There are a few things we can do to improve the performance of this workload:

1) Eliminate the repeated allocations done in the CheckBoundsStage::getNext() here. We can instead hold a buffer in the stage and re-use that each call to toBson().

2) Skip the allocation done in IndexBoundsBuilder::checkKey() (there's already a TODO about this)

3) WTIndexCursorBase::nextKeyString() makes a copy of the key string which requires allocating. We could instead keep a buffer up front in the IXscan stage and then pass a pointer to it. This would be a bit invasive, but it's worth trying to see how much benefit there is.