[core][oom] Use retriable lifo policy for dask 3x nightly test (ray-…

…project#32361) Signed-off-by: Clarence Ng <[email protected]> 3x nightly dask test is failing, due to enabling of group-by-owner oom killer policy This switches the test to use the previous policy Signed-off-by: Edward Oakes <[email protected]>
edoakes · Mar 22, 2023 · acbee19 · acbee19
1 parent d382c02
commit acbee19
Showing 1 changed file with 5 additions and 0 deletions.
diff --git a/release/nightly_tests/dask_on_ray/dask_on_ray_app_config.yaml b/release/nightly_tests/dask_on_ray/dask_on_ray_app_config.yaml
@@ -1,4 +1,9 @@
 base_image: {{ env["RAY_IMAGE_NIGHTLY_CPU"] | default("anyscale/ray:nightly-py37") }}
+# We use retriable_lifo as the workload can crash due to multiple tasks from different
+# callers running on the same node, we also observed raylet memory leak that would
+# trigger the group-by-policy to fail the workload.
+# https://github.com/ray-project/ray/issues/32195
+env_vars: {"RAY_worker_killing_policy": "retriable_lifo"}
 debian_packages: []
 
 python: