PUT /node/shutdown and PUT /node/config have a timeout parameter that allows to gracefully wait for timeout seconds before actual shutdown happens to allow mutating DML or bulk insertion operations to converge.
As of now OSOperations.apply() just sleeps for timeout seconds before it executes operations. This is an inappropriate approach to be replaced with more dynamic one, e.g. Agent sets cluster to read-only and continiously asks DMLProc, DDLProc (at the main node) and WriteEngine (at all nodes) if there are any operations on the fly before it proceeds with shutdown.