Dec 172018
 

The OpenShift community produces a lot of interesting tutorials about how to try new solutions and configurations but unfortunately they are mostly based on a minimal setup such as MiniShift, which is definitely a cool gimmick, but badly resembles a real cluster setup. Often those posts only concentrate on the known good path about how something is supposed to function in the best case. They rarely mention how it could be debugged or fixed if it doesn’t work as expected. As all of us know, the more complex a system is, the more can go wrong and this technology is no exception especially when run in a real distributed setup. To give you some insight in how such procedures can go wrong, I’d like to share the experience I made when I tried to update my multi-master/multi-node OKD cluster. As an experienced Linux engineer or developer you might think that version updates are nothing special or exciting, but this experience will disabuse you. I hit many issues and here is how I did it.

IMPORTANT: This is not a guide how to upgrade OpenShift. It’s only a field report which is missing a lot of technical details for a successful upgrade. Please always investigate the official documentation.

Starting Position

At home I run a small OKD cluster consisting of three masters, each also hosting a etcd member, and four nodes, of which there are two infrastructure nodes, hosting the routers and registry, and two compute nodes, hosting the applications. I feel that’s the minimal setup required to resemble a production-like cluster. To make the setup a bit more interesting the persistent storage is served by a container-native storage (CNS) configuration were three GlusterFS pods are distributed on the masters. This definitely deviates from how a production cluster should be setup, but as I’m running this locally at home, I don’t have enough resources available for separate storage nodes.

My masters and nodes are running atop of CentOS Atomic Host which I updated to the latest 7.1811 release just a few days ago. As identity provider for OpenShift I’m using a FreeIPA server on a separate CentOS box. Since I installed this cluster with OpenShift Origin 3.9 five months ago, it was running stable and I had a lot of fun with it. After the recently published security advisories time has come to finally take a chance to upgrade OpenShift.

What will change?

The first thing, you should always do before starting an OpenShift update, is to carefully read the release notes. I explicitly linked the Red Hat OpenShift Container Platform (OCP) release notes here, because OKD unfortunately doesn’t nicely touch up theirs. For the initial release update they are mostly congruent. Make sure to study it carefully, as this might be the primary source of information once something starts going down. For the 3.10 update, an important information is the new handling of the containerized master controller and API services. Eventually we now have a basic idea about what to expect.

Updating the Ansible Inventory

It would be nice, if there was a fool-proof command to run the update and it seems that OpenShift 4 with its Cluster Version Operator is heading there. But until then we need to carefully study and follow the official OKD 3.10 Upgrade Guide. It’s important to get the documentation for the correct release because the involved adjustments to the Ansible inventory are different from release to release. For those not knowing how OpenShift 3.x release upgrades work, it’s done via Ansible playbooks which are using the same inventory (definition on how everything will be configured) as the initial OpenShift cluster installation.

In my inventory file, I first added the Node Group assignments. E.g. the infrastructure nodes are no longer defined via openshift_node_labels variable, but via a dedicated openshift_node_group_name variable which references a node group definition from the openshift_node_groups configuration. The same changes have to be done also for the master and the compute nodes:

  • OpenShift 3.9:
    [infra-nodes:vars]
    # Set region to be dedicated for infrastructure pods
    openshift_node_labels={'region': 'infra', 'zone': 'default'}
    
  • OpenShift 3.10:
    [infra-nodes:vars]
    # Set infra node group
    openshift_node_group_name='node-config-infra'
    

Note that although the openshift_node_labels variable is no longer effective, no labels will be removed during the upgrade. So if you don’t get the label definition right at the beginning you don’t have to worry that after an in-place upgrade some workload is suddenly not scheduled anymore.

I had some custom openshift_node_kubelet_args defined in my OpenShift 3.9 inventory but this variable is also not respected any longer. With 3.10 the correct way to customize the node configuration is to define a edits argument in the corresponding node group definition, which is then applied to a ConfigMap resource by the custom yedit Ansible module. While writing such a definition itself is already not super intuitive, it can only be done by re-defining the entire openshift_node_groups variable, possibly also blowing up every other node group definition if done wrong. For the moment, I chose to drop my custom node configuration entirely to make the inventory less error prone for now.

Before running the upgrade playbooks it’s also important that every manual configuration update done in the past (e.g. in the master-config.yml) has to be reflected somewhere in the Ansible inventory. Otherwise the change might be lost after the upgrade. In my setup I still had to add the LDAP authenticator to the openshift_master_identity_providers variable because I added it manually after the initial cluster installation.

The section about Special Considerations When Using Containerized GlusterFS gave me a bit of a bad feeling as my GlusterFS pods are running on the control-plane hosts. But it’s not an easy task to change that now, so I chose to still go on with the upgrade and hope for a work-around in case something should break.

