_else(|err| err.into_inner()) = Some((Instant::now(), snapshot.clone()));
+    snapshot
+}
+
+fn next_tier_registry_generation() -> u64 {
+    TIER_REGISTRY_GENERATION
+        .fetch_update(Ordering::AcqRel, Ordering::Relaxed, |current| Some(current.saturating_add(1)))
+        .unwrap_or(u64::MAX)
+}
+
+fn validate_tier_registry_names(mut names: Vec<String>) -> Result<Arc<[String]>, ()> {
+    if names.iter().any(|name| {
+        name.is_empty()
+            || name.len() > MAX_TIER_REGISTRY_NAME_BYTES
+            || name.bytes().any(|byte| byte.is_ascii_control())
+            || name == UNKNOWN_TIER
+            || name == storageclass::STANDARD
+            || name == storageclass::RRS
+    }) {
+        return Err(());
+    }
+    names.sort_unstable();
+    if names.windows(2).any(|pair| pair[0] == pair[1]) {
+        return Err(());
+    }
+    Ok(names.into())
+}
 
 /// Tier names currently registered in the tier configuration, cached for
 /// `TIER_NAME_CACHE_TTL`.
 pub(crate) async fn runtime_tier_names() -> Arc<[String]> {
+    runtime_tier_registry().await.names
+}
+
+/// Return the immutable tier registry for one scanner cycle/leader pair.
+/// Different buckets and disks belonging to the same cycle share this entry,
+/// so a TTL refresh cannot split one published cycle across generations.
+pub(crate) async fn runtime_tier_registry_for_cycle(cycle: u64, leader_epoch: u64) -> TierRegistrySnapshot {
+    let key = (cycle, leader_epoch);
+    prune_inactive_tier_cycle_snapshots(cycle, leader_epoch);
     {
-        let cached = TIER_NAME_CACHE.read().unwrap_or_else(|err| err.into_inner()).clone();
-        if let Some((refreshed_at, names)) = cached
-            && refreshed_at.elapsed() < TIER_NAME_CACHE_TTL
-        {
-            return names;
+        let cached = TIER_CYCLE_SNAPSHOTS.read().unwrap_or_else(|err| err.into_inner());
+        if let Some(snapshot) = cached.get(&key) {
+            return snapshot.clone();
         }
     }
 
-    let tiers = ecstore_get_global_tier_config_mgr().read().await.list_tiers();
-    let names: Arc<[String]> = tiers.iter().map(|tier| tier.name.clone()).collect::<Vec<_>>().into();
-    *TIER_NAME_CACHE.write().unwrap_or_else(|err| err.into_inner()) = Some((Instant::now(), Arc::clone(&names)));
-    names
+    let mut snapshot = runtime_tier_registry().await;
+    // The registry generation describes the configuration snapshot, not the
+    // scan that consumed it. Keep it stable across cycles so a healthy cache
+    // can be reused; cycle and leader fencing are carried separately by the
+    // cache metadata and scan plan.
+    snapshot.generation = tier_registry_generation(&snapshot.names);
+    let mut cached = TIER_CYCLE_SNAPSHOTS.write().unwrap_or_else(|err| err.into_inner());
+    if let Some(existing) = cached.get(&key) {
+        return existing.clone();
+    }
+    cached.insert(key, snapshot.clone());
+    snapshot
+}
+
+fn prune_inactive_tier_cycle_snapshots(cycle: u64, leader_epoch: u64) {
+    let active = TIER_ACTIVE_CYCLES.read().unwrap_or_else(|err| err.into_inner());
+    let mut snapshots = TIER_CYCLE_SNAPSHOTS.write().unwrap_or_else(|err| err.into_inner());
+    snapshots.retain(|(entry_cycle, entry_epoch), _| {
+        active.contains_key(&(*entry_cycle, *entry_epoch))
+            || *entry_epoch > leader_epoch
+            || (*entry_epoch == leader_epoch && *entry_cycle >= cycle)
+    });
+}
+
+pub(crate) struct TierRegistryCycleGuard {
+    key: (u64, u64),
+}
+
+impl Drop for TierRegistryCycleGuard {
+    fn drop(&mut self) {
+        let mut active = TIER_ACTIVE_CYCLES.write().unwrap_or_else(|err| err.into_inner());
+        if let Some(count) = active.get_mut(&self.key) {
+            *count = count.saturating_sub(1);
+            if *count == 0 {
+                active.remove(&self.key);
+            }
+        }
+    }
+}
+
+pub(crate) fn begin_tier_registry_cycle(cycle: u64, leader_epoch: u64) -> TierRegistryCycleGuard {
+    let mut active = TIER_ACTIVE_CYCLES.write().unwrap_or_else(|err| err.into_inner());
+    let count = active.entry((cycle, leader_epoch)).or_default();
+    *count = count.saturating_add(1);
+    TierRegistryCycleGuard {
+        key: (cycle, leader_epoch),
+    }
+}
+
+fn tier_registry_generation(names: &[String]) -> u64 {
+    let mut hasher = Sha256::new();
+    hasher.update(b"rustfs-tier-registry-v1");
+    for name in names {
+        hasher.update(u64::try_from(name.len()).unwrap_or(u64::MAX).to_le_bytes());
+        hasher.update(name.as_bytes());
+    }
+    let digest = hasher.finalize();
+    let mut prefix = [0_u8; 8];
+    prefix.copy_from_slice(&digest[..8]);
+    u64::from_le_bytes(prefix)
+}
+
+/// Drop cycle snapshots only after the scanner has finished publishing a
+/// cycle. In-flight or retryable cycles must retain their original registry;
+/// TTL/capacity eviction could make a later bucket in the same cycle refresh
+/// to a different generation.
+pub(crate) fn complete_tier_registry_cycle(cycle: u64, leader_epoch: u64) {
+    let active = TIER_ACTIVE_CYCLES.read().unwrap_or_else(|err| err.into_inner());
+    let mut cached = TIER_CYCLE_SNAPSHOTS.write().unwrap_or_else(|err| err.into_inner());
+    cached.retain(|(entry_cycle, entry_epoch), _| {
+        active.contains_key(&(*entry_cycle, *entry_epoch))
+            || *entry_epoch > leader_epoch
+            || (*entry_epoch == leader_epoch && *entry_cycle > cycle)
+    });
 }
 
 /// Test-only cache reset; the production cache has no invalidation hook
@@ -410,6 +621,9 @@ pub(crate) async fn runtime_tier_names() -> Arc<[String]> {
 #[cfg(test)]
 fn reset_tier_name_cache_for_test() {
     *TIER_NAME_CACHE.write().unwrap_or_else(|err| err.into_inner()) = None;
+    TIER_ACTIVE_CYCLES.write().unwrap_or_else(|err| err.into_inner()).clear();
+    TIER_CYCLE_SNAPSHOTS.write().unwrap_or_else(|err| err.into_inner()).clear();
+    TIER_REGISTRY_GENERATION.store(0, Ordering::Relaxed);
 }
 
 pub(crate) async fn enqueue_runtime_free_version(oi: ScannerObjectInfo) {
@@ -742,6 +956,68 @@ mod tests {
         assert!(Arc::ptr_eq(&first, &second));
     }
 
+    #[tokio::test]
+    async fn tier_registry_cycle_snapshot_stays_fixed_while_active() {
+        reset_tier_name_cache_for_test();
+        let cycle = 9_000_001;
+        let leader_epoch = 9_000_002;
+        let guard = begin_tier_registry_cycle(cycle, leader_epoch);
+        let first = runtime_tier_registry_for_cycle(cycle, leader_epoch).await;
+
+        // Simulate a TTL refresh observing a different configuration while the
+        // original cycle is still scanning. The active cycle entry must win.
+        *TIER_NAME_CACHE.write().unwrap_or_else(|err| err.into_inner()) = Some((
+            Instant::now() - TIER_NAME_CACHE_TTL - Duration::from_secs(1),
+            TierRegistrySnapshot {
+                generation: u64::MAX,
+                names: Arc::from(["COLD".to_string()]),
+                refresh_failed: false,
+            },
+        ));
+        let second = runtime_tier_registry_for_cycle(cycle, leader_epoch).await;
+        assert_eq!(second.generation, first.generation);
+        assert_eq!(second.names, first.names);
+
+        drop(guard);
+        complete_tier_registry_cycle(cycle, leader_epoch);
+        reset_tier_name_cache_for_test();
+    }
+
+    #[tokio::test]
+    async fn tier_registry_generation_survives_new_cycle_with_same_names() {
+        reset_tier_name_cache_for_test();
+        let first_cycle = 9_000_011;
+        let second_cycle = first_cycle + 1;
+        let leader_epoch = 9_000_012;
+
+        let first_guard = begin_tier_registry_cycle(first_cycle, leader_epoch);
+        let first = runtime_tier_registry_for_cycle(first_cycle, leader_epoch).await;
+        drop(first_guard);
+        complete_tier_registry_cycle(first_cycle, leader_epoch);
+
+        let second_guard = begin_tier_registry_cycle(second_cycle, leader_epoch);
+        let second = runtime_tier_registry_for_cycle(second_cycle, leader_epoch).await;
+        assert_eq!(first.names, second.names);
+        assert_eq!(first.generation, second.generation);
+
+        drop(second_guard);
+        complete_tier_registry_cycle(second_cycle, leader_epoch);
+        reset_tier_name_cache_for_test();
+    }
+
+    #[test]
+    fn invalid_tier_registry_names_fail_closed_for_refresh() {
+        assert!(validate_tier_registry_names(vec!["COLD\n".to_string()]).is_err());
+        assert!(validate_tier_registry_names(vec![UNKNOWN_TIER.to_string()]).is_err());
+        assert!(validate_tier_registry_names(vec!["COLD".to_string(), "COLD".to_string()]).is_err());
+        assert_eq!(
+            validate_tier_registry_names(vec!["WARM".to_string(), "COLD".to_string()])
+                .expect("valid registry names")
+                .as_ref(),
+            ["COLD".to_string(), "WARM".to_string()]
+        );
+    }
+
     #[test]
     fn foreground_read_guard_tracks_stream_lifetime() {
         reset_foreground_read_activity_for_test();

