API
BestBefore is cluster scoped, in API group bestbefore.k-belt.io/v1alpha1.
apiVersion: bestbefore.k-belt.io/v1alpha1
kind: BestBefore
metadata:
name: default-pool
spec:
nodeClaimSelector:
matchLabels:
karpenter.sh/nodepool: default
maxAge: 504h
maxConcurrent: "10%"
taintDriftedNodes: false
spec
nodeClaimSelector
Required. A standard Kubernetes label selector, matched against NodeClaim labels. Karpenter labels
every NodeClaim with karpenter.sh/nodepool, karpenter.sh/capacity-type,
node.kubernetes.io/instance-type and the well-known topology labels, so those are usually what you
select on.
An empty selector ({}) matches every NodeClaim in the cluster.
maxAge
Required. How long a NodeClaim may live, measured from its creationTimestamp. A Go duration
string, so 504h rather than 21d, and it must be positive. The API server rejects 0s, negative
values, and anything it cannot parse.
Reaching maxAge marks the NodeClaim as drifted. Karpenter decides when it is actually replaced,
which is why maxAge is not a guaranteed maximum node lifetime. Keep expireAfter for that, and
see Sizing maxAge.
maxConcurrent
Optional. Limits the number of NodeClaims this policy can have drifted at one time, either as a
count ("3") or as a percentage of the NodeClaims its selector matches ("10%"). If undefined,
k-belt marks every stale NodeClaim at once, and the NodePool’s disruption budgets alone pace the
rollout.
If the value is a percentage, k-belt calculates the number it may mark as
allowed = roundup(matched * percentage) - already_drifted. If the value is a count, k-belt uses
that number as a static ceiling, max_concurrent - already_drifted. Percentages are calculated
from every NodeClaim the selector matches, not only the stale ones. For instance, a policy matching
6 NodeClaims with maxConcurrent: "10%" may have 1 NodeClaim drifted at a time, rounding up from
6 * .1 = 0.6.
A NodeClaim counts against the limit from the time k-belt marks it until the NodeClaim is removed, including while Karpenter drains it. Once the limit is reached, k-belt re-checks the remaining stale NodeClaims every 30 seconds and marks the oldest of them as slots free up. Karpenter otherwise replaces drifted nodes in the order it observed the drift, so a limit is also how you get the oldest nodes replaced first.
spec:
maxAge: 504h
maxConcurrent: "10%"
Use maxConcurrent with taintDriftedNodes to limit how many nodes carry
the taint at one time.
If a policy takes over a NodeClaim another policy had already marked, because that policy no longer applies, the NodeClaim counts against the new policy’s limit. k-belt does not undo the drift to fit the limit, since Karpenter is already replacing that node.
k-belt counts drifted NodeClaims from its cache, so if two NodeClaims are marked at nearly the same time a policy can briefly exceed its limit. Karpenter’s disruption budgets still apply.
taintDriftedNodes
Optional, defaults to false. Adds bestbefore.k-belt.io/drifted:PreferNoSchedule to the Node of
every NodeClaim this policy marks, so the scheduler prefers nodes that aren’t about to be replaced.
The taint is removed if the drift is undone.
Karpenter’s scheduling simulation treats a
PreferNoScheduletaint on an existing node as a hard constraint, unless one of your NodePool templates also carries aPreferNoScheduletaint. With this on, Karpenter may launch capacity for pods that would have fitted on tainted nodes, and consolidate less while a rotation runs. Adding anyPreferNoScheduletaint to the NodePool template makes Karpenter relax these back into preferences.
Use this with maxConcurrent, which limits how many NodeClaims the policy marks
at one time and therefore how many nodes are tainted. With maxConcurrent: "3", at most three
nodes are tainted, however many are stale.
status
| Field | Meaning |
|---|---|
matchedNodeClaims |
NodeClaims the selector matches |
staleNodeClaims |
Matched NodeClaims older than maxAge |
driftedNodeClaims |
NodeClaims this policy has marked that Karpenter hasn’t replaced yet |
observedGeneration |
Spec generation the status was computed from |
conditions |
Ready is False with reason InvalidSelector when the selector cannot be parsed |
A policy with an unparseable selector matches nothing, so any nodes it had marked are restored until you fix it.