From f1f30b3071254e3640292b4eb483b04f5d40d77a Mon Sep 17 00:00:00 2001 From: Yan Bondarenko <202671653+yankawai@users.noreply.github.com> Date: Tue, 20 Oct 2026 12:05:51 +0300 Subject: [PATCH 1/2] fix(etcd): require an even spread of members over nodes by default The default spread of the etcd members was ScheduleAnyway, which only adds to the score of a node, and other scores can outweigh it. A new etcd on a three-node cluster got its first two members on one node. On a node-local StorageClass a member cannot leave its node, so losing that node takes two of the three members and the quorum. DoNotSchedule with maxSkew 1 still lets an etcd with more members than nodes start, because the skew is counted between the nodes that exist. One node takes all three members and two nodes take two and one. nodeTaintsPolicy is Ignore, so a cordoned node still counts and a member waits for it instead of landing next to another member. The operator copies the constraints into an EtcdMember when it creates it and builds the Pod from that copy, so members that exist keep the old setting and only members created later get the new one. Assisted-by: LLM Signed-off-by: Yan Bondarenko <202671653+yankawai@users.noreply.github.com> --- .../extra/etcd/templates/etcd-cluster.yaml | 3 +- .../etcd/tests/topology-spread_test.yaml | 46 ++++++++++++++----- 2 files changed, 37 insertions(+), 12 deletions(-) diff --git a/packages/extra/etcd/templates/etcd-cluster.yaml b/packages/extra/etcd/templates/etcd-cluster.yaml index 2824ce371..67dd34bd9 100644 --- a/packages/extra/etcd/templates/etcd-cluster.yaml +++ b/packages/extra/etcd/templates/etcd-cluster.yaml @@ -69,7 +69,8 @@ spec: topologySpreadConstraints: - maxSkew: 1 topologyKey: "kubernetes.io/hostname" - whenUnsatisfiable: ScheduleAnyway + whenUnsatisfiable: DoNotSchedule + nodeTaintsPolicy: Ignore labelSelector: matchLabels: app.kubernetes.io/instance: etcd diff --git a/packages/extra/etcd/tests/topology-spread_test.yaml b/packages/extra/etcd/tests/topology-spread_test.yaml index 4b96fc310..11bb92243 100644 --- a/packages/extra/etcd/tests/topology-spread_test.yaml +++ b/packages/extra/etcd/tests/topology-spread_test.yaml @@ -8,7 +8,7 @@ templates: - templates/etcd-cluster.yaml tests: - - it: uses a soft per-host spread by default + - it: uses a strict per-host spread by default set: _cluster: {} documentSelector: @@ -16,17 +16,41 @@ tests: value: EtcdCluster asserts: - equal: - path: spec.topologySpreadConstraints[0].maxSkew - value: 1 + path: spec.topologySpreadConstraints[0] + value: + maxSkew: 1 + topologyKey: kubernetes.io/hostname + whenUnsatisfiable: DoNotSchedule + nodeTaintsPolicy: Ignore + labelSelector: + matchLabels: + app.kubernetes.io/instance: etcd + - lengthEqual: + path: spec.topologySpreadConstraints + count: 1 + + - it: renders a soft cluster-wide spread policy as written + set: + _cluster: + scheduling: + globalAppTopologySpreadConstraints: | + topologySpreadConstraints: + - maxSkew: 1 + topologyKey: kubernetes.io/hostname + whenUnsatisfiable: ScheduleAnyway + documentSelector: + path: kind + value: EtcdCluster + asserts: - equal: - path: spec.topologySpreadConstraints[0].topologyKey - value: kubernetes.io/hostname - - equal: - path: spec.topologySpreadConstraints[0].whenUnsatisfiable - value: ScheduleAnyway - - equal: - path: spec.topologySpreadConstraints[0].labelSelector.matchLabels["app.kubernetes.io/instance"] - value: etcd + path: spec.topologySpreadConstraints[0] + value: + maxSkew: 1 + topologyKey: kubernetes.io/hostname + whenUnsatisfiable: ScheduleAnyway + labelSelector: + matchLabels: + app.kubernetes.io/instance: etcd - lengthEqual: path: spec.topologySpreadConstraints count: 1 From 9f6fdd77c6ca964e0adcc387e50993487f2e594e Mon Sep 17 00:00:00 2001 From: Yan Bondarenko <202671653+yankawai@users.noreply.github.com> Date: Sat, 24 Oct 2026 15:52:37 +0300 Subject: [PATCH 2/2] docs(etcd): update the affinity notes for the hard member spread The affinity description and the comment above the affinity block explained the soft default by a member that would stay Pending on its node-local volume. Now that the spread is DoNotSchedule