crates/scanner/src/remote_scanner/stream.rs

@@ -16,8 +16,8 @@
 use crate::RUSTFS_META_BUCKET;
 use crate::scanner_budget::{ScannerCycleBudget, ScannerCycleBudgetConfig};
 use crate::scanner_io::{
-    DataUsageCacheScanState, ScannerDiskScanOutcome, ScannerIODisk, acquire_scanner_cache_locks, cache_root_entry_info,
-    current_cache_root_or_prepare, scanner_set_disk_inventory,
+    DataUsageCacheReuseOptions, DataUsageCacheScanState, ScannerDiskScanOutcome, ScannerIODisk, acquire_scanner_cache_locks,
+    cache_root_entry_info, current_cache_root_or_prepare_with_generation, scanner_set_disk_inventory,
 };
 use crate::storage_api::owner::NS_SCANNER_PROTOCOL_VERSION;
 use crate::{
@@ -213,6 +213,7 @@ pub(crate) struct RemoteScannerScanSpec<'a> {
     pub(crate) session_id: Uuid,
     pub(crate) session_sequence: u64,
     pub(crate) scan_plan_digest: DataUsageScanPlanDigest,
+    pub(crate) tier_registry_generation: u64,
     pub(crate) skip_healing: bool,
     pub(crate) scan_mode: HealScanMode,
 }
@@ -223,6 +224,7 @@ struct RemoteScannerResponseExpectation<'a> {
     source: DataUsageCacheSource,
     next_cycle: u64,
     scan_plan_digest: DataUsageScanPlanDigest,
+    tier_registry_generation: u64,
 }
 
 #[derive(Debug)]
@@ -669,6 +671,11 @@ async fn scan_and_persist_local_bucket(
         scan_mode,
         ..
     } = request;
+    // Keep the worker's cycle snapshot alive through cache reuse, scanning,
+    // and persistence. Without the guard, a later cycle can prune this key
+    // while this request is still running and allow a second registry to be
+    // selected for the same cycle.
+    let _tier_cycle_guard = crate::begin_tier_registry_cycle(next_cycle, leader_epoch);
     let store = resolve_scanner_object_store_handle()
         .ok_or_else(|| RemoteScannerServerError::worker("remote namespace scanner object layer is unavailable"))?;
     validate_remote_scanner_request_fence_with_store(next_cycle, leader_epoch, store.clone())
@@ -704,7 +711,25 @@ async fn scan_and_persist_local_bucket(
     let revisions = cache.load_with_revisions(set.clone(), &cache_name).await.map_err(|err| {
         RemoteScannerServerError::worker(format!("remote namespace scanner cache load or revision lookup failed: {err}"))
     })?;
-    let scan_state = current_cache_root_or_prepare(&mut cache, &bucket, source, next_cycle, leader_epoch, scan_plan_digest, true);
+    // Remote workers use the same cycle-frozen registry as `scan_data_folder`.
+    // Requiring its generation here prevents a cache snapshot classified by an
+    // older registry from being reused before the folder scan gets a chance to
+    // refresh it.
+    let tier_registry_generation = crate::runtime_tier_registry_for_cycle(next_cycle, leader_epoch)
+        .await
+        .generation;
+    let scan_state = current_cache_root_or_prepare_with_generation(
+        &mut cache,
+        &bucket,
+        source,
+        next_cycle,
+        leader_epoch,
+        scan_plan_digest,
+        DataUsageCacheReuseOptions {
+            require_source: true,
+            tier_registry_generation: Some(tier_registry_generation),
+        },
+    );
     match scan_state {
         DataUsageCacheScanState::Current(usage) => {
             if guard.is_lock_lost() {
@@ -869,6 +894,7 @@ pub(crate) async fn scan_remote_bucket(
         session_id,
         session_sequence,
         scan_plan_digest,
+        tier_registry_generation,
         skip_healing,
         scan_mode,
     } = spec;
@@ -957,6 +983,7 @@ pub(crate) async fn scan_remote_bucket(
             source: expected_source,
             next_cycle,
             scan_plan_digest,
+            tier_registry_generation,
         },
         authenticator,
         rpc_deadline,
@@ -1012,6 +1039,7 @@ where
             source: expected_source,
             next_cycle: TEST_NEXT_CYCLE,
             scan_plan_digest: expected_scan_plan_digest,
+            tier_registry_generation: 0,
         },
         authenticator,
         Instant::now() + NS_SCANNER_MAX_RPC_LIFETIME,
@@ -1111,6 +1139,11 @@ where
                         "remote namespace scanner returned usage for a different bucket plan",
                     )));
                 }
+                if complete.usage.tier_registry_generation != Some(expected.tier_registry_generation) {
+                    return Err(RemoteScannerStreamError::reconciled(StorageError::other(
+                        "remote namespace scanner returned usage for a different tier registry generation",
+                    )));
+                }
                 if !complete.usage.entry.children.is_empty() {
                     return Err(RemoteScannerStreamError::reconciled(StorageError::other(
                         "remote namespace scanner returned non-flattened bucket usage",

crates/scanner/src/remote_scanner/stream/tests.rs

@@ -195,6 +195,7 @@ fn test_usage(bucket: &str, objects: usize) -> DataUsageEntryInfo {
         name: bucket.to_string(),
         parent: crate::DATA_USAGE_ROOT.to_string(),
         entry,
+        tier_registry_generation: Some(0),
     }
 }
 
@@ -751,6 +752,43 @@ async fn complete_terminal_frame_reconciles_progress_and_usage() {
     assert_eq!(budget.progress(), (3, 2));
 }
 
+#[tokio::test]
+async fn terminal_usage_from_a_different_tier_generation_is_rejected() {
+    let request_id = Uuid::new_v4();
+    let writer_auth = FrameAuthenticator::for_test(request_id);
+    let reader_auth = FrameAuthenticator::for_test(request_id);
+    let (mut writer, reader) = tokio::io::duplex(4096);
+    tokio::spawn(async move {
+        let mut usage = test_usage("bucket", 1);
+        usage.tier_registry_generation = Some(1);
+        let mut sequence = 0;
+        write_frame(
+            &mut writer,
+            &writer_auth,
+            &mut sequence,
+            &RemoteScannerFrame::terminal(
+                RemoteScannerProgress::default(),
+                RemoteScannerFrameResult::Complete(Box::new(RemoteScannerComplete {
+                    source: TEST_SOURCE,
+                    scan_plan_digest: TEST_PLAN_DIGEST,
+                    usage,
+                    pending_maintenance_work: false,
+                })),
+            ),
+        )
+        .await
+        .expect("terminal frame should write");
+    });
+
+    let parent = CancellationToken::new();
+    let budget = ScannerCycleBudget::new(&parent, ScannerCycleBudgetConfig::default());
+    let error = consume_remote_scanner_stream(reader, parent, budget, "bucket", TEST_SOURCE, TEST_PLAN_DIGEST, reader_auth)
+        .await
+        .expect_err("generation mismatch must fail closed");
+
+    assert!(error.to_string().contains("tier registry generation"));
+}
+
 #[tokio::test]
 async fn complete_terminal_frame_after_budget_expiry_is_partial() {
     let request_id = Uuid::new_v4();

crates/scanner/src/scanner_folder.rs

@@ -59,9 +59,9 @@ use tracing::{debug, error, warn};
 use crate::{
     Disk, DiskError, DiskInfoOptions, Evaluator, Event, LcEventSrc, ListPathRawOptions, ObjectOpts, ReplicationConfig,
     ReplicationHealObject, ReplicationQueueAdmission, ReplicationStatusType, STORAGE_FORMAT_FILE, ScannerDiskExt as _,
-    ScannerLifecycleConfigExt as _, ScannerVersioningConfigExt as _, StorageError, apply_expiry_rule, apply_transition_rule,
-    enqueue_runtime_newer_noncurrent, is_reserved_or_invalid_bucket, list_path_raw, path2_bucket_object,
-    path2_bucket_object_with_base_path, queue_replication_heal, scanner_is_erasure,
+    ScannerLifecycleConfigExt as _, ScannerVersioningConfigExt as _, StorageError, TierRegistrySnapshot, apply_expiry_rule,
+    apply_transition_rule, enqueue_runtime_newer_noncurrent, is_reserved_or_invalid_bucket, list_path_raw, path2_bucket_object,
+    path2_bucket_object_with_base_path, queue_replication_heal, runtime_tier_registry_for_cycle, scanner_is_erasure,
     scanner_replication_config_for_lifecycle_eval,
 };
 use crate::{ScannerObjectInfo as ObjectInfo, ScannerObjectToDelete as ObjectToDelete};
@@ -638,6 +638,20 @@ fn apply_scanner_size_summary(into: &mut DataUsageEntry, summary: &SizeSummary)
     }
 
     into.add_tier_sizes(&summary.tier_stats);
+    into.add_unknown_tier_stats(&summary.unknown_tier_stats);
+    into.tier_accounting_proof = match (into.tier_accounting_proof, Some(summary.tier_accounting_proof)) {
+        (Some(mut current), Some(next)) => {
+            current.saturating_add(next);
+            Some(current)
+        }
+        (Some(_), None) => None,
+        (None, next) => next,
+    };
+    if into.unknown_tier_stats.as_ref().is_some_and(|stats| stats.counter_overflowed)
+        && let Some(proof) = into.tier_accounting_proof.as_mut()
+    {
+        proof.overflowed = true;
+    }
 }
 
 fn data_usage_root_has_progress(root: &DataUsageEntry) -> bool {
@@ -648,6 +662,8 @@ fn data_usage_root_has_progress(root: &DataUsageEntry) -> bool {
         || root.delete_markers > 0
         || root.failed_objects > 0
         || root.replication_stats.is_some()
+        || root.all_tier_stats.as_ref().is_some_and(|stats| !stats.is_empty())
+        || root.unknown_tier_stats.as_ref().is_some_and(|stats| !stats.is_empty())
 }
 
 fn partial_cache_is_useful(root: &DataUsageEntry, pending_heals_changed: bool) -> bool {
@@ -682,6 +698,9 @@ pub struct FolderScanner {
     budget: Arc<ScannerCycleBudget>,
     skip_heal: Arc<std::sync::atomic::AtomicBool>,
     local_disk: Arc<Disk>,
+    /// Tier registry frozen for this folder scan. A refresh applies to the
+    /// next scan and cannot mix generations in one aggregate.
+    tier_registry: TierRegistrySnapshot,
     pending_heals_changed: bool,
     pending_size_reconciliation_keys: HashSet<String>,
     pending_size_reconciliation_scopes: HashSet<String>,
@@ -1450,7 +1469,11 @@ impl FolderScanner {
                     continue;
                 }
 
