Chaos Faults for Kubernetes
Introduction
Kubernetes faults disrupt the resources running on a Kubernetes cluster. They can be categorized into pod-level faults and node-level faults.
Kubelet service kill
Kubelet service kill makes the application unreachable on the account of the node turning unschedulable (NotReady).
Node CPU hog
Node CPU hog exhausts the CPU resources on a Kubernetes node.
Node drain
Node drain drains the node of all its resources running on it.
Node IO stress
Node IO stress causes I/O stress on the Kubernetes node.
Node memory hog
Node memory hog causes memory resource exhaustion on the Kubernetes node.
Node network latency
Node network latency introduces network latency to the Kubernetes node.
Kubelet service kill
Kubelet service kill makes the application unreachable on the account of the node turning unschedulable (NotReady).
- Kubelet service is stopped (or killed) on a node to make it unschedulable for a specific duration defined by the
TOTAL_CHAOS_DURATION
environment variable. - The application node goes back to normal state and services are resumed after the chaos duration.
Details
Use cases
This fault determines the resilience of an application when a node becomes unschedulable, i.e. NotReady state.Node CPU hog
Node CPU hog exhausts the CPU resources on a Kubernetes node. The CPU chaos is injected using a helper pod running the Linux stress tool (a workload generator). The chaos affects the application for a period defined by the TOTAL_CHAOS_DURATION
environment variable.
Details
Use cases
The fault aims to verify the resiliency of applications whose replicas may be evicted on account of nodes turning unschedulable (Not Ready) or new replicas not being able to schedule due to a lack of CPU resources. The fault causes CPU stress on the target node(s). It simulates the situation of lack of CPU for processes running on the application, which degrades their performance. It also helps verify metrics-based horizontal pod autoscaling as well as vertical autoscale, i.e. demand based CPU addition. It helps scalability of nodes based on growth beyond budgeted pods. It verifies the autopilot functionality of (cloud) managed clusters. It benefits include verifying multi-tenant load issues (when the load increases on one container, it does not cause downtime in other containers).Node drain
Node drain drains the node of all its resources running on it. Due to this, services running on the target node should be rescheduled to run on other nodes.
Details
Use cases
Node drain fault drains all the resources running on a node. This fault determines the resilience of the application when the application replicas scheduled on a node are removed. It validates the application failover capabilities when a node suddenly becomes unavailable. It simulates node maintenance activity (hardware refresh, OS patching, Kubernetes upgrade). It verifies resource budgeting on cluster nodes (whether request (or limit) settings honored on available nodes), and whether topology constraints are adhered to (node selectors, tolerations, zone distribution, affinity(or anti-affinity) policies) or not.Node IO stress
Node IO stress causes I/O stress on the Kubernetes node. The amount of I/O stress is specifed as the size in percentage of the total free space available on the file system using FILESYSTEM_UTILIZATION_PERCENTAGE
environment variable or in gigabytes(GB) using FILESYSTEM_UTILIZATION_BYTES
environment variable. When both the values are provided, FILESYSTEM_UTILIZATION_PERCENTAGE
takes precendence. It tests application resiliency on replica evictions that occur due I/O stress on the available disk space.
Details
Use cases
The fault aims to verify the resilience of applications that share the disk resource for ephemeral or persistent storage purposes during high disk I/O usage. It simulates slower disk operations by the application and nosiy neighbour problems by hogging the disk bandwidth. It also verifies the disk performance on increasing I/O threads and varying I/O block sizes. It checks if the application functions under high disk latency conditions, when I/O traffic is very high and includes large I/O blocks, and when other services monopolize the I/O disks.Node memory hog
Node memory hog causes memory resource exhaustion on the Kubernetes node. It is injected using a helper pod running the Linux stress-ng tool (a workload generator). The chaos affects the application foe a duration specified by the TOTAL_CHAOS_DURATION
environment variable.
Details
Use cases
Node memory hog causes memory resource exhaustion on the Kubernetes node. The fault aims to verify resilience of applications whose replicas may be evicted on account on nodes becoming unschedulable (Not Ready) due to lack of memory resources. It simulates the situation of memory leaks in the deployment of microservices, application slowness due to memory starvation, and noisy neighbour problems due to hogging. It verifies pod priority and QoS setting for eviction purposes. It also verifies application restarts on OOM kills.Node network latency
Node network latency causes network latency on the Kubernetes node. It is injected using a helper pod running in the target node. The chaos affects the application running on the target node for a duration specified by the TOTAL_CHAOS_DURATION
environment variable.
