CCE Node Problem Detector

Introduction

CCE node problem detector (NPD) is an add-on that monitors abnormal events of cluster nodes and connects to a third-party monitoring platform. It is a daemon running on each node. It collects node issues from different daemons and reports them to the API server. The NPD add-on can run as a DaemonSet or a daemon.

For more information, see node-problem-detector.

Constraints

• When using this add-on, do not format or partition node disks.
• Each NPD process occupies 30 m CPU and 100 MB memory.
• If the NPD version is 1.18.45 or later, the EulerOS version of the host machine must be 2.5 or later.

Permissions

To monitor kernel logs, the NPD add-on needs to read the host /dev/kmsg. Therefore, the privileged mode must be enabled. For details, see privileged.

In addition, CCE mitigates risks according to the least privilege principle. Only the following privileges are available for NPD running:

• cap_dac_read_search: permission to access /run/log/journal.
• cap_sys_admin: permission to access /dev/kmsg.

Installing the Add-on

1. Log in to the CCE console and click the cluster name to access the cluster console. Choose Add-ons in the navigation pane, locate CCE Node Problem Detector on the right, and click Install.
2. On the Install Add-on page, configure the specifications.

   Table 1 NPD configuration

   Parameter	Description
   Add-on Specifications	The specifications can be Custom.
   Pods	If you select Custom, you can adjust the number of pods as required.
   Containers	If you select Custom, you can adjust the container specifications as required.

3. Configure the add-on parameters.

   Only v1.16.0 and later versions support the configurations.

   Table 2 NPD parameters

   Parameter	Description
   common.image.pullPolicy	An image pulling policy. The default value is IfNotPresent.
   feature_gates	A feature gate
   npc.maxTaintedNode	The maximum number of nodes that NPC can add taints to when a single fault occurs on multiple nodes for minimizing impact. The value can be in int or percentage format.
   npc.nodeAffinity	Node affinity of the controller

4. Configure scheduling policies for the add-on.
   

   • Scheduling policies do not take effect on add-on instances of the DaemonSet type.
   • When configuring multi-AZ deployment or node affinity, ensure that there are nodes meeting the scheduling policy and that resources are sufficient in the cluster. Otherwise, the add-on cannot run.

   Table 3 Configurations for add-on scheduling

   Parameter	Description
   Multi AZ	Preferred: Deployment pods of the add-on will be preferentially scheduled to nodes in different AZs. If all the nodes in the cluster are deployed in the same AZ, the pods will be scheduled to that AZ. Equivalent mode: Deployment pods of the add-on are evenly scheduled to the nodes in the cluster in each AZ. If a new AZ is added, you are advised to increase add-on pods for cross-AZ HA deployment. With the Equivalent multi-AZ deployment, the difference between the number of add-on pods in different AZs will be less than or equal to 1. If resources in one of the AZs are insufficient, pods cannot be scheduled to that AZ. Required: Deployment pods of the add-on will be forcibly scheduled to nodes in different AZs. If there are fewer AZs than pods, the extra pods will fail to run.
   Node Affinity	Not configured: Node affinity is disabled for the add-on. Node Affinity: Specify the nodes where the add-on is deployed. If you do not specify the nodes, the add-on will be randomly scheduled based on the default cluster scheduling policy. Specified Node Pool Scheduling: Specify the node pool where the add-on is deployed. If you do not specify the node pool, the add-on will be randomly scheduled based on the default cluster scheduling policy. Custom Policies: Enter the labels of the nodes where the add-on is to be deployed for more flexible scheduling policies. If you do not specify node labels, the add-on will be randomly scheduled based on the default cluster scheduling policy. If multiple custom affinity policies are configured, ensure that there are nodes that meet all the affinity policies in the cluster. Otherwise, the add-on cannot run.
   Toleration	Using both taints and tolerations allows (not forcibly) the add-on Deployment to be scheduled to a node with the matching taints, and controls the Deployment eviction policies after the node where the Deployment is located is tainted. The add-on adds the default tolerance policy for the node.kubernetes.io/not-ready and node.kubernetes.io/unreachable taints, respectively. The tolerance time window is 60s. For details, see Taints and Tolerations.

