Repository navigation
Fix benchmark timing, deployment safety and integrate Onggi P2 - #53
Merged
Merged
Conversation
gspark-etri
added a commit
to ai-computing/hetero-llm-d
that referenced
this pull request
Oct 23, 2026
Reinstalling the ai-rm manifests reset controller-owned placement data, and a successful Deployment patch followed by a failed KEDA patch could prevent all later retries. Bootstrap the placement ConfigMap only when absent, reconcile each object's generation and actual state independently, preserve newer generations, and abort on KEDA discovery/read failures before patching. Validation: - Linux offline checks passed: 16 placement/EPP tests, 5 installer tests, and 1 model-store discovery test; EPP configs and Bash syntax checks passed. Regressions were observed before fixes and independent review completed. - Actual Kubernetes/KEDA 2.19.0 API validation used a paused ScaledObject and zero-replica Deployment in an isolated namespace. The Deployment patch succeeded; the following KEDA patch intentionally received HTTP 403. Replanning emitted only the unfinished ScaledObject, retry reduced maxReplicaCount 4→2, the Deployment generation stayed unchanged, and the final plan was empty. - Namespace-scoped permissions and revocation were verified. Temporary test resources were deleted. The required local research report, STATUS/event updates, and sanitized API evidence are included in experiments/2026-10-23-placement-api.md and experiments/data/placement-api-20261023.json. The companion chart GET permission fix is in ai-computing/ai-rm#53. No production rollout or live inference-serving benchmark was performed.
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
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.
JSON fallback responses could fabricate zero TPOT and pass SLO checks without timing evidence, and slow response drips or multiple connection addresses could exceed the intended timeout. Preserve unknown/invalid timing, reject contradictory zero counts, and share one I/O deadline across connection attempts and streaming/JSON responses.
Chart-backed components now stop before generic kubectl mutation and print a Helm command. Controller apply mode grants namespace-scoped Deployment get/list/patch permissions, including legacy --apply values; operational flag abbreviations are rejected. Live API validation exposed kubectl patch's prerequisite GET, which is now covered by chart regressions. Helm setup precedes tools tests in CI so RBAC regressions run.
Update English/Korean READMEs, component/status docs and the Onggi pin from P0 to reviewed P2 commit d0b16db. Depends on ai-computing/onggi#67; merge that first. Other ecosystem fixes remain separate: ai-computing/hetero-llm-d#12 and ai-computing/hetero-npu-p2p#10. Their hub pins are unchanged.
Validation:
Full commands, sanitized API evidence, and boundaries: docs/superpowers/validation-2026-10-23.md and docs/superpowers/artifacts/controlplane-2026-10-23.json.
No production rollout or actual NPU-serving benchmark performed. OS DNS remains outside socket cancellation; live TLS/proxy paths were not exercised.