Fixing a failed CNS Brick Process

Once I felt confident that my inventory was in good shape, I started the control-plane upgrade playbook placed at playbooks/byo/openshift-cluster/upgrades/v3_10/upgrade.yml. Only after a few minutes it already failed for the first time. The error message said, that my hosted registry presistent volume is not healthy. That are good news, because the playbook did properly detect my unconventional CNS setup and was even able to check the healthiness. Fortunately, this issue was already familiar to me and it was easily fixed. Here is how you do this:

  1. Change to the glusterfs project (or any custom project where you are running the GlusterFS pods) as a project or cluster administrator and query the names of the GlusterFS pods:
    $ oc project glusterfs
    Now using project "glusterfs" on server "https://openshift.example.com:8443".
    $ oc get pods -n glusterfs -o wide
    NAME                      READY     STATUS    RESTARTS   AGE       IP            NODE
    glusterfs-storage-dz8qj   1/1       Running   3          2d        10.0.0.10     master01.example.com
    glusterfs-storage-jncsl   1/1       Running   3          1d        10.0.0.11     master02.example.com
    glusterfs-storage-r24wg   1/1       Running   2          11h       10.0.0.12     master03.example.com
    heketi-storage-1-w8c42    1/1       Running   1          6d        10.129.0.21   node02.example.com
    
  2. Connect to one of the GlusterFS pods and list the volume status. E.g.:
    $ oc rsh glusterfs-storage-dz8qj
    sh-4.2# gluster volume status
    Status of volume: glusterfs-registry-volume                                 
    Gluster process                             TCP Port  RDMA Port  Online  Pid  
    ------------------------------------------------------------------------------
    Brick 10.0.0.11:/var/lib/heketi/mounts/vg_61                                 
    bc5a26248cc6ea9fb7ffaae4edbe93/brick_128dfe                                 
    b5436dad25702e689d3d6f4b8a/brick            49152     0          Y       204
    Brick 10.0.0.12:/var/lib/heketi/mounts/vg_5e                                   
    cf67ed85f71cf28090d7db1acc6433/brick_d36469                                 
    4c7034276c22e264eb2576413b/brick            49152     0          N
    Brick 10.0.0.10:/var/lib/heketi/mounts/vg_49                                   
    e81b0f4c91942c7657e9b7ffff7834/brick_8717e0                                 
    992390a7c04431890ba56b7656/brick            49152     0          Y       224
    Self-heal Daemon on localhost               N/A       N/A        Y       215
    Self-heal Daemon on 10.0.0.11               N/A       N/A        Y       163  
    Self-heal Daemon on 10.0.0.12               N/A       N/A        Y       172  
                                                                                
    Task Status of Volume glusterfs-registry-volume                             
    ------------------------------------------------------------------------------
    There are no active volume tasks
    [...]
    

    According to my experience it can happen that sometimes a brick displays a N in the Online column which means that the corresponding brick process wasn’t started successfully. If multiple bricks of the same volume are down, your entire volume is down and must be properly recovered. In such a case don’t continue with the steps below!

  3. Via IP address of the brick, you can figure out which host is affected and then you can simply delete the corresponding pod:
    $ oc delete pod glusterfs-storage-r24wg
    

    The pod will be automatically restarted and the brick processes should come up this time.

Fixing the Hosted Registry Storage Definition

The second run of the control-plane playbook eventually attested that all GlusterFS volumes are healthy but again it failed only two tasks later with a rather cryptic error message, something like:

TASK [openshift_storage_glusterfs : Check for GlusterFS cluster health] **********************************************************************************************************************************************************************
task path: /usr/share/ansible/openshift-ansible/roles/openshift_storage_glusterfs/tasks/cluster_health.yml:4
Using module file /usr/share/ansible/openshift-ansible/roles/lib_utils/library/glusterfs_check_containerized.py
FAILED - RETRYING: Check for GlusterFS cluster health (120 retries left).Result was: {
    "attempts": 1, 
    "changed": false, 
    "invocation": {
        "module_args": {
            "cluster_name": "registry", 
            "exclude_node": "master01.example.com", 
            "oc_bin": "/usr/local/bin/oc", 
            "oc_conf": "/etc/origin/master/admin.kubeconfig", 
            "oc_namespace": "default"
        }

Didn’t it just said, that all the GlusterFS volumes are healthy? What the heck is "cluster_name": "registry" and what is it doing with GlusterFS in the ‘default’ namespace anyway?

The solution for this, I found after digging deep in the openshift_storage_glusterfs Ansible role and reading the CNS installation instructions again and again. I became a victim of my “simplified” CNS setup. The reference installation is meant to have two separate CNS GlusterFS clusters. One exclusively for the hosted registry volume (hinted by the [glusterfs_registry] Ansible host group) and a second cluster for any other persistent volumes (hinted by the [glusterfs] Ansible host group). As mentioned before, I’m limited in available hosts so I added the master hosts to both host groups and set the glusterfs_devices variable to the same device when installing the CNS. That’s already everything that was needed to create the hosted registry volume with OpenShift 3.9 in the “regular” CNS cluster. However the 3.10 playbook expects the registry volume to be in a different project with a different naming. Fortunately all that was needed to fix this were some additional inventory variables in the [OSEv3:vars]:

# Adjust variables for registry storage to match default converged glusterfs storage setup
openshift_storage_glusterfs_registry_name=storage
openshift_storage_glusterfs_registry_namespace=glusterfs

With the updated inventory I started the control-plane upgrade playbook once more. This time it ran for quite a while and even started to do some real stuff. It replaced the docker run command in the ‘origin-node’ systemd service with a runc command using the 3.10 image. Finally some progress. But eventually another error aborted the playbook and again it was a totally unexpected one.