5. Click Install.

Components

NPD Check Items

Check items are supported only in 1.16.0 and later versions.

Check items cover events and statuses.

• Event-related

  For event-related check items, when a problem occurs, NPD reports an event to the API server. The event type can be Normal (normal event) or Warning (abnormal event).

  Table 5 Event-related check items

  Check Item	Function	Description
  OOMKilling	Listen to the kernel logs and check whether OOM events occur and are reported. Typical scenario: When the memory usage of a process in a container exceeds the limit, OOM is triggered and the process is terminated.	Warning event Listening object: /dev/kmsg Matching rule: "Killed process \\d+ (.+) total-vm:\\d+kB, anon-rss:\\d+kB, file-rss:\\d+kB.*"
  TaskHung	Listen to the kernel logs and check whether taskHung events occur and are reported. Typical scenario: Disk I/O suspension causes process suspension.	Warning event Listening object: /dev/kmsg Matching rule: "task \\S+:\\w+ blocked for more than \\w+ seconds\\."
  ReadonlyFilesystem	Check whether the Remount root filesystem read-only error occurs in the system kernel by listening to the kernel logs. Typical scenario: A user detaches a data disk from a node by mistake on the ECS, and applications continuously write data to the mount point of the data disk. As a result, an I/O error occurs in the kernel and the disk is remounted as a read-only disk. NOTE: If the rootfs of node pods is of the device mapper type, an error will occur in the thin pool if a data disk is detached. This will affect NPD and NPD will not be able to detect node faults.	Warning event Listening object: /dev/kmsg Matching rule: Remounting filesystem read-only

• Status-related

  For status-related check items, when a problem occurs, NPD reports an event to the API server and changes the node status synchronously. This function can be used together with Node-problem-controller fault isolation to isolate nodes.

  If the check period is not specified in the following check items, the default period is 30 seconds.

  Table 6 Checking system components

  Check Item	Function	Description
  Container network component error CNIProblem	Check the status of the CNI components (container network components).	None
  Container runtime component error CRIProblem	Check the status of Docker and containerd of the CRI components (container runtime components).	Check object: Docker or containerd
  Frequent restarts of Kubelet FrequentKubeletRestart	Periodically backtrack system logs to check whether the key component Kubelet restarts frequently.	Default threshold: 10 restarts within 10 minutes If Kubelet restarts for 10 times within 10 minutes, it indicates that the system restarts frequently and a fault alarm is generated. Listening object: logs in the /run/log/journal directory
  Frequent restarts of Docker FrequentDockerRestart	Periodically backtrack system logs to check whether the container runtime Docker restarts frequently.
  Frequent restarts of containerd FrequentContainerdRestart	Periodically backtrack system logs to check whether the container runtime containerd restarts frequently.
  kubelet error KubeletProblem	Check the status of the key component Kubelet.	None
  kube-proxy error KubeProxyProblem	Check the status of the key component kube-proxy.	None

  Table 7 Checking system metrics

  Check Item	Function	Description
  Conntrack table full ConntrackFullProblem	Check whether the conntrack table is full.	Default threshold: 90% Usage: nf_conntrack_count Maximum value: nf_conntrack_max
  Insufficient disk resources DiskProblem	Check the usage of the system disk and CCE data disks (including the CRI logical disk and kubelet logical disk) on the node.	Default threshold: 90% Source: df -h Currently, additional data disks are not supported.
  Insufficient file handles FDProblem	Check if the FD file handles are used up.	Default threshold: 90% Usage: the first value in /proc/sys/fs/file-nr Maximum value: the third value in /proc/sys/fs/file-nr
  Insufficient node memory MemoryProblem	Check whether memory is used up.	Default threshold: 80% Usage: MemTotal-MemAvailable in /proc/meminfo Maximum value: MemTotal in /proc/meminfo
  Insufficient process resources PIDProblem	Check whether PID process resources are exhausted.	Default threshold: 90% Usage: nr_threads in /proc/loadavg Maximum value: smaller value between /proc/sys/kernel/pid_max and /proc/sys/kernel/threads-max.