-                let sz = match self.local_disk.get_size(item.clone()).await {
+                let sz = match self
+                    .local_disk
+                    .get_size_with_tier_names(item.clone(), &self.tier_registry.names)
+                    .await
+                {
                     Ok(sz) => sz,
                     Err(e) => {
                         let failure_action = classify_get_size_failure(&item, &e);
@@ -2326,6 +2349,10 @@ pub async fn scan_data_folder(
 
     let failed_object_ttl = rustfs_utils::get_env_u32(ENV_FAILED_OBJECT_TTL_SECS, DEFAULT_FAILED_OBJECT_TTL_SECS) as u64;
     let failed_objects_max = rustfs_utils::get_env_u32(ENV_FAILED_OBJECTS_MAX, DEFAULT_FAILED_OBJECTS_MAX) as usize;
+    let tier_registry = runtime_tier_registry_for_cycle(cache.info.next_cycle, cache.info.leader_epoch).await;
+    let mut cache = cache;
+    cache.fold_retired_tiers(&tier_registry.names);
+    cache.info.tier_registry_generation = Some(tier_registry.generation);
 
     // Create folder scanner
     let mut scanner = FolderScanner {
@@ -2354,6 +2381,7 @@ pub async fn scan_data_folder(
         budget: budget.clone(),
         skip_heal,
         local_disk,
+        tier_registry,
         pending_heals_changed: false,
         pending_size_reconciliation_keys: HashSet::new(),
         pending_size_reconciliation_scopes: HashSet::new(),

crates/scanner/src/scanner_folder/item_actions.rs

@@ -566,11 +566,62 @@ impl ScannerItem {
         item.object_path()
     }
 
+    fn effective_tier(oi: &ObjectInfo) -> &str {
+        if oi.transitioned_object.status == crate::TRANSITION_COMPLETE {
+            oi.transitioned_object.tier.as_str()
+        } else {
+            oi.storage_class.as_deref().unwrap_or(crate::storageclass::STANDARD)
+        }
+    }
+
+    fn tier_name_is_known(tier: &str, tier_names: &[String]) -> bool {
+        !tier.is_empty()
+            && tier != crate::data_usage_define::UNKNOWN_TIER
+            && (tier == crate::storageclass::STANDARD
+                || tier == crate::storageclass::RRS
+                || tier_names.iter().any(|name| name == tier))
+    }
+
+    pub(crate) fn tier_is_known(oi: &ObjectInfo, tier_names: &[String]) -> bool {
+        Self::tier_name_is_known(Self::effective_tier(oi), tier_names)
+    }
+
+    fn action_requires_known_tier(action: IlmAction) -> bool {
+        matches!(
+            action,
+            IlmAction::TransitionAction
+                | IlmAction::TransitionVersionAction
+                | IlmAction::DeleteAction
+                | IlmAction::DeleteVersionAction
+                | IlmAction::DeleteRestoredAction
+                | IlmAction::DeleteRestoredVersionAction
+                | IlmAction::DeleteAllVersionsAction
+                | IlmAction::DelMarkerDeleteAllVersionsAction
+        )
+    }
+
+    fn action_blocked_by_unknown_tier(
+        action: IlmAction,
+        oi: &ObjectInfo,
+        all_versions_known: bool,
+        tier_names: &[String],
+        target: &str,
+    ) -> bool {
+        if !Self::action_requires_known_tier(action) {
+            return false;
+        }
+        !Self::tier_is_known(oi, tier_names)
+            || (action.delete_all() && !all_versions_known)
+            || (matches!(action, IlmAction::TransitionAction | IlmAction::TransitionVersionAction)
+                && !Self::tier_name_is_known(target, tier_names))
+    }
+
     pub async fn apply_actions(
         &mut self,
         object_infos: Vec<ObjectInfo>,
         lock_retention: Option<Arc<ObjectLockConfiguration>>,
         versioning_config: VersioningConfiguration,
+        tier_names: &[String],
         size_summary: &mut SizeSummary,
     ) {
         let object_path = self.object_path();
@@ -694,6 +745,9 @@ impl ScannerItem {
         let mut noncurrent_unknown: Vec<&ObjectInfo> = Vec::new();
         let mut cumulative_size = 0;
         let mut remaining_versions = object_infos.len();
+        let all_versions_known = object_infos
+            .iter()
+            .all(|candidate| Self::tier_is_known(candidate, tier_names));
         'eventLoop: {
             for (i, event) in events.iter().enumerate() {
                 let oi = &object_infos[i];
@@ -799,6 +853,18 @@ impl ScannerItem {
                 let mut size = actual_size;
                 let mut account_now = true;
 
+                // A retired/unknown source tier may point at a remote object
+                // that cannot be safely deleted or transitioned. Lifecycle
+                // evaluation is still useful for accounting, but all
+                // side-effecting tier actions fail closed until the registry
+                // recognizes the source again.
+                if Self::action_blocked_by_unknown_tier(event.action, oi, all_versions_known, tier_names, &event.storage_class) {
+                    size = self.heal_actions(oi, actual_size, size_summary).await;
+                    size_summary.actions_accounting(oi, size, actual_size);
+                    cumulative_size += size;
+                    continue;
+                }
+
                 match event.action {
                     IlmAction::DeleteAllVersionsAction | IlmAction::DelMarkerDeleteAllVersionsAction => {
                         debug!(
@@ -1303,6 +1369,51 @@ mod tests {
         assert_eq!(item.object_path(), "object");
     }
 
+    #[test]
+    fn unknown_tier_never_triggers_transition() {
+        let object = ObjectInfo {
+            storage_class: Some("retired-tier".to_string()),
+            ..Default::default()
+        };
+        let tier_names = ["WARM".to_string()];
+        assert!(!ScannerItem::tier_is_known(&object, &tier_names));
+        assert!(ScannerItem::action_requires_known_tier(IlmAction::TransitionAction));
+        assert!(ScannerItem::action_requires_known_tier(IlmAction::DeleteVersionAction));
+        assert!(ScannerItem::action_blocked_by_unknown_tier(
+            IlmAction::TransitionAction,
+            &object,
+            false,
+            &tier_names,
+            "WARM"
+        ));
+        assert!(!ScannerItem::action_blocked_by_unknown_tier(
+            IlmAction::NoneAction,
+            &object,
+            false,
+            &tier_names,
+            "WARM"
+        ));
+
+        let known = ObjectInfo {
+            storage_class: Some(crate::storageclass::STANDARD.to_string()),
+            ..Default::default()
+        };
+        assert!(ScannerItem::action_blocked_by_unknown_tier(
+            IlmAction::DeleteAllVersionsAction,
+            &known,
+            false,
+            &tier_names,
+            "WARM"
+        ));
+        assert!(!ScannerItem::action_blocked_by_unknown_tier(
+            IlmAction::TransitionAction,
+            &known,
+            true,
+            &tier_names,
+            crate::storageclass::STANDARD
+        ));
+    }
+
     #[test]
     fn size_resolution_rejects_negative_overflow_and_unknown_compression() {
         let compressed = |actual_size: i64, declared: Option<&str>| {
@@ -1681,7 +1792,7 @@ mod tests {
             ..Default::default()
         };
         let mut summary = SizeSummary::default();
-        item.apply_actions(vec![object], None, VersioningConfiguration::default(), &mut summary)
+        item.apply_actions(vec![object], None, VersioningConfiguration::default(), &[], &mut summary)
             .await;
 
         let bounded_bucket = bounded_reconciliation_field(&item.bucket);

crates/scanner/src/scanner_folder/tests.rs

@@ -337,6 +337,11 @@ async fn build_test_scanner() -> (FolderScanner, std::path::PathBuf) {
         budget: ScannerCycleBudget::new(&CancellationToken::new(), Default::default()),
         skip_heal: Arc::new(AtomicBool::new(false)),
         local_disk: disk,
+        tier_registry: crate::TierRegistrySnapshot {
+            generation: 0,
+            names: Arc::new([]),
+            refresh_failed: false,
+        },
         pending_heals_changed: false,
         pending_size_reconciliation_keys: HashSet::new(),
         pending_size_reconciliation_scopes: HashSet::new(),

crates/scanner/src/scanner_io.rs

@@ -58,7 +58,8 @@ use crate::{
     BucketTargetSys, BucketVersioningSys, Disk, DiskError, ECStore, EcstoreError as Error, EcstoreResult as Result,
     RUSTFS_META_BUCKET, ReplicationConfig, STORAGE_FORMAT_FILE, ScannerConfigObjectDelete as _, ScannerDiskExt as _,
     ScannerLifecycleConfigExt as _, ScannerReplicationConfigExt as _, ScannerVersioningConfigExt as _, SetDisks, StorageError,
-    enqueue_runtime_free_version, get_lifecycle_config, get_object_lock_config, get_replication_config, runtime_tier_names,
+    begin_tier_registry_cycle, complete_tier_registry_cycle, enqueue_runtime_free_version, get_lifecycle_config,
+    get_object_lock_config, get_replication_config, runtime_tier_names, runtime_tier_registry_for_cycle,
     scanner_publication_admission_for_epoch, scanner_publication_epoch, storageclass,
 };
 
