diff --git a/localstack-core/localstack/services/dynamodb/server.py b/localstack-core/localstack/services/dynamodb/server.py
index 68dac757b9d82..2e41477749762 100644
--- a/localstack-core/localstack/services/dynamodb/server.py
+++ b/localstack-core/localstack/services/dynamodb/server.py
@@ -62,7 +62,6 @@ def __init__(
             self.db_path = None
 
         if self.db_path:
-            mkdir(self.db_path)
             self.db_path = os.path.abspath(self.db_path)
 
         self.heap_size = config.DYNAMODB_HEAP_SIZE
@@ -80,12 +79,18 @@ def get() -> "DynamodbServer":
     def start_dynamodb(self) -> bool:
         """Start the DynamoDB server."""
 
+        # For the v2 provider, the DynamodbServer has been made a singleton. Yet, the Server abstraction is modelled
+        # after threading.Thread, where Start -> Stop -> Start is not allowed. This flow happens during state resets.
+        # The following is a workaround that permits this flow
+        self._started.clear()
+        self._stopped.clear()
+
         # Note: when starting the server, we had a flag for wiping the assets directory before the actual start.
         # This behavior was needed in some particular cases:
         # - pod load with some assets already lying in the asset folder
         # - ...
         # The cleaning is now done via the reset endpoint
-        self._stopped.clear()
+        mkdir(self.db_path)
         started = self.start()
         self.wait_for_dynamodb()
         return started