Category: Kubernetes

K8s, resolv.conf, and ndots

I had a very strange problem when firewalld was used with nftables as the back end – rules configured properly in firewalld didn’t exist in the nftables rulesets so … didn’t exist. The most obvious failure in the k8s cluster was DNS resolution – requests to any nodes where nftables was the back end just timed out. In diagnosing the “dns queries time out” issue, I was watching the logs from the coredns pods. And I saw a lot of NXDOMAIN errors. Not because I had a hostname mistyped or anything – each pod was appending every domain in the resolv.conf search order before trying the actual hostname.

Quick solution was to update our hostnames to include the trailing dot for the root zone. It is not redishost.example.com but rather redishost.example.com.

But that didn’t explain why – I’ve got plenty of Linux boxes where there are some search domains in resolv.conf. Never once seen redishost.example.com.example.com come across the query log. There is a configuration that I’ve rarely used that is designed to speed up getting to the search list. You can configure ndots – the default is one, but you can set whatever positive integer you would like. Surely, they wouldn’t set ndots to something crazy high … right??

Oh, look –

Defaulted container "kafka-streams-app" out of: kafka-streams-app, filebeat
bash-4.4# cat /etc/resolv.conf
search kstreams.svc.cluster.local svc.cluster.local cluster.local mgmt.example.net dsys.example.net dnoc.example.net admin.example.net example.com
nameserver 10.6.0.5
options ndots:5

Yup, it’s right there in the source — and it’s been there for seven years:

What does this mean? Well, ndots is really just the number of dots in a hostname. If there are fewer than ndots dots, the resolver will try appending the search domains first and then try what you typed as a last resort. With one dot, that basically means a string with no dots will get the search domains appended. I guess if you go out and register a gTLD for your company – my hostname is literally just example. – then you’ll have a little inefficiency as the search domains are tried. But that’s a really edge case. With the k8s default, anything with fewer than five dots gets all of those search domains appended first.

So I need redishost.example.com? I see the following resolutions fail because there is no such hostname:

[INFO] 64.24.29.155:57014 - # "A IN redishost.example.com.svc.cluster.local. udp # false 512" NXDOMAIN qr,aa,rd 158 0.00019419s
[INFO] 64.24.29.155:57028 - # "AAAA IN redishost.example.com.svc.cluster.local. udp # false 512" NXDOMAIN qr,aa,rd 158 0.00019419s
[INFO] 64.24.29.155:56096 - # "A IN redishost.example.com.cluster.local. udp # false 512" NXDOMAIN qr,aa,rd 158 0.00019419s
[INFO] 64.24.29.155:56193 - # "AAAA IN redishost.example.com.cluster.local. udp # false 512" NXDOMAIN qr,aa,rd 158 0.00019419s
[INFO] 64.24.29.155:55001 - # "A IN redishost.example.com.mgmt.example.net. udp # false 512" NXDOMAIN qr,aa,rd 158 0.00019419s
[INFO] 64.24.29.155:55194 - # "AAAA IN redishost.example.com.mgmt.example.net. udp # false 512" NXDOMAIN qr,aa,rd 158 0.00019419s
[INFO] 64.24.29.155:54078 - # "A IN redishost.example.com.dsys.example.net. udp # false 512" NXDOMAIN qr,aa,rd 158 0.00019419s
[INFO] 64.24.29.155:54127 - # "AAAA IN redishost.example.com.dsys.example.net. udp # false 512" NXDOMAIN qr,aa,rd 158 0.00019419s
[INFO] 64.24.29.155:52061 - # "A IN redishost.example.com.dnoc.example.net. udp # false 512" NXDOMAIN qr,aa,rd 158 0.00019419s
[INFO] 64.24.29.155:52182 - # "AAAA IN redishost.example.com.dnoc.example.net. udp # false 512" NXDOMAIN qr,aa,rd 158 0.00019419s
[INFO] 64.24.29.155:51018 - # "A IN redishost.example.com.admin.example.net. udp # false 512" NXDOMAIN qr,aa,rd 158 0.00019419s
[INFO] 64.24.29.155:51104 - # "AAAA IN redishost.example.com.admin.example.net. udp # false 512" NXDOMAIN qr,aa,rd 158 0.00019419s
[INFO] 64.24.29.155:50052 - # "A IN redishost.example.com.example.com. udp # false 512" NXDOMAIN qr,aa,rd 158 0.00019419s
[INFO] 64.24.29.155:50189 - # "AAAA IN redishost.example.com.example.com. udp # false 512" NXDOMAIN qr,aa,rd 158 0.00019419s

Wonderful — IPv6 is enabled and it’s trying AAAA records too. Finally it resolves redishost.example.com!

Luckily, there is a quick solution. Update the deployment YAML to include a custom ndots value – I like 1. I could see where someone might want two – something.else where I need svc.cluster.local appended, maybe I don’t want to waste time looking up something.else … I don’t want to do that. But I could see why something higher than one might be desirable in k8s. Not sure I buy it’s awesome enough to be the default, though!

