I even know someone who stores invoice items in memcached. Whenever memchached evicts something that hasn't made it onwards to real storage, they lose the ability to invoice the right customer for an ad impression.
Consider two storage solutions used to gather line items to invoices. The line items are ten cents on average.
One system is properly ACID, and running it costs one cent per line item on the invoice. The other loses 1% of writes randomly, but its hardware/backup/ops requirements are lower, so it costs just 0.1c/item.
The ACID way gives you complete invoices, but you spend 10% of the invoiced amount on the invoices. The lossy way makes your invoices 1% smaller, randomly, but the loss+cost adds up to about 2% instead of 10%.
It's like returns on physical goods, really. A random percentage of customers will return goods, you can't control that but you can estimate and monitor it, and optimize if the numbers aren't good enough.
I don't understand how you can say any of this seriously.
Overinvoicing would be bad. But underinvoicing is just another cost of business. If the cost is smaller than the alternative, fine.
Relational databases place such emphasis on reliability. Mongo is generally reliable but only two or three nines, and it's simple.
For example, if you're collecting sub-cent line items for invoices, you might prefer to collect 99.x% of the line items simply over 100% at greater complication and expense, particularly if you can measure x.