Actions Required ================ * Long-term solution to 'rogue' interfaces is unclear. May require update to OpenShift 3.9 or later. However, proposed work-around should do unless execution rate grows significantly. * All other problems found in logs can be ignored. Client Connection ================= * For some reason OpenShift requests client certificates. This is ignored by majority of old browsers, but Firefox 70+ is able to offer installed user certificates. If KIT certificate selected, OpenShift fails to start. This can be easily circumvented just by pressing 'Cancel' when client-cerificate selection box pops up. Leaked resourced after node termination: Rogue network interfaces on OpenVSwitch bridge, unreclaimed IPs in pod-network, ... ======================================= Sometimes OpenShift fails to clean-up after terminated pod properly. The actual reason is unclear, but severity of the problem is increased if extreme amount of images is presented in local Docker storage. Several thousands is defenitively intensifies this problem. * The issues are discussed here: https://bugzilla.redhat.com/show_bug.cgi?id=1518684 https://bugzilla.redhat.com/show_bug.cgi?id=1518912 * And can be determined by looking into: ovs-vsctl show Problems: * As number of rogue interfaces grow, it start to have impact on performance. Operations with ovs slows down and at some point the pods schedulled to the affected node fail to start due to timeouts. This is indicated in 'oc describe' as: 'failed to create pod sandbox' * With time, the new rogue interfaces are created faster and faster. At some point, it really slow downs system and causes pod failures (if many pods are re-scheduled in paralllel) even if not so many rogue interfaces still present * Furthermore, there is a limit range of IPs allocated for pod-network at each node. Whatever it is caused by tje lost bridges or it is an unrellated resource-management problem in OpenShift, but this IPs also start to leak. As number of leaked IPs increase, it gets longer for OpenShift to find IP which is still free and pod schedulling slows down further. At some point, the complete range of IPs will get exhausted and pods will fail to start (after long waiting in Scheduling state) on the affected node. * Even if not failed, it takes several minutes to schedule the pod on the affected nodes. Cause: * Unclear, but it seems periodic ADEI cron jobs causes the issue if many images are present in docker. * Could be related to 'container kill failed' problem explained in the section bellow. Cannot kill container ###: rpc error: code = 2 desc = no such process Solutions: * According to RedHat the temporal solution is to reboot affected node (just temporarily reduces the rate how often the new spurious interfaces appear, but not preventing the problem completely in my case). The problem should go away, but may re-apper after a while. * The simplest work-around is to just remove rogue interface. They will be re-created, but performance problems only starts after hundreds accumulate. ovs-vsctl del-port br0 * Similarly, the unused IPs could be cleaned in "/var/lib/cni/networks/openshift-sdn", just check if docker image referenced in each IP file is still running with "docker ps". Afterwards, the 'orgin-node' service should be restarted. * It seems also helpful to purge unused docker images to reduce the rate of interface apperance. Status: * Cron job is installed which cleans rogue interfaces as they number hits 25. Hanged pods =========== POD processes may stuck. Normally, such processes will be detected using 'liveliness' probe and will be restarted by OpenShift if necessary. However, ocasionally processes may stuck in syscalls (such processes are marked with 'D' in ps). These processes can't be killed with SIGKILL and OpenShift will not be able to terminate them leaving indefinitely in 'Terminating' status. Problems: * Pods stuck in 'Terminating' status preventing start of new replicas. In case of 'jobs', large number of 'Terminating' pods could overload OpenShift controllers. Cause: * One reason are the spurious locks on the GlusterFS file system. On CentOS 7, it impossible to interrupt process waiting for the lock initiated by blocking 'flock' call. It gets stuck in a syscall and is indicated by state 'D' in the ps output. Sometimes, GlusterFS may kept files locked despite that processes holding these locks have already exited/crashed. I am not sure about exact conditions when this happens, but it seems for instance the crashed Docker daemon may cause effect if some of the running containers were holding locks on GFS at the moment of crash. - We can verify if this is the case by checking if process associated with the problematic pod is stuck in state 'D' and by analyzing its backtrace (/proc//stack). Solutions: * Avoid blocking flock on GlusterFS. Use polling with sleep instead. To release already stuck pods, we need to find and destroy problematic locks. GlusterFS allows to debug locks using 'statedump', check GlusterFS documentation for details. While there is also mechanism to clean such locks. It is not always working. Alternative is to remove locked files AND keep them removed for a while until all blocked 'flock' syscalls are released. Hanged MySQL connection ======================= Stale MySQL locks may prevent new clients connecting to certain tables in MySQL database. Problems: * The problem may affect either only clients trying to obtain 'write' access or for all usage patterns. In the first case, it will cause ADEI 'caching' threads to hang indefinitely and 'maintain' threads will be terminated after specified timeout leaving administrative scripts un-processed. Cause: * For whatever reason, some crashed clients may preserve the locks. I believe it could also be caused by crashed 'docker' daemon as one possibel reason. The problem can be found bt executing 'SHOW PROCESSLIST' on MySQL server. More diagnostic possibilities are discussed in MySQL notes. Solutions; * Normally, restarting MySQL pod should be enough. Orphaning / pod termination problems in the