-
Notifications
You must be signed in to change notification settings - Fork 6.8k
Open
Labels
community-backlogcompiled-graphscoreIssues that should be addressed in Ray CoreIssues that should be addressed in Ray CoreenhancementRequest for new feature and/or capabilityRequest for new feature and/or capability
Description
Description
In #49711 , when a reader reads from a list of channels and one of them returns RayTaskError
, a list of the same RayTaskError
is returned to the user.
The change already provided the following:
- Early failure detection: failure would be raised early if it happens in later channels
- Error message where the failure happened
However, we want to consider and revisit if the provided information is accurate and comprehensive enough:
- Would it be useful to indicate the exact channel that has the failure: [core][compiled-graphs] Read input channels in a round-robin manner with a short timeout, and fail fast if a RayTaskError is found #49711 (comment)
- For unfinished channels, another idea is to have a special value "RayChannelPendingValue" to indicate it's not finished, i.e., the return list has both RayTaskError and RayChannelPendingValue
- For already finished results, we should probably just return them instead of overwriting with RayTaskError from the problematic channel
- Evaluate above for multi-return case
cc: @kevin85421
Use case
No response
Metadata
Metadata
Assignees
Labels
community-backlogcompiled-graphscoreIssues that should be addressed in Ray CoreIssues that should be addressed in Ray CoreenhancementRequest for new feature and/or capabilityRequest for new feature and/or capability