diff --git a/crates/lance-context-core/src/store_base.rs b/crates/lance-context-core/src/store_base.rs
index b3f10f2..bc1b1b9 100644
--- a/crates/lance-context-core/src/store_base.rs
+++ b/crates/lance-context-core/src/store_base.rs
@@ -1233,6 +1233,11 @@ impl StorageBase {
         // master builds it before fan-out (#277), but a worker's own timer
         // or a manual merge must not depend on the master having been here.
         if self.dataset.count_fragments() > 0 && !self.has_key_btree_index().await? {
+            metrics::counter!("rollout_merge_index_built_on_demand_total").increment(1);
+            info!(
+                uri = %self.dataset.uri(),
+                "base table has no key BTree; building it before the merge"
+            );
             self.create_key_btree_index().await?;
         }
         let key_index = merge_schema.index_of(&self.key_column)?;
@@ -1747,7 +1752,7 @@ impl StorageBase {
                         // the only cure is the merge that is already behind.
                         // Fail the read fast so the caller retries after the
                         // merge instead of taking the worker down with it.
-                        metrics::counter!("rollout_reads_refused_pending_total").increment(1);
+                        note_read_refused(&uri);
                         return Err(LanceError::io(format!(
                             "{PENDING_GENERATIONS_EXCEEDED}: shard {shard_id} has {pending} \
                              flushed generations pending merge (cap {max_at}); retry after \
@@ -1791,7 +1796,7 @@ impl StorageBase {
             .map(|snapshot| snapshot.flushed_generations.len())
             .sum();
         if max_at != 0 && total > max_at {
-            metrics::counter!("rollout_reads_refused_pending_total").increment(1);
+            note_read_refused(self.dataset.uri());
             return Err(LanceError::io(format!(
                 "{PENDING_GENERATIONS_EXCEEDED}: {total} flushed generations pending merge \
                  across {} shards (cap {max_at}); retry after the merge catches up",
@@ -2198,6 +2203,21 @@ pub(crate) fn align_batch_to_schema(
     Ok(RecordBatch::try_new(target_schema, columns)?)
 }
 
+/// A read refused because the store's WAL backlog is over the cap. Labeled by
+/// store so the fleet can see *which* store is unreadable, not just that one
+/// is: a single store answered 503 for five hours today before anyone
+/// noticed. The label is the dataset directory name, bounded by the number
+/// of stores a worker touches.
+fn note_read_refused(uri: &str) {
+    let store = uri
+        .trim_end_matches('/')
+        .rsplit('/')
+        .next()
+        .unwrap_or("unknown")
+        .to_string();
+    metrics::counter!("rollout_reads_refused_pending_total", "store" => store).increment(1);
+}
+
 /// Derive the MemWAL shard UUID a server instance writes to from its stable
 /// instance id. Deterministic (UUID v5), so the same instance id always maps to
 /// the same shard across restarts and reopens — the losing/gaining semantics of
diff --git a/crates/lance-context-master/src/scanner.rs b/crates/lance-context-master/src/scanner.rs
index 806d020..6336fb4 100644
--- a/crates/lance-context-master/src/scanner.rs
+++ b/crates/lance-context-master/src/scanner.rs
@@ -360,6 +360,7 @@ async fn scan_once_inner(state: &Arc<MasterState>) -> lance::Result<usize> {
         }
     }
     snapshot.sort_by(|a, b| a.name.cmp(&b.name));
+    record_backlog_gauges(&snapshot);
 
     // Retire experiments with no writes for the configured window. Each is
     // merged, compacted and verified quiescent first; only then is its row
@@ -521,6 +522,66 @@ async fn pending_or_timeout(
     }
 }
 
