chaoskube periodically kills random pods in your Kubernetes cluster.
Test how your system behaves under arbitrary pod failures.
Running it will kill a pod in any namespace every 10 minutes by default.
$ chaoskube INFO starting up dryRun=true interval=10m0s version=v0.19.0 INFO connecting to cluster master="https://kube.you.me" serverVersion=v1.10.5+coreos.0 INFO setting pod filter annotations= labels= minimumAge=0s namespaces= INFO setting quiet times daysOfYear="" timesOfDay="" weekdays="" INFO setting timezone location=UTC name=UTC offset=0 INFO terminating pod name=kube-dns-v20-6ikos namespace=kube-system INFO terminating pod name=nginx-701339712-u4fr3 namespace=chaoskube INFO terminating pod name=kube-proxy-gke-earthcoin-pool-3-5ee87f80-n72s namespace=kube-system INFO terminating pod name=nginx-701339712-bfh2y namespace=chaoskube INFO terminating pod name=heapster-v1.2.0-1107848163-bhtcw namespace=kube-system INFO terminating pod name=l7-default-backend-v1.0-o2hc9 namespace=kube-system INFO terminating pod name=heapster-v1.2.0-1107848163-jlfcd namespace=kube-system INFO terminating pod name=nginx-701339712-bfh2y namespace=chaoskube INFO terminating pod name=nginx-701339712-51nt8 namespace=chaoskube ...
chaoskube allows to filter target pods by namespaces, labels, annotations and age as well as exclude certain weekdays, times of day and days of a year from chaos.
$ helm install stable/chaoskube
Refer to chaoskube on kubeapps.com to learn how to configure it and to find other useful Helm charts.
Refer to example manifest. Be sure to give chaoskube appropriate permissions using provided ClusterRole.
chaoskube will be friendly and not kill anything. When you validated your target cluster you may disable dry-run mode by passing the flag
--no-dry-run. You can also specify a more aggressive interval and other supported flags for your deployment.
If you're running in a Kubernetes cluster and want to target the same cluster then this is all you need to do.
If you want to target a different cluster or want to run it locally specify your cluster via the
--master flag or provide a valid kubeconfig via the
--kubeconfig flag. By default, it uses your standard kubeconfig path in your home. That means, whatever is the current context in there will be targeted.
If you want to increase or decrease the amount of chaos change the interval between killings with the
--interval flag. Alternatively, you can increase the number of replicas of your
chaoskube by default kills any pod in all your namespaces, including system pods and itself.
chaoskube provides a simple HTTP endpoint that can be used to check that it is running. This can be used for Kubernetes liveness and readiness probes. By default, this listens on port 8080. To disable, pass
However, you can limit the search space of
chaoskube by providing label, annotation, and namespace selectors, pod name include/exclude patterns, as well as a minimum age setting.
$ chaoskube --labels 'app=mate,chaos,stage!=production' ... INFO setting pod filter labels="app=mate,chaos,stage!=production"
This selects all pods that have the label
app set to
mate, the label
chaos set to anything and the label
stage not set to
production or unset.
You can filter target pods by namespace selector as well.
$ chaoskube --namespaces 'default,testing,staging' ... INFO setting pod filter namespaces="default,staging,testing"
This will filter for pods in the three namespaces
Namespaces can additionally be filtered by a namespace label selector.
$ chaoskube --namespace-labels='!integration' ... INFO setting pod filter namespaceLabels="!integration"
This will exclude all pods from namespaces with the label
You can filter pods by name:
$ chaoskube --included-pod-names 'foo|bar' --excluded-pod-names 'prod' ... INFO setting pod filter excludedPodNames=prod includedPodNames="foo|bar"
This will cause only pods whose name contains 'foo' or 'bar' and does not contain 'prod' to be targeted.
You can also exclude namespaces and mix and match with the label and annotation selectors.
$ chaoskube \ --labels 'app=mate,chaos,stage!=production' \ --annotations '!scheduler.alpha.kubernetes.io/critical-pod' \ --namespaces '!kube-system,!production' ... INFO setting pod filter annotations="!scheduler.alpha.kubernetes.io/critical-pod" labels="app=mate,chaos,stage!=production" namespaces="!kube-system,!production"
This further limits the search space of the above label selector by also excluding any pods in the
production namespaces as well as ignore all pods that are marked as critical.
The annotation selector can also be used to run
chaoskube as a cluster addon and allow pods to opt-in to being terminated as you see fit. For example, you could run
chaoskube like this:
$ chaoskube --annotations 'chaos.alpha.kubernetes.io/enabled=true' --debug ... INFO setting pod filter annotations="chaos.alpha.kubernetes.io/enabled=true" DEBU found candidates count=0 DEBU no victim found
Unless you already use that annotation somewhere, this will initially ignore all of your pods (you can see the number of candidates in debug mode). You could then selectively opt-in individual deployments to chaos mode by annotating their pods with
apiVersion: apps/v1 kind: Deployment metadata: name: my-app spec: replicas: 3 template: metadata: annotations: chaos.alpha.kubernetes.io/enabled: "true" spec: ...
