diff options
Diffstat (limited to 'docs')
| -rw-r--r-- | docs/consistency.txt | 14 | ||||
| -rw-r--r-- | docs/logs.txt | 10 | ||||
| -rw-r--r-- | docs/maintenance.txt | 55 | ||||
| -rw-r--r-- | docs/managment.txt | 3 | ||||
| -rw-r--r-- | docs/problems.txt | 9 | ||||
| -rw-r--r-- | docs/troubleshooting.txt | 103 | ||||
| -rw-r--r-- | docs/users.txt | 21 | ||||
| -rw-r--r-- | docs/vision.txt | 19 |
8 files changed, 203 insertions, 31 deletions
diff --git a/docs/consistency.txt b/docs/consistency.txt index 91a0ee7..3769a60 100644 --- a/docs/consistency.txt +++ b/docs/consistency.txt @@ -9,6 +9,10 @@ General overview oc get pvc --all-namespaces -o wide - API health check curl -k https://apiserver.kube-service-catalog.svc/healthz + - Docker status (at each node) + docker info + * Enough Data and Metadata Space is available + * The number of resident images is in check (>500-1000 - bad, >2000-3000 - terrible) Nodes ===== @@ -31,7 +35,7 @@ Storage Networking ========== - Check that correct upstream name servers are listed for both DNSMasq (host) and SkyDNS (pods). - If not fix and restart 'origin-node' and 'dnsmasq'. + If not fix and restart 'origin-node' and 'dnsmasq' (it happens that DNSMasq is just stuck). * '/etc/dnsmasq.d/origin-upstream-dns.conf' * '/etc/origin/node/resolv.conf' @@ -46,12 +50,14 @@ Networking - Ensure, we don't have override of cluster_name to first master (which we do during the provisioning of OpenShift plays) - - Sometimes OpenShift fails to clean-up after terminated pod properly. This causes rogue - network interfaces to remain in OpenVSwitch fabric. This can be determined by errors like: + - Sometimes OpenShift fails to clean-up after terminated pod properly (this problem is particularly + triggered on the systems with huge number of resident docker images). This causes rogue network + interfaces to remain in OpenVSwitch fabric. This can be determined by errors like: could not open network device vethb9de241f (No such device) reported by 'ovs-vsctl show' or present in the log '/var/log/openvswitch/ovs-vswitchd.log' which may quickly grow over 100MB quickly. If number of rogue interfaces grows too much, - the pod scheduling will start time-out on the affected node. + the pod scheduling gets even worse (compared to delays caused only be docker images) and + will start time-out on the affected node. * The work-around is to delete rogue interfaces with ovs-vsctl del-port br0 <iface> This does not solve the problem, however. The new interfaces will get abandoned by OpenShift. diff --git a/docs/logs.txt b/docs/logs.txt index e27b1ff..d33ef0a 100644 --- a/docs/logs.txt +++ b/docs/logs.txt @@ -2,6 +2,10 @@ ================= - Various RPC errors. ... rpc error: code = # desc = xxx ... + + - PLEG is not healthy: pleg was last seen active 3m0.448988393s ago; threshold is 3m0s + This is severe and indicates communication probelm (or at least high latency) with docker daemon. As result the node can be marked + temporary NotReady and cause eviction of all resident pods. - 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 @@ -25,10 +29,14 @@ There are no adverse effects to this. It is a potential kernel issue, but should be just ignored by the customer. Nothing is going to break. https://bugzilla.redhat.com/show_bug.cgi?id=1425278 - - E0625 03:59:52.438970 23953 watcher.go:210] watch chan error: etcdserver: mvcc: required revision has been compacted seems fine and can be ignored. + - E0926 09:29:50.744454 93115 mount_linux.go:172] Mount failed: exit status 1 + Output: Failed to start transient scope unit: Connection timed out + It seems caused by too many parallel mounts (about 500 per-node) may cause systemd to hang. + Details: https://github.com/kubernetes/kubernetes/issues/79194 + * Suggested to use 'setsid' to mount volumes instead of 'systemd-run' /var/log/openvswitch/ovs-vswitchd.log ===================================== diff --git a/docs/maintenance.txt b/docs/maintenance.txt new file mode 100644 index 0000000..9f52e18 --- /dev/null +++ b/docs/maintenance.txt @@ -0,0 +1,55 @@ +Unused resources +================ + ! Cleaning of images is necessary