This page explains how to configure out of resource handling with
kubelet needs to preserve node stability when available compute resources
are low. This is especially important when dealing with incompressible
compute resources, such as memory or disk space. If such resources are exhausted,
nodes become unstable.
kubelet can proactively monitor for and prevent total starvation of a
compute resource. In those cases, the
kubelet can reclaim the starved
resource by proactively failing one or more Pods. When the
a Pod, it terminates all of its containers and transitions its
kubelet supports eviction decisions based on the signals described in the following
table. The value of each signal is described in the Description column, which is based on
kubelet summary API.
Each of the above signals supports either a literal or percentage based value. The percentage based value is calculated relative to the total capacity associated with each signal.
The value for
memory.available is derived from the cgroupfs instead of tools
free -m. This is important because
free -m does not work in a
container, and if users use the node
allocatable feature, out of resource decisions
are made local to the end user Pod part of the cgroup hierarchy as well as the
root node. This
reproduces the same set of steps that the
kubelet performs to calculate
kubelet excludes inactive_file (i.e. # of bytes of
file-backed memory on inactive LRU list) from its calculation as it assumes that
memory is reclaimable under pressure.
kubelet supports only two filesystem partitions.
nodefsfilesystem that kubelet uses for volumes, daemon logs, etc.
imagefsfilesystem that container runtimes uses for storing images and container writable layers.
imagefs is optional.
kubelet auto-discovers these filesystems using
kubelet does not care about any other filesystems. Any other types
of configurations are not currently supported by the kubelet. For example, it is
not OK to store volumes and logs in a dedicated
In future releases, the
kubelet will deprecate the existing garbage
support in favor of eviction in response to disk pressure.
kubelet supports the ability to specify eviction thresholds that trigger the
kubelet to reclaim resources.
Each threshold has the following form:
eviction-signalis an eviction signal token as defined in the previous table.
operatoris the desired relational operator, such as
quantityis the eviction threshhold quantity, such as
1Gi. These tokens must match the quantity representation used by Kubernetes. An eviction threshold can also be expressed as a percentage using the
For example, if a node has
10Gi of total memory and you want trigger eviction if
the available memory falls below
1Gi, you can define the eviction threshold as
memory.available<1Gi. You cannot use both.
A soft eviction threshold pairs an eviction threshold with a required
administrator-specified grace period. No action is taken by the
to reclaim resources associated with the eviction signal until that grace
period has been exceeded. If no grace period is provided, the
returns an error on startup.
In addition, if a soft eviction threshold has been met, an operator can
specify a maximum allowed Pod termination grace period to use when evicting
pods from the node. If specified, the
kubelet uses the lesser value among
pod.Spec.TerminationGracePeriodSeconds and the max allowed grace period.
If not specified, the
kubelet kills Pods immediately with no graceful
To configure soft eviction thresholds, the following flags are supported:
eviction-softdescribes a set of eviction thresholds (e.g.
memory.available<1.5Gi) that if met over a corresponding grace period would trigger a Pod eviction.
eviction-soft-grace-perioddescribes a set of eviction grace periods (e.g.
memory.available=1m30s) that correspond to how long a soft eviction threshold must hold before triggering a Pod eviction.
eviction-max-pod-grace-perioddescribes the maximum allowed grace period (in seconds) to use when terminating pods in response to a soft eviction threshold being met.
A hard eviction threshold has no grace period, and if observed, the
will take immediate action to reclaim the associated starved resource. If a
hard eviction threshold is met, the
kubelet kills the Pod immediately
with no graceful termination.
To configure hard eviction thresholds, the following flag is supported:
eviction-harddescribes a set of eviction thresholds (e.g.
memory.available<1Gi) that if met would trigger a Pod eviction.
kubelet has the following default hard eviction threshold:
kubelet evaluates eviction thresholds per its configured housekeeping interval.
housekeeping-intervalis the interval between container housekeepings.
kubelet maps one or more eviction signals to a corresponding node condition.
If a hard eviction threshold has been met, or a soft eviction threshold has been met
independent of its associated grace period, the
kubelet reports a condition that
reflects the node is under pressure.
The following node conditions are defined that correspond to the specified eviction signal.
|Node Condition||Eviction Signal||Description|
||Available memory on the node has satisfied an eviction threshold|
||Available disk space and inodes on either the node’s root filesytem or image filesystem has satisfied an eviction threshold|
kubelet continues to report node status updates at the frequency specified by
--node-status-update-frequency which defaults to
If a node is oscillating above and below a soft eviction threshold, but not exceeding its associated grace period, it would cause the corresponding node condition to constantly oscillate between true and false, and could cause poor scheduling decisions as a consequence.
To protect against this oscillation, the following flag is defined to control how
kubelet must wait before transitioning out of a pressure condition.
eviction-pressure-transition-periodis the duration for which the
kubelethas to wait before transitioning out of an eviction pressure condition.
kubelet would ensure that it has not observed an eviction threshold being met
for the specified pressure condition for the period specified before toggling the
condition back to
If an eviction threshold has been met and the grace period has passed,
kubelet initiates the process of reclaiming the pressured resource
until it has observed the signal has gone below its defined threshold.
kubelet attempts to reclaim node level resources prior to evicting end-user Pods. If
disk pressure is observed, the
kubelet reclaims node level resources differently if the
machine has a dedicated
imagefs configured for the container runtime.
nodefs filesystem has met eviction thresholds,
kubelet frees up disk space by deleting the dead Pods and their containers.
imagefs filesystem has met eviction thresholds,
kubelet frees up disk space by deleting all unused images.
nodefs filesystem has met eviction thresholds,
kubelet frees up disk space in the following order:
kubelet is unable to reclaim sufficient resource on the node,
kubelet begins evicting Pods.
kubelet ranks Pods for eviction first by their quality of service, and then by the consumption
of the starved compute resource relative to the Pods’ scheduling requests.