etcd Backup Failure

Before updating the etcd cluster, there is a task which would backup the etcd database and this failed miserably. It couldn’t run docker exec etcd_container etcdctl backup [...]. When executing the command manually on a master host, I received the same error message:

rpc error: code = 2 desc = oci runtime error: exec failed: container_linux.go:247: starting container process caused "process_linux.go:110: decoding init error from pipe caused \"read parent: connection reset by peer\""

My first suspicion was the Ansible role. Maybe the backup command is wrong? But I couldn’t find any radical changes in the commit history regarding the etcd backup and a blocking issue like this is unlikely to stay unnoticed for such a long time. Maybe something wrong with the image? The OpenShift Origin 3.9 setup was using a rather atypical image at that time, the only one from the Fedora image registry (registry.fedoraproject.org/latest/etcd:latest). When I see the ‘latest’ tag being used with containers I’m instantly suspicious that bugs might sneak in unnoticed as different users may get different images depending on when they are pulling them. Maybe they mistakenly pushed an image without a shell or without the etcdctl binary? So I tried to ask in the #openshift IRC channel on Freenode if someone experienced the same issue before but didn’t get any reply. Suddenly I had an idea: Only a few hours before I was using the etcdctl tool from the Atomic host to do my own etcd backup. I just need to find a way to make Ansible use the etcdctl from the host and everything would be fine. So I was digging a bit in the Ansible etcd role and a few minutes later I set r_etcd_common_etcdctl_command to "etcdctl" in my inventory, being confident that this would fix my issue. It won’t, but I won’t find out anytime soon…

The Master API cannot find the LDAP CA Certificate

In the next attempt, the playbook happily ran the etcd backup, upgraded the etcd images, converted them to be a static pod on all masters and did the same for the other two control plane services, the API service and the controller service, starting on the first master. Eventually the ‘origin-master-api’ and ‘origin-master-controller’ services were shutdown and the corresponding pods should be started, so the playbook was waiting for the API service to come up and waited and waited… The pod didn’t come up. Hmpf. Time has already come to use the new debugging command I read about in the release notes to see what’s going on:

# /usr/local/bin/master-logs api api

That is an alternative for the corresponding oc command that I’m also able to run from my client machine:

$ oc logs master-api-node01.example.com -n kube-system

But the latter one was behaving weird. Sometimes it hung although the API services of the two other nodes were still up. There is definitely something wrong.

When checking the logs locally, I saw an error that my FreeIPA CA certificate which should be used to validate the LDAPS connections cannot be found. That’s strange. I explicitly configured the ca key in the openshift_master_identity_providers variable pointing it to the correct CA certificate. I did this in other OpenShift cluster inventories before and there it was working… But those were not running OpenShift 3.10 or later. With 3.10 the playbook developers removed the possibility to custom-name the CA certificate so the ca key from the inventory was silently ignored. Only after checking the installation instructions regarding Configuring identity providers with Ansible, I found an inconspicuous comment that the CA certificate destination path now follows a given naming convention. When I was adding the identity provider configuration to the inventory before the update, I didn’t specify the openshift_master_openid_ca or openshift_master_openid_ca_file variables which will ensure that the CA certificate is copied to the correct place. After all the certificates were already on the master hosts and the identity provider was working, so I didn’t want the upgrade playbook to touch the certificate. Now that’s the result: My mistake. Still, I like issues that are clear and can be fixed so easily. A quick rename of the certificate on the master hosts made the API service successfully start again.

How a Docker Bug broke the etcd Cluster

All API servers are running again, although only the first one in the final configuration, but the oc command invocation still feel sluggish. Sometimes it even hangs completely. When checking the process list it attracted my attention that the etcd processes are only a few minutes old and sometimes they are not running at all. So I was checking the etcd cluster health and here it is: Two cluster members are down and one is in the state unhealthy. That is bad… Immediately, I started manually triggering etcd restarts. But only a few minutes later they shutdown again. I was checking the log files and there were errors, but I couldn’t figure out a single reason what might cause this mess. Then I found that the /etc/etcd/etcd.conf was updated during the playbook run, so I restored the backup, but again it wouldn’t fix the issue. Eventually I started to accept the thought that I might need to completely restore the etcd database from a backup because the database might already be so corrupt in the mean time that it is not able to find a stable state anymore.

The OKD documentation for Restoring etcd quorum would be the correct guide that you need to follow in this situation, but for a reason I landed at Restoring etcd. That confronted me with yet another issue: This guide was not yet properly updated for OpenShift 3.10. Some parts of the documentation still reference etcd as systemd service. But in my setup it’s a pod. Trying to pass the --force-new-cluster parameter to the etcd process via systemd override obviously doesn’t have any effect. Eventually I found out about the /etc/origin/node/pod/etcd.yaml file which contains the pod definition. And here we are able to correctly pass the parameter so that it is picked up by the pod startup command. But again, even with an empty database, a few minutes later my pod would die again. Something is badly broken here. In the YAML definition I also found the liveness probe. So once the pod was started once more, I tried to execute the liveness probe to see what it returns and the result looked familiar, in a bad way:

rpc error: code = 2 desc = oci runtime error: exec failed: container_linux.go:247: starting container process caused "process_linux.go:110: decoding init error from pipe caused \"read parent: connection reset by peer\""

Ouch! Now I understand why the pods keep restarting. It’s again the same error that also caused the etcd backup failure before. But now I’m using the new quay.io/coreos/etcd:v3.2.22 image. This disproved my theory that a buggy image might be the reason. For the moment, I ran out of ideas… Until I remembered that I recently read a post about a docker bug (#1655214) that affected CentOS 7. Thanks for that! After checking the docker version on Atomic host 7.1811 (docker-1.13.1-84.git07f3374.el7.centos) it’s confirmed. That’s the root cause for so much trouble so far.