if amount of resident images grow above 1000. Everything else has not caused problems yet and could + be ignored unless blocking other actions (e.g. clean-up of old images) + + - Deployments. As is this hasn't caused problems yet, but old versions of 'rc' may block removal of the old images and this may + have negative impact on performance. + oc adm prune deployments --orphans --keep-complete=3 --keep-failed=1 --keep-younger-than=60m --confirm + oc adm prune builds --orphans --keep-complete=3 --keep-failed=1 --keep-younger-than=60m --confirm + * This is, however, does not clean old 'rc' controllers which are allowed by 'revisionHistoryLimit' (and may be something else as + well). There is a script included to clean such controllers 'prunerc.sh' + + - OpenShift sometimes fails to clean stopped containers. This containers again may block removal of images (and likely on itself also + can use Docker performance penalties if accumulated). + * The lost containers can be identified by looking into the /var/log/messages. + PodSandbox "aa28e9c7605cae088838bb4c9b92172083680880cd4c085d93cbc33b5b9e8910" from runtime service failed: ... + * We can find and remove the corresponding container (the short id is just first letters of the long id) + docker ps -a | grep aa28e9c76 + docker rm <id> + * But in general any not-running container which is for a long time remains in stopped state could be considered lost. We can remove + all of them or just ones related to the specific image (if we are cleaning images and something blocks deletion of an old version) + docker rm $(docker ps -a | grep Exited | grep adei | awk '{ print $1 }') + + - If cleaning containers manually or/and forcing termination of pods, some remnants could be left in '/var/lib/origin/openshift.local.volumes/pods' + * Probably, it is also could happen in other cases. This can be detected by looking in /var/log/messages for something like + Orphaned pod "212074ca-1d15-11e8-9de3-525400225b53" found, but volume paths are still present on disk. + * If unknown, the location for the pod in question could be found with 'find . -name heketi*' or something like (the containers names will be listed + under this subdirectory, so they can be used in search)... + * There could be problematic mounts which can be freed with lazy umount + * The folders for removed pods may (and should) be removed. + + - Prunning unused images (this is required as if large amount is accumulated, the additional latencies in communication with docker + daemon will be inrtoduced and result in severe penalties to scheduling performance). Official way to clean unused images is + oc adm prune images --keep-tag-revisions=3 --keep-younger-than=60m --confirm + * This is, however, will keep all images referenced by exisitng bc, dc, rc, and pods (see above). So, it could be worth cleaning OpenShift resources + before before proceeding with images. If images doesn't go, it worth also tryig to clean orphaned containers. + * Some images could be also orphanned by OpenShift infrastructure. OpenShift supports 'hard' prunning to handle such images. + https://docs.openshift.com/container-platform/3.7/admin_guide/pruning_resources.html + First check if something needs to be done: + oc -n default exec -i -t "$(oc -n default get pods -l deploymentconfig=docker-registry -o jsonpath=$'{.items[0].metadata.name}\n')" -- /usr/bin/dockerregistry -prune=check + If there is many orphans, the hard pruning can be executed. This requires additional permissions + for service account running docker-registry + service_account=$(oc get -n default -o jsonpath=$'system:serviceaccount:{.metadata.namespace}:{.spec.template.spec.serviceAccountName}\n' dc/docker-registry) + oc adm policy add-cluster-role-to-user system:image-pruner ${service_account} + and should be done with docker registry in read-only mode (requires restart of default/docker-registry containers) + oc env -n default dc/docker-registry 'REGISTRY_STORAGE_MAINTENANCE_READONLY={"enabled":true}' # wait until new pods rolled out + oc -n default exec -i -t "$(oc -n default get pods -l deploymentconfig=docker-registry -o