by default, the spread keeps such a member Pending as well, once its node holds another member and some other node holds none. Both texts now explain the soft default by what is left, that a required rule forbids two members on one node even when every node has one, and state the node-local case for both rules. The description also says that the operator copies affinity into each EtcdMember when it creates it, so a recreated Pod keeps the rule its member was created with. README, values.schema.json and the ApplicationDefinition are regenerated from it. Assisted-by: LLM Signed-off-by: Yan Bondarenko <202671653+yankawai@users.noreply.github.com> --- packages/extra/etcd/README.md | 20 +++++++++---------- .../extra/etcd/templates/etcd-cluster.yaml | 14 ++++++++----- packages/extra/etcd/values.schema.json | 2 +- packages/extra/etcd/values.yaml | 2 +- packages/system/etcd-rd/cozyrds/etcd.yaml | 2 +- 5 files changed, 22 insertions(+), 18 deletions(-) diff --git a/packages/extra/etcd/README.md b/packages/extra/etcd/README.md index e96c63832..91341ff32 100644 --- a/packages/extra/etcd/README.md +++ b/packages/extra/etcd/README.md @@ -34,14 +34,14 @@ coordinates). ### Common parameters -| Name | Description | Type | Value | -| ------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------- | -------- | -| `version` | etcd version (X.Y.Z) to run, set on the EtcdCluster's spec.version; the operator pulls etcd from its own configured image. | `string` | `3.6.11` | -| `size` | Persistent Volume size. | `quantity` | `4Gi` | -| `storageClass` | StorageClass used to store the data. | `string` | `""` | -| `replicas` | Number of etcd replicas. | `int` | `3` | -| `resources` | Resource configuration for etcd. | `object` | `{}` | -| `resources.cpu` | Number of CPU cores allocated. | `quantity` | `1000m` | -| `resources.memory` | Amount of memory allocated. | `quantity` | `512Mi` | -| `affinity` | Affinity rules set verbatim on the EtcdCluster's spec.affinity, replacing the chart default. When left empty the chart applies a soft (preferred) podAntiAffinity that spreads members across nodes by kubernetes.io/hostname on a best-effort basis, so a single node failure is less likely to take out more than one member. The default is intentionally soft: a hard rule would leave a member Pending when there are fewer schedulable nodes than replicas, or when node-local storage pins a member to a node. Set a hard requiredDuringSchedulingIgnoredDuringExecution rule here to forbid co-location outright (only on clusters with enough nodes and non-node-pinned storage). Note that the operator applies affinity only when it creates or replaces a member Pod and does not reschedule running members, so on a cluster that already has co-located members the rule takes effect only as each member is replaced, not retroactively. | `object` | `{}` | +| Name | Description | Type | Value | +| ------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------- | -------- | +| `version` | etcd version (X.Y.Z) to run, set on the EtcdCluster's spec.version; the operator pulls etcd from its own configured image. | `string` | `3.6.11` | +| `size` | Persistent Volume size. | `quantity` | `4Gi` | +| `storageClass` | StorageClass used to store the data. | `string` | `""` | +| `replicas` | Number of etcd replicas. | `int` | `3` | +| `resources` | Resource configuration for etcd. | `object` | `{}` | +| `resources.cpu` | Number of CPU cores allocated. | `quantity` | `1000m` | +| `resources.memory` | Amount of memory allocated. | `quantity` | `512Mi` | +| `affinity` | Affinity rules set verbatim on the EtcdCluster's spec.affinity, replacing the chart default. When left empty the chart applies a soft (preferred) podAntiAffinity by kubernetes.io/hostname on top of the topologySpreadConstraints, which by default require an even spread of members over nodes (whenUnsatisfiable DoNotSchedule, maxSkew 1) unless the platform sets scheduling.globalAppTopologySpreadConstraints. The anti-affinity stays soft because a required rule forbids two members on one node even when every node already has one, so an etcd with more members than schedulable