Summary

  • Cloud Run minimum instances keep a target number of service containers warm for request latency and incur charges while idle. Google says those instances can restart at any time; the minimum is a best-effort target.
  • Instance-based billing gives CPU outside requests and supports short background work, but an idle service container can still be shut down. Google documents ten seconds of SIGTERM grace and recommends retry-capable task handling for asynchronous work.

A paid chair is not a named employee

Minimum instances is an attractive button for a service with an awkward first-request delay. Google keeps a chosen number of containers ready, even when no one calls them. That readiness costs money: under request-based billing, idle warm instances have a lower billed rate; under instance-based billing, the entire container lifecycle uses the default rate. The buyer is paying for available capacity.

The process occupying that capacity has a different contract. Google says minimum instances can be restarted at any time. The configured count is best effort, and capacity shortages, infrastructure rebalancing, crashes, quota limits or billing limits can briefly pull it below target. Three minimum instances are Google's high-availability recommendation, not an individual container's lifetime warranty. A dashboard showing the target count does not tell a background task that its original process is still running.

Instance-based billing complicates the temptation. It allocates CPU outside request handling, so a container can perform short asynchronous work after sending a response. Yet the same documentation says an idle container, including one kept warm, may be shut down at any time. A SIGTERM handler can take ten seconds to finish or record a checkpoint. Work that cannot fit or recover within that window needs a retry and persistence plan. Google points to Cloud Tasks for retried asynchronous execution, and offers a separate Cloud Run instance resource for a single long-lived process with restart policies and runs of up to seven days.

The market question is what the budget line actually purchases. A service with minimum instances buys faster requests and a best-effort warm floor. A durable job needs a record of what finished, what can be retried, and who pays for the chosen execution model. Google says a mostly idle singleton Cloud Run instance can cost less than a service with minimum instances and instance-based billing, but the cheaper option depends on the workload and SKU. For a particular buyer, instance restarts, task duration, recovery records, latency and charges would settle that comparison. The word “minimum” alone cannot.

Sources