There should be a connection pool implementation that always uses a new connection per request. Note: This is already supported technically, but harder than it needs to be for library consumers.