jsonpath=$'{.items[0].metadata.name}\n')" -- /usr/bin/dockerregistry -prune=delete + oc env -n default dc/docker-registry REGISTRY_STORAGE_MAINTENANCE_READONLY- + + - Cleaning old images which doesn't want to go. + * Investigating image streams and manually deleting the old versions of the images + oc get is adei -o yaml + oc delete image sha256:04afd4d4a0481e1510f12d6d071f1dceddef27416eb922cf524a61281257c66e + * Cleaning old dangling images using docker (on all nodes). Tried and as it seems caused no issues to the operation of the cluster. + docker rmi $(docker images --filter "dangling=true" -q --no-trunc) diff --git a/docs/managment.txt b/docs/managment.txt index cfc6aff..96ae559 100644 --- a/docs/managment.txt +++ b/docs/managment.txt @@ -43,6 +43,9 @@ DOs and DONTs - Few administrative tools could cause troubles. Don't run * oc adm diagnostics + - Old docker 1.12 has many problems. RedHat recommends updating. Don't do! OpenShift 3.7 is only supportingh + docker 1.12 and will not work with 1.13 or later. + Failures / Immidiate ======== diff --git a/docs/problems.txt b/docs/problems.txt index 1d729cd..e616fe4 100644 --- a/docs/problems.txt +++ b/docs/problems.txt @@ -7,7 +7,9 @@ Actions Required Rogue network interfaces on OpenVSwitch bridge ============================================== - Sometimes OpenShift fails to clean-up after terminated pod properly. The actual reason is unclear. + 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 issue is discussed here: https://bugzilla.redhat.com/show_bug.cgi?id=1518684 * And can be determined by looking into: @@ -23,7 +25,8 @@ Rogue network interfaces on OpenVSwitch bridge * 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. + * 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 @@ -35,6 +38,8 @@ Rogue network interfaces on OpenVSwitch bridge * 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 <iface> + * It seems 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. diff --git a/docs/troubleshooting.txt b/docs/troubleshooting.txt index ea987b5..2290901 100644 --- a/docs/troubleshooting.txt +++ b/docs/troubleshooting.txt @@ -134,9 +134,53 @@ etcd (and general operability) pods (failed pods, rogue namespaces, etc...) ==== - - The 'pods' scheduling may fail on one (or more) of the nodes after long waiting with 'oc logs' reporting - timeout. The 'oc describe' reports 'failed to create pod sandbox'. This can be caused by failure to clean-up - after terminated pod properly. It causes rogue network interfaces to remain in OpenVSwitch fabric. + - OpenShift has numerous problems with clean-up resources after the pods. The problems are more likely to happen on the + heavily loaded systems: cpu, io, interrputs, etc. + * This may be indicated in the logs with various errors reporting inability to stop containers/processes, free network + and storage resources. A few examples (not complete) + dockerd-current: time="2019-09-30T18:46:12.298297013Z" level=warning msg="container kill failed because of 'container not found' or 'no such process': Cannot kill container 00a456097fcf8d70a0461f05813e5a1f547446dd10b3b43ebc1f0bb09e841d1b: rpc error: code = 2 desc = no such process" + origin-node: W0930 18:46:11.286634 2497 util.go:87] Warning: Unmount skipped because path does not exist: /var/lib/origin/openshift.local.volumes/pods/6aecbed1-e3b2-11e9-bbd6-0cc47adef0e6/volumes/kubernetes.io~glusterfs/adei-tmp + Error syncing pod 1ed138cd-e2fc-11e9-bbd6-0cc47adef0e6 ("adei-smartgrid-maintain-1569790800-pcmdp_adei(1ed138cd-e2fc-11e9-bbd6-0cc47adef0e6)"), skipping: failed to "CreatePodSandbox" for "adei-smartgrid-maintain-1569790800-pcmdp_adei(1ed138cd-e2fc-11e9-bbd6-0cc47adef0e6)" with CreatePodSandboxError: "CreatePodSandbox for pod \"adei-smartgrid-maintain-1569790800-pcmdp_adei(1ed138cd-e2fc-11e9-bbd6-0cc47adef0e6)\" failed: rpc error: code = 2 desc = NetworkPlugin cni failed to set up pod \"adei-smartgrid-maintain-1569790800-pcmdp_adei\" network: CNI request failed with status 400: 'failed to Statfs \"/proc/28826/ns/net\": no such file or directory\n'" + * A more severe form is then PLEG (POD Lifecycle Event Generator) errors