Move supervision polling task to PythonActorMeshImpl constructor #1628
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
meta-cla
bot
added
the
CLA Signed
|
@dulinriley has exported this pull request. If you are a Meta employee, you can view the originating Diff in D85093139. |
dulinriley
added a commit
to dulinriley/monarch
that referenced
this pull request
Oct 21, 2025
…a-pytorch#1628) Summary: Part of meta-pytorch#1209 Supervision was not working in cases where there was no "await" of an endpoint. This didn't provide error messages when an endpoint was broadcasted to without waiting for a result. This can happen with "broadcast", but also with implicit messages like the `__init__` constructor as well. To fix this, move the creation of the state polling loop from the first call to "supervision_event" to closer to construction time. It cannot be done in the constructor due to needing an "Instance" which is not availble, particularly when unpickling an ActorMeshRef. Each "supervision_event" call clones a tokio::watch::Receiver so it gets the same most recent event as every other endpoint. This also allows a mesh to report an error to its owner much earlier if that mesh was idle. Also improve the stringification of MeshFailure to not use the Debug format which is too verbose for ActorSupervisionEvent. Differential Revision: D85093139
dulinriley
force-pushed
the
export-D85093139
branch
from
b7869e1 to
f058629
Compare
…a-pytorch#1628) Summary: Part of meta-pytorch#1209 Supervision was not working in cases where there was no "await" of an endpoint. This didn't provide error messages when an endpoint was broadcasted to without waiting for a result. This can happen with "broadcast", but also with implicit messages like the `__init__` constructor as well. To fix this, move the creation of the state polling loop from the first call to "supervision_event" to closer to construction time. It cannot be done in the constructor due to needing an "Instance" which is not availble, particularly when unpickling an ActorMeshRef. Each "supervision_event" call clones a tokio::watch::Receiver so it gets the same most recent event as every other endpoint. This also allows a mesh to report an error to its owner much earlier if that mesh was idle. Also improve the stringification of MeshFailure to not use the Debug format which is too verbose for ActorSupervisionEvent. Reviewed By: mariusae Differential Revision: D85093139
dulinriley
force-pushed
the
export-D85093139
branch
from
f058629 to
62727c5
Compare
|
This pull request has been merged in 2c811fd. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
Part of #1209
Supervision was not working in cases where there was no "await" of an endpoint.
This didn't provide error messages when an endpoint was broadcasted to without waiting
for a result. This can happen with "broadcast", but also with implicit messages like the
__init__constructor as well.
To fix this, move the creation of the state polling loop from the first call to "supervision_event" to closer
to construction time. It cannot be done in the constructor due to needing an "Instance" which is
not availble, particularly when unpickling an ActorMeshRef. Each "supervision_event" call clones
a tokio::watch::Receiver so it gets the same most recent event as every other endpoint.
This also allows a mesh to report an error to its owner much earlier if that mesh was idle.
Also improve the stringification of MeshFailure to not use the Debug format which is too verbose
for ActorSupervisionEvent.
Differential Revision: D85093139