Details
Use cases
Node network latency introduces a delay in the network communication of a Kubernetes node. The fault aims to verify the resilience of applications when faced with increased network response times. It is designed to test the behavior of applications under delayed network conditions, especially in systems where timely data transfer and communication are crucial.It simulates the scenarios of high-latency network connections, such as cross-continental data transfers, or situations where a service is communicating with a slow or overburdened external data source. The fault tests the application's ability to maintain service quality and responsiveness in sub-optimal network conditions. It verifies how well the application handles increased response times, timeouts, and the potential for increased queue lengths or backlogs due to network delays. It can also be used to confirm the correct functioning of timeout settings and retry mechanisms in applications.
Node network loss
Node network loss causes network loss on the Kubernetes node. It is injected using a helper pod running the target node. The chaos affects the application running on the target node for a duration specified by the TOTAL_CHAOS_DURATION
environment variable.
Details
Use cases
Node network loss simulates packet loss in the network communication of a Kubernetes node. The fault aims to verify the resilience of applications when faced with disrupted network communication, reflecting real-world scenarios such as unstable connections, network partitions, or infrastructure outages.It mimics situations where the network becomes unreliable, leading to potential data transmission failures, retries, and extended communication delays. The fault challenges applications by hindering their ability to communicate with other services, data stores, or external APIs effectively. It verifies the robustness of applications in handling network interruptions, ensuring data integrity in the face of packet loss, and the effectiveness of error-handling mechanisms under network failures. Additionally, it can be used to test failover strategies, data synchronization policies, and the efficiency of retry logic in applications.
Node restart
Node restart disrupts the state of the node by restarting it. It tests deployment sanity (replica availability and uninterrupted service) and recovery workflows of the application pod.
Details
Use cases
This fault determines the deployment sanity (replica availability and uninterrupted service) and recovery workflows of the application pod in the event of an unexpected node restart. It simulates loss of critical services (or node-crash). It verifies resource budgeting on cluster nodes (whether request(or limit) settings honored on available nodes), and whether topology constraints are adhered to (node selectors, tolerations, zone distribution, affinity(or anti-affinity) policies) or not.Node taint
Node taint taints (contaminates) the node by applying the desired effect. The resources that contain the corresponding tolerations only can bypass the taints.
Details
Use cases
The fault aims to verify the resiliency of applications when a certain taint is added to a node. It simulates loss of critical services (or node-crash). It verifies resource budgeting on cluster nodes (whether request(or limit) settings honored on available nodes), and whether topology constraints are adhered to (node selectors, tolerations, zone distribution, affinity(or anti-affinity) policies) or not.Container kill
Container kill is a Kubernetes pod-level chaos fault that causes container failure on specific (or random) replicas of an application resource.
- It tests an application's deployment sanity (replica availability and uninterrupted service) and recovery workflow.
- It tests the recovery of pods that possess sidecar containers.
Details
Use cases
It tests an application's deployment sanity (replica availability and uninterrupted service) and recovery workflow when certain replicas are not available.Disk fill
Disk fill is a Kubernetes pod-level chaos fault that applies disk stress by filling the pod's ephemeral storage on a node.
- It evicts the application pod if its capacity exceeds the pod's ephemeral storage limit.
- It tests the ephemeral storage limits and ensures that the parameters are sufficient.
- It evaluates the application's resilience to disk stress (or replica) evictions.
Details
Use cases
This fault tests the ephemeral storage limits and determines the resilience of the application to unexpected storage exhaustions.FS fill
FS fill is a Kubernetes pod-level chaos fault that applies fs stress by filling the pod's ephemeral storage of the pod.
Use cases
- Tests the ephemeral storage limits and ensures that the parameters are sufficient.
- Determines the resilience of the application to unexpected storage exhaustion.
- Evaluates the application's resilience to FS stress or replica evictions.
- Verifies file system performance, and thin-provisioning support.
- Verifies space reclamation (UNMAP) capabilities on storage.
Pod autoscaler
Pod autoscaler is a Kubernetes pod-level chaos fault that determines whether nodes can accomodate multiple replicas of a given application pod.
