I have encountered some mysterious network errors in our kubernetes cluster. Although I originally encountered these errors using ingress, there are even more errors when I bypass our load balancer, bypass kube-proxy and bypass nginx-ingress. The most errors are present when going directly to services and straight to the pod IPs. I believe this is because the load balancer and nginx have some better error handling than the raw iptable routing.
To test the error I use apache benchmark from VM on same subnet, any concurrency level, no keep-alive, connect to the pod IP and use a high enough request number to give me time to either scale up or scale down a deployment. Odd thing is it doesn't matter at all which deployment I modify since it always causes the same sets of errors even when its not related to the pod I am modifying. ANY additions or removals of pods will trigger apache benchmark errors. Manual deletions, scaling up/down, auto-scaling all trigger errors. If there are no pod changes while the ab test is running then no errors get reported. Note keep-alive does seem to greatly reduce if not eliminate the errors, but I only tested that a handful of times and never saw an error.
Other than some bizarre iptable conflict, I really don't see how deleting pod A can effect network connections of pod B. Since the errors are brief and go away within seconds it seems more like a brief network outage.
Sample ab test: ab -n 5000 -c 2 https://10.112.0.24/
Errors when using HTTPS:
SSL handshake failed (5).
SSL read failed (5) - closing connectionErrors when using HTTP:
apr_socket_recv: Connection reset by peer (104)
apr_socket_recv: Connection refused (111)Example ab output. I ctl-C after encountering first errors:
$ ab -n 5000 -c 2 https://10.112.0.24/
This is ApacheBench, Version 2.3 <$Revision: 1826891 gt;
Copyright 1996 Adam Twiss, Zeus Technology Ltd, http://www.zeustech.net/
Licensed to The Apache Software Foundation, http://www.apache.org/
Benchmarking 10.112.0.24 (be patient)
Completed 500 requests
Completed 1000 requests
SSL read failed (5) - closing connection
Completed 1500 requests
^C
Server Software: nginx
Server Hostname: 10.112.0.24
Server Port: 443
SSL/TLS Protocol: TLSv1.2,ECDHE-RSA-AES256-GCM-SHA384,2048,256
Document Path: /
Document Length: 2575 bytes
Concurrency Level: 2
Time taken for tests: 21.670 seconds
Complete requests: 1824
Failed requests: 2
(Connect: 0, Receive: 0, Length: 1, Exceptions: 1)
Total transferred: 5142683 bytes
HTML transferred: 4694225 bytes
Requests per second: 84.17 [#/sec] (mean)
Time per request: 23.761 [ms] (mean)
Time per request: 11.881 [ms] (mean, across all concurrent requests)
Transfer rate: 231.75 [Kbytes/sec] received
Connection Times (ms)
min mean[+/-sd] median max
Connect: 5 15 9.8 12 82
Processing: 1 9 9.0 6 130
Waiting: 0 8 8.9 6 129
Total: 7 23 14.4 19 142
Percentage of the requests served within a certain time (ms)
50% 19
66% 24
75% 28
80% 30
90% 40
95% 54
98% 66
99% 79
100% 142 (longest request)Current sysctl settings that may be relevant:
net.netfilter.nf_conntrack_tcp_be_liberal = 1
net.nf_conntrack_max = 131072
net.netfilter.nf_conntrack_buckets = 65536
net.netfilter.nf_conntrack_count = 1280
net.ipv4.ip_local_port_range = 27050 65500I didn't see any conntrack "full" errors. Best I could tell there isn't packet loss. We recently upgraded from 1.14 and didn't notice the issue but I can't say for certain it wasn't there. I believe we will be forced to migrate away from romana soon since it doesn't seem to be maintained anymore and as we upgrade to kube 1.16.x we are encountering problems with it starting up.
I have searched the internet all day today looking for similar problems and the closest one that resembles our problem is https://tech.xing.com/a-reason-for-unexplained-connection-timeouts-on-kubernetes-docker-abd041cf7e02 but I have no idea how to implement the iptable masquerade --random-fully option given we use romana and I read (https://github.com/kubernetes/kubernetes/pull/78547#issuecomment-527578153) that random-fully is the default for linux kernel 5 which we are using. Any ideas?
\====== Update Nov 5, 2019 ======
It has been suggested to test an alternate CNI. I chose calico since we used that in an older Debian based kube cluster. I rebuilt a VM with our most basic Centos 7 template (vSphere) so there is a little baggage coming from our customizations. I can't list everything we customized in our template but the most notable change is the kernel 5 upgrade yum --enablerepo=elrepo-kernel -y install kernel-ml.