Updating Docker on Atomic Host

I didn’t need to dig too much into Atomic Host so far, as most of the stuff was simply working or was easily fixed with an update in the past. But this time it didn’t look that there is an imminent update. Release 7.1811 was only a few days old. I could roll back, but the previous version is 7.1808. That’s three months back and somehow defeats the purpose of my update, to get the latest security fixes. Fortunately CentOS already released new docker packages where this bug is fixed. Now I just need to find a way to update the docker packages independently from the ostree image? This time the documentation gods were on my side. I quickly found Dusty Mabe’s Atomic Host 101 Lab Part 4: Package Layering, Experimental Features.

Here my guide for quickly working around Bug #1655214 by updating the docker packages to release 1.13.1-88.git07f3374.el7.centos on CentOS Atomic Host:

  1. Create a temporary directory and download the corresponding RPM packages from a mirror of your choice:
    # mkdir /tmp/docker-1.13.1-88
    # cd /tmp/docker-1.13.1-88
    # for pkg in docker docker-client docker-common docker-lvm-plugin docker-novolume-plugin ; do \
        curl -O https://mirror.init7.net/centos/7/extras/x86_64/Packages/$pkg-1.13.1-88.git07f3374.el7.centos.x86_64.rpm ; \
      done
    
  2. From within the directory run rpm-ostree override replace to replace the docker packages from the ostree layer with the new RPMs:
    # rpm-ostree override replace docker*
    Checking out tree ee5a6f2... done
    Inactive requests:
      docker (already provided by docker-2:1.13.1-84.git07f3374.el7.centos.x86_64)
    Enabled rpm-md repositories: base updates extras
    Updating metadata for 'base': [=============] 100%
    rpm-md repo 'base'; generated: 2018-11-25 16:00:34
    Updating metadata for 'updates': [=============] 100%
    rpm-md repo 'updates'; generated: 2018-12-10 15:34:27
    Updating metadata for 'extras': [=============] 100%
    rpm-md repo 'extras'; generated: 2018-12-10 16:00:03
    Importing metadata [=============] 100%
    Resolving dependencies... done
    Relabeling (5/5) [=============] 100%
    Applying 5 overrides
    Processing packages (10/10) [=============] 100%
    Running pre scripts... 1 done
    Running post scripts... 5 done
    Writing rpmdb... done
    Writing OSTree commit... done
    Copying /etc changes: 42 modified, 5 removed, 613 added
    Transaction complete; bootconfig swap: no; deployment count change: 0
    Freed: 580.5 kB (pkgcache branches: 0)
    Upgraded:
      docker 2:1.13.1-84.git07f3374.el7.centos -> 2:1.13.1-88.git07f3374.el7.centos
      docker-client 2:1.13.1-84.git07f3374.el7.centos -> 2:1.13.1-88.git07f3374.el7.centos
      docker-common 2:1.13.1-84.git07f3374.el7.centos -> 2:1.13.1-88.git07f3374.el7.centos
      docker-lvm-plugin 2:1.13.1-84.git07f3374.el7.centos -> 2:1.13.1-88.git07f3374.el7.centos
      docker-novolume-plugin 2:1.13.1-84.git07f3374.el7.centos -> 2:1.13.1-88.git07f3374.el7.centos
    Run "systemctl reboot" to start a reboot
    
  3. Reboot the host.

I carefully did this on one master server after the other and surprisingly all the services (except etcd) were started normally. Even my GlusterFS pods came up again as nothing had happened. But still, the etcd cluster was offline and with it the Master API was inaccessible. No oc commands were possible.

Fixing the etcd Cluster

With the docker issue being fixed, I now had to bring up the etcd cluster again. The database was likely in a confused state because of all the failed attempts before, so I decided to restore a known good state. As briefly mentioned before, to do so, you actually create a new cluster with the database from a backup. Because the OpenShift documentation on how to do this cannot be followed easily, I list the exact steps below how I manged to do it:

  1. Make sure that all the etcd processes are down and not coming up again automatically. On an OpenShift 3.10 cluster, you prevent automatic startup by moving the /etc/origin/node/pod/etcd.yaml definition to a backup location e.g. /etc/origin/node/pod/disabled/ on every etcd host.
  2. First create a new one node etcd cluster on the first etcd host. To do so, we need some preparation:
    • The /etc/etcd/etcd.conf configuration must not contain any previous configurations regarding the INITIAL_CLUSTER or INITIAL_CLUSTER_STATE. I was able to simply use the etcd.conf generated by the upgrade playbook which already set those two variables to the correct value:
      ETCD_INITIAL_CLUSTER=
      ETCD_INITIAL_CLUSTER_STATE=new
      

      Also make sure, the ETCD_INITIAL_ADVERTISE_PEER_URLS only contains the URL of the first host itself and no other peers:

