Skip to content

Task arguments are retrieved serially from the object store. #2891

@robertnishihara

Description

@robertnishihara

@stephanie-wang noticed that when we do ray.get, we fetch objects in parallel from the object store:

ray/python/ray/worker.py

Lines 2510 to 2511 in e05baed

if isinstance(object_ids, list):
values = worker.get_object(object_ids)

But when we get the arguments for a task, we get them serially:

ray/python/ray/worker.py

Lines 850 to 853 in e05baed

for (i, arg) in enumerate(serialized_args):
if isinstance(arg, ray.ObjectID):
# get the object from the local object store
argument = self.get_object([arg])[0]

For settings when many object IDs are passed into a remote function, this can slow things down.

Metadata

Metadata

Assignees

No one assigned

    Labels

    good-first-issueGreat starter issue for someone just starting to contribute to Ray

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions