Why non-atomic delete and upsert?

[aglebov]

Question

The implementation of the delete and upsert operation make an deliberate choice to create multiple snapshots, breaking the atomicity expectation.

For example, if I start a delete operation to remove some entries, I do not expect consumers of the data to see a snapshot with some but not all of the entries gone.

Given that the implementation choice looks very deliberate, I'm curious about the rationale for it?

It seems that other clients of the library expect atomic operation as well, for example #3076