retry on connection errors

eval_protocol/pytest/exception_config.py

@@ -23,6 +23,7 @@ def get_default_retryable_exceptions() -> Set[Type[Exception]]:
         return _default_retryable_exceptions
 
     # Lazy imports (these are expensive)
+    import aiohttp
     import httpx
     import litellm
     import requests
@@ -32,6 +33,9 @@ def get_default_retryable_exceptions() -> Set[Type[Exception]]:
         ConnectionError,  # type: ignore[assignment]
         TimeoutError,  # type: ignore[assignment]
         OSError,  # type: ignore[assignment]  # Covers network-related OS errors
+        # aiohttp library exceptions
+        aiohttp.ClientConnectionError,
+        aiohttp.ServerDisconnectedError,
         # Requests library exceptions
         requests.exceptions.ConnectionError,
         requests.exceptions.Timeout,

eval_protocol/pytest/remote_rollout_processor.py

@@ -104,11 +104,27 @@ async def _process_row(row: EvaluationRow) -> EvaluationRow:
             try:
                 session = self._get_or_create_session()
                 async with session.post(init_url, json=init_payload.model_dump(), timeout=timeout_init) as resp:
+                    if resp.status >= 500:
+                        body = await resp.text()
+                        raise ConnectionError(f"Remote /init returned server error (HTTP {resp.status}): {body}")
                     if resp.status >= 400:
                         body = await resp.text()
                         raise RuntimeError(f"Remote /init failed (HTTP {resp.status}): {body}")
                     resp.raise_for_status()
                     await resp.read()  # Drain the response body and release the connection back to the pool
+            except asyncio.CancelledError:
+                # Distinguish intentional cancellation (Ctrl+C, test teardown) from
+                # aiohttp-internal cancellation caused by a poisoned DNS resolver
+                # after a server disconnect.  Task.cancelling() returns the number
+                # of pending cancel() calls; > 0 means someone explicitly cancelled
+                # this task.
+                current = asyncio.current_task()
+                if current is not None and current.cancelling() > 0:  # pyright: ignore[reportAttributeAccessIssue]
+                    raise  # Intentional cancellation — propagate immediately
+                # Network-level failure; discard the session so retries get a
+                # fresh connection pool.
+                self._session = None
+                raise ConnectionError("Remote server connection lost (request cancelled)")
             except asyncio.TimeoutError:
                 raise TimeoutError(
                     f"The /init endpoint tried {init_url} with {init_payload.model_dump()} but timed out after 300 seconds."