How to use K8S node_problem_detector? - kubernetes

Question

Question
node-problem-detector is mentioned in Monitor Node Health documentation if K8S. How do we use it if it is not in GCE? Does it feed information to Dashboard or provide API metrics?

Answer 1

"This tool aims to make various node problems visible to the upstream layers in cluster management stack. It is a daemon which runs on each node, detects node problems and reports them to apiserver."
Err Ok but... What does that actually mean? How can I tell if it went to the api server?
What does the before and after look like? Knowing that would help me understand what it's doing.
Before installing Node Problem Detector I see:
Bash# kubectl describe node ip-10-40-22-166.ec2.internal | grep -i condition -A 20 | grep Ready -B 20
Conditions:
Type Status LastHeartbeatTime LastTransitionTime Reason Message
---- ------ ----------------- ------------------ ------ -------
NetworkUnavailable False Thu, 20 Jun 2019 12:30:05 -0400 Thu, 20 Jun 2019 12:30:05 -0400 WeaveIsUp Weave pod has set this
OutOfDisk False Thu, 20 Jun 2019 18:27:39 -0400 Thu, 20 Jun 2019 12:29:44 -0400 KubeletHasSufficientDisk kubelet has sufficient disk space available
MemoryPressure False Thu, 20 Jun 2019 18:27:39 -0400 Thu, 20 Jun 2019 12:29:44 -0400 KubeletHasSufficientMemory kubelet has sufficient memory available
DiskPressure False Thu, 20 Jun 2019 18:27:39 -0400 Thu, 20 Jun 2019 12:29:44 -0400 KubeletHasNoDiskPressure kubelet has no disk pressure
PIDPressure False Thu, 20 Jun 2019 18:27:39 -0400 Thu, 20 Jun 2019 12:29:44 -0400 KubeletHasSufficientPID kubelet has sufficient PID available
Ready True Thu, 20 Jun 2019 18:27:39 -0400 Thu, 20 Jun 2019 12:30:14 -0400 KubeletReady kubelet is posting ready status
After installing Node Problem Detector I see:
Bash# helm upgrade --install npd stable/node-problem-detector -f node-problem-detector.values.yaml
Bash# kubectl rollout status daemonset npd-node-problem-detector #(wait for up)
Bash# kubectl describe node ip-10-40-22-166.ec2.internal | grep -i condition -A 20 | grep Ready -B 20
Conditions:
Type Status LastHeartbeatTime LastTransitionTime Reason Message
---- ------ ----------------- ------------------ ------ -------
DockerDaemon False Thu, 20 Jun 2019 22:06:17 -0400 Thu, 20 Jun 2019 22:04:14 -0400 DockerDaemonHealthy Docker daemon is healthy
EBSHealth False Thu, 20 Jun 2019 22:06:17 -0400 Thu, 20 Jun 2019 22:04:14 -0400 NoVolumeErrors Volumes are attaching successfully
KernelDeadlock False Thu, 20 Jun 2019 22:06:17 -0400 Thu, 20 Jun 2019 22:04:14 -0400 KernelHasNoDeadlock kernel has no deadlock
ReadonlyFilesystem False Thu, 20 Jun 2019 22:06:17 -0400 Thu, 20 Jun 2019 22:04:14 -0400 FilesystemIsNotReadOnly Filesystem is not read-only
NetworkUnavailable False Thu, 20 Jun 2019 12:30:05 -0400 Thu, 20 Jun 2019 12:30:05 -0400 WeaveIsUp Weave pod has set this
OutOfDisk False Thu, 20 Jun 2019 22:07:10 -0400 Thu, 20 Jun 2019 12:29:44 -0400 KubeletHasSufficientDisk kubelet has sufficient disk space available
MemoryPressure False Thu, 20 Jun 2019 22:07:10 -0400 Thu, 20 Jun 2019 12:29:44 -0400 KubeletHasSufficientMemory kubelet has sufficient memory available
DiskPressure False Thu, 20 Jun 2019 22:07:10 -0400 Thu, 20 Jun 2019 12:29:44 -0400 KubeletHasNoDiskPressure kubelet has no disk pressure
PIDPressure False Thu, 20 Jun 2019 22:07:10 -0400 Thu, 20 Jun 2019 12:29:44 -0400 KubeletHasSufficientPID kubelet has sufficient PID available
Ready True Thu, 20 Jun 2019 22:07:10 -0400 Thu, 20 Jun 2019 12:30:14 -0400 KubeletReady kubelet is posting ready status
Note I asked for help coming up with a way to see this for all nodes, Kenna Ofoegbu came up with this super useful and readable gem:
