diff --git a/src/Common/ThreadStatus.h b/src/Common/ThreadStatus.h
index 03b99a1800b1..f89585d04f89 100644
--- a/src/Common/ThreadStatus.h
+++ b/src/Common/ThreadStatus.h
@@ -116,6 +116,16 @@ class ThreadGroup
     void attachQueryForLog(const String & query_, UInt64 normalized_hash = 0);
     void attachInternalProfileEventsQueue(const InternalProfileEventsQueuePtr & profile_queue);
 
+    /// Override the cancellation predicate. All threads that subsequently attach to this
+    /// group via ThreadGroupSwitcher inherit the predicate in their local_data, making
+    /// isQueryCanceled() reflect task-level cancellation without a process-list entry.
+    /// Required for part and partition export cancellation during S3 outage.
+    void setCancelPredicate(QueryIsCanceledPredicate predicate)
+    {
+        std::lock_guard lock(mutex);
+        shared_data.query_is_canceled_predicate = std::move(predicate);
+    }
+
     /// When new query starts, new thread group is created for it, current thread becomes master thread of the query
     static ThreadGroupPtr createForQuery(ContextPtr query_context_, FatalErrorCallback fatal_error_callback_ = {});
 
diff --git a/src/Storages/MergeTree/ExportPartTask.cpp b/src/Storages/MergeTree/ExportPartTask.cpp
index 20e8d52f03a1..83e8e28daeb1 100644
--- a/src/Storages/MergeTree/ExportPartTask.cpp
+++ b/src/Storages/MergeTree/ExportPartTask.cpp
@@ -175,6 +175,28 @@ bool ExportPartTask::executeStep()
         manifest.query_id,
         local_context);
 
+    /*
+        This is a hack to fix the issue where S3 is out, ClickHouse keeps retrying S3 requests deep
+        in the AWS SDK and never check for the `isCancelled()` flag. That prevents the task from being killed / cancelled. It also prevents the table from being dropped.
+
+        Merges and mutations don't suffer from this problem because they don't make requests to S3 :). Select statements
+        do make requests to S3, but the cancel predicate is properly setup for regular queries.
+
+        I think this is the first time we have a background operation that makes requests to S3, so we need to connect the dots.
+
+        The simples way is this one, and given the release timeline, I am opting for it.
+    */
+    (*exports_list_entry)->thread_group->setCancelPredicate(
+        [weak_this = weak_from_this()]() -> bool
+        {
+            if (auto shared_this = weak_this.lock())
+            {
+                return shared_this->isCancelled();
+            }
+
+            return true;
+        });
+
     SinkToStoragePtr sink;
 
     const auto new_file_path_callback = [&exports_list_entry](const std::string & file_path)
@@ -185,6 +207,8 @@ bool ExportPartTask::executeStep()
 
     try
     {
+        ThreadGroupSwitcher switcher((*exports_list_entry)->thread_group, ThreadName::EXPORT_PART);
+
         const auto filename = buildDestinationFilename(manifest, storage.getStorageID(), local_context);
 
         sink = destination_storage->import(
@@ -232,8 +256,6 @@ bool ExportPartTask::executeStep()
             local_context,
             getLogger("ExportPartition"));
 
-        ThreadGroupSwitcher switcher((*exports_list_entry)->thread_group, ThreadName::EXPORT_PART);
-
         /// We need to support exporting materialized and alias columns to object storage. For some reason, object storage engines don't support them.
         /// This is a hack that materializes the columns before the export so they can be exported to tables that have matching columns
         materializeSpecialColumns(plan_for_part.getCurrentHeader(), metadata_snapshot, local_context, plan_for_part);
diff --git a/src/Storages/MergeTree/ExportPartTask.h b/src/Storages/MergeTree/ExportPartTask.h
index a3f1635c4902..1596f2bf23c9 100644
--- a/src/Storages/MergeTree/ExportPartTask.h
+++ b/src/Storages/MergeTree/ExportPartTask.h
@@ -7,7 +7,7 @@
 namespace DB
 {
 
-class ExportPartTask : public IExecutableTask
+class ExportPartTask : public IExecutableTask, public std::enable_shared_from_this<ExportPartTask>
 {
 public:
     explicit ExportPartTask(