      ETCD_INITIAL_ADVERTISE_PEER_URLS=https://10.0.0.10:2380
      
    • Restore the etcd database from a backup. Fortunately the upgrade playbook automatically created a backup after the etcd upgrade, so I’m going to restore to that state:
      # mv /var/lib/etcd/member /var/lib/etcd/member.orig
      # cp -rP /var/lib/etcd/openshift-backup-post-3.0-20181214022846/member /var/lib/etcd/
      
    • When starting the first etcd member for the first time, we need to pass the --force-new-cluster argument to the process. This will override the cluster definition from the database files. To do so, the etcd.yaml file has to be adjusted. Here the important snippet (everything else should be kept as it is):
      spec:
        containers:
        - args:
          - '#!/bin/sh
      
            set -o allexport
      
            source /etc/etcd/etcd.conf
      
            exec etcd --force-new-cluster
      
            '
      
    • If everything is ready to start the etcd process, move the altered etcd.yaml file back to the /etc/origin/node/pod directory. Within a few moments, the pod should startup and create a new cluster.
  3. Check the initial cluster state via:
    # etcdctl2 cluster-health
    member 67aa8b8cc701 is healthy: got healthy result from https://10.0.0.10:2379
    

    If something went wrong, you might want to check the logs via:

    # /usr/local/bin/master-logs etcd etcd
    
  4. Initially the first member still advertises a PeerURL pointing to ‘localhost’:
    # etcdctl2 member list
    67aa8b8cc701: name=master01.example.com peerURLs=http://localhost:2380 clientURLs=https://10.0.0.10:2379 isLeader=true
    

    This must be updated by the correct host URL pointing to itself:

    # etcdctl2 member update 67aa8b8cc701 https://10.0.0.10:2380
    Updated member with ID 67aa8b8cc701 in cluster
    

    Then it correctly shows:

    # etcdctl2 member list
    67aa8b8cc701: name=master01.example.com peerURLs=https://10.0.0.10:2380 clientURLs=https://10.0.0.10:2379 isLeader=true
    
  5. This configuration was automatically saved in the database. So the --force-new-cluster argument can be removed again. Edit the etcd.yaml in-place to restore the original configuration. After doing so, restart the etcd process with:
    # /usr/local/bin/master-restart etcd
    

    If it comes up again and shows healthy, we can continue the add the other two cluster members.

  6. The following steps to add another cluster member obviously have to be done on for both other etcd hosts:
    1. Add the new host to the cluster by executing the following command on the first etcd host:
      # etcdctl2 member add master02.example.com https://10.0.0.11:2380
      Added member name master02.example.com with ID a6b2e8d0d392083b to cluster
      
      ETCD_NAME="member02.example.com"
      ETCD_INITIAL_CLUSTER="member01.example.com=https://10.0.0.10:2380,member02.example.com=https://10.0.0.11:2380"
      ETCD_INITIAL_CLUSTER_STATE="existing"
      

      The new member will then be displayed as ‘unstarted’ in the member list.

    2. Prepare the /etc/etcd/etcd.conf file on the new etcd host by defining the variables as shown in the output of the etcdctl2 member add command above. The ETCD_INITIAL_CLUSTER value will automatically be extended with each new member added to the cluster.
    3. Delete the old database on the new etcd host. It will automatically be synced from the other cluster members once the new node has joined:
      # mv /var/lib/etcd/member /var/lib/etcd/member.orig
      
    4. Enable the etcd pod by moving the etcd.yaml back to /etc/origin/node/pod. Within a few minutes the etcd process should be started and eventually join the etcd cluster.

Once the etcd cluster was restored, the oc command was finally working again and I could check the state of the etcd pods also via OpenShift client:

$ oc get pods -n kube-system | grep etcd
master-etcd-master01.example.com          1/1       Running   5          1h
master-etcd-master02.example.com          1/1       Running   0          47m
master-etcd-master03.example.com          1/1       Running   0          2m

During the entire time the etcd cluster was down the OpenShift cluster continued running. The registry, routers and applications such as my Gitea setup were online all the time and even the CNS cluster running on the master hosts handled the debugging and restart session with bravery. Fortunately I had a super static setup during that time and so no deployments or replica count enforcement needed to be executed which would have been impossible anyway. Still I feel it’s a positive fact that shows the resiliency the platform has gained over time.

Finishing the Control Plane Upgrade

After a longer detour, I was finally back at the point were I could start another run of the control plane upgrade playbook. Remember, when the playbook aborted before it did so after upgrading the control plane services on the first master node, there are still two to go. So I started the playbook once again.

By now I have a really good feeling about the state of the playbook in this release. As you can see above, it failed on me many times in all different stages of the update, but it always had a good reason and it was always able to pick up where it left. My experience with initial upgrade attempts of earlier OpenShift releases was unfortunately not always that good. For example it happened to me that I had to restore a master host from a snapshot, because the playbook failed to correctly detect the upgrade state in the second run, after it aborted the first run due to a syntax error in a post-upgrade task.