Redeployed and instantly cut the DNS traffic by about 90% — and reduced application latency as each DNS call no longer has to have fourteen failures before the final success.

Kubernetes / Containerd Image Pull Failure

We are in the process of moving our k8s environment from CentOS 7 to RHEL 8.8 hosts — which means the version of pretty much everything involved is being updated. All of the pods that use images from an internal registry fail to load. At first, we were thinking DNS resolution … but the test pods we spun up all resolved names appropriately. 

2023-09-13 13:48:34 [root@k8s ~/]# kubectl describe pod data-sync-app-deployment-78d58f7cd4-4mlsb -n streams
Events:
  Type     Reason            Age                 From               Message
  ----     ------            ----                ----               -------
  Normal   Scheduled         15m                 default-scheduler  Successfully assigned kstreams/data-sync-app-deployment-78d58f7cd4-4mlsb to ltrkarkvm1593-uos
  Normal   Pulled            15m                 kubelet            Container image "docker.elastic.co/beats/filebeat:7.9.1" already present on machine
  Normal   Created           15m                 kubelet            Created container filebeat
  Normal   Started           15m                 kubelet            Started container filebeat
  Normal   BackOff           15m (x3 over 15m)   kubelet            Back-off pulling image "imageregistry.example.net:5000/myapp/app_uat"
  Warning  Failed            15m (x3 over 15m)   kubelet            Error: ImagePullBackOff
  Normal   Pulling           14m (x3 over 15m)   kubelet            Pulling image "imageregistry.example.net:5000/myapp/app_uat"
  Warning  Failed            14m (x3 over 15m)   kubelet            Failed to pull image "imageregistry.example.net:5000/myapp/app_uat": rpc error: code = Unknown desc = failed to pull and unpack image "imageregistry.example.net:5000/myapp/app_uat:latest": failed to resolve reference "imageregistry.example.net:5000/npm/app_uat:latest": get registry endpoints: parse endpoint url: parse " http://imageregistry.example.net:5000": first path segment in URL cannot contain colon
  Warning  Failed            14m (x3 over 15m)   kubelet            Error: ErrImagePull
  Warning  DNSConfigForming  31s (x73 over 15m)  kubelet            Search Line limits were exceeded, some search paths have been omitted, the applied search line is: kstreams.s            vc.cluster.local svc.cluster.local cluster.local mgmt.windstream.net dsys.windstream.net dnoc.windstream.net

I have found “first path segment in URL cannot contain colon” in reference to Go — and some previous versions at that. There were all sorts of suggestions for working around the issue — escaping the colon, starting with “//”, adding single or double quotes around the string, downgrading to a version of Go not impacted by the problem. Nothing worked.

A few hours with no progress, I thought some time investigating “how can I work around this?” was in order. Kubernetes is using containerd … so it should be feasible to pre-stage the image in containerd and then set our imagePullPolicy values to IfNotPresent or Never

To pre-seed the images in containerd so that they are available for kubernetes run:

ctr -n=k8s.io image pull -u $REGISTRYUSER:$REGISTRYPASSWORD --plain-http imageregistry.example.net:5000/myapp/app_uat:latest

This must be run on every k8s worker in the environment — if a pod tries to spin up on server2 but you’ve only seeded the image file on server1 … the pod will fail to load. We need to update this staged image every time we make changes to the application. Better than not using the new servers, so that’ll just be the process for a while.

Ultimately, the problem ended up being that a few of the workers had a leading space in the TOML file for the repo — how that got there, I have no idea. But once there was no longer extraneous white-space, we could deploy the pods without issue. Now that it’s working “as designed”, we deleted the pre-seeded image using:

ctr -n=k8s.io images rm ImageNameHere

K8s 1.24.12 Upgrade

Trying to upgrade our dev Kubernetes environment to 1.24.12 … and we encountered what seems to be a fairly common error — unknown service runtime.v1alpha2.RuntimeService

kubeserver:~ # kubeadm init
I0323 13:53:26.492921   55320 version.go:256] remote version is much newer: v1.26.3; falling back to: stable-1.24
[init] Using Kubernetes version: v1.24.12
[preflight] Running pre-flight checks
        [WARNING Firewalld]: firewalld is active, please ensure ports [6443 10250] are open or your cluster may not function correctly
error execution phase preflight: [preflight] Some fatal errors occurred:
        [ERROR FileAvailable--etc-kubernetes-manifests-kube-apiserver.yaml]: /etc/kubernetes/manifests/kube-apiserver.yaml already exists
        [ERROR FileAvailable--etc-kubernetes-manifests-kube-controller-manager.yaml]: /etc/kubernetes/manifests/kube-controller-manager.yaml already exists
        [ERROR FileAvailable--etc-kubernetes-manifests-kube-scheduler.yaml]: /etc/kubernetes/manifests/kube-scheduler.yaml already exists
        [ERROR FileAvailable--etc-kubernetes-manifests-etcd.yaml]: /etc/kubernetes/manifests/etcd.yaml already exists
        [ERROR CRI]: container runtime is not running: output: E0323 13:53:26.741684   55340 remote_runtime.go:948] "Status from runtime service failed" err="rpc error: code = Unimplemented desc = unknown service runtime.v1alpha2.RuntimeService"
time="2023-03-23T13:53:26-05:00" level=fatal msg="getting status of runtime: rpc error: code = Unimplemented desc = unknown service runtime.v1alpha2.RuntimeService"
, error: exit status 1
        [ERROR DirAvailable--var-lib-etcd]: /var/lib/etcd is not empty
[preflight] If you know what you are doing, you can make a check non-fatal with `--ignore-preflight-errors=...`
To see the stack trace of this error execute with --v=5 or higher