@@ -144,6 +145,7 @@ pub struct ScannerBucketScanPlan {
     all_buckets: Arc<Vec<BucketInfo>>,
     digest: DataUsageScanPlanDigest,
     leader_epoch: u64,
+    tier_registry_generation: u64,
     /// Epoch captured once for the whole scanner cycle.  `None` is retained
     /// for unfenced test implementations; production plans always carry the
     /// admission token captured before bucket enumeration.
@@ -357,12 +359,20 @@ pub(crate) fn cache_root_entry_info(cache: &DataUsageCache) -> std::result::Resu
         name: cache.info.name.clone(),
         parent: DATA_USAGE_ROOT.to_string(),
         entry,
+        tier_registry_generation: cache.info.tier_registry_generation,
     })
 }
 
-fn apply_bucket_result_to_cache(cache: &mut DataUsageCache, result: DataUsageEntryInfo, update_time: SystemTime) {
+fn apply_bucket_result_to_cache(cache: &mut DataUsageCache, result: DataUsageEntryInfo, update_time: SystemTime) -> bool {
+    if cache.info.tier_registry_generation != result.tier_registry_generation {
+        // A result from another registry generation must never be folded into
+        // this cycle. Leaving it unapplied makes the cycle incomplete and
+        // forces the caller to re-account it under one frozen registry.
+        return false;
+    }
     cache.replace(&result.name, &result.parent, result.entry);
     cache.info.last_update = Some(update_time);
+    true
 }
 
 fn should_publish_completed_snapshot(completed_count: usize, total_count: usize, budget_elapsed: bool, cancelled: bool) -> bool {
@@ -514,6 +524,9 @@ pub trait ScannerIODisk: Send + Sync + Debug + 'static {
     ) -> Result<ScannerDiskScanOutcome>;
 
     async fn get_size(&self, item: ScannerItem) -> Result<SizeSummary>;
+
+    /// Read one object using a registry snapshot captured at scan start.
+    async fn get_size_with_tier_names(&self, item: ScannerItem, tier_names: &[String]) -> Result<SizeSummary>;
 }
 
 #[derive(Debug)]
@@ -676,7 +689,10 @@ use cache::*;
 use dirty_usage::*;
 use guards::*;
 
-pub(crate) use cache::{DataUsageCacheScanState, acquire_scanner_cache_locks, current_cache_root_or_prepare};
+pub(crate) use cache::{
+    DataUsageCacheReuseOptions, DataUsageCacheScanState, acquire_scanner_cache_locks,
+    current_cache_root_or_prepare_with_generation,
+};
 pub use dirty_usage::{
     ScannerDirtyUsageAckError, ScannerDirtyUsageState, acknowledge_dirty_usage_generation, clear_dirty_usage_bucket,
     record_dirty_usage_bucket, record_scanner_maintenance_change, scanner_activity_epoch, scanner_dirty_usage_state,

crates/scanner/src/scanner_io/cache.rs

@@ -100,13 +100,14 @@ where
     let _ = tokio::time::timeout(SCANNER_CACHE_LOCK_LOSS_SHUTDOWN_TIMEOUT, scan).await;
 }
 
-pub(crate) fn current_cache_root_entry(
+pub(crate) fn current_cache_root_entry_with_generation(
     cache: &DataUsageCache,
     name: &str,
     source: DataUsageCacheSource,
     next_cycle: u64,
     leader_epoch: u64,
     scan_plan_digest: DataUsageScanPlanDigest,
+    tier_registry_generation: Option<u64>,
 ) -> std::result::Result<Option<DataUsageEntryInfo>, ScannerError> {
     let metadata_is_current = cache.info.name == name
         && cache.info.source == Some(source)
@@ -115,7 +116,8 @@ pub(crate) fn current_cache_root_entry(
         && cache.info.last_update.is_some()
         && cache.info.next_cycle == next_cycle
         && cache.info.leader_epoch == leader_epoch
-        && cache.info.cache_key_format == DATA_USAGE_CACHE_KEY_FORMAT;
+        && cache.info.cache_key_format == DATA_USAGE_CACHE_KEY_FORMAT
+        && tier_registry_generation.is_none_or(|generation| cache.info.tier_registry_generation == Some(generation));
     if !metadata_is_current {
         return Ok(None);
     }
@@ -131,6 +133,13 @@ pub(crate) enum DataUsageCacheScanState {
     },
 }
 
+#[derive(Clone, Copy, Debug, Default)]
+pub(crate) struct DataUsageCacheReuseOptions {
+    pub(crate) require_source: bool,
+    pub(crate) tier_registry_generation: Option<u64>,
+}
+
+#[cfg(test)]
 pub(crate) fn current_cache_root_or_prepare(
     cache: &mut DataUsageCache,
     name: &str,
@@ -140,11 +149,51 @@ pub(crate) fn current_cache_root_or_prepare(
     scan_plan_digest: DataUsageScanPlanDigest,
     require_source: bool,
 ) -> DataUsageCacheScanState {
-    match current_cache_root_entry(cache, name, source, next_cycle, leader_epoch, scan_plan_digest) {
+    current_cache_root_or_prepare_with_generation(
+        cache,
+        name,
+        source,
+        next_cycle,
+        leader_epoch,
+        scan_plan_digest,
+        DataUsageCacheReuseOptions {
+            require_source,
+            tier_registry_generation: None,
+        },
+    )
+}
+
+pub(crate) fn current_cache_root_or_prepare_with_generation(
+    cache: &mut DataUsageCache,
+    name: &str,
+    source: DataUsageCacheSource,
+    next_cycle: u64,
+    leader_epoch: u64,
+    scan_plan_digest: DataUsageScanPlanDigest,
+    options: DataUsageCacheReuseOptions,
+) -> DataUsageCacheScanState {
+    if options.tier_registry_generation.is_some_and(|generation| {
+        cache.info.next_cycle <= next_cycle
+            && cache.info.leader_epoch <= leader_epoch
+            && cache.info.tier_registry_generation != Some(generation)
+    }) {
+        // Make prepare_for_scan take its reset path so an entry classified by
+        // an older registry cannot be reused under the new cycle generation.
+        cache.info.scan_plan_digest = None;
+    }
+    match current_cache_root_entry_with_generation(
+        cache,
+        name,
+        source,
+        next_cycle,
+        leader_epoch,
+        scan_plan_digest,
+        options.tier_registry_generation,
+    ) {
         Ok(Some(root)) => DataUsageCacheScanState::Current(Box::new(root)),
         current => DataUsageCacheScanState::Prepared {
             invalid_current: current.err(),
-            outcome: cache.prepare_for_scan(name, next_cycle, leader_epoch, source, scan_plan_digest, require_source),
+            outcome: cache.prepare_for_scan(name, next_cycle, leader_epoch, source, scan_plan_digest, options.require_source),
         },
     }
 }
@@ -159,7 +208,7 @@ pub(super) fn cache_snapshot_is_current(
     scan_plan_digest: DataUsageScanPlanDigest,
 ) -> bool {
     matches!(
-        current_cache_root_entry(cache, name, source, next_cycle, leader_epoch, scan_plan_digest),
+        current_cache_root_entry_with_generation(cache, name, source, next_cycle, leader_epoch, scan_plan_digest, None),
         Ok(Some(_))
     )
 }
@@ -168,6 +217,7 @@ pub(super) fn completed_data_usage_info(
     results: &[DataUsageCache],
     expected_sources: &HashSet<DataUsageCacheSource>,
     all_buckets: &[String],
+    tier_registry_names: &[String],
     bucket_plan_complete: bool,
     budget_elapsed: bool,
     cancelled: bool,
@@ -183,6 +233,19 @@ pub(super) fn completed_data_usage_info(
         return None;
     }
 
+    // A generation is comparable across nodes because it is derived from the
+    // frozen registry names. Cycle and leader fencing remain separate cache
+    // metadata. Legacy peers omit the generation; an all-legacy result remains
+    // readable, but mixing legacy and new (or two new generations) would make
+    // the per-tier accounting ambiguous.
+    let registry_generation = results.first()?.info.tier_registry_generation;
+    if results.iter().any(|result| match registry_generation {
+        Some(generation) => result.info.tier_registry_generation != Some(generation),
+        None => result.info.tier_registry_generation.is_some(),
+    }) {
+        return None;
+    }
+
     if results.iter().any(|result| result.root().is_none()) {
         return None;
     }
@@ -200,9 +263,16 @@ pub(super) fn completed_data_usage_info(
         if !total.checked_merge(&merged) {
             return None;
         }
+        if !tier_accounting_proof_is_publishable(&merged, registry_generation, tier_registry_names) {
+            return None;
+        }
         bucket_entries.insert(bucket.clone(), merged);
     }
 
+    if !tier_accounting_proof_is_publishable(&total, registry_generation, tier_registry_names) {
+        return None;
+    }
+
     let merged_last_update = results.iter().filter_map(|result| result.info.last_update).max()?;
     let buckets_usage = bucket_entries
         .iter()
@@ -220,6 +290,7 @@ pub(super) fn completed_data_usage_info(
         delete_markers_total_count: u64::try_from(total.delete_markers).ok()?,
         objects_total_size: u64::try_from(total.size).ok()?,
         tier_stats: total.all_tier_stats.filter(|tiers| !tiers.is_empty()),
+        unknown_tier_stats: total.unknown_tier_stats.filter(|stats| !stats.is_empty()),
         buckets_count: u64::try_from(all_buckets.len()).ok()?,
         bucket_sizes,
         buckets_usage,
@@ -229,6 +300,109 @@ pub(super) fn completed_data_usage_info(
     Some((data_usage_info, merged_last_update))
 }
 