- It examines the node auto-scaling feature by determining whether the pods were successfully rescheduled within a specified time frame if the existing nodes are running at the specified limits.
Details
Use cases
This fault determines how an application accomodates multiple replicas of a given application pod at unexpected point in time.Pod CPU hog exec
Pod CPU hog exec is a Kubernetes pod-level chaos fault that consumes excess CPU resources of the application container.
- It simulates conditions where the application pods experience CPU spikes due to expected (or undesired) processes thereby testing the behaviour of application stack.
Details
Use cases
Disk pressure or CPU hog affects Kubernetes applications which result in the eviction of the application replica and impacts its delivery. These issues are referred to as "noisy neighbour" problems. The fault causes CPU stress on the target pod(s). It simulates the situation of lack of CPU for processes running on the application, which degrades their performance. It also helps verify metrics-based horizontal pod autoscaling as well as vertical autoscale, i.e. demand based CPU addition. It helps scalability of nodes based on growth beyond budgeted pods. It verifies the autopilot functionality of (cloud) managed clusters. Injecting a rogue process into a target container starves the main microservice (typically pid 1) of the resources allocated to it (where limits are defined). This slows down the application traffic or exhausts the resources leading to eviction of all pods. These faults helps build immunity to such stress cases. It benefits include verifying multi-tenant load issues (when the load increases on one container, it does not cause downtime in other containers).Pod CPU hog
Pod CPU hog is a Kubernetes pod-level chaos fault that excessively consumes CPU resources, resulting in a significant increase in the CPU resource usage of a pod.
- Simulates a situation where an application's CPU resource usage unexpectedly spikes.
Details
Use cases
Disk pressure or CPU hog affects Kubernetes applications which result in the eviction of the application replica and impacts its delivery. These issues are referred to as "noisy neighbour" problems. The fault causes CPU stress on the target pod(s). It simulates the situation of lack of CPU for processes running on the application, which degrades their performance. It also helps verify metrics-based horizontal pod autoscaling as well as vertical autoscale, i.e. demand based CPU addition. It helps scalability of nodes based on growth beyond budgeted pods. It verifies the autopilot functionality of (cloud) managed clusters. Injecting a rogue process into a target container starves the main microservice (typically pid 1) of the resources allocated to it (where limits are defined). This slows down the application traffic or exhausts the resources leading to eviction of all pods. These faults helps build immunity to such stress cases. It benefits include verifying multi-tenant load issues (when the load increases on one container, it does not cause downtime in other containers).Pod delete
Pod delete is a Kubernetes pod-level chaos fault that causes specific (or random) replicas of an application resource to fail forcibly (or gracefully).
- It tests an application's deployment sanity (replica availability and uninterrupted service) and recovery workflow.
Details
Use cases
In distributed systems like Kubernetes, your application replicas may not be sufficient to manage the traffic (indicated by SLIs) when some of the replicas are unavailable due to failures. It is important to ensure that the applications have minimum number of available replicas. One of the common application failures is when the pressure on other replicas increases, and how the horizontal pod autoscaler scales based on the observed resource utilization. It is also important to understand how much time it takes for persistent volume to after rescheduling. This fault helps reproduce such a situation with forced (or graceful) pod failure on specific (or random) replicas of an application resource. It checks the deployment sanity (replica availability and uninterrupted service) and recovery workflow of the application.Pod DNS error
Pod DNS error is a Kubernetes pod-level chaos fault that injects chaos to disrupt DNS resolution in pods.
- It removes access to services by blocking the DNS resolution of host names (or domains).
Details
Use cases
This fault determines the resilience of an application to DNS errors. It determines how quickly an application can resolve the host names and recover from the failure.Pod DNS spoof
Pod DNS spoof is a Kubernetes pod-level chaos fault that injects chaos into pods to mimic DNS resolution.
- It resolves DNS target host names (or domains) to other IPs as specified in the
SPOOF_MAP
environment variable in the chaosengine configuration.
Details
Use cases
This fault determines the resilience of an application when host names are resolved incorrectly. It determines how quickly an application can resolve the host names and recover from the failure. It simulates custom responses from a spoofed upstream service.Pod HTTP latency
Pod HTTP latency is a Kubernetes pod-level chaos fault that injects HTTP response latency by starting proxy server and redirecting the traffic through it.