zsh# nodes=$(kubectl get nodes | sed '1d' | awk '{print $1}') && for node in $nodes; do; kubectl describe node | sed -n '/Conditions/,/Ready/p' ; done
Bash# (same command, gives errors)
Ok so now I know what Node Problem Detector does but... what good is adding a condition to the node, how do I use the condition to do something useful?
Question: How to use Kubernetes Node Problem Detector?
Use Case #1: Auto heal borked nodes
Step 1.) Install Node Problem Detector, so it can attach new condition metadata to nodes.
Step 2.) Leverage Planetlabs/draino to cordon and drain nodes with bad conditions.
Step 3.) Leverage https://github.com/kubernetes/autoscaler/tree/master/cluster-autoscaler to auto heal. (When the node is cordon and drained it'll be marked unscheduleable, this will trigger a new node to be provisioned, and then the bad node's resource utilization will be super low which will cause the bad node to get deprovisioned)
Source: https://github.com/kubernetes/node-problem-detector#remedy-systems
Use Case #2: Surface the unhealthy node event so that it can be detected by Kubernetes, and then injested into your monitoring stack so you have an auditable historic record that the event occurred and when.
These unhealthy node events are logged somewhere on the host node, but usually, the host node is generating so much noisy/useless log data that these events aren't usually collected by default.
Node Problem Detector knows where to look for these events on the host node and filters out the noise when it sees the signal of a negative outcome it'll post it to its pod log, which isn't noisy.
The pod log is likely getting ingested into an ELK and Prometheus Operator stack, where it can be detected, alerted on, stored, and graphed.
Also, note that nothing is stopping you from implementing both use cases.
Update, added a snippet of node-problem-detector.helm-values.yaml file per request in comment:
log_monitors:
#https://github.com/kubernetes/node-problem-detector/tree/master/config contains the full list, you can exec into the pod and ls /config/ to see these as well.
- /config/abrt-adaptor.json #Adds ABRT Node Events (ABRT: automatic bug reporting tool), exceptions will show up under "kubectl describe node $NODENAME | grep Events -A 20"
- /config/kernel-monitor.json #Adds 2 new Node Health Condition Checks "KernelDeadlock" and "ReadonlyFilesystem"
- /config/docker-monitor.json #Adds new Node Health Condition Check "DockerDaemon" (Checks if Docker is unhealthy as a result of corrupt image)
# - /config/docker-monitor-filelog.json #Error: "/var/log/docker.log: no such file or directory", doesn't exist on pod, I think you'd have to mount node hostpath to get it to work, gain doesn't sound worth effort.
# - /config/kernel-monitor-filelog.json #Should add to existing Node Health Check "KernelDeadlock", more thorough detection, but silently fails in NPD pod logs for me.
custom_plugin_monitors: #[]
# Someone said all *-counter plugins are custom plugins, if you put them under log_monitors, you'll get #Error: "Failed to unmarshal configuration file "/config/kernel-monitor-counter.json""
- /config/kernel-monitor-counter.json #Adds new Node Health Condition Check "FrequentUnregisteredNetDevice"
- /config/docker-monitor-counter.json #Adds new Node Health Condition Check "CorruptDockerOverlay2"
- /config/systemd-monitor-counter.json #Adds 3 new Node Health Condition Checks "FrequentKubeletRestart", "FrequentDockerRestart", and "FrequentContainerdRestart"

Answer 2

Considering node-problem-detector is a Kubernetes addon, you would need to install that addon on your own Kubernetes server.
A Kubernetes CLuster has an addon-manager that will use it.

Answer 3

Do you mean:how to install it?
kubectl create -f https://github.com/kubernetes/node-problem-detector.yaml