* Use docker to run the `kubelet.service` container
* Update Kubelet mounts to match Fedora CoreOS
* Remove unused `/etc/ssl/certs` mount (see
https://github.com/poseidon/typhoon/pull/810)
* Remove unused `/usr/share/ca-certificates` mount
* Remove `/etc/resolv.conf` mount, Docker default is ok
* Change `delete-node.service` to use docker instead of rkt
and inline ExecStart, as was done on Fedora CoreOS
* Fix permission denied on shutdown `delete-node`, caused
by the kubeconfig mount changing with the introduction of
node TLS bootstrap
Background
* podmand, rkt, and runc daemonless container process runners
provide advantages over the docker daemon for system containers.
Docker requires workarounds for use in systemd units where the
ExecStart must tail logs so systemd can monitor the daemonized
container. https://github.com/moby/moby/issues/6791
* Why switch then? On Flatcar Linux, podman isn't shipped. rkt
works, but isn't developing while container standards continue
to move forward. Typhoon has used runc for the Kubelet runner
before in Fedora Atomic, but its more low-level. So we're left
with Docker, which is less than ideal, but shipped in Flatcar
* Flatcar Linux appears to be shifting system components to
use docker, which does provide some limited guards against
breakages (e.g. Flatcar cannot enable docker live restore)
* Originally, poseidon/terraform-render-bootstrap generated
TLS certificates, manifests, and cluster "assets" written
to local disk (`asset_dir`) during terraform apply cluster
bootstrap
* Typhoon v1.17.0 introduced bootstrapping using only Terraform
state to store cluster assets, to avoid ever writing sensitive
materials to disk and improve automated use-cases. `asset_dir`
was changed to optional and defaulted to "" (no writes)
* Typhoon v1.18.0 deprecated the `asset_dir` variable, removed
docs, and announced it would be deleted in future.
* Add Terraform output `assets_dir` map
* Remove the `asset_dir` variable
Cluster assets are now stored in Terraform state only. For those
who wish to write those assets to local files, this is possible
doing so explicitly.
```
resource local_file "assets" {
for_each = module.yavin.assets_dist
filename = "some-assets/${each.key}"
content = each.value
}
```
Related:
* https://github.com/poseidon/typhoon/pull/595
* https://github.com/poseidon/typhoon/pull/678
* CoreOS Container Linux was deprecated in v1.18.3 (May 2020)
in favor of Fedora CoreOS and Flatcar Linux. CoreOS Container
Linux references were kept to give folks more time to migrate,
but AMIs have now been deleted. Time is up.
Rel: https://coreos.com/os/eol/
* seccomp graduated to GA in Kubernetes v1.19. Support for
seccomp alpha annotations will be removed in v1.22
* Replace seccomp annotations with the GA seccompProfile
field in the PodTemplate securityContext
* Switch profile from `docker/default` to `runtime/default`
(no effective change, since docker is the runtime)
* Verify with docker inspect SecurityOpt. Without the profile,
you'd see `seccomp=unconfined`
Related: https://github.com/poseidon/terraform-render-bootstrap/pull/215
* Fix race condition for bootstrap-secrets SELinux context on non-bootstrap controllers in multi-controller FCOS clusters
* On first boot from disk on non-bootstrap controllers, adding bootstrap-secrets races with kubelet.service starting, which can cause the secrets assets to have the wrong label until kubelet.service restarts (service, reboot, auto-update)
* This can manifest as `kube-apiserver`, `kube-controller-manager`, and `kube-scheduler` pods crashlooping on spare controllers on first cluster creation
* Fedora CoreOS now ships systemd-udev's `default.link` while
Flannel relies on being able to pick its own MAC address for
the `flannel.1` link for tunneled traffic to reach cni0 on
the destination side, without being dropped
* This change first appeared in FCOS testing-devel 32.20200624.20.1
and is the behavior going forward in FCOS since it was added
to align FCOS network naming / configs with the rest of Fedora
and address issues related to the default being missing
* Flatcar Linux (and Container Linux) has a specific flannel.link
configuration builtin, so it was not affected
* https://github.com/coreos/fedora-coreos-tracker/issues/574#issuecomment-665487296
Note: Typhoon's recommended and default CNI provider is Calico,
unless `networking` is set to flannel directly.
