The key issue is that versioned state graphs make it harder to do causality analysis than raw logs would allow. Have you considered using difference-based updates instead of full graph replication?

Or do you find that the similarity filtering overhead justifies the reduced noise propagation?