We found a lot of people online with the same issue who (1) removed the config.toml and tried again, (2) changed the SystemdCGroup setting in the config, or uninstalled and reinstalled some/all of the components until it worked. Unfortunately, removing or modifying the config didn’t help. And removing and reinstalling everything wasn’t particularly appealing. However, we noticed that the same error was reported directly from containerd:

kubeserver:~ # crictl ps
E0323 13:53:07.061777   55228 remote_runtime.go:557] "ListContainers with filter from runtime service failed" err="rpc error: code = Unimplemented desc = unknown service runtime.v1alpha2.RuntimeService" filter="&ContainerFilter{Id:,State:&ContainerStateValue{State:CONTAINER_RUNNING,},PodSandboxId:,LabelSelector:map[string]string{},}"
FATA[0000] listing containers: rpc error: code = Unimplemented desc = unknown service runtime.v1alpha2.RuntimeService

Looking at the plugins, there were some in an error state

kubeserver:~ # ctr plugins ls
TYPE                                  ID                       PLATFORMS      STATUS
io.containerd.content.v1              content                  -              ok
io.containerd.snapshotter.v1          aufs                     linux/amd64    skip
io.containerd.snapshotter.v1          btrfs                    linux/amd64    skip
io.containerd.snapshotter.v1          devmapper                linux/amd64    error
io.containerd.snapshotter.v1          native                   linux/amd64    ok
io.containerd.snapshotter.v1          overlayfs                linux/amd64    error
io.containerd.snapshotter.v1          zfs                      linux/amd64    skip

So … it seemed reasonable to look for errors in the messages log from containerd. And, yeah, we had all sorts of errors. Including a rather scary one about reformatting the file system!

Mar 23 13:24:51 kubeserver containerd: time="2023-03-23T13:24:51.726984260-05:00" level=warning msg="failed to load plugin io.containerd.snapshotter.v1.overlayfs" error="/var/lib/containerd/io.containerd.snapshotter.v1.overlayfs does not support d_type. If the backing filesystem is xfs, please reformat with ftype=1 to enable d_type support"

That would do it — we have a dedicated partition for the k8s stuff … and that volume is formatted the right way — xfs_info confirmed ftype=1

kubeserver:~ # xfs_info /kubernetes/
meta-data=/dev/mapper/kubernetes-kubernetes isize=512    agcount=4, agsize=131071744 blks
         =                       sectsz=512   attr=2, projid32bit=1
         =                       crc=1        finobt=0 spinodes=0
data     =                       bsize=4096   blocks=524286976, imaxpct=5
         =                       sunit=0      swidth=0 blks
naming   =version 2              bsize=4096   ascii-ci=0 ftype=1
log      =internal               bsize=4096   blocks=255999, version=2
         =                       sectsz=512   sunit=0 blks, lazy-count=1
realtime =none                   extsz=4096   blocks=0, rtextents=0

However containerd doesn’t really know anything about this volume, does it? The default location that containerd wants to use isn’t set up to support d_type. Editing /etc/containerd/config.toml, root now tells containerd to use our special partition for ‘stuff’ …

And we were able to run kubeadm init without error. Everything came up as it should have, and our k8s server was upgraded!

Reverse engineering Kubernetes YAML’s

Ideally, the definitions for Kubernetes objects are all safely stored in your code repository — you can easily revert back to the previous, working iteration, you can see who changed what, and you’ve got a copy of it all available if super electromagnet man takes a stroll through the data center. Ideally.

Here, in the real world, we took over management of a k8s implementation that’s been in service for about a year now. And, fortunately, the production YAMLs are all in the repo. The development system, on the other hand, isn’t. Logic dictates that the config would be similar, but it’s always good to check.

I wrote a quick script to dump YAML files for all of the configmaps, cron jobs, deployments, horizontal pod autoscaling, jobs, persistent volumes, persistent volume claims, secrets, service accounts, services, and stateful sets. 

#!/bin/bash
nsbase="namespace/"
for ns in $(kubectl get namespaces -o=name)
do
        ns=${ns#${nsbase}}
        for n in $(kubectl get --namespace=$ns -o=name configmap,cronjob,deployment,hpa,job,pv,pvc,secret,serviceaccount,service,statefulset)
        do
            yamlfile="${ns}/${n}.yaml"
            mkdir -p $(dirname ${ns}/${n})
            kubectl get --namespace=$ns -o=yaml $n > $yamlfile
        done
done