I am not sure that the scanning heap works very well with large roxie queries at the moment
- roxie doesn't use unique allocators, so lots of allocators get commoned up
- roxie has large differences between the sizes of buckets - which means large number of allocators share the same underlying heap.
- rows are often buffered between subgraphs or in splitters.
It would be safer to use the linked-list implementation until some of the items above are addressed.