+fn tier_accounting_proof_is_publishable(
+    entry: &DataUsageEntry,
+    registry_generation: Option<u64>,
+    tier_registry_names: &[String],
+) -> bool {
+    let has_scalar_usage = entry.size > 0
+        || entry.objects > 0
+        || entry.versions > 0
+        || entry.delete_markers > 0
+        || entry.failed_objects > 0
+        || !entry.obj_sizes.is_empty()
+        || !entry.obj_versions.is_empty()
+        || entry.replication_stats.as_ref().is_some_and(|stats| !stats.is_empty());
+    let has_tier_accounted_data = entry
+        .all_tier_stats
+        .as_ref()
+        .is_some_and(|stats| stats.tiers.values().any(|tier| !tier.is_empty()))
+        || entry.unknown_tier_stats.as_ref().is_some_and(|stats| {
+            stats.counter_overflowed
+                || stats.unknown_bytes > 0
+                || stats.unknown_physical_bytes > 0
+                || stats.unknown_objects > 0
+                || stats.unknown_versions > 0
+        });
+
+    let Some(proof) = entry.tier_accounting_proof else {
+        return !has_scalar_usage && !has_tier_accounted_data;
+    };
+    if proof.overflowed
+        || entry
+            .unknown_tier_stats
+            .as_ref()
+            .is_some_and(|stats| stats.counter_overflowed)
+        || u64::try_from(entry.size).ok() != Some(proof.logical_total)
+    {
+        return false;
+    }
+    let unknown_logical = entry.unknown_tier_stats.as_ref().map_or(0, |stats| stats.unknown_bytes);
+    let unknown_physical = entry
+        .unknown_tier_stats
+        .as_ref()
+        .map_or(0, |stats| stats.unknown_physical_bytes);
+    if proof
+        .logical_known
+        .checked_add(unknown_logical)
+        .is_none_or(|total| total != proof.logical_total)
+        || proof
+            .physical_known
+            .checked_add(unknown_physical)
+            .is_none_or(|total| total != proof.physical_total)
+    {
+        return false;
+    }
+    if registry_generation.is_some()
+        && entry.all_tier_stats.as_ref().is_some_and(|stats| {
+            stats.tiers.keys().any(|tier| {
+                tier != crate::UNKNOWN_TIER
+                    && tier != crate::storageclass::STANDARD
+                    && tier != crate::storageclass::RRS
+                    && !tier_registry_names.iter().any(|allowed| allowed == tier)
+            })
+        })
+    {
+        return false;
+    }
+
+    if !has_tier_accounted_data {
+        return true;
+    }
+
+    let Some(tiers) = entry.all_tier_stats.as_ref() else {
+        return false;
+    };
+    let map_unknown_physical = tiers.tiers.get(crate::UNKNOWN_TIER).map_or(0, |stats| stats.total_size);
+    let companion_unknown_physical = entry
+        .unknown_tier_stats
+        .as_ref()
+        .map_or(0, |stats| stats.unknown_physical_bytes);
+    if map_unknown_physical != companion_unknown_physical {
+        return false;
+    }
+
+    // A no-configuration scan intentionally stores only UNKNOWN_TIER after
+    // the first unknown object; STANDARD/RRS remain absent to preserve the
+    // historical empty-map shape. In that shape the scalar proof is the sole
+    // source of known physical bytes. Configured registries seed at least one
+    // non-UNKNOWN key, whose map total must match the proof.
+    let has_known_tier_map = tiers.tiers.keys().any(|tier| tier.as_str() != crate::UNKNOWN_TIER);
+    if !has_known_tier_map {
+        return true;
+    }
+    let Some(known_tier_physical_total) = tiers
+        .tiers
+        .iter()
+        .filter(|(tier, _)| tier.as_str() != crate::UNKNOWN_TIER)
+        .map(|(_, stats)| stats)
+        .try_fold(0_u64, |total, stats| total.checked_add(stats.total_size))
+    else {
+        return false;
+    };
+    proof.physical_known == known_tier_physical_total
+}
+
 /// Build a non-authoritative view from the set snapshots that completed this
 /// cycle plus compatible per-set last-known-good caches.  The caller must
 /// persist this result only on the observational object; a missing set is
@@ -238,6 +412,7 @@ pub(super) fn observational_data_usage_info(
     results: &[DataUsageCache],
     expected_sources: &HashSet<DataUsageCacheSource>,
     all_buckets: &[String],
+    tier_registry_names: &[String],
     expected_plan_digest: DataUsageScanPlanDigest,
     scanner_cycle: u64,
     leader_epoch: u64,
@@ -316,6 +491,13 @@ pub(super) fn observational_data_usage_info(
     if usable.is_empty() {
         return None;
     }
+    let registry_generation = usable.first()?.0.info.tier_registry_generation;
+    if usable.iter().any(|(result, _)| match registry_generation {
+        Some(generation) => result.info.tier_registry_generation != Some(generation),
+        None => result.info.tier_registry_generation.is_some(),
+    }) {
+        return None;
+    }
 
     let mut total = DataUsageEntry::default();
     let mut bucket_entries = HashMap::with_capacity(all_buckets.len());
@@ -337,6 +519,15 @@ pub(super) fn observational_data_usage_info(
             }
         }
     }
+    if bucket_entries
+        .values()
+        .any(|entry| !tier_accounting_proof_is_publishable(entry, registry_generation, tier_registry_names))
+    {
+        return None;
+    }
+    if !tier_accounting_proof_is_publishable(&total, registry_generation, tier_registry_names) {
+        return None;
+    }
     let merged_last_update = merged_last_update?;
     let buckets_usage = bucket_entries
         .iter()
@@ -352,6 +543,7 @@ pub(super) fn observational_data_usage_info(
             delete_markers_total_count: u64::try_from(total.delete_markers).ok()?,
             objects_total_size: u64::try_from(total.size).ok()?,
             tier_stats: total.all_tier_stats.filter(|tiers| !tiers.is_empty()),
+            unknown_tier_stats: total.unknown_tier_stats.filter(|stats| !stats.is_empty()),
             buckets_count: u64::try_from(buckets_usage.len()).ok()?,
             bucket_sizes: buckets_usage
                 .iter()
@@ -463,13 +655,14 @@ pub(super) async fn persist_and_publish_cache_snapshot(
         return None;
     }
     if matches!(
-        current_cache_root_entry(
+        current_cache_root_entry_with_generation(
             &persisted,
             DATA_USAGE_ROOT,
             source,
             cache_snapshot.info.next_cycle,
             cache_snapshot.info.leader_epoch,
             scan_plan_digest,
+            cache_snapshot.info.tier_registry_generation,
         ),
         Ok(Some(_))
     ) {

crates/scanner/src/scanner_io/io_cache.rs

@@ -31,6 +31,7 @@ impl ScannerIOCache for SetDisks {
             all_buckets,
             digest: scan_plan_digest,
             leader_epoch,
+            tier_registry_generation,
             publication_epoch,
             dirty_usage_buckets,
             bucket_failures,
@@ -70,6 +71,7 @@ impl ScannerIOCache for SetDisks {
                     next_cycle: want_cycle,
                     last_update: Some(now),
                     leader_epoch,
+                    tier_registry_generation: Some(tier_registry_generation),
                     source: Some(source),
                     snapshot_complete: true,
                     scan_plan_digest: Some(scan_plan_digest),
@@ -267,7 +269,14 @@ impl ScannerIOCache for SetDisks {
         record_disk_bucket_scans_active(0, &pool_label, &set_label);
         let _reset_disk_bucket_scan_gauges = DiskBucketScanGaugeReset::new(pool_label.clone(), set_label.clone());
 
-        let old_lkg = old_cache.info.snapshot_complete.then(|| {
+        // Fence a stale set aggregate before copying entries into per-bucket work caches.
+        if old_cache.info.next_cycle <= want_cycle
+            && old_cache.info.leader_epoch <= leader_epoch
+            && old_cache.info.tier_registry_generation != Some(tier_registry_generation)
+        {
+            old_cache.info.scan_plan_digest = None;
+        }
+        let old_lkg = old_cache.info.snapshot_complete.then_some({
             (
                 old_cache.info.next_cycle,
                 old_cache.info.last_update,
@@ -333,6 +342,7 @@ impl ScannerIOCache for SetDisks {
                 name: DATA_USAGE_ROOT.to_string(),
                 next_cycle: want_cycle,
                 leader_epoch,
+                tier_registry_generation: Some(tier_registry_generation),
                 source: Some(source),
                 snapshot_complete: false,
                 scan_plan_digest: Some(scan_plan_digest),
@@ -394,8 +404,9 @@ impl ScannerIOCache for SetDisks {
                         };
 
