Even then, it’s not _that_ much extra complexity. It’s more data, but ultimately it’s still largely shuffling opaque binary blobs between, at most, small groups.

Eh, everytime a product manager says 'add this new feature, it will be easy and low impact' I cringe. Try as you might, you cannot fully measure complexity impact before a change occurs, only after. You can never know how much one of those little changes will drastically change user behavior. The product I work with had one of those small changes. We implemented better PDF report generation. The new report generation was even more efficient per file created than the old version.

It lead to performance issues at almost all of our large customers.....

See, before that point customers would just export the XML statistics and generate their own reports, but now with the reports being good, they pushed report generation back to our product. Report generation had always been coupled with another service before that point so a team had to be split off to separate these services and API controllers so they could run standalone. This required a fast feedback loop with the customers to ensure we understood their storage requirements and the ability to access secondary SQL datastores instead of the primary. Now we have a much more scalable, but far more complicated product so we had to train the support team handle these issues.

> Considering Facebook paid a lot of $ for the service they can't let it slide into obscurity.

Fallacy: https://wikipedia.org/wiki/Sunk_cost

It is hard to learn to avoid biased thinking in our own lives! I think that cutting one’s losses is a very important skill to learn.

You don't "cut losses" by axing a wildly successful acquisition while swimming in money. That sort of thing isn't unbiased but asinine and you "can't" do that because making asinine decisions tends to be received badly.