Skip to content

[DataFlow runtime · M5 2/4] Control-plane backpressure (watermarks, caps, starvation)#607

Merged
jiapingW merged 1 commit into
mainfrom
dataflow-up-10-m5-backpressure
Jun 29, 2026
Merged

[DataFlow runtime · M5 2/4] Control-plane backpressure (watermarks, caps, starvation)#607
jiapingW merged 1 commit into
mainfrom
dataflow-up-10-m5-backpressure

Conversation

@maocheng23

Copy link
Copy Markdown
Collaborator

Adds BackpressureController (hysteresis high/low watermarks, per-worker inflight caps, train-lease caps, separate rollout/trainer starvation) wired optionally into DataFlowController.

Part of the DataFlow runtime M5/M6 stacked series (continues the M1–M4 work in #594#601 / #603). Stacked PRs — merge bottom-up (up-9 first). Lint (pre-commit) + runtime CPU test suite green.

🤖 Generated with Claude Code

…tion)

BackpressureController is pure policy: it reads data-plane capacity through a
narrow CapacityReporter (FeatureStore.health, ints only) and decides when the
controller pauses prompt leasing. The store reports capacity; the controller
decides scheduling (ownership per control_plane/backpressure.md).

- high/low byte watermarks with a latched hysteresis band (no flapping)
- per-rollout-worker in-flight prompt cap; per-call trainer-lease cap
- rollout-starvation vs trainer-starvation counted separately
- DataFlowController.status() now exposes train_backlog + backpressure snapshot
- default (no policy) preserves exact M1-M4 lease behavior

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@maocheng23 maocheng23 requested a review from FrankLeeeee as a code owner June 28, 2026 00:33
@gemini-code-assist

Copy link
Copy Markdown
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-9-m5-dataplane to main June 29, 2026 13:36
@jiapingW jiapingW merged commit 204299a into main Jun 29, 2026
1 check passed
@jiapingW jiapingW deleted the dataflow-up-10-m5-backpressure branch June 29, 2026 13:39
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants