In order to survive server restart during an index build's constraint violations checking phase, each entry in the constraints table must be read and then checked, and only then deleted. Currently (I believe?), each entry is read, deleted then checked, which is not recoverable.

Simultaneous index builds on secondaries can potentially wait a long while between building the index (scanning, writing) and performing constraint checking. It's therefore desirable to make the constraint checking phase resumable after server restart, to avoid restarting the entire index build.

This is technically an optimization. It will be required, however, if we ever make index builds entirely recoverable.