nodes could not start all of them. Set a requiredDuringSchedulingIgnoredDuringExecution rule here only on clusters with at least as many schedulable nodes as members. With node-local storage a recreated member Pod can only return to its own node, so a required rule keeps it Pending while that node holds another member, and the default spread does the same once some other node holds none. The operator copies affinity into each EtcdMember when it creates the member and builds every Pod of that member from that copy, so a change does not reschedule running members and takes effect only as members are replaced by new ones. | `object` | `{}` | diff --git a/packages/extra/etcd/templates/etcd-cluster.yaml b/packages/extra/etcd/templates/etcd-cluster.yaml index 67dd34bd9..3676f53cb 100644 --- a/packages/extra/etcd/templates/etcd-cluster.yaml +++ b/packages/extra/etcd/templates/etcd-cluster.yaml @@ -79,11 +79,15 @@ spec: affinity: {{- toYaml . | nindent 4 }} {{- else }} - # Soft anti-affinity: spread etcd members across nodes on a best-effort - # basis so a single node failure is unlikely to take out more than one - # member. Preferred (not required) because a hard rule would leave a member - # Pending when replicas exceed schedulable nodes, or when node-local storage - # pins a member to a node. Override via .Values.affinity for a hard rule. + # Soft anti-affinity, a preference on top of the spread constraint above, + # which by default is the hard rule that keeps two members off one node + # while another node has none. Preferred, not required, because a required + # rule also forbids two members on one node when every node already has + # one, so an etcd with more members than schedulable nodes could not start + # all of them. With node-local storage a recreated member Pod can only + # return to its own node, and the default spread keeps it Pending there + # while that node holds another member and some other node holds none. + # Override via .Values.affinity for a hard rule. affinity: podAntiAffinity: preferredDuringSchedulingIgnoredDuringExecution: diff --git a/packages/extra/etcd/values.schema.json b/packages/extra/etcd/values.schema.json index b438ca9d7..c33219f4f 100644 --- a/packages/extra/etcd/values.schema.json +++ b/packages/extra/etcd/values.schema.json @@ -67,7 +67,7 @@ } }, "affinity": { - "description": "Affinity rules set verbatim on the EtcdCluster's spec.affinity, replacing the chart default. When left empty the chart applies a soft (preferred) podAntiAffinity that spreads members across nodes by kubernetes.io/hostname on a best-effort basis, so a single node failure is less likely to take out more than one member. The default is intentionally soft: a hard rule would leave a member Pending when there are fewer schedulable nodes than replicas, or when node-local storage pins a member to a node. Set a hard requiredDuringSchedulingIgnoredDuringExecution rule here to forbid co-location outright (only on clusters with enough nodes and non-node-pinned storage). Note that the operator applies affinity only when it creates or replaces a member Pod and does not reschedule running members, so on a cluster that already has co-located members the rule takes effect only as each member is replaced, not retroactively.", + "description": "Affinity rules set verbatim on the EtcdCluster's spec.affinity, replacing the chart default. When left empty the chart applies a soft (preferred) podAntiAffinity by kubernetes.io/hostname on top of the topologySpreadConstraints, which by default require an even spread of members over nodes (whenUnsatisfiable DoNotSchedule, maxSkew 1) unless the platform sets scheduling.globalAppTopologySpreadConstraints. The anti-affinity stays soft because a required rule forbids two members on one node even when every node already has one, so an etcd with more members than schedulable nodes could not start all of them. Set a requiredDuringSchedulingIgnoredDuringExecution rule here only on clusters with at least as many schedulable nodes as members. With node-local storage a recreated member Pod can only return to its own node, so a required rule keeps it