  Table 8 Checking the storage

  Check Item	Function	Description
  Disk read-only DiskReadonly	Periodically perform write tests on the system disk and CCE data disks (including the CRI logical disk and Kubelet logical disk) of the node to check the availability of key disks.	Detection paths: /mnt/paas/kubernetes/kubelet/ /var/lib/docker/ /var/lib/containerd/ /var/paas/sys/log/cceaddon-npd/ The temporary file npd-disk-write-ping is generated in the detection path. Currently, additional data disks are not supported.
  emptyDir storage pool error EmptyDirVolumeGroupStatusError	Check whether the ephemeral volume group on the node is normal. Impact: Pods that depend on the storage pool cannot write data to the temporary volume. The temporary volume is remounted as a read-only file system by the kernel due to an I/O error. Typical scenario: When creating a node, a user configures two data disks as a temporary volume storage pool. Some data disks are deleted by mistake. As a result, the storage pool becomes abnormal.	Detection period: 30s Source: vgs -o vg_name, vg_attr Principle: Check whether the VG (storage pool) is in the P state. If yes, some PVs (data disks) are lost. Joint scheduling: The scheduler can automatically identify a PV storage pool error and prevent pods that depend on the storage pool from being scheduled to the node. Exceptional scenario: The NPD add-on cannot detect the loss of all PVs (data disks), resulting in the loss of VGs (storage pools). In this case, kubelet automatically isolates the node, detects the loss of VGs (storage pools), and updates the corresponding resources in nodestatus.allocatable to 0. This prevents pods that depend on the storage pool from being scheduled to the node. The damage of a single PV cannot be detected by this check item, but by the ReadonlyFilesystem check item.
  PV storage pool error LocalPvVolumeGroupStatusError	Check the PV group on the node. Impact: Pods that depend on the storage pool cannot write data to the persistent volume. The persistent volume is remounted as a read-only file system by the kernel due to an I/O error. Typical scenario: When creating a node, a user configures two data disks as a persistent volume storage pool. Some data disks are deleted by mistake.
  Mount point error MountPointProblem	Check the mount point on the node. Exceptional definition: You cannot access the mount point by running the cd command. Typical scenario: Network File System (NFS), for example, obsfs and s3fs is mounted to a node. When the connection is abnormal due to network or peer NFS server exceptions, all processes that access the mount point are suspended. For example, during a cluster upgrade, a kubelet is restarted, and all mount points are scanned. If the abnormal mount point is detected, the upgrade fails.	Alternatively, you can run the following command: for dir in `df -h | grep -v "Mounted on" | awk "{print \\$NF}"`;do cd $dir; done && echo "ok"
  Suspended disk I/O DiskHung	Check whether I/O suspension occurs on all disks on the node, that is, whether I/O read and write operations are not responded. Definition of I/O suspension: The system does not respond to disk I/O requests, and some processes are in the D state. Typical scenario: Disks cannot respond due to abnormal OS hard disk drivers or severe faults on the underlying network.	Check object: all data disks Source: /proc/diskstat Alternatively, you can run the following command: iostat -xmt 1 Threshold: Average usage: ioutil >= 0.99 Average I/O queue length: avgqu-sz >= 1 Average I/O transfer volume: iops (w/s) + ioth (wMB/s) <= 1 NOTE: In some OSs, no data changes during I/O. In this case, calculate the CPU I/O time usage. The value of iowait should be greater than 0.8.
  Slow disk I/O DiskSlow	Check whether all disks on the node have slow I/Os, that is, whether I/Os respond slowly. Typical scenario: EVS disks have slow I/Os due to network fluctuation.	Check object: all data disks Source: /proc/diskstat Alternatively, you can run the following command: iostat -xmt 1 Default threshold: Average I/O latency: await >= 5000 ms NOTE: If I/O requests are not responded and the await data is not updated, this check item is invalid.

