Retry transient Docling readiness connections

backend/app/services/readiness.py

@@ -16,6 +16,7 @@
 from app.services.runtime_paths import resolve_binary
 
 _DOCLING_HEALTH_TIMEOUT_SECONDS = 5.0
+_DOCLING_HEALTH_CONNECT_RETRIES = 2
 
 
 def _check_payload(
@@ -150,7 +151,13 @@ async def _docling_check(settings: Any, diagnostics: dict[str, Any]) -> dict[str
             headers = {"Authorization": f"Bearer {token}"} if token else {}
             try:
                 timeout = httpx.Timeout(_DOCLING_HEALTH_TIMEOUT_SECONDS)
-                async with httpx.AsyncClient(timeout=timeout) as client:
+                transport = httpx.AsyncHTTPTransport(
+                    retries=_DOCLING_HEALTH_CONNECT_RETRIES
+                )
+                async with httpx.AsyncClient(
+                    timeout=timeout,
+                    transport=transport,
+                ) as client:
                     response = await client.get(health_url, headers=headers)
             except httpx.ConnectError:
                 return _check_payload(

backend/tests/test_readiness.py

@@ -40,9 +40,14 @@ def _remote_docling_settings(tmp_path, *, token=""):
 def _patch_async_client(monkeypatch, *, response=None, error=None):
     calls = []
 
+    class FakeTransport:
+        def __init__(self, *, retries):
+            self.retries = retries
+
     class FakeAsyncClient:
-        def __init__(self, *, timeout):
+        def __init__(self, *, timeout, transport):
             self.timeout = timeout
+            self.transport = transport
 
         async def __aenter__(self):
             return self
@@ -51,11 +56,19 @@ async def __aexit__(self, _exc_type, _exc_value, _traceback):
             return None
 
         async def get(self, url, *, headers):
-            calls.append({"timeout": self.timeout, "url": url, "headers": headers})
+            calls.append(
+                {
+                    "timeout": self.timeout,
+                    "transport": self.transport,
+                    "url": url,
+                    "headers": headers,
+                }
+            )
             if error is not None:
                 raise error
             return response
 
+    monkeypatch.setattr(readiness.httpx, "AsyncHTTPTransport", FakeTransport)
     monkeypatch.setattr(readiness.httpx, "AsyncClient", FakeAsyncClient)
     return calls
 
@@ -138,6 +151,7 @@ async def test_remote_docling_health_check_uses_canonical_endpoint_and_timeout(
     assert calls[0]["headers"] == {}
     assert calls[0]["timeout"].connect == 5
     assert calls[0]["timeout"].read == 5
+    assert calls[0]["transport"].retries == 2
 
 
 @pytest.mark.asyncio