Pending while that node holds another member, and the default spread does the same once some other node holds none. The operator copies affinity into each EtcdMember when it creates the member and builds every Pod of that member from that copy, so a change does not reschedule running members and takes effect only as members are replaced by new ones.", "type": "object", "default": {}, "x-kubernetes-preserve-unknown-fields": true diff --git a/packages/extra/etcd/values.yaml b/packages/extra/etcd/values.yaml index 61b994d7c..c5be94909 100644 --- a/packages/extra/etcd/values.yaml +++ b/packages/extra/etcd/values.yaml @@ -23,5 +23,5 @@ resources: cpu: 1000m memory: 512Mi -## @param {object} [affinity] - Affinity rules set verbatim on the EtcdCluster's spec.affinity, replacing the chart default. When left empty the chart applies a soft (preferred) podAntiAffinity that spreads members across nodes by kubernetes.io/hostname on a best-effort basis, so a single node failure is less likely to take out more than one member. The default is intentionally soft: a hard rule would leave a member Pending when there are fewer schedulable nodes than replicas, or when node-local storage pins a member to a node. Set a hard requiredDuringSchedulingIgnoredDuringExecution rule here to forbid co-location outright (only on clusters with enough nodes and non-node-pinned storage). Note that the operator applies affinity only when it creates or replaces a member Pod and does not reschedule running members, so on a cluster that already has co-located members the rule takes effect only as each member is replaced, not retroactively. +## @param {object} [affinity] - Affinity rules set verbatim on the EtcdCluster's spec.affinity, replacing the chart default. When left empty the chart applies a soft (preferred) podAntiAffinity by kubernetes.io/hostname on top of the topologySpreadConstraints, which by default require an even spread of members over nodes (whenUnsatisfiable DoNotSchedule, maxSkew 1) unless the platform sets scheduling.globalAppTopologySpreadConstraints. The anti-affinity stays soft because a required rule forbids two members on one node even when every node already has one, so an etcd with more members than schedulable nodes could not start all of them. Set a requiredDuringSchedulingIgnoredDuringExecution rule here only on clusters with at least as many schedulable nodes as members. With node-local storage a recreated member Pod can only return to its own node, so a required rule keeps it Pending while that node holds another member, and the default spread does the same once some other node holds none. The operator copies affinity into each EtcdMember when it creates the member and builds every Pod of that member from that copy, so a change does not reschedule running members and takes effect only as members are replaced by new ones. affinity: {} diff --git a/packages/system/etcd-rd/cozyrds/etcd.yaml b/packages/system/etcd-rd/cozyrds/etcd.yaml index 90ed8a80a..1ce88e8ae 100644 --- a/packages/system/etcd-rd/cozyrds/etcd.yaml +++ b/packages/system/etcd-rd/cozyrds/etcd.yaml @@ -8,7 +8,7 @@ spec: plural: etcds singular: etcd openAPISchema: |- - {"title":"Chart Values","type":"object","properties":{"version":{"description":"etcd version (X.Y.Z) to run, set on the EtcdCluster's spec.version; the operator pulls etcd from its own configured image.","type":"string","default":"3.6.11"},"size":{"description":"Persistent Volume size.","default":"4Gi","pattern":"^(\\+|-)?(([0-9]+(\\.[0-9]*)?)|(\\.[0-9]+))(([KMGTPE]i)|[numkMGTPE]|([eE](\\+|-)?(([0-9]+(\\.[0-9]*)?)|(\\.[0-9]+))))?$","anyOf":[{"type":"integer"},{"type":"string"}],"x-kubernetes-int-or-string":true},"storageClass":{"description":"StorageClass used to store the data.","type":"string","default":""},"replicas":{"description":"Number of etcd replicas.","type":"integer","default":3},"resources":{"description":"Resource configuration for etcd.","type":"object","default":{},"properties":{"cpu":{"description":"Number of CPU cores allocated.","default":"1000m","pattern":"^(\\+|-)?(([0-9]+(\\.[0-9]*)?)|(\\.[0-9]+))(([KMGTPE]i)|[numkMGTPE]|([eE](\\+|-)?