mirror of
https://github.com/cozystack/cozystack.git
synced 2026-11-22 12:43:32 +09:00
The scope of the default LimitRange is every namespace a Package targets whose name does not start with `tenant-`, and cozystack-scheduler installs into kube-system, so a namespace owned by the underlying Kubernetes distribution is defaulted like any other. Neither the flag help nor the installer value said so, and both are what somebody setting the knob reads. docs/operations/system-memory-limits.md already covers this, including the part that surprises: containers the kubelet builds from an on-disk manifest never pass through admission, so a Talos control-plane static pod gets no memory.max from this feature in the first place and the mirror pod is the only thing the LimitRange reaches. Both notes now carry the same statement, so neither reading of the knob leaves it out. Assisted-By: LLM Signed-off-by: Myasnikov Daniil <myasnikovdaniil2001@gmail.com>
111 lines
7.1 KiB
YAML
111 lines
7.1 KiB
YAML
# Render a bare `Namespace cozy-system` resource (with PSA + identity labels).
|
|
# Off by default: helm install/upgrade callers create the namespace via
|
|
# `--create-namespace` and let the pre-install labeler hook stamp the labels.
|
|
# The top-level Makefile turns this on for the kubectl-apply release artifacts
|
|
# (_out/assets/cozystack-operator-*.yaml).
|
|
bareNamespace: false
|
|
cozystackOperator:
|
|
# Deployment variant: talos, generic, hosted
|
|
variant: talos
|
|
# Platform version this install reports (telemetry, observability). Supplied
|
|
# as deploy-time data, not baked into the image — so a release-candidate image
|
|
# can be promoted to stable by retagging alone. Injected as the
|
|
# COZYSTACK_VERSION env var on the operator; when empty the operator falls back
|
|
# to the value baked at build time. The Makefile sets this for the
|
|
# kubectl-apply release artifacts (_out/assets/cozystack-operator-*.yaml).
|
|
platformVersion: ""
|
|
# Stop the operator sending cluster telemetry (nodes, storage, packages) to
|
|
# https://telemetry.cozystack.io. This covers the OPERATOR only. The second
|
|
# collector, cozystack-controller, reports application counts and is
|
|
# configured through the platform Package — set telemetry.disabled there as
|
|
# well, or the opt-out is partial. Declared here rather than left implicit
|
|
# because the documented opt-out is a `--set` against this key.
|
|
disableTelemetry: false
|
|
image: ghcr.io/cozystack/cozystack/cozystack-operator:v1.6.0@sha256:93b305f78823b070c59ca631b213e39bd9d8a256d5c520f156a03f5f72f57a5d
|
|
platformSourceUrl: 'oci://ghcr.io/cozystack/cozystack/cozystack-packages'
|
|
platformSourceRef: 'digest=sha256:bf68208730860fa8e47f378a0260e79a0262e0783fa63ba4f7f97f57a48b2d23'
|
|
# When non-empty, overrides the operator's --helmrelease-interval flag
|
|
# (operator default: 5m). E2E sets this to 30s; production should leave empty.
|
|
helmReleaseInterval: ""
|
|
# When non-empty, overrides the operator's --helmrelease-retry-interval flag
|
|
# (operator default: 30s). Controls how fast failed install/upgrade attempts
|
|
# are retried. Decoupled from helmReleaseInterval (healthy reconcile cadence).
|
|
helmReleaseRetryInterval: ""
|
|
# When non-empty, overrides the operator's --helmrelease-install-timeout flag
|
|
# (operator default: 10m). Bounds individual k8s operations (job/hook/wait)
|
|
# during Helm install. Charts with long-running install hooks may need longer.
|
|
helmReleaseInstallTimeout: ""
|
|
# When non-empty, overrides the operator's --helmrelease-upgrade-timeout flag
|
|
# (operator default: 10m). Same semantics as helmReleaseInstallTimeout, for
|
|
# the Helm upgrade action.
|
|
helmReleaseUpgradeTimeout: ""