                         let mut cache = cache_mutex_clone.lock().await;
-                        apply_bucket_result_to_cache(&mut cache, result, SystemTime::now());
-                        completed_bucket_count_clone.fetch_add(1, Ordering::Relaxed);
+                        if apply_bucket_result_to_cache(&mut cache, result, SystemTime::now()) {
+                            completed_bucket_count_clone.fetch_add(1, Ordering::Relaxed);
+                        }
                     }
                 }
             }
@@ -527,6 +538,7 @@ impl ScannerIOCache for SetDisks {
                                 session_id: remote_session_id,
                                 session_sequence: request_sequence,
                                 scan_plan_digest: bucket_scan_plan_digest,
+                                tier_registry_generation,
                                 skip_healing: healing,
                                 scan_mode,
                             },
@@ -742,14 +754,17 @@ impl ScannerIOCache for SetDisks {
                             continue;
                         }
                     };
-                    let scan_state = current_cache_root_or_prepare(
+                    let scan_state = current_cache_root_or_prepare_with_generation(
                         &mut cache,
                         &bucket.name,
                         source,
                         want_cycle,
                         leader_epoch,
                         bucket_scan_plan_digest,
-                        require_cache_source,
+                        DataUsageCacheReuseOptions {
+                            require_source: require_cache_source,
+                            tier_registry_generation: Some(tier_registry_generation),
+                        },
                     );
                     let outcome = match scan_state {
                         DataUsageCacheScanState::Current(root) => {
@@ -1237,6 +1252,7 @@ impl ScannerIOCache for SetDisks {
             incomplete_scope.info.next_cycle = want_cycle;
             incomplete_scope.info.last_update = None;
             incomplete_scope.info.leader_epoch = leader_epoch;
+            incomplete_scope.info.tier_registry_generation = Some(tier_registry_generation);
             incomplete_scope.info.source = Some(source);
             incomplete_scope.info.snapshot_complete = false;
             incomplete_scope.info.scan_plan_digest = Some(scan_plan_digest);

crates/scanner/src/scanner_io/io_cycle.rs

@@ -48,6 +48,7 @@ impl ScannerIOCycle for ECStore {
         scan_mode: HealScanMode,
     ) -> Result<ScannerCycleResult> {
         let child_token = ctx.child_token();
+        let _tier_cycle_guard = begin_tier_registry_cycle(want_cycle, leader_epoch);
 
         // Check the local pool metadata before listing buckets. A failed or
         // canceled decommission remains suspended after its worker exits, so
@@ -140,6 +141,8 @@ impl ScannerIOCycle for ECStore {
             scanner_bucket_plan_digest(&all_buckets, crate::scanner::scanner_activity_snapshot_digest(&activity_before));
         let dirty_usage_snapshot = Arc::new(snapshot_dirty_usage_buckets(&all_buckets, dirty_generation_before_bucket_list));
         let cache_cycle_floor = Arc::new(AtomicU64::new(want_cycle));
+        let tier_registry = runtime_tier_registry_for_cycle(want_cycle, leader_epoch).await;
+        let tier_registry_generation = tier_registry.generation;
 
         if all_buckets.is_empty() {
             reset_set_scan_gauges();
@@ -172,6 +175,9 @@ impl ScannerIOCycle for ECStore {
             {
                 return Ok(ScannerCycleResult::new(status, None).with_publication_epoch(publication_epoch));
             }
+            if status == ScannerCycleStatus::Complete {
+                complete_tier_registry_cycle(want_cycle, leader_epoch);
+            }
             let dirty_usage_clear =
                 (status == ScannerCycleStatus::Complete).then(|| dirty_usage_snapshot.buckets.as_ref().clone());
             let remote_dirty_usage_acknowledgements = if status == ScannerCycleStatus::Complete {
@@ -266,6 +272,7 @@ impl ScannerIOCycle for ECStore {
                     all_buckets: Arc::clone(&all_buckets),
                     digest: scan_plan_digest,
                     leader_epoch,
+                    tier_registry_generation,
                     publication_epoch,
                     dirty_usage_buckets: dirty_usage_snapshot.buckets.clone(),
                     bucket_failures: bucket_failures.clone(),
@@ -399,6 +406,7 @@ impl ScannerIOCycle for ECStore {
             &results,
             &expected_sources,
             &all_bucket_names,
+            &tier_registry.names,
             bucket_plan_complete,
             budget_elapsed,
             ctx.is_cancelled(),
@@ -410,6 +418,7 @@ impl ScannerIOCycle for ECStore {
                     &results,
                     &expected_sources,
                     &all_bucket_names,
+                    &tier_registry.names,
                     scan_plan_digest,
                     want_cycle,
                     leader_epoch,
@@ -441,6 +450,9 @@ impl ScannerIOCycle for ECStore {
             &failed_buckets,
         );
         result?;
+        if cycle_status == ScannerCycleStatus::Complete {
+            complete_tier_registry_cycle(want_cycle, leader_epoch);
+        }
         let remote_dirty_usage_acknowledgements = if cycle_status == ScannerCycleStatus::Complete {
             crate::scanner::scanner_dirty_usage_acknowledgements(&activity_before)
         } else {

crates/scanner/src/scanner_io/io_disk.rs

@@ -13,29 +13,38 @@
 // limitations under the License.
 /// ScannerIODisk implementation for Disk: get_size and the per-disk bucket scan.
 use super::*;
+use crate::UNKNOWN_TIER;
 
 ///
 /// Seed [`SizeSummary::tier_stats`] from the cached tier-name list.
 ///
-/// Preserves the original seeding semantics: with no tiers configured the map
-/// stays completely empty (STANDARD/RRS are not seeded either); otherwise the
-/// standard storage classes are seeded alongside every configured tier so
-/// per-object accounting always finds its tier key.
+/// Preserves the original no-tier shape: with no tiers configured the map
+/// stays completely empty (STANDARD/RRS/UNKNOWN are not seeded either).
+/// Otherwise the standard storage classes and one fixed unknown bucket are
+/// seeded alongside every configured tier so per-object accounting never
+/// inserts an untrusted metadata key.
 pub(super) fn tier_stats_template(tier_names: &[String]) -> HashMap<String, TierStats> {
-    let mut tier_stats = HashMap::with_capacity(tier_names.len() + 2);
+    let mut tier_stats = HashMap::with_capacity(tier_names.len() + 3);
     for tier_name in tier_names {
-        tier_stats.insert(tier_name.clone(), TierStats::default());
+        if tier_name != UNKNOWN_TIER {
+            tier_stats.insert(tier_name.clone(), TierStats::default());
+        }
     }
     if !tier_stats.is_empty() {
         tier_stats.insert(storageclass::STANDARD.to_string(), TierStats::default());
         tier_stats.insert(storageclass::RRS.to_string(), TierStats::default());
+        tier_stats.insert(UNKNOWN_TIER.to_string(), TierStats::default());
     }
     tier_stats
 }
 
 #[async_trait::async_trait]
 impl ScannerIODisk for Disk {
-    async fn get_size(&self, mut item: ScannerItem) -> Result<SizeSummary> {
+    async fn get_size(&self, item: ScannerItem) -> Result<SizeSummary> {
+        self.get_size_with_tier_names(item, &runtime_tier_names().await).await
+    }
+
+    async fn get_size_with_tier_names(&self, mut item: ScannerItem, tier_names: &[String]) -> Result<SizeSummary> {
         let done_object = Metrics::time(Metric::ScanObject);
 
         if !is_xl_meta_path(&item.path) {
@@ -105,12 +114,13 @@ impl ScannerIODisk for Disk {
             .map(|v| ObjectInfo::from_file_info(v, item.bucket.as_str(), object_path.as_str(), versioned))
             .collect::<Vec<ObjectInfo>>();
 
-        let mut size_summary = SizeSummary::default();
-
-        // Tier names come from the process-wide TTL cache; seeding from them
-        // replaces the per-object clone of every full TierConfig.
-        let tier_names = runtime_tier_names().await;
-        size_summary.tier_stats = tier_stats_template(&tier_names);
+        // The caller supplies one registry snapshot for the whole folder scan;
+        // seeding from it prevents a TTL refresh from mixing generations in a
+        // single result.
+        let mut size_summary = SizeSummary {
+            tier_stats: tier_stats_template(tier_names),
+            ..Default::default()
+        };
 
         let lock_config = object_lock_config_for_scanner_item(&item).await;
 
@@ -120,12 +130,14 @@ impl ScannerIODisk for Disk {
         // `object_infos`.
         global_metrics().record_scanner_versions_scanned(object_infos.len() as u64);
 
-        item.apply_actions(object_infos, lock_config, versioning_config, &mut size_summary)
+        item.apply_actions(object_infos, lock_config, versioning_config, tier_names, &mut size_summary)
             .await;
 
         if !free_version_infos.is_empty() {
             for oi in free_version_infos {
-                enqueue_runtime_free_version(oi).await;
+                if ScannerItem::tier_is_known(&oi, tier_names) {
+                    enqueue_runtime_free_version(oi).await;
+                }
             }
         }
 

crates/scanner/src/scanner_io/publish_gate_tests.rs

@@ -13,7 +13,8 @@
 // limitations under the License.
 
 use super::*;
-use rustfs_data_usage::{ReplicationAllStats, ReplicationTargetUsage};
+use crate::data_usage_define::{UNKNOWN_TIER, UnknownTierStats, hash_path};
+use rustfs_data_usage::{ReplicationAllStats, ReplicationTargetUsage, TierAccountingProof};
 
 const TEST_PLAN_DIGEST: DataUsageScanPlanDigest = DataUsageScanPlanDigest([7; 32]);
 
