Key generation for wildcard indexes generates these “multikey metadata keys” for each document. These keys essentially says which paths in each document are multikey. Since the number of multikey paths is unbounded, instead of storing them in the catalog like a regular index, those keys are stored in the index itself.

The keys are kept in memory which consumes the quota allocated for the index build, until the very end of the index build. On large collections there's less and less memory available to the sorter which has to spill more frequently, until it reaches a point where each key has to be spilled, making the index build extremely slow.