Unfortunately, there will always be some cases where OpenShift fails to install properly. In these events, it is helpful to understand the likely failure modes as well as how to troubleshoot the failure.
The installer doesn't provision worker nodes directly, like it does with master nodes. Instead, the cluster relies on the Machine API Operator, which is able to scale up and down nodes on supported platforms. If more than fifteen to twenty minutes (depending on the speed of the cluster's Internet connection) have elapsed without any workers, the Machine API Operator needs to be investigated.
The status of the Machine API Operator can be checked by running the following command from the machine used to install the cluster:
When the Kubernetes API is unavailable, the master nodes will need to checked to ensure that they are running the correct components. This requires SSH access so it is necessary to include an administrator's SSH key during the installation.
The first thing to check is to make sure that etcd is running on each of the masters. The etcd logs can be viewed by running the following on each master node:
For added security, SSH isn't available from the Internet by default. There are several options for enabling this functionality:
- Create a bastion host that is accessible from the Internet and has access to the cluster. If the bootstrap machine hasn't been automatically destroyed yet, it can double as a temporary bastion host since it is given a public IP address.
- Configure network peering or a VPN to allow remote access to the private network.
In order to SSH into the master nodes as user `core`, it is necessary to include an administrator's SSH key during the installation.
The selected key, if any, will be added to the `core` user's `~/.ssh/authorized_keys` via [Ignition](https://github.com/coreos/ignition) and is not configured via platform-specific approaches like [AWS key pairs][aws-key-pairs].
See [here][machine-config-daemon-ssh-keys] for information about managing SSH keys via the machine-config daemon.
If SSH isn't able to connect to the nodes, they may be waiting on the bootstrap node before they can boot. The initial set of master nodes fetch their boot configuration (the Ignition Config) from the bootstrap node and will not complete until they successfully do so. Check the console output of the nodes to determine if they have successfully booted or if they are waiting for Ignition to fetch the remote config.
Master nodes waiting for Ignition is indicative of problems on the bootstrap node. SSH into the bootstrap node to [investigate further](#troubleshooting-the-bootstrap-node).
### Troubleshooting the Bootstrap Node
If the bootstrap node isn't available, first double check that it hasn't been automatically removed by the installer. If it's not being created in the first place, the installer will need to be [troubleshot](#installer-fails-to-create-resources).
1. If SSH is available, the following command can be run on the bootstrap node: `journalctl --unit=bootkube.service`
2. Regardless of whether or not SSH is available, the following command can be run: `curl --insecure --cert ${INSTALL_DIR}/tls/journal-gatewayd.crt --key ${INSTALL_DIR}/tls/journal-gatewayd.key 'https://${BOOTSTRAP_IP}:19531/entries?follow&_SYSTEMD_UNIT=bootkube.service'`
The installer can also gather a log bundle from the bootstrap host using SSH as describe in [troubleshooting bootstrap](./troubleshootingbootstrap.md) document.
Error signing CSR provided in request from agent: error parsing profile: invalid organization
```
This is safe to ignore and merely indicates that the etcd bootstrapping is still in progress. etcd makes use of the CSR APIs provided by Kubernetes to issue and rotate its TLS assets, but these facilities aren't available before etcd has formed quorum. In order to break this dependency loop, a CSR service is run on the bootstrap node which only signs CSRs for etcd. When the Kubelet attempts to go through its TLS bootstrap, it is initially denied because the service it is communicating with only respects CSRs from etcd. After etcd starts and the control plane begins bootstrapping, an approver is scheduled and the Kubelet CSR requests will succeed.
The easiest way to get more debugging information from the installer is to check the log file (`.openshift_install.log`) in the install directory. Regardless of the logging level specified, the installer will write its logs in case they need to be inspected retroactively.