You can exclude pods that have recently started by using the
$ chaoskube --minimum-age 6h ... INFO setting pod filter minimumAge=6h0m0s
Limit the Chaos
You can limit the time when chaos is introduced by weekdays, time periods of a day, day of a year or all of them together.
Add a comma-separated list of abbreviated weekdays via the
--excluded-weekdays options, a comma-separated list of time periods via the
--excluded-times-of-day option and/or a comma-separated list of days of a year via the
--excluded-days-of-year option and specify a
--timezone by which to interpret them.
$ chaoskube \ --excluded-weekdays=Sat,Sun \ --excluded-times-of-day=22:00-08:00,11:00-13:00 \ --excluded-days-of-year=Apr1,Dec24 \ --timezone=Europe/Berlin ... INFO setting quiet times daysOfYear="[Apr 1 Dec24]" timesOfDay="[22:00-08:00 11:00-13:00]" weekdays="[Saturday Sunday]" INFO setting timezone location=Europe/Berlin name=CET offset=1
Local or pick a timezone name from the (IANA) tz database. If you're testing
chaoskube from your local machine then
Local makes the most sense. Once you deploy
chaoskube to your cluster you should deploy it with a specific timezone, e.g. where most of your team members are living, so that both your team and
chaoskube have a common understanding when a particular weekday begins and ends, for instance. If your team is spread across multiple time zones it's probably best to pick
UTC which is also the default. Picking the wrong timezone shifts the meaning of a particular weekday by a couple of hours between you and the server.
||interval between pod terminations||10m|
||label selector to filter pods by||(matches everything)|
||annotation selector to filter pods by||(matches everything)|
||namespace selector to filter pods by||(all namespaces)|
||label selector to filter namespaces and its pods by||(all namespaces)|
||regular expression pattern for pod names to include||(all included)|
||regular expression pattern for pod names to exclude||(none excluded)|
||weekdays when chaos is to be suspended, e.g. "Sat,Sun"||(no weekday excluded)|
||times of day when chaos is to be suspended, e.g. "22:00-08:00"||(no times of day excluded)|
||days of a year when chaos is to be suspended, e.g. "Apr1,Dec24"||(no days of year excluded)|
||timezone from tz database, e.g. "America/New_York", "UTC" or "Local"||(UTC)|
||Minimum age to filter pods by||0s (matches every pod)|
||don't kill pods, only log what would have been done||true|
||specify the format of the log messages. Options are text and json||text|
||include the calling function name and location in the log messages||false|
||The address of the slack webhook for notifications||disabled|
There are several other projects that allow you to create some chaos in your Kubernetes cluster.
- kube-monkey is a sophisticated pod-based chaos monkey for Kubernetes. Each morning it compiles a schedule of pod terminations that should happen throughout the day. It allows to specify a mean time between failures on a per-pod basis, a feature that
chaoskubelacks. It can also be made aware of groups of pods forming an application so that it can treat them specially, e.g. kill all pods of an application at once.
kube-mokeyallows filtering targets globally via configuration options as well allows pods to opt-in to chaos via annotations,it allows individual apps to opt-in in their own unique way, as an example, app-a can request to kill him each week day one pod, while app-b which more couragues can request to kill 50% of pods. It understands a similar configuration file used by Netflix's ChaosMonkey.
- PowerfulSeal is indeed a powerful tool to trouble your Kubernetes setup. Besides killing pods it can also take out your Cloud VMs or kill your Docker daemon. It has a vast number of configuration options to define what can be killed and when. It also has an interactive mode that allows you to kill pods easily.
- fabric8's chaos monkey: A chaos monkey that comes bundled as an app with fabric8's Kubernetes platform. It can be deployed via a UI and reports any actions taken as a chat message and/or desktop notification. It can be configured with an interval and a pod name pattern that possible targets must match.
- k8aos: An interactive tool that can issue a series of random pod deletions across an entire Kubernetes cluster or scoped to a namespace.
- pod-reaper kills pods based on an interval and a configurable chaos chance. It allows to specify possible target pods via a label selector and namespace. It has the ability successfully shutdown itself after a while and therefore might be suited to work well with Kubernetes Job objects. It can also be configured to kill every pod that has been running for longer than a configurable duration.
- kubernetes-pod-chaos-monkey: A very simple random pod killer using
kubectlwritten in a couple lines of bash. Given a namespace and an interval it kills a random pod in that namespace at each interval. Pretty much like
chaoskubeworked in the beginning.
This project wouldn't be where it is with the ideas and help of several awesome contributors:
- Thanks to @twildeboer and @klautcomputing who sparked the idea of limiting chaos during certain times, such as business hours or holidays as well as the first implementations of this feature in #54 and #55.
- Thanks to @klautcomputing for the first attempt to solve the missing percentage feature as well as for providing the RBAC config files.
- Thanks to @j0sh3rs for bringing the Helm chart to the latest version.
- Thanks to @klautcomputing, @grosser, @twz123, @hchenxa and @bavarianbidi for improvements to the Dockerfile and docs in #31, #40 and #58.
- Thanks to @bakins for adding the minimum age filter in #86.
- Thanks to @bakins for adding a health check and Prometheus metrics in #94 and #97.
Feel free to create issues or submit pull requests.