We encounter a class of bugs in WiredTiger where traversing the tree structure doesn't work correctly. It is difficult to guarantee the correctness of the system, since the tree is lock free it is not practical to add assertions about the state of the tree.

It would be useful to introduce a method of testing that increases the window for race conditions in split code. Doing that coupled with end-result order guarantees in an application should uncover bugs earlier in the development cycle.