The installer uses the [cluster-version-operator] to create all the components of an OpenShift cluster. When the installer fails to initialize the cluster, the most important information can be fetched by looking at the [ClusterVersion][clusterversion] and [ClusterOperator][clusteroperator] objects:
Some of most important [conditions][cluster-operator-conditions] to take note are `Failing`, `Available` and `Progressing`. You can look at the conditions using:
$ oc --kubeconfig=${INSTALL_DIR}/auth/kubeconfig get clusterversion version -o=jsonpath='{range .status.conditions[*]}{.type}{" "}{.status}{" "}{.message}{"\n"}{end}'
To get detailed information on why an individual cluster operator is `Failing` or not yet `Available`, you can check the status of that individual operator, for example `monitoring`:
Again, the cluster operators also publish [conditions][cluster-operator-conditions] like `Failing`, `Available` and `Progressing` that can help user provide information on the current state of the operator:
**NOTE:** Failing to initialize the cluster is usually not a fatal failure in terms of cluster creation as the user can look at the failures from `ClusterOperator` to debug failures for a cluster operator and take actions which can allow `cluster-version-operator` to make progress.
### Installer Fails to Fetch Console URL
The installer fetches the URL for OpenShift console using the [route][route-object] in `openshift-console` namespace. If the installer fails the fetch the URL for the console:
1. Check if the console router is `Available` or `Failing`
The installer adds the default ingress certificate to the list of trusted client certificate authorities in `${INSTALL_DIR}/auth/kubeconfig`. If the installer fails to add the ingress certificate to `kubeconfig`, you can fetch the certificate from the cluster using the following command:
This usually means that, for whatever reason, networking is not available on the node. You can confirm this by looking at the detailed output of the node:
The first thing to determine is the status of the SDN. The SDN deploys three daemonsets:
- *sdn-controller*, a control-plane component
- *sdn*, the node-level networking daemon
- *ovs*, the OpenVSwitch management daemon
All 3 must be healthy (though only a single `sdn-controller` needs to be running). `sdn` and `ovs` must be running on every node, and DESIRED should equal AVAILABLE. On a healthy 2-node cluster you would see:
```console
$ kubectl -n openshift-sdn get daemonsets
NAME DESIRED CURRENT READY UP-TO-DATE AVAILABLE NODE SELECTOR AGE
This means the network-operator didn't run. Skip ahead [to that section](#debugging-the-cluster-network-operator). Otherwise, let's debug the SDN.
#### Debugging the openshift-sdn
On the NotReady node, you need to find out which pods, if any, are in a bad state. Be sure to substitute in the correct `spec.nodeName` (or just remove it).
```console
$ kubectl -n openshift-sdn get pod --field-selector "spec.nodeName=ip-10-0-27-9.ec2.internal"
NAME READY STATUS RESTARTS AGE
ovs-dk8bh 1/1 Running 1 52m
sdn-8nl47 1/1 CrashLoopBackoff 3 52m
```
Then, retrieve the logs for the SDN (and the OVS pod, if that is failed):
```sh
kubectl -n openshift-sdn logs sdn-8nl47
```
Some common error messages:
-`Cannot fetch default cluster network`: This means the `sdn-controller` has failed to run to completion. Retrieve its logs with `kubectl -n openshift-sdn logs -l app=sdn-controller`.
-`warning: Another process is currently listening on the CNI socket, waiting 15s`: Something has gone wrong, and multiple SDN processes are running. SSH to the node in question, capture the out of `ps -faux`. If you just need the cluster up, reboot the node.
- Error messages about ovs or OpenVSwitch: Check that the `ovs-*` pod on the same node is healthy. Retrieve its logs with `kubectl -n openshift-sdn logs ovs-<name>`. Rebooting the node should fix it.
- Any indication that the control plane is unavailable: Check to make sure the apiserver is reachable from the node. You may be able to find useful information via `journalctl -f -u kubelet`.
If you think it's a misconfiguration, file a [network operator](https://github.com/openshift/cluster-network-operator) issue. RH employees can also try #forum-sdn.
#### Debugging the cluster-network-operator
The cluster network operator is responsible for deploying the networking components. It does this in response to a special object created by the installer.
From a deployment perspective, the network operator is often the "canary in the coal mine." It runs very early in the installation process, after the master nodes have come up but before the bootstrap control plane has been torn down. It can be indicative of more subtle installer issues, such as long delays in bringing up master nodes or apiserver communication issues. Nevertheless, it can have other bugs.
First, determine that the network configuration exists:
