[DataFlow runtime · online] Two-axis staleness gate for online disaggregated training#623
Closed
maocheng23 wants to merge 1 commit into
Closed
[DataFlow runtime · online] Two-axis staleness gate for online disaggregated training#623maocheng23 wants to merge 1 commit into
maocheng23 wants to merge 1 commit into
Conversation
…raining
Online training keeps producing fresh draft weights while the rollout pool keeps
generating samples; without hot-swapping the rollout (out of scope -- no
hot-update), the pool drifts behind and its samples go stale. Training on stale
features hurts, so the consumer must drop them. This adds the staleness half of
the weight lifecycle (no hot-update / rollback / serving accept-length gate):
* contracts: WeightVersion (+ WeightStatus) -- metadata-only published draft
version; publish order is the staleness axis.
* control_plane/version_policy.py:
- WeightRegistry: publish-ordered history; draft_lag(sample) = distance from
the newest published version (None = unknown -> maximally stale). Optional
durable metadata_store backing.
- StalenessPolicy: two axes -- draft (max_draft_lag) + target
(require_target_match) -- returning an accept/reasons assessment.
- DriftMonitor: rolling mean/max draft lag; drifting() for the orchestrator.
- StalenessGatedQueue: wraps a ref queue so the loader only sees fresh refs;
stale ones are acked (backpressure clears) and their features aborted (freed)
instead of trained on.
* build_disagg_online_consumer gains optional staleness_policy / weight_registry
/ current_target_version / drift_monitor: when set, the streamed queue is gated
so a drifted rollout can't poison training. Off by default (every ref trained).
11 CPU tests (registry lag, both policy axes, unknown-version, drift, gate
drop+abort+passthrough).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
Base automatically changed from
dataflow-up-17-online-disagg
to
dataflow-up-16-zerocopy
June 30, 2026 13:29
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
Two-axis staleness for the online consumer so it drops features made under a stale draft/target version. Adds
contracts.WeightVersion/WeightStatus+control_plane/version_policy.py(WeightRegistry publish/latest/draft_lag, StalenessPolicy two-axis, DriftMonitor, StalenessGatedQueue) + optional gate params onbuild_disagg_online_consumer(off by default). NO hot_update/rollback/accept-length.Stacked on the online-disagg PR above. Recreated (rebased) from the closed #615.
🤖 Generated with Claude Code