If they actually do this across their code they will never have to run "real service calls", because those services are also tested the same way. It's mocking all the way down.
An issue I see with this is that there's a potential window when tests pass with bad data. For example, tests using a mock and the mock is periodically verified with a service call. The mock could become bad data, but won't be marked as bad until the next service call verification. Until that happens, the tests using the mock will all pass. It's not clear from the article how they address this.
That's how I would do it, at least.
If you have task-oriented build system with dependencies (like Google does with Blaze), it'd fit right in.
Though depending on how you deploy, you may have old and new services at different versions. Assuming a level backwards/forwards compatibility may be reasonable.
You record your expectations towards the service mock, and then later verify that these expectations hold against the real service. This decoupling allows you to run your "end-to-end" test suite very often, say on every check in, and then only verify the contract say every hour or day, depending on how stable your test environments are. The theory is that this gives quicker feedback, reduce flakyness and avoid combinatorial explosion of number of tests needed to run.
Mocks are usually (inevitably?) a per-test structure which makes them outrageously fragile and tempting to over mock to get a passing test
I prefer building something that acts like the mocked thing (ie a database) but does So both ends (it looks like a database to the web service and a web service to the database. Then the stub is my contract
It's awkward
Well it's still better because you only have to run the services you are actually testing right?
This very much neglects the probably significant overhead of writing, maintaining, understanding the M mocks on top of the base of N tests, but since, as you said, this is "at Google scale" the trade-off seemingly becomes worthwhile
However, that would be the best case, which is the mock data is sufficient for all N tests. That is not likely to be true, but probably a M number of mocks can still cover all N tests. And M < N in most cases so there's still savings in number of service calls. In the worst case, N tests will need N mocks and it will be the same as before.