+/// Fleet-wide view of the WAL backlog, refreshed every scan. The read cap
+/// (`ROLLOUT_WAL_PENDING_MAX_GENERATIONS`, 4,096 by default) turns pending
+/// generations into refused reads, so `stores_over_read_cap` is the number of
+/// stores whose reads are failing right now; it is the alert that was
+/// missing when a store sat at 17k for hours.
+fn record_backlog_gauges(snapshot: &[StatRow]) {
+    let b = Backlog::summarize(snapshot);
+    metrics::gauge!("master_wal_pending_generations_total").set(b.total as f64);
+    metrics::gauge!("master_wal_pending_generations_max").set(b.max as f64);
+    metrics::gauge!("master_stores_pending_over_read_cap").set(b.over_read_cap as f64);
+    metrics::gauge!("master_stores_pending_over_1k").set(b.over_1k as f64);
+}
+
+#[derive(Debug, Default, PartialEq, Eq)]
+struct Backlog {
+    total: i64,
+    max: i64,
+    over_read_cap: i64,
+    over_1k: i64,
+}
+
+impl Backlog {
+    const READ_CAP: i64 = lance_context_core::DEFAULT_PENDING_GENERATIONS_MAX as i64;
+
+    fn summarize(snapshot: &[StatRow]) -> Self {
+        let mut b = Self::default();
+        for row in snapshot {
+            let p = row.pending_wal_generations.max(0);
+            b.total += p;
+            b.max = b.max.max(p);
+            if p > Self::READ_CAP {
+                b.over_read_cap += 1;
+            }
+            if p > 1000 {
+                b.over_1k += 1;
+            }
+        }
+        b
+    }
+}
+
+/// Pending generations that moved while the base-table version did not.
+/// This is exactly the state that hid a 17k-generation backlog from the
+/// merge sweep before #281: the version-unchanged shortcut reused the stale
+/// count. Counted so it can be alerted on; the row itself is corrected.
+fn note_pending_drift(name: &str, previous: i64, current: usize) {
+    let current = current as i64;
+    if current != previous {
+        metrics::counter!("master_stats_pending_recounted_total").increment(1);
+        if current > previous.max(0) + 256 {
+            tracing::warn!(
+                store = %name,
+                previous,
+                current,
+                "pending WAL generations grew while the base version stood still"
+            );
+        }
+    }
+}
+
 /// Observe one generic store. Only the fields the WAL-merge sweep and the UI
 /// need: version (for the skip-if-unchanged shortcut), base row count,
 /// fragment count and pending MemWAL generations. Compaction counters stay at
@@ -553,6 +614,7 @@ async fn observe_generic(
     if let Some(prev) = previous {
         if prev.version != StatRow::UNKNOWN_VERSION && prev.version == current_version {
             let pending = pending_or_timeout(name, store.pending_wal_generations()).await?;
+            note_pending_drift(name, prev.pending_wal_generations, pending);
             let mut row = prev.clone();
             row.uri = uri.to_string();
             row.pending_wal_generations = pending as i64;
@@ -622,6 +684,7 @@ async fn observe_one(
     if let Some(prev) = previous {
         if prev.version != StatRow::UNKNOWN_VERSION && prev.version == current_version {
             let pending = pending_or_timeout(name, store.pending_wal_generations()).await?;
+            note_pending_drift(name, prev.pending_wal_generations, pending);
             let mut row = prev.clone();
             row.uri = uri.to_string();
             row.pending_wal_generations = pending as i64;
@@ -1320,6 +1383,25 @@ mod retirement_tests {
     /// experiment is invisible to them forever. Dropping one with pending
     /// generations would strand them: never merged, read amplification never
     /// recovers, `_mem_wal/` never reclaimed, and no process left to notice.
+    #[test]
+    fn backlog_summary_counts_stores_over_the_read_cap() {
+        let mut a = row("a", "u", 0);
+        a.pending_wal_generations = 17_254;
+        let mut b = row("b", "u", 0);
+        b.pending_wal_generations = 1_200;
+        let mut c = row("c", "u", 0);
+        c.pending_wal_generations = 3;
+        assert_eq!(
+            Backlog::summarize(&[a, b, c]),
+            Backlog {
+                total: 18_457,
+                max: 17_254,
+                over_read_cap: 1,
+                over_1k: 2,
+            }
+        );
+    }
+
     #[tokio::test]
     async fn retirement_drains_the_wal_before_dropping_the_row() {
         let dir = TempDir::new().unwrap();