are reported: + origin-node: I0925 07:52:00.422291 93115 kubelet.go:1796] skipping pod synchronization - [PLEG is not healthy: pleg was last seen active 3m0.448988393s ago; threshold is 3m0s] + This indicates a severe delays in communication with docker daemon (can be checked with 'docker info') and may result in node marked + temporarily NotReady causing 'pod' eviction. As pod eviction causes extensive load on the other nodes (which may also be affected of the + same problem), the initial single-node issue may render all cluster unusable. + * With mass evictions, the things could get even worse causing faults in etcd communication. This is reported like: + etcd: lost the TCP streaming connection with peer 2696c5f68f35c672 (stream MsgApp v2 reader) + * Apart from overloaded nodes (max cpu%, io, interrupts), PLEG issues can be caused by + 1. Excessive amount of resident docker images on the node (see bellow) + 2. This can cause and will be further amplified by the spurious interfaces on OpenVSwich (see bellow) + x. Nuanced issues between kubelet, docker, logging, networking and so on, with remediation of the issue sometimes being brutal (restarting all nodes etc, depending on the case). + https://github.com/kubernetes/kubernetes/issues/45419#issuecomment-496818225 + * The problem is not bound to CronJobs, but having regular scheduled jobs make it presence significantly more visible. + Furthermore, CronJobs especially scheduling fat containers, like ADEI, significantly add to the I/O load on the system + and may cause more severe form. + + - After a while, the 'pods' schedulling may get more-and-more sluggish, in general or if assigned to a specific node. + * The docker images are accumulating on the nodes over time. After a threshold it will start adding the latency to the + operation of docker daemon, slow down the pod scheduling (on the affected nodes), and may cause other sever side effects. + The problems will start appearing at around 500-1000 images accumulated at a specific node. With 2000-3000, it will get + severe and almost unusable (3-5 minutes to start a pod). So, eventually the unused images should be cleaned + oc adm prune images --keep-tag-revisions=3 --keep-younger-than=60m --confirm + or alternatively per-node: + docker rmi $(docker images --filter "dangling=true" -q --no-trunc) + * Some images could be orphanned by OpenShift infrastructure (there was not a major number of orphaned images on KaaS yet). + OpenShift supports 'hard' prunning to handle such images. + https://docs.openshift.com/container-platform/3.7/admin_guide/pruning_resources.html + * Even afterwards, a significant number of images may stay resident. There is two inter-related problems: + 1. Docker infrastructure relies on the intermediate images. Consequently, very long Dockerfiles will create a LOT of images. + 2. OpenShift keeps history of 'rc' which may refence several versions of old docker images. This will be not cleaned by the + described approach. Furthermore, stopped containers lost by OpenShift infrastructure (see above) also prevent clean-up of + the images + Currenly, a dozen KDB pods produce about 200-300 images. In some cases, optimization of dockerfiles and, afterwards, a trough + cleanup of old images may become necessity. The intermediate images can be found with 'docker images -a' (all images with + <none> as repository and the name), but there is no easy way to find pod populating them. One, but not very convinient is the following + project (press F5 on startup): https://github.com/TomasTomecek/sen + + - In a more sever4 form, the 'pods' scheduling may fail all together on one (or more) of the nodes. After a long waiting, + the 'oc logs' will report timeout. The 'oc describe' reports 'failed to create pod sandbox'. This can be caused by failure + to clean-up after terminated pod properly. It causes rogue network interfaces to remain in OpenVSwitch fabric. * This can be determined by errors reported using 'ovs-vsctl show' or present in the log '/var/log/openvswitch/ovs-vswitchd.log' which may quickly grow over 100MB quickly. could not open network device