  Table 9 Other check items

  Check Item	Function	Description
  Abnormal NTP NTPProblem	Check whether the node clock synchronization service ntpd or chronyd is running properly and whether a system time drift is caused.	Default clock offset threshold: 8000 ms
  Process D error ProcessD	Check whether there is a process D on the node.	Default threshold: 10 abnormal processes detected for three consecutive times Source: /proc/{PID}/stat Alternately, you can run the ps aux command. Exceptional scenario: The ProcessD check item ignores the resident D processes (heartbeat and update) on which the SDI driver on the BMS node depends.
  Process Z error ProcessZ	Check whether the node has processes in Z state.
  ResolvConf error ResolvConfFileProblem	Check whether the ResolvConf file is lost. Check whether the ResolvConf file is normal. Exceptional definition: No upstream domain name resolution server (nameserver) is included.	Object: /etc/resolv.conf
  Existing scheduled event ScheduledEvent	Check whether scheduled live migration events exist on the node. A live migration plan event is usually triggered by a hardware fault and is an automatic fault rectification method at the IaaS layer. Typical scenario: The host is faulty. For example, the fan is damaged or the disk has bad sectors. As a result, live migration is triggered for VMs.	Source: http://169.254.169.254/meta-data/latest/events/scheduled This check item is an Alpha feature and is disabled by default.

  The kubelet component has the following default check items, which have bugs or defects. You can fix them by upgrading the cluster or using NPD.

  Table 10 Default kubelet check items

  Check Item	Function	Description
  Insufficient PID resources PIDPressure	Check whether PIDs are sufficient.	Interval: 10 seconds Threshold: 90% Defect: In community version 1.23.1 and earlier versions, this check item becomes invalid when over 65535 PIDs are used. For details, see issue 107107. In community version 1.24 and earlier versions, thread-max is not considered in this check item.
  Insufficient memory MemoryPressure	Check whether the allocable memory for the containers is sufficient.	Interval: 10 seconds Threshold: max. 100 MiB Allocable = Total memory of a node – Reserved memory of a node Defect: This check item checks only the memory consumed by containers, and does not consider that consumed by other elements on the node.
  Insufficient disk resources DiskPressure	Check the disk usage and inodes usage of the kubelet and Docker disks.	Interval: 10 seconds Threshold: 90%

Node-problem-controller Fault Isolation

Fault isolation is supported only by add-ons of 1.16.0 and later versions.

By default, if multiple nodes become faulty, NPC adds taints to up to 10% of the nodes. You can set npc.maxTaintedNode to increase the threshold.

The open source NPD plugin provides fault detection but not fault isolation. CCE enhances the node-problem-controller (NPC) based on the open source NPD. This component is implemented based on the Kubernetes node controller. For faults reported by NPD, NPC automatically adds taints to nodes for node fault isolation.

Collecting Prometheus Metrics

The NPD daemon pod exposes Prometheus metric data on port 19901. By default, the NPD pod is added with the annotation metrics.alpha.kubernetes.io/custom-endpoints: '[{"api":"prometheus","path":"/metrics","port":"19901","names":""}]'. You can build a Prometheus collector to identify and obtain NPD metrics from http://{{NpdPodIP}}:{{NpdPodPort}}/metrics.

If the NPD add-on version is earlier than 1.16.5, the exposed port of Prometheus metrics is 20257.

Currently, the metric data includes problem_counter and problem_gauge, as shown below.

# HELP problem_counter Number of times a specific type of problem have occurred.
# TYPE problem_counter counter
problem_counter{reason="DockerHung"} 0
problem_counter{reason="DockerStart"} 0
problem_counter{reason="EmptyDirVolumeGroupStatusError"} 0
...
# HELP problem_gauge Whether a specific type of problem is affecting the node or not.
# TYPE problem_gauge gauge
problem_gauge{reason="CNIIsDown",type="CNIProblem"} 0
problem_gauge{reason="CNIIsUp",type="CNIProblem"} 0
problem_gauge{reason="CRIIsDown",type="CRIProblem"} 0
problem_gauge{reason="CRIIsUp",type="CRIProblem"} 0
..