Update Retryable Exceptions

aana/sdk.py

@@ -45,15 +45,19 @@ def __init__(
             name (str, optional): The name of the application. Defaults to "app".
             migration_func (Callable | None): The migration function to run. Defaults to None.
             retryable_exceptions (list[Exception, str] | None): The exceptions that can be retried in the task queue.
-                                                                Defaults to ['InferenceException'].
+                                                                Defaults to ['InferenceException', 'ActorDiedError', 'OutOfMemoryError'].
         """
         self.name = name
         self.migration_func = migration_func
         self.endpoints: dict[str, Endpoint] = {}
         self.deployments: dict[str, Deployment] = {}
 
         if retryable_exceptions is None:
-            self.retryable_exceptions = [InferenceException]
+            self.retryable_exceptions = [
+                "InferenceException",
+                "ActorDiedError",
+                "OutOfMemoryError",
+            ]
         else:
             self.retryable_exceptions = retryable_exceptions
         # Convert exceptions to string if they are not already

aana/utils/core.py

@@ -93,7 +93,9 @@ async def sleep_exponential_backoff(
         attempts (int): The number of attempts so far.
         jitter (bool): Whether to add jitter to the delay. Default is True.
     """
-    delay = min(initial_delay * (2**attempts), max_delay)
+    # Prevent overflow by using min(attempt, 32) since 2^32 is already huge
+    capped_attempt = min(attempts, 32)
+    delay = min(initial_delay * (2**capped_attempt), max_delay)
     # Full jitter
     delay_with_jitter = random.uniform(0, delay) if jitter else delay  # noqa: S311
     await asyncio.sleep(delay_with_jitter)