vethb9de241f (No such device) @@ -149,7 +193,7 @@ pods (failed pods, rogue namespaces, etc...) * The issue is discussed here: https://bugzilla.redhat.com/show_bug.cgi?id=1518684 https://bugzilla.redhat.com/show_bug.cgi?id=1518912 - + - After crashes / upgrades some pods may end up in 'Error' state. This is quite often happen to * kube-service-catalog/controller-manager * openshift-template-service-broker/api-server @@ -180,26 +224,24 @@ pods (failed pods, rogue namespaces, etc...) * OpenShift upgrade, the namespaces are gone (but there could be a bunch of new problems). * ... i don't know if install, etc. May cause the trouble... - - There is also rogue pods (mainly due to some problems with unmounting lost storage), etc. If 'oc delete' does not - work for a long time. It worth - * Determining the host running failed pod with 'oc get pods -o wide' - * Going to the pod and killing processes and stopping the container using docker command - * Looking in the '/var/lib/origin/openshift.local.volumes/pods' for the remnants of the container - - This can be done with 'find . -name heketi*' or something like... - - There could be problematic mounts which can be freed with lazy umount - - The folders for removed pods may (and should) be removed. - - - Looking into the '/var/log/messages', it is sometimes possible to spot various erros like - * Orphaned pod "212074ca-1d15-11e8-9de3-525400225b53" found, but volume paths are still present on disk. - The volumes can be removed in '/var/lib/origin/openshift.local.volumes/pods' on the corresponding node - * PodSandbox "aa28e9c7605cae088838bb4c9b92172083680880cd4c085d93cbc33b5b9e8910" from runtime service failed: ... - - We can find and remove the corresponding container (the short id is just first letters of the long id) - docker ps -a | grep aa28e9c76 - docker rm <id> - - We further can just destroy all containers which are not running (it will actually try to remove all, - but just error message will be printed for running ones) - docker ps -aq --no-trunc | xargs docker rm + - There is also rogue pods (mainly due to some problems with unmounting lost storage) remaining "Deleting" state, etc. + There are two possible situations: + * The containers are actually already terminated, but OpenShift is not aware of it for some reason. + * The containers are actually still running, but OpenShift is not able to terminate them for some reason. + It is relatively easy to find out which is the case: + * Finding the host running the failed pod with 'oc get pods -o wide' + * Checking if associated containers are still running on the host with 'docker ps' + The first case it relatively easy to handle, - one can simply enforce pod removal with + oc delete --grace-period=0 --force + In the second case we need + * To actually stop containers before proceeding (enforcing will just leave them running forever). This can + be done directly using 'docker' commands. + * It also may be worth trying to clean associated resources. Check 'maintenace' documentation for details. + - Permission problems will arise if non-KaaS namespace (using high range supplemental-group for GlusterFS mounts) is converted + to KaaS (gid ranges within 1000 - 10,000 at the moment). The allowed gids should be configured in the namespace specification + and the pods should be allowed to access files. Possible errors: + unable to create pods: pods "mongodb-2-" is forbidden: no providers available to validate pod request @@ -219,6 +261,14 @@ Storage Particularly there is a big problem for ansible-ran virtual machines. The system disk is stored under '/root/VirtualBox VMs' and is not cleaned/destroyed unlike second hard drive on 'vagrant destroy'. So, it should be cleaned manually. + + - Too many parallel mounts (above 500 per node) may cause systemd slow-down/crashes. It is indicated by + the following messages in the log: + E0926 09:29:50.744454 93115 mount_linux.go:172] Mount failed: exit status 1 + Output: Failed to start transient scope unit: Connection timed out + * Solution is unclear, there are some suggestions to use 'setsid' in place of 'systemd-run' to do mounting, + but