* Observed frequent kube-scheduler and controller-manager
restarts with Calico as the CNI provider. Root cause was
unclear since control plane was functional and tests of
pod to pod network connectivity passed
* Root cause: Calico sets up cali* and tunl* network interfaces
for containers on hosts. NetworkManager tries to manage these
interfaces. It periodically disconnected veth pairs. Logs did
not surface this issue since its not an error per-se, just Calico
and NetworkManager dueling for control. Kubernetes correctly
restarted pods failing health checks and ensured 2 replicas were
running so the control plane functioned mostly normally. Pod to
pod connecitivity was only affected occassionally. Pain to debug.
* Solution: Configure NetworkManager to ignore the Calico ifaces
per Calico's recommendation. Cloud-init writes files after
NetworkManager starts, so a restart is required on first boot. On
subsequent boots, the file is present so no restart is needed
* (containerized) kube-proxy warns that it is unable to
load the ip_vs kernel module despite having the correct
mounts. Atomic uses an xz compressed module and modprobe
in the container was not compiled with compression support
* Workaround issue for now by always loading ip_vs on-host
* https://github.com/kubernetes/kubernetes/issues/60
* http://www.projectatomic.io/blog/2018/04/fedora-atomic-20-apr-18/
* Atomic publishes nightly AMIs which sometimes don't boot
or have issues. Until there is a source of reliable AMIs,
pin the best known working AMI
* Rel 66a66f0d18544591ffdbf8fae9df790113c93d72
* Use the upstream bootkube image packaged with the
required metadata to be usable as a system container
under systemd
* Run bootkube with runc so no host level components
use Docker any more. Docker is still the runtime
* Remove bootkube script and old systemd unit
* Change kubelet system image to use --cgroups-per-qos=true
(default) instead of false
* Change kubelet system image to use --enforce-node-allocatable=pods
instead of an empty string
* Fix kubelet port-forward on Google Cloud / Fedora Atomic
* Mount the host's /etc/hosts in kubelet system containers
* Problem: kubelet runc system containers on Atomic were not
mounting the host's /etc/hosts, like rkt-fly does on Container
Linux. `kubectl port-forward` calls socat with localhost. DNS
servers on AWS, DO, and in many bare-metal environments resolve
localhost to the caller as a convenience. Google Cloud notably
does not nor is it required to do so and this surfaced the
missing /etc/hosts in runc kubelet namespaces.
* Enable etcd v3.3 metrics to expose metrics for
scraping by Prometheus
* Use k8s.gcr.io instead of gcr.io/google_containers
* Add flexvolume plugin mount to controller manager
* Update kube-dns from v1.14.8 to v1.14.9
* Mount /opt/cni/bin in kubelet system container so
CNI plugin binaries can be found. Before, flannel
worked because the kubelet falls back to flannel
plugin baked into the hyperkube (undesired)
* Move the CNI bin install location later, since /opt
changes may be lost between ostree rebases
* Use the upstream hyperkube image packaged with the
required metadata to be usable as a system container
under systemd
* Fix port-forward since socat is included
* Several known hacks and broken areas
* Download v1.10 Kubelet from release tarball
* Install flannel CNI binaries to /opt/cni
* Switch SELinux to Permissive
* Disable firewalld service
* port-forward won't work, socat missing
* Update manifests for Kubernetes v1.10.0
* Update etcd from v3.3.2 to v3.3.3
* Add disk_type optional variable on AWS
* Remove redundant kubeconfig copy on AWS
* Distribute etcd secres only to controllers
* Organize module variables and ssh steps
* Allow multi-controller clusters on Google Cloud
* GCP regional network load balancers have a long open
bug in which requests originating from a backend instance
are routed to the instance itself, regardless of whether
the health check passes or not. As a result, only the 0th
controller node registers. We've recommended just using
single master GCP clusters for a while
* https://issuetracker.google.com/issues/67366622
* Workaround issue by switching to a GCP TCP Proxy load
balancer. TCP proxy lb routes traffic to a backend service
(global) of instance group backends. In our case, spread
controllers across 3 zones (all regions have 3+ zones) and
organize them in 3 zonal unmanaged instance groups that
serve as backends. Allows multi-controller cluster creation
* GCP network load balancers only allowed legacy HTTP health
checks so kubelet 10255 was checked as an approximation of
controller health. Replace with TCP apiserver health checks
to detect unhealth or unresponsive apiservers.
* Drawbacks: GCP provision time increases, tailed logs now
timeout (similar tradeoff in AWS), controllers only span 3
zones instead of the exact number in the region
* Workaround in Typhoon has been known and posted for 5 months,
but there still appears to be no better alternative. Its
probably time to support multi-master and accept the downsides
* Terraform v0.11.4 introduced changes to remote-exec
that mean Typhoon bare-metal clusters require multiple
runs of terraform apply to ssh and bootstrap.
* Bare-metal installs PXE boot a live instance to install
to disk and then reboot from disk as controllers/workers.
Terraform remote-exec has no way to "know" to wait until
the reboot has occurred to kickoff Kubernetes bootstrap.
Previously Typhoon created a "debug" user during this
install phase to allow an admin to SSH, but remote-exec
would hang, trying to connect as user "core". Terraform
v0.11.4 changes this behavior so remote-exec fails and
a user must re-run terraform apply until succeeding.
* A new way to "trick" remote-exec into waiting for the
reboot into the disk install is to run SSH on a non-standard
port during the disk install. This retains the ability
for an admin to SSH during install (most distros don't have
this) and fixes the issue so only a single run of terraform
apply is needed.
* https://github.com/hashicorp/terraform/pull/17359#issuecomment-376415464
Dalton Hubble
@luke
Matt Dorn