diff --git a/control_plane/attachment_service/src/compute_hook.rs b/control_plane/attachment_service/src/compute_hook.rs
index bebc62ac2f..1a8dc6b86d 100644
--- a/control_plane/attachment_service/src/compute_hook.rs
+++ b/control_plane/attachment_service/src/compute_hook.rs
@@ -14,7 +14,6 @@ use utils::{
 
 use crate::service::Config;
 
-const BUSY_DELAY: Duration = Duration::from_secs(1);
 const SLOWDOWN_DELAY: Duration = Duration::from_secs(5);
 
 pub(crate) const API_CONCURRENCY: usize = 32;
@@ -280,11 +279,10 @@ impl ComputeHook {
                 Err(NotifyError::SlowDown)
             }
             StatusCode::LOCKED => {
-                // Delay our retry if busy: the usual fast exponential backoff in backoff::retry
-                // is not appropriate
-                tokio::time::timeout(BUSY_DELAY, cancel.cancelled())
-                    .await
-                    .ok();
+                // We consider this fatal, because it's possible that the operation blocking the control one is
+                // also the one that is waiting for this reconcile.  We should let the reconciler calling
+                // this hook fail, to give control plane a chance to un-lock.
+                tracing::info!("Control plane reports tenant is locked, dropping out of notify");
                 Err(NotifyError::Busy)
             }
             StatusCode::SERVICE_UNAVAILABLE
@@ -306,7 +304,12 @@ impl ComputeHook {
         let client = reqwest::Client::new();
         backoff::retry(
             || self.do_notify_iteration(&client, url, &reconfigure_request, cancel),
-            |e| matches!(e, NotifyError::Fatal(_) | NotifyError::Unexpected(_)),
+            |e| {
+                matches!(
+                    e,
+                    NotifyError::Fatal(_) | NotifyError::Unexpected(_) | NotifyError::Busy
+                )
+            },
             3,
             10,
             "Send compute notification",
diff --git a/control_plane/attachment_service/src/service.rs b/control_plane/attachment_service/src/service.rs
index cceecebb7f..fe2358abae 100644
--- a/control_plane/attachment_service/src/service.rs
+++ b/control_plane/attachment_service/src/service.rs
@@ -3936,9 +3936,6 @@ impl Service {
     /// Helper for methods that will try and call pageserver APIs for
     /// a tenant, such as timeline CRUD: they cannot proceed unless the tenant
     /// is attached somewhere.
-    ///
-    /// TODO: this doesn't actually ensure attached unless the PlacementPolicy is
-    /// an attached policy.  We should error out if it isn't.
     fn ensure_attached_schedule(
         &self,
         mut locked: std::sync::RwLockWriteGuard<'_, ServiceState>,
@@ -3947,10 +3944,26 @@ impl Service {
         let mut waiters = Vec::new();
         let (nodes, tenants, scheduler) = locked.parts_mut();
 
-        for (_tenant_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
+        for (tenant_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
             shard.schedule(scheduler)?;
 
+            // The shard's policies may not result in an attached location being scheduled: this
+            // is an error because our caller needs it attached somewhere.
+            if shard.intent.get_attached().is_none() {
+                return Err(anyhow::anyhow!(
+                    "Tenant {tenant_id} not scheduled to be attached"
+                ));
+            };
+
+            if shard.stably_attached().is_some() {
+                // We do not require the shard to be totally up to date on reconciliation: we just require
+                // that it has been attached on the intended node.   Other dirty state such as unattached secondary
+                // locations, or compute hook notifications can be ignored.
+                continue;
+            }
+
             if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
+                tracing::info!("Waiting for shard {tenant_shard_id} to reconcile, in order to ensure it is attached");
                 waiters.push(waiter);
             }
         }
diff --git a/test_runner/regress/test_sharding_service.py b/test_runner/regress/test_sharding_service.py
index c33d2ca0da..5a86e03d2b 100644
--- a/test_runner/regress/test_sharding_service.py
+++ b/test_runner/regress/test_sharding_service.py
@@ -433,10 +433,13 @@ def test_sharding_service_compute_hook(
     # Set up fake HTTP notify endpoint
     notifications = []
 
+    handle_params = {"status": 200}
+
     def handler(request: Request):
-        log.info(f"Notify request: {request}")
+        status = handle_params["status"]
+        log.info(f"Notify request[{status}]: {request}")
         notifications.append(request.json)
-        return Response(status=200)
+        return Response(status=status)
 
     httpserver.expect_request("/notify", method="PUT").respond_with_handler(handler)
 
@@ -504,6 +507,24 @@ def test_sharding_service_compute_hook(
 
     wait_until(10, 1, received_split_notification)
 
+    # If the compute hook is unavailable, that should not block creating a tenant and
+    # creating a timeline.  This simulates a control plane refusing to accept notifications
+    handle_params["status"] = 423
+    degraded_tenant_id = TenantId.generate()
+    degraded_timeline_id = TimelineId.generate()
+    env.storage_controller.tenant_create(degraded_tenant_id)
+    env.storage_controller.pageserver_api().timeline_create(
+        PgVersion.NOT_SET, degraded_tenant_id, degraded_timeline_id
+    )
+
+    # Ensure we hit the handler error path
+    env.storage_controller.allowed_errors.append(
+        ".*Failed to notify compute of attached pageserver.*tenant busy.*"
+    )
+    env.storage_controller.allowed_errors.append(".*Reconcile error.*tenant busy.*")
+    assert notifications[-1] is not None
+    assert notifications[-1]["tenant_id"] == str(degraded_tenant_id)
+
     env.storage_controller.consistency_check()