logs ================================================ There is several classes of problems reported with unknow reprecursions in the system log. Currently, I don't see any negative side effects except some of these issues may trigger "rogue interfaces" problem. ! container kill failed because of 'container not found' or 'no such process': Cannot kill container ###: rpc error: code = 2 desc = no such process" Despite the errror, the containers are actually killed and pods destroyed. However, this error likely triggers problem with rogue interfaces staying on the OpenVSwitch bridge. Scenario: * happens with short-living containers - containerd: unable to save f7c3e6c02cdbb951670bc7ff925ddd7efd75a3bb5ed60669d4b182e5337dec23:d5b9394468235f7c9caca8ad4d97e7064cc49cd59cadd155eceae84545dc472a starttime: read /proc/81994/stat: no such process containerd: f7c3e6c02cdbb951670bc7ff925ddd7efd75a3bb5ed60669d4b182e5337dec23:d5b9394468235f7c9caca8ad4d97e7064cc49cd59cadd155eceae84545dc472a (pid 81994) has become an orphan, killing it Scenario: This happens every couple of minutes and attributed to perfectely alive and running pods. * For instance, ipekatrin1 was complaining some ADEI pod. * After I removed this pod, it immidiately started complaining on 'glusterfs' replica. * If 'glusterfs' pod re-created, the problem persist. * It seems only a single pod is affected at each given moment (at least this was always true on ipekatrin1 & ipekatrin2 while I was researching the problem) Relations: * This problem is not aligned with the next 'container not found' problem. One happens with short-living containers which actually get destroyed. This one is triggered for persistent container which keep going. And in fact this problem is triggered significantly more frequently. Cause: * Seems related to docker health checks due to a bug in docker 1.12* which is resolved in 1.13.0rc2 https://github.com/moby/moby/issues/28336 Problems: * It seems only extensive logging, according to the discussion in the issue Solution: Ignore for now * docker-1.13 had some problems with groups (I don't remember exactly) and it was decided to not run it with current version of KaaS. * Only update docker after extensive testing on the development cluster or not at all. - W0625 03:49:34.231471 36511 docker_sandbox.go:337] failed to read pod IP from plugin/docker: NetworkPlugin cni failed on the status hook for pod "...": Unexpected command output nsenter: cannot open /proc/63586/ns/net: No such file or directory - W0630 21:40:20.978177 5552 docker_sandbox.go:337] failed to read pod IP from plugin/docker: NetworkPlugin cni failed on the status hook for pod "...": CNI failed to retrieve network namespace path: Cannot find network namespace for the terminated container "..." Scenario: * It seems can be ignored, see RH bug. * Happens with short-living containers (adei cron jobs) Relations: * This is also not aligned with 'container not found'. The time in logs differ significantly. * It is also not aligned with 'orphan' problem. Cause: ? https://bugzilla.redhat.com/show_bug.cgi?id=1434950 - E0630 14:05:40.304042 5552 glusterfs.go:148] glusterfs: failed to get endpoints adei-cfg[an empty namespace may not be set when a resource name is provided] E0630 14:05:40.304062 5552 reconciler.go:367] Could not construct volume information: MountVolume.NewMounter failed for volume "kubernetes.io/glusterfs/4 I guess some configuration issue.... Probably can be ignored... Scenario: * Reported on long running pods with persistent volumes (katrin, adai-db) * Also seems an unrelated set of the problems. Evicted Pods ============ Pods are evicted if node running pod becomes unavailable or have not enough resources to run the pod. - It is possible to lookup which resource is likely triggering event by > oc describe node ipekatrin2.ipe.kit.edu Type Status LastHeartbeatTime LastTransitionTime Reason Message ---- ------ ----------------- ------------------ ------ ------- OutOfDisk False Tue, 05 Apr 2022 03:24:54 +0200 Tue, 21 Dec 2021 19:09:33 +0100 KubeletHasSufficientDisk kubelet has sufficient disk space available MemoryPressure False Tue, 05 Apr 2022 03:24:54 +0200 Tue, 21 Dec 2021 19:09:33 +0100 KubeletHasSufficientMemory kubelet has sufficient memory available DiskPressure False Tue, 05 Apr 2022 03:24:54 +0200 Mon, 04 Apr 2022 10:00:23 +0200 KubeletHasNoDiskPressure kubelet has no disk pressure Ready True Tue, 05 Apr 2022 03:24:54 +0200 Tue, 21 Dec 2021 19:09:43 +0100 KubeletReady kubelet is posting ready status The latest transition is 'DiskPressure' happened on Apr 04. So, likely disk is an issue. - DiskPressure eviction * This might happen because the pod writting to much output to the logs (standard ouput). This logs are stored under '/var/lib/origin/openshift.local.volumes/pods/...' and if growing large might use all the space in '/var' file system. OpenShift is not rotating these logs and have no other mechanisms to prevent large output eventually causing space issues. So, pods have to rate-limit output to stdout. Otherwise, we need to find misbehaving pods which writes too much... * Another problem is 'inode' pressure. This can be checked with 'df' and anything above 80% is definitively a sign of a problem df -i The particular folder with lots of inodes can be found with the following command: { find / -xdev -printf '%h\n' | sort | uniq -c | sort -k 1 -n; } 2>/dev/null Likely there would be some openshift-related volume logs in '/var/lib/origin/openshift.local.volumes/plugins/kubernetes.io/' Check particularly cronJob logs mounting volumes, e.g. various 'adei' stuff. Can be cleaned with find /var/lib/origin/openshift.local.volumes/plugins/kubernetes.io/ -name '*.log' -delete - If resource is not available for long time, node will become NotReady and all pods will be evicted. However, short term problems caused by pod itself are likely cause only eviction of this particular pod itself only (once pod evicted disk/memory-space is recalimed, logs are deleted). So, it is possible to find the problematic pod by looking which pod was evicted most frequently.