Google ADK 2.9.0 reruns failed workflow nodes, side effects too

Google's ADK changed how it resumes a failed workflow. Before 2.9.0, a node that had already finished replayed on resume as if nothing happened. Now it runs again, including whatever it did the first time.

A node that sends an email or calls a payment API fires that side effect a second time. That happens whenever the workflow resumes after a failure partway through. Google's 2.9.0 release notes tell developers to make node bodies idempotent, putting the burden on whoever wrote the workflow, not on the resume logic itself.

The same release adds a FallbackModel for model failover, YAML-defined graph workflows instead of code-defined ones, and opt-in support for MCP SDK 2.x. Those are the features that will get attention. The resume change is the one that breaks something in production if nobody reads that far down the changelog.

Anyone running ADK workflows with nodes that touch anything outside the process needs to audit every one of them before upgrading. A node that calls a payment processor or a notification service needs a check that it hasn't already run before it runs again. The release notes don't say whether ADK gives any built-in way to tell a rerun from a first run. Building that guard falls on the developer.