This time the playbook finished successfully and my control plane was finally at release 3.10:

# /usr/local/bin/oc version
oc v3.10.0+c99b16a-90
kubernetes v1.10.0+b81c8f8
features: Basic-Auth GSSAPI Kerberos SPNEGO

Server https://openshift.example.com:8443
openshift v3.10.0+c99b16a-90
kubernetes v1.10.0+b81c8f8

Running the Node Upgrade Playbook

After the control-plane was done, I had to upgrade the infrastructure and compute nodes. A separate playbook placed at playbooks/byo/openshift-cluster/upgrades/v3_10/upgrade_nodes.yml is available. Initially I only wanted to run it on a single node to make sure everything works as expected. This can be done by passing the -e openshift_upgrade_nodes_label=kubernetes.io/hostname=node03.example.com argument, where the given host name is obviously the node that should be upgraded, to the playbook execution command. The playbook completed without error already on the first attempt. So I continued with the other nodes.

One fact is super important when upgrading the nodes to OpenShift 3.10. The /etc/origin/node/node-config.yaml is completely regenerated based on the settings in the corresponding node group (and/or the defaults) and so any prior adjustment not reflected in the inventory is lost. Therefore make sure that you perfectly understand the Node Group concept and how it affects your node layout and configuration.

To give you an example how to customize the upgrade behavior on the infrastructure nodes, I added the following arguments to the playbook execution: -e openshift_upgrade_nodes_label=region=infra -e openshift_upgrade_nodes_serial=50%.

Fixing the Infrastructure Node Selector

It confused me that the NodeSelector of the infrastructure components such as the registry and routers were not updated to the new defaults. In the inventory I explicitly defined the new node selector:

openshift_hosted_registry_selector='node-role.kubernetes.io/infra=true'

But when checking the DeploymentConfig of the registry, I can still find the old NodeSelector:

$ oc get dc docker-registry -n default -o json | jq .spec.template.spec.nodeSelector
{
  "region": "infra"
}

So I manually triggered an update of the NodeSelector property in all the DeploymentConfigs using it. E.g.:

$ oc patch dc docker-registry -n default --type json --patch '[{"op":"replace","path":"/spec/template/spec/nodeSelector","value":{"node-role.kubernetes.io/infra":"true"}}]'
deploymentconfig "docker-registry" patched

NodeSelectors can also be set in DaemonSets, as annotations in projects or even globally via master-config.yaml. Therefore make sure to update them all, when required, before removing any labels from the nodes.

After checking that all the pods are up and running again, I was finally able to remove the old infrastructure labels from the nodes:

$ oc label node node01.example.com region- zone-

Summarizing

This was not my first OpenShift update ever, but my first update from 3.9 to 3.10. This obviously means that I made some mistakes and had wrong assumptions from which I did learn a lot. I hope I could share some insights and useful hints for those of you that haven’t done this before. Otherwise it will at least help me in the future to run this update an other cluster much smoother.

At the end some advice for those of you who also need to do such an upgrade:

  • You need to have a test cluster where you can practice such updates. It doesn’t need to be big but the Ansible inventory variables should be structurally as similar as possible to those of the production cluster. As you saw above, a lot of errors just happened due to wrong inventory variables. Ideally the test cluster should have some workload so that you experience how the applications behave during the update and and so that you can test if everything still works after an upgrade.
  • Emphasis your Ansible inventory. Everything of your configuration that can fit into the Ansible inventory must be defined there and must be maintained there. It can cost you a lot of time debugging or even result in application downtime during an upgrade if you manually updated the cluster configuration without adjusting the configuration in the inventory. Even when it sometimes feels like it’s more work than benefit it’s always worth it.
  • Preparation is key. Carefully read through the upstream documentation available. Most likely you also have some internal documentation where your infrastructure specifics are written down. Run the upgrade on a test cluster before you do it in production. If it doesn’t work on the first attempt, update your notes and try it again. Try to gain as much experience as possible on the test infrastructure so that you already know what to do if something goes wrong in production.
  • Plan a lot of time. Doing such an upgrade is a lot of work! Give yourself enough time to do a proper preparation and also the actual upgrade window itself should give you enough time to fix issues when they arise. Plan in the scale of hours or better days. Ansible is slow. If you have to restart the playbook because of an error after 15 minutes this will eat up your time fast.

Thanks for reading. As always I’d welcome some feedback or critics in the comments.

Dec 122018
 

The first thing that someone would need when operating or playing around with OKD (better known as OpenShift) is a git version control service. Personally I’m a fan of Gitea and that’s why I’d like to show a way how to run Gitea in a OpenShift environment. Gitea upstream already provides a great container image which I’m are going to use. But as some of you may have already experienced, running an image on docker and running it in OpenShift is two different pairs of shoes. The fact that the Gitea image runs an integrated SSH server means that it doesn’t simply match the widely discussed Web application pattern. Therefore I’m trying to explain some of the difficulties that one might encounter when moving such an application to OpenShift.

My environment consists of a multi-node OpenShift cluster. Obviously Gitea should be high available so that if a node goes down, one would still be able to access the git repositories. One pod is no pod, so Gitea must be deployed with a replica count of at least two. Accessing the pods over HTTP is already solved by the OpenShift default infrastructure via redundant HAProxy routers. I’d probably explain how to achieve a redundant router setup in one of my next blog posts but this time I’d like to emphasis on the Gitea SSH access via NodePort service feature. The following graphic shows a communication overview of such a setup:

NodePort Service

In OpenShift the Kubernetes Service resource is responsible for directing the traffic (TCP, UDP or SCTP) to the individual application pods. It maps the service name (e.g. ‘gitea’) via SkyDNS to a so-called ClusterIP. This is a virtual IP address that is not assigned to any host or container network interface but still used as packet destination within the cluster SDN (software defined network). After receiving a packet to this so-called ClusterIP the Linux kernel of an OpenShift node rewrites the packet destination to an IP address of an actual application pod and acts as a virtual network load-balancer.

