f

be/src/runtime/memory/mem_tracker_limiter.cpp

@@ -88,6 +88,8 @@ std::shared_ptr<MemTrackerLimiter> MemTrackerLimiter::create_shared(MemTrackerLi
     auto tracker = std::make_shared<MemTrackerLimiter>(type, label, byte_limit);
     // Write tracker is only used to tracker the size, so limit == -1
     auto write_tracker = std::make_shared<MemTrackerLimiter>(type, "Memtable" + label, -1);
+    // Memtable has a separate logic to deal with memory flush, so that should not check the limit in memtracker.
+    write_tracker->set_enable_reserve_memory(true);
     tracker->_write_tracker.swap(write_tracker);
 #ifndef BE_TEST
     DCHECK(ExecEnv::tracking_memory());

be/src/runtime/memory/mem_tracker_limiter.h

@@ -381,7 +381,11 @@ inline Status MemTrackerLimiter::check_limit(int64_t bytes) {
         return Status::OK();
     }
 
-    // check limit should ignore memtable size, because it is treated as a cache
+    // If reserve not enabled, then should check limit here to kill the query when limit exceed.
+    // For insert into select or pure load job, its memtable is accounted in a seperate memtracker limiter,
+    // and its reserve is set to true. So that it will not reach this logic.
+    // Only query and load job has exec_mem_limit and the _limit > 0, other memtracker limiter's _limit is -1 so
+    // it will not take effect.
     if (_limit > 0 && consumption() + bytes > _limit) {
         return Status::MemoryLimitExceeded(fmt::format("failed alloc size {}, {}",
                                                        MemCounter::print_bytes(bytes),