mirror of
https://github.com/cozystack/cozystack.git
synced 2026-11-02 18:16:53 +09:00
Myasnikov Daniil
c9ba7a3cca
The Talos userspace OOM handler (v1.12+) ranks cgroups with
memory_max.hasValue() ? 0.0 :
{Besteffort: 1.0, Burstable: 0.5, ...}[class] * memory_current
and discards every cgroup scoring zero, so only a memory limit takes a
pod out of the victim set. Selection is decoupled from the trigger: a
tenant workload thrashing against its own limit drives node memory PSI,
and a limit-free system DaemonSet is killed for it. metallb-speaker and
linstor-satellite were the observed victims.
Set requests and limits on the node DaemonSets that had none. Requests
come from 14d peak working set measured on a 3-node dev cluster; limits
are deliberately loose, sized as a runaway backstop rather than a fitted
ceiling, because several of these scale with cluster size that a dev
cluster understates and because undersizing converts a rare
pressure-driven kill into a deterministic one. A limit costs nothing
until it is hit -- only the request reserves.
memory.max lands on the POD cgroup only when EVERY container in the pod
carries a limit, so sidecars and init containers are covered too; one
limit-free container returns the whole pod to the victim set.
Two DaemonSets are not chart-rendered and needed their own seam:
linstor-satellite via the LinstorSatelliteConfiguration podTemplate, and
virt-handler via KubeVirt customizeComponents.patches, since virt-operator
owns that DaemonSet. Note linstor-controller stays deliberately limit-free
-- it is a JVM that must not be OOM-killed mid-operation, and cluster_test
pins that.
Two things the multus manifest needed beyond the values change.
customize-deployment.patch is the source for that template, so the
limits are carried there rather than only in the generated file, which
`make update` would discard. And the reason kube-multus had no memory
limit was that this patch DELETED upstream's 50Mi rather than raising
it -- 50Mi being well under the 230Mi measured peak -- which is what
removed memory.max from the pod cgroup in the first place.
Not covered, because upstream hardcodes the containers with no resources
knob: the four frr-k8s cp-* init containers, the kube-ovn hostpath-init
and install-cni init containers, and cozy-proxy, whose vendored chart
exposes no resources value at all. hami and kilo are left alone as
optional components with no usage measurements behind them. All of these
still get a limit from the default LimitRange the operator maintains in
every system namespace.
Every limit is pinned by a helm-unittest assertion, and every assertion
was mutation-tested: changing the rendered value turns its suite red.
Assisted-By: Claude <noreply@anthropic.com>
Signed-off-by: Myasnikov Daniil <myasnikovdaniil2001@gmail.com>