- It injects the latency into the service whose port is specified using the
TARGET_SERVICE_PORT
environment variable. - It evaluates the application's resilience to lossy (or flaky) HTTP responses.
Details
Use cases
This fault evaluates the application's resilience to lossy (or flaky) HTTP responses.Pod HTTP modify body
Pod HTTP modify body is a Kubernetes pod-level chaos fault that injects chaos on the service whose port is provided using the TARGET_SERVICE_PORT
environment variable.
- This is done by starting the proxy server and redirecting the traffic through the proxy server.
- Can be used to overwrite the HTTP response body by providing the new body value as
RESPONSE_BODY
. - It can test the application's resilience to error or incorrect HTTP response body.
Details
Use cases
It can test the application's resilience to error or incorrect HTTP response body.Pod HTTP modify header
Pod HTTP modify header is a Kubernetes pod-level chaos fault that injects chaos on the service whose port is provided using the TARGET_SERVICE_PORT
environment variable.
- This is done by starting the proxy server and redirecting the traffic through the proxy server.
- It can cause modification of headers of requests and responses of the service. This can be used to test service resilience towards incorrect or incomplete headers.
Details
Use cases
This can be used to test service resilience towards incorrect or incomplete headers.Pod HTTP reset peer
Pod HTTP reset peer is a Kubernetes pod-level chaos fault that injects chaos on the service whose port is specified using the TARGET_SERVICE_PORT
environment variable.
- This stops the outgoing HTTP requests by resetting the TCP connection by starting the proxy server and redirecting the traffic through the proxy server.
- It can test the application's resilience to lossy/flaky HTTP connection.
Details
Use cases
It can test the application's resilience to lossy/flaky HTTP connection.Pod HTTP status code
Pod HTTP status code is a Kubernetes pod-level fault injects chaos inside the pod by modifying the status code of the response from the application server to the desired status code provided by the user.
- The port for the service is specified using the
TARGET_SERVICE_PORT
environment variable by starting the proxy server and redirecting the traffic through the proxy server. - It tests the application's resilience to error code HTTP responses from the provided application server.
Details
Use cases
It tests the application's resilience to error code HTTP responses from the provided application server.Pod IO stress
Pod I/O stress is a Kubernetes pod-level chaos fault that causes IO stress on the application pod by spiking the number of input and output requests.
- Aims to verify the resiliency of applications that share this disk resource for ephemeral (or persistent) storage.
Details
Use cases
Disk pressure or CPU hog affects Kubernetes applications that results in the eviction of the application replica and impacts its delivery. These issues are referred to as "noisy neighbour" problems. Stressing the disk with continuous and heavy I/O can degrade the reads and writes with respect to the microservices. Scratch space consumed on a node may lead to lack of memory for new containers to be scheduled. These faults helps build immunity to such stress cases.Pod memory hog exec
Pod memory hog exec is a Kubernetes pod-level chaos fault that consumes memory resources on the application container in megabytes.
- It simulates conditions where app pods experience Memory spikes either due to expected/undesired processes thereby testing how the overall application stack behaves when this occurs.
Details
Use cases
Memory usage within containers is subject to various constraints in Kubernetes. If the limits are specified in their spec, exceeding them results in termination of the container (due to OOMKill of the primary process, often pid 1). This restarts container dependng on policy specified. For containers with no limits on memory, node can be killed based on their oom_score. This results in a bigger blast radius.This fault causes stress within the target container, which may result in the primary process in the container to be constrained or eat up the available system memory on the node.
Pod memory hog
Pod memory hog is a Kubernetes pod-level chaos fault that consumes memory resources in excess, resulting in a significant spike in the memory usage of a pod.
- Simulates a condition where the memory usage of an application spikes up unexpectedly.
Details
Use cases
Memory usage within containers is subject to various constraints in Kubernetes. If the limits are specified in their spec, exceeding them results in termination of the container (due to OOMKill of the primary process, often pid 1). This restarts container dependng on policy specified. For containers with no limits on memory, node can be killed based on their oom_score. This results in a bigger blast radius.This fault causes stress within the target container, which may result in the primary process in the container to be constrained or eat up the available system memory on the node.
Pod network corruption
Pod network corruption is a Kubernetes pod-level chaos fault that injects corrupted packets of data into the specified container by starting a traffic control (tc) process with netem rules to add egress packet corruption.
- Tests the application's resilience to lossy (or flaky) network.