feat(chart): make the apihost HPA usable (memory target, behavior, no scale-to-1)

Three gaps in the existing autoscaling scaffolding:

- targetMemoryUtilizationPercentage (0 = off) so the memory metric can be
  declared at all. It is the metric that matters here: apihost limits 1Gi
  against a 256Mi request, and the oxocache L1 grows inside the process,
  so CPU alone would let a pod reach its limit and be OOM-killed before
  HPA ever reacted.
- behavior passthrough, so a release can tune stabilization windows and
  per-period policies instead of living with the cluster default.
- the Deployment now renders replicas = minReplicas when autoscaling is on
  instead of omitting the field. Omitting it is a trap: helm's three-way
  merge turns the absent field into a null patch, the API server re-applies
  its default of 1, and every upgrade briefly drops the main API to a
  single replica before the HPA notices.

ZhenYi committed 1 month ago

parent 20a153a commit 7b1d475 Browse files

File tree 2

apihost.yaml Modified
values.yaml Modified
No files match.

chart/templates/apihost.yaml +19 −1 View

@@ -7,7 +7,13 @@

77 {{- include "gitdata.labels" . | nindent 4 }}

88 app.kubernetes.io/component: apihost

99spec:

10 {{- if not .Values.apihost.autoscaling.enabled }}

10 {{- if .Values.apihost.autoscaling.enabled }}

11 # 开着 HPA 时也要写 replicas，且写 minReplicas：省略该字段会让 helm 的

12 # three-way merge 把它 patch 成 null，API server 随即按默认值收成 1，

13 # 升级瞬间主 API 少一个副本。写 minReplicas 则升级前后副本数不变，

14 # 之后完全由 HPA 接管（超出 min 的扩容不会被这张清单拉回）。

15 replicas: {{ .Values.apihost.autoscaling.minReplicas }}

16 {{- else }}

1117 replicas: {{ .Values.apihost.replicaCount }}

1218 {{- end }}

1319 revisionHistoryLimit: 3

@@ -138,6 +144,18 @@

138144 target:

139145 type: Utilization

140146 averageUtilization: {{ .Values.apihost.autoscaling.targetCPUUtilizationPercentage }}

147 {{- if gt (int (.Values.apihost.autoscaling.targetMemoryUtilizationPercentage | default 0)) 0 }}

148 - type: Resource

149 resource:

150 name: memory

151 target:

152 type: Utilization

153 averageUtilization: {{ .Values.apihost.autoscaling.targetMemoryUtilizationPercentage }}

154 {{- end }}

155 {{- with .Values.apihost.autoscaling.behavior }}

156 behavior:

157 {{- toYaml . | nindent 4 }}

158 {{- end }}

141159{{- end }}

142160{{- if .Values.apihost.pdb.enabled }}

143161---

chart/values.yaml +5 −0 View

@@ -109,6 +109,11 @@

109109 minReplicas: 2

110110 maxReplicas: 8

111111 targetCPUUtilizationPercentage: 70

112 # 0 disables the memory metric. The limit is 1Gi while the request is 256Mi,

113 # so a memory target is what actually keeps the L1 cache from being OOM-killed.

114 targetMemoryUtilizationPercentage: 0

115 # Empty map renders no `behavior:` block (k8s default scaling policy).

116 behavior: {}

112117 maxSurge: 1

113118 maxUnavailable: 0

114119 podAntiAffinity: soft # soft | hard | ""