Repository navigation
break(framework): Generate error replies for unavailable nodes in Strategy.start() - #8426
Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 23a31bc78b
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
🟡 Changes recommended
The shared polling interval overrides simulation configuration and can omit replies that arrive within the timeout.
1 open issue
What changed in this PR
Updates built-in strategies to stop waiting for unavailable SuperNodes, following the destination-oriented delivery changes in #8409.
Changes:
- Shares reply collection between training and evaluation, preserving received replies while skipping unavailable nodes.
- Rejects empty replies in differential-privacy validation.
- Adds partial-reply tests and documents unavailable-node handling.
The changes are cohesive and do not need splitting. No separate readability or consistency concerns remain; approval is blocked by a polling regression.
| File | Description |
|---|---|
framework/py/flwr/serverapp/strategy/strategy.py |
Uses shared reply collection and updates timeout documentation. |
framework/py/flwr/serverapp/strategy/strategy_utils.py |
Adds availability-aware message polling. |
framework/py/flwr/serverapp/strategy/strategy_utils_test.py |
Tests unavailable nodes and timeout with partial replies. |
framework/py/flwr/serverapp/strategy/dp_fixed_clipping.py |
Rejects empty reply sets before aggregation. |
Review effort: Balanced
github-actions
Bot
added
the
Maintainer
panh99
changed the title
Strategy.start()
Issue
Description
When a SuperNode goes offline, the SuperLink no longer generates an error reply. Built-in strategies therefore keep waiting for that node until the round timeout expires.
Related issues/PRs
Follow-up to #8409.
Proposal
Explanation
Use a shared strategy helper for training and evaluation to push messages once, collect replies, and stop waiting for outstanding messages whose destinations disappear from
Grid.get_node_ids(). Keep replies received before a node becomes unavailable and preserve the timeout for nodes that remain available.Reject empty replies in the shared DP validator so adaptive clipping skips empty rounds without dividing by zero.