diff --git a/dbms/src/DataStreams/ParallelAggregatingBlockInputStream.cpp b/dbms/src/DataStreams/ParallelAggregatingBlockInputStream.cpp
index 447dfdfbed5..9d9d73b3631 100644
--- a/dbms/src/DataStreams/ParallelAggregatingBlockInputStream.cpp
+++ b/dbms/src/DataStreams/ParallelAggregatingBlockInputStream.cpp
@@ -195,6 +195,9 @@ void ParallelAggregatingBlockInputStream::Handler::onFinish()
 void ParallelAggregatingBlockInputStream::Handler::onException(std::exception_ptr & exception, size_t thread_num)
 {
     parent.exceptions[thread_num] = exception;
+    Int32 old_value = -1;
+    parent.first_exception_index.compare_exchange_strong(old_value, static_cast<Int32>(thread_num), std::memory_order_seq_cst, std::memory_order_relaxed);
+
     /// can not cancel parent inputStream or the exception might be lost
     if (!parent.executed)
         /// kill the processor so ExchangeReceiver will be closed
@@ -220,7 +223,8 @@ void ParallelAggregatingBlockInputStream::execute()
     processor.process();
     processor.wait();
 
-    rethrowFirstException(exceptions);
+    if (first_exception_index != -1)
+        std::rethrow_exception(exceptions[first_exception_index]);
 
     if (isCancelledOrThrowIfKilled())
         return;
diff --git a/dbms/src/DataStreams/ParallelAggregatingBlockInputStream.h b/dbms/src/DataStreams/ParallelAggregatingBlockInputStream.h
index b96fe3ad0e5..b7e9c3fe9a1 100644
--- a/dbms/src/DataStreams/ParallelAggregatingBlockInputStream.h
+++ b/dbms/src/DataStreams/ParallelAggregatingBlockInputStream.h
@@ -102,6 +102,7 @@ class ParallelAggregatingBlockInputStream : public IProfilingBlockInputStream
 
     ManyAggregatedDataVariants many_data;
     Exceptions exceptions;
+    std::atomic<Int32> first_exception_index{-1};
 
     struct ThreadData
     {