I don't have access to that code currently, since it's for a prior employer. However, if I remember it correctly, it ended up using the TopParentID mentioned in the article to first find a simple count by that column, did a "running total" type calculation on that (I use a CLR function for that, blindingly fast) to get top level range start and stop values. All of that was very, very fast, like milliseconds. I think it then repeated that for each level till it got zero for @@rowcount, but the lower levels weren't as fast because they had to actually crawl the hierarchy to get the number of nodes beneath each, instead of just a count on TopParentID.

I tried the update method mentioned in Joe's article and found that it was WAY too slow for an in-use database. My update solution was more complex, but much, much faster.

I could improve the process immensely with the difference between what I know now and what I knew when I built it, but that's pretty much true of any code I wrote more than about a month ago. Just some simple Cross Apply inline queries would make the thing much more efficient.

The 2,700 rows was for one hierarchy within a multi-million row table. 11 seconds to resolve anything on a 2,700-row table would imply that I was running it on, maybe a 286 CPU with 2 Meg of RAM? TSR-80? Timex/Sinclair 1000? Not sure how far back I'd have to go to get that bad of performance, even on an adjacency crawl. If I remember correctly, the table had somewhere around 2-million total rows, and 2,700 nodes, 6 levels deep, was the biggest single hierarchy within it.