From a7f01e7f05cf97796fa49ae53b86154013116c15 Mon Sep 17 00:00:00 2001 From: David Igou Date: Wed, 9 Sep 2026 12:44:27 +0000 Subject: [PATCH] fix(ace-step): preserve manually scaled replicas ArgoCD must honor the replica ignore rule during sync so manual GPU workload scaling remains durable. Assisted-by: GPT-5.6 Luna --- applications/ace-step/README.md | 5 +++-- applications/ace-step/kustomization.yaml | 2 +- clusters/ocp/values.yaml | 4 +++- 3 files changed, 7 insertions(+), 4 deletions(-) diff --git a/applications/ace-step/README.md b/applications/ace-step/README.md index fe3aa2a2..08cf6320 100644 --- a/applications/ace-step/README.md +++ b/applications/ace-step/README.md @@ -16,9 +16,10 @@ pod also carries the burst nodeSelector and `workload=burst` toleration. Claiming both GPUs means ComfyUI and casval-hosted llmkube models (`qwen3-35b`, `qwen38-27b`) cannot run concurrently. Scale those to `0` -before leaving ACE-Step at `1`. +before starting ACE-Step. -ArgoCD ignores `/spec/replicas` so a live scale-down is not reverted. +ArgoCD ignores `/spec/replicas` and uses `RespectIgnoreDifferences=true`, so +a live scale-down is not reverted during sync. ## Persistence diff --git a/applications/ace-step/kustomization.yaml b/applications/ace-step/kustomization.yaml index 6eacac67..636ce1b9 100644 --- a/applications/ace-step/kustomization.yaml +++ b/applications/ace-step/kustomization.yaml @@ -16,7 +16,7 @@ helmCharts: controllers: ace-step: type: deployment - replicas: 1 + replicas: 0 # Exclusive nvidia.com/gpu: 2 — a rolling update would deadlock # waiting for a second pair of GPUs that casval does not have. strategy: Recreate diff --git a/clusters/ocp/values.yaml b/clusters/ocp/values.yaml index 31038934..8dc39ef7 100644 --- a/clusters/ocp/values.yaml +++ b/clusters/ocp/values.yaml @@ -609,7 +609,9 @@ applications: source: path: applications/ace-step # Replicas are scaled manually. Leave them alone so ArgoCD does not - # revert a scale-down back to the 1 checked into git. + # revert a scale-down back to the 0 checked into git. + syncOptions: + - RespectIgnoreDifferences=true ignoreDifferences: - group: apps kind: Deployment