Interesting. I am curious about the Tiny Tree/Linked Tree performance comparisons (the DOM doesn't surprise me). The design goal was improved memory usage for Tiny Tree, yet - the memory usage is the metric where Tiny Tree shows only a very small improvement over Linked Tree (370Mb vs 327Mb) - the build time for Linked Tree is much worse than Tiny Tree; the only reason I can think of is Java object allocation overhead. Is there anything else going on here? - the query time for Linked Tree is more than 6 times worse than for Tiny Tree. Why is that? Do you store a document order index in the linked tree nodes? I would also be interested to hear more about what schema information you use, both from the PSVI and directly from the schema. Obviously there is the simple type information for element and attribute values. You also mentioned inferring more explicit paths. I'm wondering whether it would be practical to generate this from RELAX NG.