Flaky integration test: ServiceToServiceScopeConcurrency class init times out in exposeHostPort #21

Summary

The Features integration test job (.github/workflows/integration.yaml) sometimes fails even though every test passes. One test class fails during beforeAll because the sdk-test-suite harness times out while setting up host-port forwarding for the Restate container. The failure is in test infrastructure, not in the SDK or the code under test.

Seen on PR #20 (run 41477981376, commit 817791f). That PR only changes .github/workflows/verify-lockfile.yml and actions.lock, and the next integration runs passed (PR #19 at 2649807, main at 6a63deb). The 12 most recent scheduled and push runs on main all passed.

Failure

* Succeeded tests: 321 / 321
* Failed classes initialization: 1
== 'default' FAILURES
default => ServiceToServiceScopeConcurrency
  => org.rnorth.ducttape.TimeoutException: org.rnorth.ducttape.TimeoutException: java.util.concurrent.TimeoutException
     org.rnorth.ducttape.unreliables.Unreliables.retryUntilSuccess(Unreliables.java:56)
     dev.restate.sdktesting.infra.HostPortForwardingKt.exposeHostPort(hostPortForwarding.kt:55)
     dev.restate.sdktesting.infra.RestateDeployer.deployAll(RestateDeployer.kt:284)
     dev.restate.sdktesting.infra.RestateDeployerExtension.beforeAll(RestateDeployerExtension.kt:41)

Failing test: scopeAndLimitKeyArePropagatedOnServiceToServiceCalls ("Concurrency limit is enforced on service-to-service calls carrying scope and limit key"). Because the error happens in beforeAll, the test body never runs. The default suite then sat idle for about 4 minutes (16:06:55 → 16:11:05) until the forwarding retry gave up. That wait accounts for most of the job's 6m30s duration.

Root cause

exposeHostPort in restatedev/sdk-test-suite (dev.restate.sdktesting.infra.hostPortForwarding.kt:55) wraps Testcontainers host-port exposure in Unreliables.retryUntilSuccess. When that setup stalls on a CI runner, the whole test class fails during initialization.

Suggested fix

• Report it upstream in restatedev/sdk-test-suite. Options there: make exposeHostPort more robust, e.g. a shorter per-attempt timeout with more attempts, or recreating the Testcontainers SSHd/port-forwarding container on failure.
• Until then, consider having the restatedev/e2e action / test runner retry failed class initializations once, or exclude ServiceToServiceScopeConcurrency in integration.yaml if it keeps failing.