@@ -89,6 +90,11 @@ fn completed_root_cache(bucket: &str, objects: usize, update_secs: u64, source:
         DataUsageEntry {
             objects,
             size: objects.saturating_mul(10),
+            tier_accounting_proof: Some(TierAccountingProof {
+                logical_total: u64::try_from(objects.saturating_mul(10)).unwrap_or(u64::MAX),
+                logical_known: u64::try_from(objects.saturating_mul(10)).unwrap_or(u64::MAX),
+                ..Default::default()
+            }),
             ..Default::default()
         },
     );
@@ -102,7 +108,7 @@ fn completed_data_usage_info_for_test(
     cancelled: bool,
 ) -> Option<(DataUsageInfo, SystemTime)> {
     let expected_sources = results.iter().filter_map(|result| result.info.source).collect::<HashSet<_>>();
-    completed_data_usage_info(results, &expected_sources, all_buckets, true, budget_elapsed, cancelled)
+    completed_data_usage_info(results, &expected_sources, all_buckets, &[], true, budget_elapsed, cancelled)
 }
 
 fn lkg_root_cache(bucket: &str, objects: usize, source: DataUsageCacheSource) -> DataUsageCache {
@@ -130,9 +136,10 @@ fn partial_usage_is_observational_not_authoritative_for_quota() {
     let expected = HashSet::from([current_source, stalled_source]);
 
     assert!(
-        completed_data_usage_info(&[current.clone(), stalled.clone()], &expected, &all_buckets, true, false, false).is_none()
+        completed_data_usage_info(&[current.clone(), stalled.clone()], &expected, &all_buckets, &[], true, false, false)
+            .is_none()
     );
-    let (observed, _) = observational_data_usage_info(&[current, stalled], &expected, &all_buckets, TEST_PLAN_DIGEST, 8, 3)
+    let (observed, _) = observational_data_usage_info(&[current, stalled], &expected, &all_buckets, &[], TEST_PLAN_DIGEST, 8, 3)
         .expect("a completed set should produce an observational view");
     assert!(observed.usage_snapshot_partial);
     assert!(!observed.usage_snapshot_complete);
@@ -157,7 +164,7 @@ fn stale_quota_uses_complete_baseline_plus_positive_deltas() {
     current.info.next_cycle = 8;
     current.info.leader_epoch = 3;
     let expected = HashSet::from([source]);
-    let (observed, _) = observational_data_usage_info(&[current], &expected, &all_buckets, TEST_PLAN_DIGEST, 8, 3)
+    let (observed, _) = observational_data_usage_info(&[current], &expected, &all_buckets, &[], TEST_PLAN_DIGEST, 8, 3)
         .expect("complete set data is a valid observational baseline");
     assert_eq!(observed.objects_total_size, 30);
     assert_eq!(observed.usage_snapshot_set_states[0].complete, true);
@@ -170,7 +177,7 @@ fn negative_delta_waits_for_set_reconciliation() {
     let mut stalled = lkg_root_cache("bucket", 4, source);
     stalled.info.lkg_scan_plan_digest = Some(DataUsageScanPlanDigest([9; 32]));
     let expected = HashSet::from([source]);
-    assert!(observational_data_usage_info(&[stalled], &expected, &all_buckets, TEST_PLAN_DIGEST, 8, 3).is_none());
+    assert!(observational_data_usage_info(&[stalled], &expected, &all_buckets, &[], TEST_PLAN_DIGEST, 8, 3).is_none());
 }
 
 #[test]
@@ -220,7 +227,7 @@ fn old_set_completion_cannot_overwrite_new_aggregate() {
     old.info.next_cycle = 7;
     old.info.leader_epoch = 2;
     let expected = HashSet::from([source]);
-    assert!(observational_data_usage_info(&[old], &expected, &all_buckets, TEST_PLAN_DIGEST, 8, 3).is_none());
+    assert!(observational_data_usage_info(&[old], &expected, &all_buckets, &[], TEST_PLAN_DIGEST, 8, 3).is_none());
 }
 
 #[test]
@@ -231,7 +238,7 @@ fn usage_aggregate_survives_restart_and_leader_failover() {
     lkg.info.lkg_leader_epoch = Some(4);
     lkg.info.lkg_next_cycle = Some(9);
     let expected = HashSet::from([source]);
-    let (observed, _) = observational_data_usage_info(&[lkg], &expected, &all_buckets, TEST_PLAN_DIGEST, 10, 5)
+    let (observed, _) = observational_data_usage_info(&[lkg], &expected, &all_buckets, &[], TEST_PLAN_DIGEST, 10, 5)
         .expect("compatible LKG should survive a leader change");
     assert_eq!(observed.usage_snapshot_set_states[0].scanner_epoch, Some(4));
     assert_eq!(observed.objects_total_size, 50);
@@ -250,11 +257,11 @@ fn usage_aggregate_cost_is_linear_in_set_count() {
         cache.info.leader_epoch = 3;
         results.push(cache);
     }
-    let (observed, _) = observational_data_usage_info(&results, &expected, &all_buckets, TEST_PLAN_DIGEST, 8, 3)
+    let (observed, _) = observational_data_usage_info(&results, &expected, &all_buckets, &[], TEST_PLAN_DIGEST, 8, 3)
         .expect("all set snapshots should aggregate");
     assert_eq!(observed.objects_total_count, 32);
     let reversed = results.iter().rev().cloned().collect::<Vec<_>>();
-    let (reversed_observed, _) = observational_data_usage_info(&reversed, &expected, &all_buckets, TEST_PLAN_DIGEST, 8, 3)
+    let (reversed_observed, _) = observational_data_usage_info(&reversed, &expected, &all_buckets, &[], TEST_PLAN_DIGEST, 8, 3)
         .expect("reordered set snapshots should aggregate");
     assert_eq!(observed.usage_snapshot_set_states, reversed_observed.usage_snapshot_set_states);
 }
@@ -276,11 +283,21 @@ fn completed_data_usage_info_publishes_tier_stats_across_sets() {
     let mut first_set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
     let mut tiered = DataUsageEntry::default();
     tiered.add_tier_sizes(&warm(100, 2, 1));
+    tiered.tier_accounting_proof = Some(TierAccountingProof {
+        physical_total: 100,
+        physical_known: 100,
+        ..Default::default()
+    });
     first_set.replace("bucket-b", DATA_USAGE_ROOT, tiered);
 
     let mut second_set = completed_root_cache("bucket-b", 2, 20, DataUsageCacheSource::new(1, 0));
     let mut tiered = DataUsageEntry::default();
     tiered.add_tier_sizes(&warm(50, 1, 1));
+    tiered.tier_accounting_proof = Some(TierAccountingProof {
+        physical_total: 50,
+        physical_known: 50,
+        ..Default::default()
+    });
     second_set.replace("bucket-a", DATA_USAGE_ROOT, tiered);
 
     let (data_usage_info, _) = completed_data_usage_info_for_test(&[first_set, second_set], &all_buckets, false, false)
@@ -299,6 +316,269 @@ fn completed_data_usage_info_publishes_tier_stats_across_sets() {
     );
 }
 
+#[test]
+fn completed_data_usage_info_rejects_logical_proof_mismatch() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.add_tier_sizes(&HashMap::from([(
+        "WARM".to_string(),
+        TierStats {
+            total_size: 10,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    )]));
+    entry.tier_accounting_proof = Some(TierAccountingProof {
+        logical_total: 10,
+        logical_known: 9,
+        physical_total: 10,
+        physical_known: 10,
+        ..Default::default()
+    });
+
+    assert!(completed_data_usage_info_for_test(&[set], &all_buckets, false, false).is_none());
+}
+
+#[test]
+fn completed_data_usage_info_rejects_logical_total_size_mismatch() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.size = 11;
+    entry.add_tier_sizes(&HashMap::from([(
+        "WARM".to_string(),
+        TierStats {
+            total_size: 10,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    )]));
+    entry.tier_accounting_proof = Some(TierAccountingProof {
+        logical_total: 10,
+        logical_known: 10,
+        physical_total: 10,
+        physical_known: 10,
+        ..Default::default()
+    });
+
+    assert!(completed_data_usage_info_for_test(&[set], &all_buckets, false, false).is_none());
+}
+
+#[test]
+fn completed_data_usage_info_rejects_physical_proof_mismatch() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.add_tier_sizes(&HashMap::from([(
+        "WARM".to_string(),
+        TierStats {
+            total_size: 10,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    )]));
+    entry.tier_accounting_proof = Some(TierAccountingProof {
+        logical_total: 10,
+        logical_known: 10,
+        physical_total: 9,
+        physical_known: 9,
+        ..Default::default()
+    });
+
+    assert!(completed_data_usage_info_for_test(&[set], &all_buckets, false, false).is_none());
+}
+
+#[test]
+fn completed_data_usage_info_rejects_unknown_physical_double_accounting() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.add_tier_sizes(&HashMap::from([(
+        UNKNOWN_TIER.to_string(),
+        TierStats {
+            total_size: 10,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    )]));
+    entry.add_unknown_tier_stats(&UnknownTierStats {
+        unknown_physical_bytes: 9,
+        ..Default::default()
+    });
+    entry.tier_accounting_proof = Some(TierAccountingProof {
+        logical_total: 10,
+        logical_known: 10,
+        physical_total: 10,
+        physical_known: 10,
+        ..Default::default()
+    });
+
+    assert!(completed_data_usage_info_for_test(&[set], &all_buckets, false, false).is_none());
+}
+
+#[test]
+fn completed_data_usage_info_rejects_unknown_counter_overflow() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.add_unknown_tier_stats(&UnknownTierStats {
+        counter_overflowed: true,
+        ..Default::default()
+    });
+
+    assert!(completed_data_usage_info_for_test(&[set], &all_buckets, false, false).is_none());
+}
+
+#[test]
+fn completed_data_usage_info_accepts_no_tier_standard_empty_map_with_proof() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+
+    let (info, _) = completed_data_usage_info_for_test(&[set], &all_buckets, false, false)
+        .expect("no-tier STANDARD/RRS usage does not require a tier map");
+    assert!(info.tier_stats.is_none());
+}
+
+#[test]
+fn completed_data_usage_info_accepts_no_tier_unknown_and_standard_shape() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.size = 13;
+    entry.add_tier_sizes(&HashMap::from([(
+        UNKNOWN_TIER.to_string(),
+        TierStats {
+            total_size: 3,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    )]));
+    entry.add_unknown_tier_stats(&UnknownTierStats {
+        unknown_bytes: 9,
+        unknown_physical_bytes: 3,
+        unknown_objects: 1,
+        unknown_versions: 1,
+        ..Default::default()
+    });
+    entry.tier_accounting_proof = Some(TierAccountingProof {
+        logical_total: 13,
+        logical_known: 4,
+        physical_total: 7,
+        physical_known: 4,
+        ..Default::default()
+    });
+
+    let (info, _) = completed_data_usage_info_for_test(&[set], &all_buckets, false, false)
+        .expect("no-tier STANDARD plus UNKNOWN should remain publishable");
+    assert_eq!(
+        info.tier_stats.expect("unknown bucket should be retained").tiers[UNKNOWN_TIER].total_size,
+        3
+    );
+}
+
+#[test]
+fn completed_data_usage_info_accepts_unknown_only_with_current_registry_generation() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    set.info.tier_registry_generation = Some(7);
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.size = 13;
+    entry.add_tier_sizes(&HashMap::from([(
+        UNKNOWN_TIER.to_string(),
+        TierStats {
+            total_size: 3,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    )]));
+    entry.add_unknown_tier_stats(&UnknownTierStats {
+        unknown_bytes: 9,
+        unknown_physical_bytes: 3,
+        unknown_objects: 1,
+        unknown_versions: 1,
+        ..Default::default()
+    });
+    entry.tier_accounting_proof = Some(TierAccountingProof {
+        logical_total: 13,
+        logical_known: 4,
+        physical_total: 7,
+        physical_known: 4,
+        ..Default::default()
+    });
+
+    let expected_sources = HashSet::from([DataUsageCacheSource::new(0, 0)]);
+    assert!(
+        completed_data_usage_info(&[set], &expected_sources, &all_buckets, &["WARM".to_string()], true, false, false,).is_some()
+    );
+}
+
+#[test]
+fn completed_data_usage_info_rejects_non_registry_tier_in_current_generation() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    set.info.tier_registry_generation = Some(7);
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.add_tier_sizes(&HashMap::from([
+        (
+            "WARM".to_string(),
+            TierStats {
+                total_size: 4,
+                num_versions: 1,
+                num_objects: 1,
+            },
+        ),
+        (
+            "RETIRED".to_string(),
+            TierStats {
+                total_size: 6,
+                num_versions: 1,
+                num_objects: 1,
+            },
+        ),
+    ]));
+    entry.tier_accounting_proof = Some(TierAccountingProof {
+        logical_total: 10,
+        logical_known: 10,
+        physical_total: 10,
+        physical_known: 10,
+        ..Default::default()
+    });
+
+    let expected_sources = HashSet::from([DataUsageCacheSource::new(0, 0)]);
+    assert!(
+        completed_data_usage_info(&[set], &expected_sources, &all_buckets, &["WARM".to_string()], true, false, false,).is_none()
+    );
+}
+
+#[test]
+fn completed_data_usage_info_rejects_legacy_proof_missing_when_tier_accounted() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.add_tier_sizes(&HashMap::from([(
+        "WARM".to_string(),
+        TierStats {
+            total_size: 10,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    )]));
+    entry.tier_accounting_proof = None;
+
+    assert!(completed_data_usage_info_for_test(&[set], &all_buckets, false, false).is_none());
+}
+
+#[test]
+fn completed_data_usage_info_rejects_legacy_proof_missing_for_scalar_usage() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.tier_accounting_proof = None;
+
+    assert!(completed_data_usage_info_for_test(&[set], &all_buckets, false, false).is_none());
+}
+
 #[test]
 fn completed_data_usage_info_omits_tier_stats_without_tiered_objects() {
     let all_buckets = vec!["bucket-a".to_string()];
@@ -310,6 +590,51 @@ fn completed_data_usage_info_omits_tier_stats_without_tiered_objects() {
     assert!(data_usage_info.tier_stats.is_none());
 }
 
