Endpoint outages: one log line per outage, run wait markers, and outerloop status - #448
Merged
Merged
Conversation
… and a read-only outerloop status A run whose on-prem server is unavailable now records when it started waiting, the fleet logs one line when an endpoint goes down and one when it recovers, and outerloop status (text or --json) lists active runs and outages without touching state. A server record may name its model and an expiry; a mismatch or an expired record counts as unavailable.
# Conflicts: # CHANGELOG.md # docs/endpoints.md # src/outerloop/harness.py
There was a problem hiding this comment.
Round 1 — reviewed head 1b695d7a — reviewer summarizer:hermes/gpt-5.6-terra over coverage+credentials+deployment+general+lifecycle+prose.
terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.
Verdict: nothing blocking — 1 advisory note.
1 finding attached to the lines below.
Merged verdict: one non-blocking concurrency finding in endpoint outage recovery. Rejected findings: none.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
When an on-prem endpoint is down, runs that need it park and wait. Today the only trace is a warning repeated on every tick. This makes outages visible to the operator, and only to the operator: the GitHub board stays about agents and results.
stage.endpoint_wait(the endpoint and when it started waiting). A successful session probe clears it.endpoint-waits/<profile>.json. No notifications, no extra probes.outerloop status [--root] [--json]is read-only: active runs with their author route, override flag, phase, GPU-hours used and budget, and endpoint waits; plus current outages. No GitHub, scheduler, probes, locks or writes.modeland anexpires_at(Unix seconds). A model mismatch or an expired record counts as unavailable, so a run never talks to the wrong model after a server swap. Records are read with a 64 KiB bound.Compatibility: state read across versions gains two additive pieces,
stage.endpoint_waitin run records and theendpoint-waits/journal. Records and roots without them mean no recorded wait; no backfill is needed. Ended runs are never touched. The legacy author-route fixture covers repeated reads and a retry after an interrupted write; unknown stage and journal keys are preserved. Rolling back to the previous kernel ignores both: wait visibility is lost, run and PR lifecycle are unchanged. TheUpgrading:line is in the CHANGELOG.Gate: pytest (2789 passed), ruff check, ruff format --check, mypy.