not clear how. Discussion: https://github.com/kubernetes/kubernetes/issues/79194 + * Can we do some rate-limiting? - Problems with pvc's can be evaluated by running oc -n openshift-ansible-service-broker describe pvc etcd @@ -271,7 +321,12 @@ MySQL The remedy is to restart slave MySQL with 'slave_parallel_workers=0', give it a time to go, and then restart back in the standard multithreading mode. - +Administration +============== + - Some management tasks may require to login on ipekatrin* nodes. Thereafter, the password-less execution of + 'oc' may fail on master nodes complaining on invalid authentication token. To fix it, it is necessary to check + /root/.kube/config and remove references on logged users keeping only 'system:admin/kaas-kit-edu:8443' alkso check + listed contexts and current-context. Performance =========== diff --git a/docs/users.txt b/docs/users.txt new file mode 100644 index 0000000..c28f400 --- /dev/null +++ b/docs/users.txt @@ -0,0 +1,21 @@ +Dockerfiles +=========== + - Long (many-layer) Dockerfiles may cause a significant disruption to the OpenShift cluster (far behind just performance penalty + of working with this layers). Currently, it is imperative to reduce the number of intermediate images resident on the OpenShift nodes. + * The better approach is to optimize Dockerfiles: All ENV defined at once, as few ARG as possible, signle COPY, and a single RUN to + setup everything. + * Alternatively, the final image can be squashed with 'docker build --squash ...' (enable experimental features in docker daemon). + However, this is incompatible with OpenShift build process. + If optimizing already running applications, it is not enough just to re-build images. Old images could be referenced by the old + 'rc' left in the system or even stopped containers lost by OpenShift infrastructure. + * Check if old images still present (https://github.com/TomasTomecek/sen is only application I am aware of capable of showing it) + * See maintenance section how to get rid if old images are still present + +Deployments +=========== + - CronJobs is currently a bit problematic and periodically cause some lost resources, etc. They can be used if necessarily, but it better + to minimize. In any case, it is crucial to minimize size of frequently scheduled Job containers (otherwise large I/O). + + +Storage +======= diff --git a/docs/vision.txt b/docs/vision.txt new file mode 100644 index 0000000..bfef287 --- /dev/null +++ b/docs/vision.txt @@ -0,0 +1,19 @@ +Ands v.2 +======== + - Try overlay2 storage driver (LVM is used in Ands v.1). Check also further docker configuration options: 'cgroup-driver', ... + - Integrate fast Ethernet and use conteiner native networking. OpenVSwitch is slow and causes problems. + - Do not run pods on Master nodes, but Gluster and a few databases pods (MySQL) are OK (multiple reasons, especially mounting a lot of Gluster Volumes) + - Object Storage should be integrated, either Gluster Block is ready for production or we have to use Ceph as well + - Automatic provisioning would be much better then handling volumes trough Ands. Basically, this will render Ands redundant. We can switch to Helm, etc. + But, we need ability to easily understand which volume belong to which pod/namespace and automatically kill redundant volumes. + +Questions +========= + - Updates to cluster configuration (evaluate current load, etc.)? Non-scheduling masters? Something with storage? Specify appropriate node parameters + - Shall we switch to plain Kubernetes or keep using OpenShift. Discussion (just take a not about security - it is right to ban containers running as + root, otherwise hazard to our data model): + https://cloudowski.com/articles/10-differences-between-openshift-and-kubernetes/ + - Can we find a good distributed storage for data-intensive databases. Current, slave-master model requires too much manual attention. + - Can we find a way to run GUI applications in containers? Kind of having CUDA profiller would be nice. + - Think about monitoring. Probably SNMP + it would be nice to have some kind of SQL database with perofrmance metrics. + |