+#[test]
+fn completed_data_usage_info_rejects_legacy_and_new_tier_generations_mixed() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let legacy = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let mut current = completed_root_cache("bucket-a", 1, 20, DataUsageCacheSource::new(1, 0));
+    current.info.tier_registry_generation = Some(42);
+
+    assert!(
+        completed_data_usage_info_for_test(&[legacy, current], &all_buckets, false, false).is_none(),
+        "legacy and generation-tagged sets must not publish a mixed snapshot"
+    );
+}
+
+#[test]
+fn current_cache_root_with_new_tier_generation_resets_old_cache() {
+    let source = DataUsageCacheSource::new(0, 0);
+    let mut cache = completed_root_cache("bucket-a", 1, 10, source);
+    cache.info.tier_registry_generation = Some(1);
+
+    let state = current_cache_root_or_prepare_with_generation(
+        &mut cache,
+        DATA_USAGE_ROOT,
+        source,
+        0,
+        0,
+        TEST_PLAN_DIGEST,
+        DataUsageCacheReuseOptions {
+            require_source: false,
+            tier_registry_generation: Some(2),
+        },
+    );
+
+    assert!(matches!(
+        state,
+        DataUsageCacheScanState::Prepared {
+            outcome: DataUsageCachePrepareOutcome::Reset,
+            ..
+        }
+    ));
+    assert!(cache.cache.is_empty(), "old-generation entries must not be reused");
+    assert_eq!(cache.info.tier_registry_generation, None);
+    assert_eq!(cache.info.scan_plan_digest, Some(TEST_PLAN_DIGEST));
+    assert!(!cache.info.snapshot_complete);
+}
+
 #[test]
 fn completed_data_usage_info_requires_every_set_before_publish() {
     let all_buckets = vec!["bucket-a".to_string(), "bucket-b".to_string(), "bucket-empty".to_string()];
@@ -434,6 +759,13 @@ fn completed_data_usage_info_flattens_nested_bucket_entries() {
             replica_size: 2048,
             replica_count: 2,
         }),
+        tier_accounting_proof: Some(TierAccountingProof {
+            logical_total: 2048,
+            logical_known: 2048,
+            physical_total: 2048,
+            physical_known: 2048,
+            ..Default::default()
+        }),
         ..Default::default()
     };
     nested.obj_sizes.add(2048);
@@ -501,7 +833,8 @@ fn completed_data_usage_info_requires_exact_topology_sources() {
     let expected_sources = HashSet::from([DataUsageCacheSource::new(0, 0), DataUsageCacheSource::new(1, 0)]);
 
     assert!(
-        completed_data_usage_info(&[first_set, unexpected_set], &expected_sources, &all_buckets, true, false, false).is_none()
+        completed_data_usage_info(&[first_set, unexpected_set], &expected_sources, &all_buckets, &[], true, false, false)
+            .is_none()
     );
 }
 
@@ -511,7 +844,7 @@ fn completed_data_usage_info_rejects_incomplete_bucket_plan() {
     let set = completed_root_cache("bucket", 2, 10, DataUsageCacheSource::new(0, 0));
     let expected_sources = HashSet::from([DataUsageCacheSource::new(0, 0)]);
 
-    assert!(completed_data_usage_info(&[set], &expected_sources, &all_buckets, false, false, false).is_none());
+    assert!(completed_data_usage_info(&[set], &expected_sources, &all_buckets, &[], false, false, false).is_none());
 }
 
 #[test]

crates/scanner/src/scanner_io/tests.rs

@@ -23,7 +23,7 @@ use crate::storage_api::owner::{
 use crate::storage_api::scan::{BucketOperations as _, DeleteBucketOptions, MakeBucketOptions, ObjectIO as _};
 use crate::{
     DiskOption, ECStore, Endpoint, EndpointServerPools, Endpoints, InstanceContext, PoolEndpoints, ScannerObjectOptions,
-    ScannerPutObjReader, init_bucket_metadata_sys_for_scanner_tests, init_ecstore_config_for_scanner_tests,
+    ScannerPutObjReader, UNKNOWN_TIER, init_bucket_metadata_sys_for_scanner_tests, init_ecstore_config_for_scanner_tests,
     init_local_disks_with_instance_ctx, new_disk, path2_bucket_object_with_base_path,
 };
 use rustfs_filemeta::FileInfo;
@@ -1030,8 +1030,8 @@ fn is_xl_meta_path_accepts_forward_separator() {
 fn tier_stats_template_seeds_tiers_and_standard_classes() {
     let template = tier_stats_template(&["WARM".to_string(), "COLD".to_string()]);
 
-    assert_eq!(template.len(), 4);
-    for tier in ["WARM", "COLD", storageclass::STANDARD, storageclass::RRS] {
+    assert_eq!(template.len(), 5);
+    for tier in ["WARM", "COLD", storageclass::STANDARD, storageclass::RRS, UNKNOWN_TIER] {
         assert_eq!(template.get(tier), Some(&TierStats::default()), "missing seed for tier {tier}");
     }
 }
@@ -1372,7 +1372,7 @@ fn apply_bucket_result_to_cache_updates_bucket_entry() {
     );
 
     let update_time = SystemTime::now();
-    apply_bucket_result_to_cache(
+    assert!(apply_bucket_result_to_cache(
         &mut cache,
         DataUsageEntryInfo {
             name: "bucket".to_string(),
@@ -1382,12 +1382,51 @@ fn apply_bucket_result_to_cache_updates_bucket_entry() {
                 objects: 2,
                 ..Default::default()
             },
+            tier_registry_generation: None,
         },
         update_time,
-    );
+    ));
 
     assert_eq!(cache.info.last_update, Some(update_time));
     let entry = cache.find("bucket").expect("bucket entry should remain present");
     assert_eq!(entry.size, 10);
     assert_eq!(entry.objects, 2);
 }
+
+#[test]
+fn apply_bucket_result_to_cache_rejects_a_different_tier_generation() {
+    let mut cache = DataUsageCache {
+        info: DataUsageCacheInfo {
+            name: DATA_USAGE_ROOT.to_string(),
+            tier_registry_generation: Some(7),
+            ..Default::default()
+        },
+        ..Default::default()
+    };
+    cache.replace(
+        "bucket",
+        DATA_USAGE_ROOT,
+        DataUsageEntry {
+            size: 3,
+            ..Default::default()
+        },
+    );
+
+    let applied = apply_bucket_result_to_cache(
+        &mut cache,
+        DataUsageEntryInfo {
+            name: "bucket".to_string(),
+            parent: DATA_USAGE_ROOT.to_string(),
+            entry: DataUsageEntry {
+                size: 11,
+                ..Default::default()
+            },
+            tier_registry_generation: Some(8),
+        },
+        SystemTime::now(),
+    );
+
+    assert!(!applied);
+    assert_eq!(cache.find("bucket").map(|entry| entry.size), Some(3));
+    assert!(cache.info.last_update.is_none());
+}