I'll summarize here my state of studying concurrent methods of page unlinking.
GIN B-tree does not have "high key". That means, that rightmost key on a page is maximal for the non-leaf page. But I do not see anything theoretical in a way of implementation of Lanin and Shasha`s methods of page merging, with slight modifications. Their paper does not even mention high key(high fence key in papers by Goetz Graefe).
But it's not a simple task due to large portions of shared code between entry tree and posting tree.
Also, I do not see a reason why this method can be practically superior to proposed patch.
Currently, I do not have resources to implement a proof of concept for fully concurrent page unlinking to make benchmarking.
I am marking this patch as returned with feedback.
Michael, sorry, but why? If I understood everything right, the main question from Jeff was why is it implemented in such way? And Jeff wanted to see other ways of solving the problem. Andrew wrote about them above and it seems that implementing them would be quite expensive and without any obvious win. I would rather expect some reaction from Jeff or may be someone else, so shouldn’t it be marked as «Ready for committer» or at least «Moved to next CF»?