In our example there is a ‘gitea’ service managing the HTTP traffic to port 3000 of the Gitea pod and a ‘gitea-ssh’ service managing the SSH traffic to port 22 of the Gitea pod. Because we can’t use the OpenShift Router as ingress for SSH, the ‘gitea-ssh’ service defines a special type called NodePort. This means that a packet sent to this port (e.g. 30022) on any OpenShift node will be received by the corresponding service and therefore forwarded to a Gitea pod. This is the simplest way how to direct non-HTTP traffic from outside of OpenShift to an application pod and can also be used for e.g. database protocols or Java RMI. Here the corresponding resource definition for the Gitea SSH service:

apiVersion: v1
kind: Service
metadata:
  name: gitea-ssh
spec:
  ports:
    - name: ssh
      nodePort: 30022
      port: 22
      protocol: TCP
      targetPort: 22
  selector:
    app: gitea
    deploymentconfig: gitea
  sessionAffinity: ClientIP
  type: NodePort

The sessionAffinity: ClientIP setting defines “sticky sessions” to avoid distributing multiple requests of the same client to different pods. I didn’t test yet, how SSH would behave without it, but I think it generally makes sense. In a running setup the service additionally shows the discussed ClusterIP which is statically assigned and the endpoints (pod IPs) which may change when pods are started and stopped:

$ oc describe service gitea-ssh
Name:                     gitea-ssh
Namespace:                vcs
Labels:                   app=gitea
                          template=gitea-persistent-template
Annotations:              
Selector:                 app=gitea,deploymentconfig=gitea
Type:                     NodePort
IP:                       172.30.8.9
Port:                     ssh  22/TCP
TargetPort:               22/TCP
NodePort:                 ssh  30022/TCP
Endpoints:                10.129.2.44:22,10.130.3.71:22
Session Affinity:         ClientIP
External Traffic Policy:  Cluster
Events:                   

From within the cluster, the Gitea SSH service can be reached via service name (extended with OpenShift project name, here ‘vcs’) DNS entry or directly via ClusterIP:

$ host gitea-ssh.vcs.svc
gitea-ssh.vcs.svc has address 172.30.8.9

$ ssh git@gitea-ssh.vcs.svc
PTY allocation request failed on channel 0
Hi there, You've successfully authenticated, but Gitea does not provide shell access.
If this is unexpected, please log in with password and setup Gitea under another user.
Connection to gitea-ssh.vcs.svc closed.

From outside the cluster, the Gitea SSH service can be reached via NodePort on any OpenShift node. To avoid a dependency on a single node in the git repository URL you can define multiple DNS entries using the same name (e.g. services.example.com) to all OpenShift node addresses:

$ host services.example.com
services.example.com has address 10.0.0.2
services.example.com has address 10.0.0.3

$ ssh -p 30022 git@services.example.com
PTY allocation request failed on channel 0
Hi there, You've successfully authenticated, but Gitea does not provide shell access.
If this is unexpected, please log in with password and setup Gitea under another user.
Connection to services.example.com closed.

Issues with NodePort

Port Assignment
The NodePort mechanism is allocating the corresponding port on each OpenShift node. To avoid a clash with node services such as the DNS resolver or the OpenShift node service the port range is restricted. It can be configured in the /etc/origin/master/master-config.yml with the option servicesNodePortRange and defaults to 30000-32767. Obviously multiple applications in the same cluster cannot use the same port and traffic to the chosen port must be allowed by the host firewall on the OpenShift nodes.

Node Groups
NodePorts are always allocated on every OpenShift cluster host running the node service which also includes the OpenShift master servers. OpenShift doesn’t provide a way to restrict the involved hosts to a subset. In my example I choose to restrict the hosts receiving traffic by only adding a limited number of nodes to the service DNS entry and block access on the others via iptables. If you don’t use an application load-balancer in front of the OpenShift routers you could also re-use the wildcard DNS entry defined for the HTTP traffic. The NodePort traffic would then follow the same path as the normal Web traffic.

Node Failure
If one OpenShift node goes down a client trying to access the Gitea SSH service might still try to connect to the unreachable host. Fortunately, the default SSH implementation used by the git command line client is quite tolerant and simply retries with another IP address. When testing this case I therefore haven’t experienced a major issue except a slight connection delay. The failure behavior might be different for other git clients or other application protocols altogether and is definitely not ideal but simple instead.

One way to improve this failure scenario would be to add a real TCP load-balancer in front of the NodePort but then there would be another piece of infrastructure that must be managed synchronously with the OpenShift infrastructure and which might be a new single point of failure.

Container with root Permissions

When starting the upstream Gitea container image in OpenShift you might likely encounter a startup failure with the following error message in the log:

s6-svscan: fatal: unable to mkfifo .s6-svscan/control: Permission denied

The Gitea image, and many other docker images not optimized for the pod concept introduced by Kubernetes, doesn’t start a single application process but a supervisor process (in this case s6) which then spawns multiple different application processes defined in /etc/s6. To do so it wants to create a FIFO in the /etc/s6/.s6-svscan directory which is only writable by the root user which fails as by default processes are started with a random unprivileged account.

