Fix bug in ray.errors and update its default behavior (#5576)

python/ray/tests/test_stress.py

@@ -11,6 +11,7 @@ import time
 
 import ray
 from ray.tests.cluster_utils import Cluster
+from ray.tests.utils import flat_errors
 import ray.ray_constants as ray_constants
 
 
@@ -397,13 +398,13 @@ def wait_for_errors(error_check):
     errors = []
     time_left = 100
     while time_left > 0:
-        errors = ray.errors()
+        errors = flat_errors()
         if error_check(errors):
             break
         time_left -= 1
         time.sleep(1)
 
-        # Make sure that enough errors came through.
+    # Make sure that enough errors came through.
     assert error_check(errors)
     return errors
 

python/ray/tests/utils.py

@@ -75,8 +75,15 @@ def run_string_as_driver_nonblocking(driver_script):
             [sys.executable, f.name], stdout=subprocess.PIPE)
 
 
+def flat_errors():
+    errors = []
+    for job_errors in ray.errors(all_jobs=True).values():
+        errors.extend(job_errors)
+    return errors
+
+
 def relevant_errors(error_type):
-    return [info for info in ray.errors() if info["type"] == error_type]
+    return [error for error in flat_errors() if error["type"] == error_type]
 
 
 def wait_for_errors(error_type, num_errors, timeout=10):