Summary

  • Cloud Run starts revision-level minimum instances when the revision appears in a traffic split or carries a traffic tag. Google warns that a tagged revision can remain active and bill without requests.
  • Service-level minimum capacity follows the traffic split. When revision-level floors join it, the warm fleet can exceed the service-level number.

The preview link has a power cord

An engineer leaves a tag on a Cloud Run revision so a team can open its preview URL. The main service sends that revision no customer requests. The dashboard's traffic percentage looks like zero, yet a revision-level minimum tells Cloud Run to keep containers warm. Google's documentation is blunt about the result: tagged revisions with that setting stay active and incur charges while waiting.

The distinction lives in the scope of the minimum. A service-level floor is divided among revisions serving traffic in proportion to their split. A revision-level floor belongs to that revision, and a tag can wake it independently of customer traffic. A preview label that looks like a routing convenience can therefore carry a standing capacity order. Google advises removing tags that are no longer needed or using service-level minimum instances to avoid the tagged-revision cost.

The two floors do not always collapse into one neat total. Google gives a worked configuration: service minimum 10, a 50/50 traffic split, revision A minimum 6 and revision B minimum 0. The service split allocates five to each revision, but A's own minimum raises its count to six. The result is eleven warm instances, one above the service-level target. That is Google's configuration example, not a customer's invoice.

Those seats have different meters. Request-based billing charges kept-warm idle minimum instances at a lower idle rate; instance-based billing charges the default rate across each instance's lifecycle. The exact bill depends on resources, region, time and workload. The operational audit starts with a shorter question: which revisions have tags, which have revision-level minima, and how many instances are being held outside the live traffic plan?

Sources