After starting up the VM these are the minimal steps to install kubernetes and run the test:
yum-config-manager --add-repo https://download.docker.com/linux/centos/docker-ce.repo
yum -y install docker-ce-3:18.09.6-3.el7.x86_64
systemctl start docker
cat <<EOF > /etc/yum.repos.d/kubernetes.repo
[kubernetes]
name=Kubernetes
baseurl=https://packages.cloud.google.com/yum/repos/kubernetes-el7-x86_64
enabled=1
gpgcheck=1
repo_gpgcheck=1
gpgkey=https://packages.cloud.google.com/yum/doc/yum-key.gpg https://packages.cloud.google.com/yum/doc/rpm-package-key.gpg
EOF
# Set SELinux in permissive mode (effectively disabling it)
setenforce 0
sed -i 's/^SELINUX=enforcing$/SELINUX=permissive/' /etc/selinux/config
echo '1' > /proc/sys/net/bridge/bridge-nf-call-iptables
yum install -y kubeadm-1.15.5-0 kubelet-1.15.5-0 kubectl-1.15.5-0
systemctl enable --now kubelet
kubeadm init --pod-network-cidr=192.168.0.0/16
mkdir -p $HOME/.kube
sudo cp -i /etc/kubernetes/admin.conf $HOME/.kube/config
kubectl taint nodes --all node-role.kubernetes.io/master-
kubectl apply -f https://docs.projectcalico.org/v3.8/manifests/calico.yaml
cat <<EOF > /tmp/test-deploy.yml
apiVersion: apps/v1 # for versions before 1.9.0 use apps/v1beta2
kind: Deployment
metadata:
name: test
spec:
selector:
matchLabels:
app: test
replicas: 1
template:
metadata:
labels:
app: test
spec:
containers:
- name: nginx
image: nginxdemos/hello
ports:
- containerPort: 80
EOF
# wait for control plane to become healthy
kubectl apply -f /tmp/test-deploy.yml
Now the setup is ready and this is the ab test:
$ docker run --rm jordi/ab -n 100 -c 1 http://192.168.4.4/
This is ApacheBench, Version 2.3 <$Revision: 1826891 gt;
Copyright 1996 Adam Twiss, Zeus Technology Ltd, http://www.zeustech.net/
Licensed to The Apache Software Foundation, http://www.apache.org/
Benchmarking 192.168.4.4 (be patient)...apr_pollset_poll: The timeout specified has expired (70007)
Total of 11 requests completedThe ab test gives up after this error. If I decrease the number of requests to see avoid the timeout this is what you would see:
$ docker run --rm jordi/ab -n 10 -c 1 http://192.168.4.4/
This is ApacheBench, Version 2.3 <$Revision: 1826891 gt;
Copyright 1996 Adam Twiss, Zeus Technology Ltd, http://www.zeustech.net/
Licensed to The Apache Software Foundation, http://www.apache.org/
Benchmarking 192.168.4.4 (be patient).....done
Server Software: nginx/1.13.8
Server Hostname: 192.168.4.4
Server Port: 80
Document Path: /
Document Length: 7227 bytes
Concurrency Level: 1
Time taken for tests: 0.029 seconds
Complete requests: 10
Failed requests: 0
Total transferred: 74140 bytes
HTML transferred: 72270 bytes
Requests per second: 342.18 [#/sec] (mean)
Time per request: 2.922 [ms] (mean)
Time per request: 2.922 [ms] (mean, across all concurrent requests)
Transfer rate: 2477.50 [Kbytes/sec] received
Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 1 0.8 1 3
Processing: 1 2 1.2 1 4
Waiting: 0 1 1.3 0 4
Total: 1 3 1.4 3 5
Percentage of the requests served within a certain time (ms)
50% 3
66% 3
75% 4
80% 5
90% 5
95% 5
98% 5
99% 5
100% 5 (longest request)This issue is technically different than the original issue I reported but this is a different CNI and there are still network issues. It does have the timeout error in common when I run the same test in the kube/romana cluster: run the ab test on the same node as the pod. Both encountered the same timeout error but in romana I could get a few thousand requests to finish before hitting the timeout. Calico encounters the timeout error before reaching a dozen requests.
Other variants or notes: - net.netfilter.nf_conntrack_tcp_be_liberal=0/1 doesn't seem to make a difference - higher -n values sometimes work but it is largely random. - running the 'ab' test at low -n values several times in a row can sometimes trigger the timeout
At this point I am pretty sure it is some issue with our centos installation but I can't even guess what it could be. Are there any other limits, sysctl or other configs that could cause this?
\====== Update Nov 6, 2019 ======
I observer that we had an older kernel installed in so I upgraded my kube/calico test VM with the same newer kernel 5.3.8-1.el7.elrepo.x86_64. After the update and a few reboots I can no longer reproduce the "apr_pollset_poll: The timeout specified has expired (70007)" timout errors.
Now that the timeout error is gone I was able to repeat the original test where I load test pod A and kill pod B on my vSphere VMs. On the romana environments the problem still existed but only when the load test is on a different host than where the pod A is located. If I run the test on the same host, no errors at all. Using Calico instead of romana, there are no load test errors on either host so the problem was gone. There may still be some setting to tweak that can help romana but I think this is "strike 3" for romana so I will start transitioning a full environment to Calico and do some acceptance testing there to ensure there are no hidden gotchas.
You mentioned that if there are no pod changes while the ab test is running, then no errors get reported. So it means that errors occur when you add pod or delete one.
This is normal behaviour as when pod gets deleted; it takes time for iptable rules changes to propagate. It may happen that container got removed, but iptable rules haven't got changed yet end packets are being forwarded to the nonexisting container, and this causes errors (it is sort of like a race condition).
The first thing you can do is always to create readiness probe as it will make sure that traffic will not be forwarded to the container until it is ready to handle requests.
The second thing to do is to handle deleting the container properly. This is a bit harder task because it may be handled at many levels, but the easiest thing you can do is adding PreStop hook to your container like this:
lifecycle:
preStop:
exec:
command:
- sh
- -c
- "sleep 5"PreStop hook gets executed at the moment of the pod deletion request. From this moment, k8s start changing iptable rules and it should stop forwarding new traffic to the container that's about to get deleted. While _sleep_ing you give some time for k8s to propagate iptable changes in the cluster while not interrupting already existing connections. After PreStop handle exits, the container will receive SIGTERM signal.
My suggestion would be to apply both of these mechanisms together and check if it helps.
You also mentioned that bypassing ingress is causing more errors. I would assume that this is due to the fact that ingress has implemented retries mechanism. If it's unable to open a connection to a container, it will try several times, and hopefully will get to a container that can handle its request.