|
|
# When non-empty, overrides the operator's --helmrelease-max-history flag
|
|
# (operator default: 5). Number of release revisions Helm keeps per HR.
|
|
# Use 0 for unlimited; lower values reduce Secret accumulation in test clusters.
|
|
helmReleaseMaxHistory: ""
|
|
# When non-empty, overrides the operator's --system-namespace-memory-limit flag
|
|
# (operator default: 4Gi). The operator maintains a LimitRange in every system
|
|
# namespace defaulting container memory for anything that declares none.
|
|
#
|
|
# This exists for the Talos userspace OOM handler (v1.12+), which ranks cgroups
|
|
# with `memory_max.hasValue() ? 0.0 : {Besteffort: 1.0, ...}[class] *
|
|
# memory_current` and discards everything scoring zero. Only a memory *limit*
|
|
# sets memory.max and buys immunity — a request merely demotes BestEffort to
|
|
# Burstable, which under strictCgroupClassOrdering still means "killed second".
|
|
# Without this, system components are the victims whenever a tenant workload
|
|
# drives node memory PSI, because the tenant workload has limits and they do not:
|
|
# packages/apps/tenant ships tenant-range-limits, which defaults container memory
|
|
# in tenant namespaces that declare resourceQuotas. Tenant namespaces are left to
|
|
# that chart; this closes the same gap for system namespaces.
|
|
#
|
|
# Scope is every namespace a Package targets whose name does not start with `tenant-`.
|
|
# That includes kube-system, because cozystack-scheduler is installed there, so a namespace
|
|
# owned by the underlying Kubernetes distribution is defaulted like any other. Containers
|
|
# the kubelet builds from an on-disk manifest rather than from the API are unaffected in
|
|
# both directions: a Talos control-plane static pod gets no memory.max from this, and the
|
|
# mirror pod representing it is the only thing admission sees.
|
|
#
|
|
# A ceiling, not a reservation: pick it well above real usage. It should stay above
|
|
# the largest memory request in any system namespace, because a defaulted limit
|
|
# below a container's own request is rejected at admission. Setting it too low
|
|
# costs precision rather than availability: before applying, the operator scans
|
|
# each system namespace for a container that requests more than this and declares
|
|
# no limit of its own, and where it finds one it raises that namespace's ceiling to
|
|
# the request instead, logging the namespace, workload, container and both limits.
|
|
# A raised namespace still gets memory.max on its pod cgroups, which is the whole
|
|
# point — withholding the LimitRange would leave it with none — and it drops back to
|
|
# this value on its own once the oversized request is gone. Grep the operator log
|
|
# for "raising the default container memory limit" to find them.
|
|
#
|
|
# A container that declares its own memory limit is never touched, so a component
|
|
# that has to sit above this ceiling states a limit in its own chart rather than
|
|
# relying on the knob; packages/system/linstor does exactly that for its JVM.
|
|
#
|
|
# Set to 0 to disable, which also removes the LimitRanges the operator previously
|
|
# created.
|
|
systemNamespaceMemoryLimit: ""
|
|
# When non-empty, overrides the operator's --system-namespace-memory-request flag
|
|
# (operator default: 32Mi). Deliberately small and always set: Kubernetes defaults
|
|
# an unset request to the limit, which would have every system container reserve
|
|
# the full limit at schedule time.
|
|
systemNamespaceMemoryRequest: ""
|
|
# Name of the Secret in cozy-system namespace containing credentials for the
|
|
# platform source. When non-empty, sets spec.secretRef on the generated
|
|
# OCIRepository or GitRepository resource. Secret type depends on source kind:
|
|
# OCI (oci://...): kubernetes.io/dockerconfigjson with registry credentials.
|
|
# Git over HTTPS: Opaque secret with `username`+`password`, or `bearerToken`.
|
|
# Git over SSH: Opaque secret with `identity` (PEM private key) and `known_hosts`.
|
|
# The Secret must already exist in cozy-system before the operator reconciles;
|
|
# otherwise source-controller will report a failed condition until it is created.
|
|
# See FluxCD source-controller docs for the exact key layout per scheme.
|
|
platformSourceSecret: ''
|
|
# Generic variant configuration (only used when cozystackOperator.variant=generic)
|
|
cozystack:
|
|
# Kubernetes API server host (IP only, no protocol/port)
|
|
# Must be the INTERNAL IP of the control-plane node
|
|
# (the IP visible on the node's network interface, not a public/NAT IP)
|
|
# Used by the operator and networking components (cilium, kube-ovn)
|
|
apiServerHost: ""
|
|
# Kubernetes API server port
|
|
apiServerPort: "6443"
|