They deprecate internal infrastructure stuff zealously and tell teams they need to be off of such and such by this date.

But it's worse than that because they'll bring up whole new datacenters without ever bringing the deprecated service up, and they also retire datacenters with some regularity. So if you run a service that depends on deprecated services you could quickly find yourself in a situation where you have to migrate to maintain N+2 redundancy but there's hardly any datacenter with capacity available in the deprecated service you depend on.

Also, how many man years of engineering do you want to spend on keeping goo.gl running. If you were an engineer would you want to be assigned this project? What are you going to put in your perf packet? "Spent 6 months of my time and also bothered engineers in other teams to keep this service that makes us no money running"?

A lot of Google infra services are built around the understanding that clients will be re-built to pick up library changes pretty often, and that you can make breaking API changes from time to time (with lots of notice).

But if you don't downgrade the old, then you're endlessly supporting systems, forever. At some point, it does become cheaper to migrate everything to the new.