As a result,
kubectl ranks and evicts Pods in the following order:
BestEffortPods consume the most of the starved resource are failed first. Local disk is a
BurstablePods consume the greatest amount of the starved resource relative to their request for that resource are killed first. If no Pod has exceeded its request, the strategy targets the largest consumer of the starved resource.
GuaranteedPods are guaranteed only when requests and limits are specified for all the containers and they are equal. A
GuaranteedPod is guaranteed to never be evicted because of another Pod’s resource consumption. If a system daemon (such as
journald) is consuming more resources than were reserved via
kube-reservedallocations, and the node only has
GuaranteedPods remaining, then the node must choose to evict a
GuaranteedPod in order to preserve node stability and to limit the impact of the unexpected consumption to other
kubelet evicts Pods one at a time to reclaim disk when
is encountered. If the
kubelet is responding to
inode starvation, it reclaims
inodes by evicting Pods with the lowest quality of service first. If the
is responding to lack of available disk, it ranks Pods within a quality of service
that consumes the largest amount of disk and kill those first.
nodefs is triggering evictions,
kubelet sorts Pods based on the usage on
imagefs is triggering evictions,
kubelet sorts Pods based on the writable layer usage of all its containers.
nodefs is triggering evictions,
kubelet sorts Pods based on their total disk usage
In certain scenarios, eviction of Pods could result in reclamation of small amount of resources. This can result in
kubelet hitting eviction thresholds in repeated successions. In addition to that, eviction of resources like
is time consuming.
To mitigate these issues,
kubelet can have a per-resource
kubelet attempts to reclaim at least
minimum-reclaim amount of resource below
the configured eviction threshold.
For example, with the following configuration:
If an eviction threshold is triggered for
kubelet works to ensure
memory.available is at least
to ensure that
nodefs.available is at least
1.5Gi, and for
works to ensure that
imagefs.available is at least
102Gi before no longer reporting pressure
on their associated resources.
0 for all resources.
The node reports a condition when a compute resource is under pressure. The scheduler views that condition as a signal to dissuade placing additional pods on the node.
|Node Condition||Scheduler Behavior|
||No new Pods are scheduled to the node.|
If the node experiences a system OOM (out of memory) event prior to the
kubelet is able to reclaim memory,
the node depends on the oom_killer to respond.
kubelet sets a
oom_score_adj value for each container based on the quality of service for the Pod.
|Quality of Service||oom_score_adj|
||min(max(2, 1000 - (1000 * memoryRequestBytes) / machineMemoryCapacityBytes), 999)|
kubelet is unable to reclaim memory prior to a node experiencing system OOM, the
oom_score based on the percentage of memory it’s using on the node, and then add the
oom_score_adj to get an
oom_score for the container, and then kills the container with the highest score.
The intended behavior should be that containers with the lowest quality of service that are consuming the largest amount of memory relative to the scheduling request should be killed first in order to reclaim memory.
Unlike Pod eviction, if a Pod container is OOM killed, it may be restarted by the
kubelet based on its
The following sections describe best practices for out of resource handling.
Consider the following scenario:
To facilitate this scenario, the
kubelet would be launched as follows:
Implicit in this configuration is the understanding that “System reserved” should include the amount of memory covered by the eviction threshold.
To reach that capacity, either some Pod is using more than its request, or the system is using more than
This configuration ensures that the scheduler does not place Pods on a node that immediately induce memory pressure and trigger eviction assuming those Pods use less than their configured request.
It is never desired for
kubelet to evict a
DaemonSet Pod, since the Pod is
immediately recreated and rescheduled back to the same node.
At the moment, the
kubelet has no ability to distinguish a Pod created
DaemonSet versus any other object. If/when that information is
kubelet could pro-actively filter those Pods from the
candidate set of Pods provided to the eviction strategy.
In general, it is strongly recommended that
BestEffort Pods to avoid being identified as a candidate Pod
for eviction. Instead
DaemonSet should ideally launch
kubelet has been freeing up disk space on demand to keep the node stable.
As disk based eviction matures, the following
kubelet flags are marked for deprecation
in favor of the simpler configuration supported around eviction.
|Existing Flag||New Flag|
The following sections describe known issues related to out of resource handling.
kubelet currently polls
cAdvisor to collect memory usage stats at a regular interval. If memory usage
increases within that window rapidly, the
kubelet may not observe
MemoryPressure fast enough, and the
will still be invoked. We intend to integrate with the
memcg notification API in a future release to reduce this
latency, and instead have the kernel tell us when a threshold has been crossed immediately.
If you are not trying to achieve extreme utilization, but a sensible measure of overcommit, a viable workaround for this issue is to set eviction thresholds at approximately 75% capacity. This increases the ability of this feature to prevent system OOMs, and promote eviction of workloads so cluster state can rebalance.
The Pod eviction may evict more Pods than needed due to stats collection timing gap. This can be mitigated by adding the ability to get root container stats on an on-demand basis (https://github.com/google/cadvisor/issues/1247) in the future.
At this time, it is not possible to know how many inodes were consumed by a particular container. If the
inode exhaustion, it evicts Pods by ranking them by quality of service. The following issue has been opened in cadvisor
to track per container inode consumption (https://github.com/google/cadvisor/issues/1422) which would allow us to rank Pods
by inode consumption. For example, this would let us identify a container that created large numbers of 0 byte files, and evict that Pod over others.