Security Context Constraints

Unlike docker, OpenShift controls the actions a pod can do and access with a tight set of rules called Security Context Constraints (SCC). By default the ‘default’ ServiceAccount used to run the application pods is a member of the ‘restricted’ SCC which among other things defines the previously mentioned randomized UID. As Gitea won’t work like this, a less restrictive SCC must be used. After reading the documentation we find, that there is already a predefined SCC which grants us just enough permissions to start our container process as root user without weakening too many other restrictions. The SCC we are heading for is ‘anyuid’. Below I’ll present different approaches how this SCC can now be assigned to the Gitea deployment:

  • The OpenShift cluster administrator can add the ‘default’ ServiceAccount of a project to the list of users in the SCC definition. This doesn’t need any special configuration in the DeploymentConfig of the application but also grants every deployment in the corresponding project ‘anyuid’ privileges. In our setup this would be done with the following command, when assuming Gitea should be deployed in the ‘vcs’ project:
    $ oc adm policy add-scc-to-user anyuid system:serviceaccount:vcs:default
    

    I’m not in favor of this approach as it “hides” the additional permissions in the default ServiceAccount and prone to break the principle of least privilege by assigning the SCC to potentially more applications than necessary.

  • Another approach is using a dedicated ServiceAccount for the Gitea deployment and only adding that to the ‘anyuid’ SCC. The project owner can create a ServiceAccount with:
    $ oc create serviceaccount gitea
    

    The cluster administrator then has to add it to the SCC as before:

    $ oc adm policy add-scc-to-user anyuid system:serviceaccount:vcs:gitea
    

    In the DeploymentConfig the ServiceAccount must be referenced with a entry under the spec.template.spec key:

    $ oc patch dc/gitea --patch '{"spec":{"template":{"spec":{"serviceAccountName": "gitea"}}}}'
    

    The dedicated ServiceAccount used in this approach already hints that there might be special privileges connected to it and is in my opinion easier to audit. The disadvantage however is the more complex configuration.

  • Instead of adding every user account individually to the SCC a dedicated user group could be created having the SCC assigned to this group. Individual ServiceAcccountss would then be added to the group and therefore inherit the SCC. This would follow common practice in identity management to assign permissions to users via privilege groups. Additionally a group management role could be created which then would permit dedicated users not having the ‘cluster-admin’ privilege to manage the group membership.
  • Unfortunately I couldn’t figure out a true self-service model where a responsible project admin could expand the necessary permissions without the possibility to interfere with other projects. In the documentation of OpenShift (<=3.7) I found a hint that it is/was(?) possible to extend the default ServiceAccounts available after creating a new project by adding the account name (e.g. ‘anyuid-service-account’) to the serviceAccountConfig.managedNames list in /etc/origin/master/master-config.yml. While this configuration is still present in newer master-config.yml, the documentation is gone and I also didn’t find a way how to automatically add a user created like this to the ‘anyuid’ SCC. Maybe it’s possible by somehow modifying the project template. If you have done this before or at least have an idea how this could be done, please drop me a line.

At the end, the way how the ‘anyuid’ SCC is assigned to the Gitea application is unimportant as long as the application pod is allowed to start the s6 supervisor process with root permissions.

Gitea Application Template

The way how OpenShift administrators can provide an application setup ready for instantiation by OpenShift project owners is through Templates. Inspired by the My journey through Openshift blog post, I wanted to create my own Gitea template fixing some issues found in the original template and extending it with the opinionated configuration presented above. You can download it from here.

The template is able to automatically setup Gitea with exception of the ‘anyuid’ SCC configuration. It requires a persistent volume (PV) for storing the git repositories and some static configurations such as the SSH authorized_keys file. By default it will use a SQLite database backend which is also stored in the PV. Optionally you can also give the connection string and credentials of a PostgreSQL or MariaDB backend which can run on OpenShift or externally.

If you want the template to be available in the Service Catalog the YAML file has to be applied to the ‘openshift’ project by a cluster administrator:

$ oc create -f gitea-persistent-template.yaml -n openshift

Afterwards it can be instantiated by any project admin via Service Catalog Web-UI or from the command line via:

$ oc new-project vcs
$ oc new-app --template=gitea-persistent -p HTTP_DOMAIN=git.example.com -p SSH_DOMAIN=services.example.com

Alternatively, if no Service Catalog is available, or the template shouldn’t be loaded to OpenShift, the application can also be created directly from the YAML file via:

$ oc new-app -f gitea-persistent-template.yaml -p HTTP_DOMAIN=git.example.com -p SSH_DOMAIN=services.example.com

IMPORTANT: The template will configure Gitea to use a ServiceAccount named according to the parameter APPLICATION_NAME (defaults to ‘gitea’). It must be added to the ‘anyuid’ SCC as described above. E.g.:

$ oc adm policy add-scc-to-user anyuid system:serviceaccount:vcs:gitea

If you have some feedback regarding the template or troubles using it, please open a Github issue. Comments, corrections or general feedback to my article can be posted below. Thanks for reading.