Azure: Retry transient Azurite image-fetch failures in integration tests

[wombatu-kun]

Summary

Closes #16530.

The iceberg-azure integration tests (TestADLSFileIO, TestADLSInputStream, TestADLSOutputStream) start a shared Azurite container in AzuriteTestBase.@BeforeAll. They intermittently fail at startup with a ContainerFetchException caused by a NotFoundException (HTTP 404) while testcontainers pulls mcr.microsoft.com/azure-storage/azurite:3.35.0. The image tag is already pinned, so this is a transient registry/CDN failure (rate-limiting or CDN inconsistency), not a missing image — re-running usually succeeds, which is exactly what makes these tests flaky.

What changed

AzuriteContainer now overrides start() to retry transient image-fetch failures before giving up: up to 5 attempts with exponential backoff (2s → 4s → 8s → 16s). Each retry re-invokes start(), which re-triggers the image pull, because testcontainers' RemoteDockerImage does not cache a failed resolution.

Notably, testcontainers' own withStartupAttempts(...) does not help here: in testcontainers 2.0.5 the image is fetched (getDockerImageName() → RemoteDockerImage) before the startupAttempts retry loop in GenericContainer.doStart(), so a fetch failure short-circuits that loop. Retrying start() itself is what actually re-attempts the pull.

The retry is deliberately narrow: it only retries when the failure (or its cause chain) is a ContainerFetchException. Genuine launch failures — e.g. a wait-strategy timeout surfaced as a ContainerLaunchException without a fetch cause — are rethrown immediately so real problems are never masked.

Tests

• New TestAzuriteContainer exercises the retry helper without Docker: succeeds without retry, retries until success, rethrows after exhausting attempts, does not retry non-fetch launch failures, and retries a fetch failure wrapped in a ContainerLaunchException.
• Confirmed the existing Azurite integration tests still pass through the overridden start() (TestADLSFileIO, TestADLSInputStream, TestADLSOutputStream) with Docker.

Note

This reduces the flakiness by tolerating transient fetch hiccups; it cannot mask a registry outage that persists across all retries.