Cloud Run scalability

This episode runs a short synthetic Cloud Run scaling demonstration. Its result is evidence about one test setup, not a capacity promise for another service or its dependencies.

Key points

• The handler is only an approximation: It waits for one second and computes prime numbers to represent backend wait plus CPU work; the presenters explicitly exclude database scaling [00:00:00]-[00:02:08].
• The dashboard visualizes a custom test harness: The service emits Pub/Sub signals when an instance is ready and about to stop, so a new instance does not appear until startup completes [00:02:08]-[00:03:14].
• Scale-up can add latency: When the harness moves from 10 to 2,000 request loops, requests queue while Cloud Run adds instances [00:03:14]-[00:04:16].
• The headline result is tightly bounded: At 30,000 loops the display still reports 100% success, but the presenters warn that a real database must sustain the traffic independently and that more active instances cost money [00:04:16]-[00:06:26].

The December 2023 demo gives no request rate, latency distribution, instance count, region, concurrency configuration, or meaningful test duration. A displayed success percentage does not prove an SLO. Reproduce the test with representative dependencies, quotas, traffic, latency objectives, failure modes, and current Cloud Run behavior.

Full video: https://www.youtube.com/watch?v=8YtsV8BNTqY(opens in a new tab)