(([0-9]+(\\.[0-9]*)?)|(\\.[0-9]+))))?$","anyOf":[{"type":"integer"},{"type":"string"}],"x-kubernetes-int-or-string":true},"memory":{"description":"Amount of memory allocated.","default":"512Mi","pattern":"^(\\+|-)?(([0-9]+(\\.[0-9]*)?)|(\\.[0-9]+))(([KMGTPE]i)|[numkMGTPE]|([eE](\\+|-)?(([0-9]+(\\.[0-9]*)?)|(\\.[0-9]+))))?$","anyOf":[{"type":"integer"},{"type":"string"}],"x-kubernetes-int-or-string":true}}},"affinity":{"description":"Affinity rules set verbatim on the EtcdCluster's spec.affinity, replacing the chart default. When left empty the chart applies a soft (preferred) podAntiAffinity that spreads members across nodes by kubernetes.io/hostname on a best-effort basis, so a single node failure is less likely to take out more than one member. The default is intentionally soft: a hard rule would leave a member Pending when there are fewer schedulable nodes than replicas, or when node-local storage pins a member to a node. Set a hard requiredDuringSchedulingIgnoredDuringExecution rule here to forbid co-location outright (only on clusters with enough nodes and non-node-pinned storage). Note that the operator applies affinity only when it creates or replaces a member Pod and does not reschedule running members, so on a cluster that already has co-located members the rule takes effect only as each member is replaced, not retroactively.","type":"object","default":{},"x-kubernetes-preserve-unknown-fields":true}}} + {"title":"Chart Values","type":"object","properties":{"version":{"description":"etcd version (X.Y.Z) to run, set on the EtcdCluster's spec.version; the operator pulls etcd from its own configured image.","type":"string","default":"3.6.11"},"size":{"description":"Persistent Volume size.","default":"4Gi","pattern":"^(\\+|-)?(([0-9]+(\\.[0-9]*)?)|(\\.[0-9]+))(([KMGTPE]i)|[numkMGTPE]|([eE](\\+|-)?(([0-9]+(\\.[0-9]*)?)|(\\.[0-9]+))))?$","anyOf":[{"type":"integer"},{"type":"string"}],"x-kubernetes-int-or-string":true},"storageClass":{"description":"StorageClass used to store the data.","type":"string","default":""},"replicas":{"description":"Number of etcd replicas.","type":"integer","default":3},"resources":{"description":"Resource configuration for etcd.","type":"object","default":{},"properties":{"cpu":{"description":"Number of CPU cores allocated.","default":"1000m","pattern":"^(\\+|-)?(([0-9]+(\\.[0-9]*)?)|(\\.[0-9]+))(([KMGTPE]i)|[numkMGTPE]|([eE](\\+|-)?(([0-9]+(\\.[0-9]*)?)|(\\.[0-9]+))))?$","anyOf":[{"type":"integer"},{"type":"string"}],"x-kubernetes-int-or-string":true},"memory":{"description":"Amount of memory allocated.","default":"512Mi","pattern":"^(\\+|-)?(([0-9]+(\\.[0-9]*)?)|(\\.[0-9]+))(([KMGTPE]i)|[numkMGTPE]|([eE](\\+|-)?(([0-9]+(\\.[0-9]*)?)|(\\.[0-9]+))))?$","anyOf":[{"type":"integer"},{"type":"string"}],"x-kubernetes-int-or-string":true}}},"affinity":{"description":"Affinity rules set verbatim on the EtcdCluster's spec.affinity, replacing the chart default. When left empty the chart applies a soft (preferred) podAntiAffinity by kubernetes.io/hostname on top of the topologySpreadConstraints, which by default require an even spread of members over nodes (whenUnsatisfiable DoNotSchedule, maxSkew 1) unless the platform sets scheduling.globalAppTopologySpreadConstraints. The anti-affinity stays soft because a required rule forbids two members on one node even when every node already has one, so an etcd with more members than schedulable nodes could not start all of them. Set a requiredDuringSchedulingIgnoredDuringExecution rule here only on clusters with at least as many schedulable nodes as members. With node-local storage a recreated member Pod can only return to its own node, so a required rule keeps it Pending while that node holds another member, and the default spread does the same once some other node holds none. The operator copies affinity into each EtcdMember when it creates the member and builds every Pod of that member from that copy, so a change does not reschedule running members and takes effect only as members are replaced by new ones.","type":"object","default":{},"x-kubernetes-preserve-unknown-fields":true}}} release: prefix: "" labels: