π OpenHands: AI-Driven Development
Fix: Use container StartedAt for Docker sandbox status grace period calculation (#13841)
Tim O'Farrell39f3b293f55c
diff --git a/openhands/app_server/sandbox/docker_sandbox_service.py b/openhands/app_server/sandbox/docker_sandbox_service.pyindex e99f60e080..bc8a6446c2 100644--- a/openhands/app_server/sandbox/docker_sandbox_service.py+++ b/openhands/app_server/sandbox/docker_sandbox_service.py@@ -253,8 +253,16 @@ class DockerSandboxService(SandboxService):except asyncio.CancelledError:raiseexcept Exception as exc:+ # Get the started_at from the docker container info and fallback to sandbox created_at+ try:+ state = container.attrs['State']+ started_at = datetime.fromisoformat(state['StartedAt'])+ except Exception:+ _logger.debug('Error getting container start time')+ started_at = sandbox_info.created_at+# If the server has exceeded the startup grace period, it's an error- if sandbox_info.created_at < utc_now() - timedelta(+ if started_at < utc_now() - timedelta(seconds=self.startup_grace_seconds):_logger.info(diff --git a/tests/unit/app_server/test_docker_sandbox_service.py b/tests/unit/app_server/test_docker_sandbox_service.pyindex f6ae716eef..858352446a 100644--- a/tests/unit/app_server/test_docker_sandbox_service.py+++ b/tests/unit/app_server/test_docker_sandbox_service.py@@ -9,7 +9,7 @@ This module tests the Docker sandbox service implementation, focusing on:- Edge cases with malformed container data"""-from datetime import datetime+from datetime import datetime, timezonefrom unittest.mock import AsyncMock, MagicMock, patchimport httpx@@ -1649,3 +1649,60 @@ class TestDockerSandboxServiceHostNetwork:# Verify no warning was logged about port collisionmock_logger.warning.assert_not_called()++ @patch('openhands.app_server.sandbox.docker_sandbox_service.utc_now')+ async def test_container_to_checked_sandbox_info_uses_container_started_at(+ self, mock_utc_now, service+ ):+ """Test that health check uses container's StartedAt for grace period calculation instead of sandbox created_at.++ This tests the fix for the bug where resuming a stopped container incorrectly+ marked the sandbox as ERROR instead of STARTING because it was using+ sandbox_info.created_at (when the sandbox record was created) instead of+ the actual container start time.+ """+ # Setup - create a fresh container with State that includes StartedAt+ container = MagicMock()+ container.name = 'oh-test-abc123'+ container.status = 'running'+ container.image.tags = ['spec456']+ now = datetime(2024, 1, 15, 12, 0, 0, tzinfo=timezone.utc)+ mock_utc_now.return_value = now++ # Container started 4 seconds ago (within 15s grace period)+ container_started_within_grace_period = datetime(+ 2024, 1, 15, 11, 59, 56, tzinfo=timezone.utc+ )+ # Sandbox was created 5 days ago (way outside grace period)+ sandbox_created_long_ago = datetime(2024, 1, 10, 10, 0, 0, tzinfo=timezone.utc)++ container.attrs = {+ 'Created': '2024-01-15T10:30:00.000000000Z',+ 'State': {'StartedAt': container_started_within_grace_period.isoformat()},+ 'Config': {+ 'Env': [+ 'OH_SESSION_API_KEYS_0=session_key_123',+ 'OTHER_VAR=other_value',+ ],+ 'WorkingDir': '/workspace',+ },+ 'NetworkSettings': {+ 'Ports': {+ '8000/tcp': [{'HostPort': '12345'}],+ '8001/tcp': [{'HostPort': '12346'}],+ }+ },+ }++ # Create a sandbox_info with old created_at (simulates old sandbox record)+ sandbox_info = await service._container_to_sandbox_info(container)+ sandbox_info.created_at = sandbox_created_long_ago++ # Health check fails but container was started recently (within 15s grace period)+ service.httpx_client.get.side_effect = httpx.HTTPError('Health check failed')++ result = await service._container_to_checked_sandbox_info(container)++ # Verify - should be STARTING because container started within grace period+ assert result is not None+ assert result.status == SandboxStatus.STARTING