* Set Kubelet cgroup driver to systemd when Flatcar Linux edge
is chosen
Note: Typhoon module status assumes use of the stable variant of
an OS channel/stream. Its possible to use earlier variants and
those are sometimes tested or developed against, but stable is
the recommendation
* Accept experimental CNI `networking` mode "cilium"
* Run Cilium v1.8.0-rc4 with overlay vxlan tunnels and a
minimal set of features. We're interested in:
* IPAM: Divide pod_cidr into /24 subnets per node
* CNI networking pod-to-pod, pod-to-external
* BPF masquerade
* NetworkPolicy as defined by Kubernetes (no L7 Policy)
* Continue using kube-proxy with Cilium probe mode
* Firewall changes:
* Require UDP 8472 for vxlan (Linux kernel default) between nodes
* Optional ICMP echo(8) between nodes for host reachability
(health)
* Optional TCP 4240 between nodes for endpoint reachability (health)
Known Issues:
* Containers with `hostPort` don't listen on all host addresses,
these workloads must use `hostNetwork` for now
https://github.com/cilium/cilium/issues/12116
* Erroneous warning on Fedora CoreOS
https://github.com/cilium/cilium/issues/10256
Note: This is experimental. It is not listed in docs and may be
changed or removed without a deprecation notice
Related:
* https://github.com/poseidon/terraform-render-bootstrap/pull/192
* https://github.com/cilium/cilium/issues/12217
* Change the Calico VXLAN interface for MTU from 1450 to 1410
* VXLAN on Azure should support MTU 1450. However, there is
history where performance measures have shown that 1410 is
needed to have expected performance. Flatcar Linux has the
same MTU 1410 override and note
* FCOS 31.20200323.3.2 was known to perform fine with 1450, but
now in 31.20200517.3.0 the right value seems to be 1410
* Remove node label `node.kubernetes.io/master` from controller nodes
* Use `node.kubernetes.io/controller` (present since v1.9.5,
[#160](https://github.com/poseidon/typhoon/pull/160)) to node select controllers
* Rename controller NoSchedule taint from `node-role.kubernetes.io/master` to
`node-role.kubernetes.io/controller`
* Tolerate the new taint name for workloads that may run on controller nodes
and stop tolerating `node-role.kubernetes.io/master` taint
* Kubelet `--lock-file` and `--exit-on-lock-contention` date
back to usage of bootkube and at one point running Kubelet
in a "self-hosted" style whereby an on-host Kubelet (rkt)
started pods, but then a Kubelet DaemonSet was scheduled
and able to take over (hence self-hosted). `lock-file` and
`exit-on-lock-contention` flags supported this pivot. The
pattern has been out of favor (in bootkube too) for years
because of dueling Kubelet complexity
* Typhoon runs Kubelet as a container via an on-host systemd
unit using podman (Fedora CoreOS) or rkt (Flatcar Linux). In
fact, Typhoon no longer uses bootkube or control plane pivot
(let alone Kubelet pivot) and uses static pods since v1.16.0
* https://github.com/poseidon/typhoon/pull/536
* Generated Kubelet TLS certificate and key are not longer
used or distributed to machines since Kubelet TLS bootstrap
is used instead. Remove the certificate and key from state
* Enable terraform-provider-ct `strict` mode for parsing
Container Linux Configs and snippets
* Fix Container Linux Config systemd unit syntax `enable`
(old) to `enabled`
* Align with Fedora CoreOS which uses strict mode already
* Build Kubelet container images internally and publish
to Quay and Dockerhub (new) as an alternative in case of
registry outage or breach
* Use our infra to provide single and multi-arch (default)
Kublet images for possible future use
* Docs: Show how to use alternative Kubelet images via
snippets and a systemd dropin (builds on #737)
Changes:
* Update docs with changes to Kubelet image building
* If you prefer to trust images built by Quay/Dockerhub,
automated image builds are still available with unique
tags (albeit with some limitations):
* Quay automated builds are tagged `build-{short_sha}`
(limit: only amd64)
* Dockerhub automated builts are tagged `build-{tag}`
and `build-master` (limit: only amd64, no shas)
Links:
* Kubelet: https://github.com/poseidon/kubelet
* Docs: https://typhoon.psdn.io/topics/security/#container-images
* Registries:
* quay.io/poseidon/kubelet
* docker.io/psdn/kubelet
* Write the systemd kubelet.service to use `KUBELET_IMAGE`
as the Kubelet. This provides a nice way to use systemd
dropins to temporarily override the image (e.g. during a
registry outage)
Note: Only Typhoon Kubelet images and registries are supported.
* Set a consistent MCS level/range for Calico install-cni
* Note: Rebooting a node was a workaround, because Kubelet
relabels /etc/kubernetes(/cni/net.d)
Background:
* On SELinux enforcing systems, the Calico CNI install-cni
container ran with default SELinux context and a random MCS
pair. install-cni places CNI configs by first creating a
temporary file and then moving them into place, which means
the file MCS categories depend on the containers SELinux
context.
* calico-node Pod restarts creates a new install-cni container
with a different MCS pair that cannot access the earlier
written file (it places configs every time), causing the
init container to error and calico-node to crash loop
* https://github.com/projectcalico/cni-plugin/issues/874
```
mv: inter-device move failed: '/calico.conf.tmp' to
'/host/etc/cni/net.d/10-calico.conflist'; unable to remove target:
Permission denied
Failed to mv files. This may be caused by selinux configuration on
the
host, or something else.
```
Note, this isn't a host SELinux configuration issue.
Related:
* https://github.com/poseidon/terraform-render-bootstrap/pull/186
* Enable bootstrap token authentication on kube-apiserver
* Generate the bootstrap.kubernetes.io/token Secret that
may be used as a bootstrap token
* Generate a bootstrap kubeconfig (with a bootstrap token)
to be securely distributed to nodes. Each Kubelet will use
the bootstrap kubeconfig to authenticate to kube-apiserver
as `system:bootstrappers` and send a node-unique CSR for
kube-controller-manager to automatically approve to issue
a Kubelet certificate and kubeconfig (expires in 72 hours)
* Add ClusterRoleBinding for bootstrap token subjects
(`system:bootstrappers`) to have the `system:node-bootstrapper`
ClusterRole
* Add ClusterRoleBinding for bootstrap token subjects
(`system:bootstrappers`) to have the csr nodeclient ClusterRole
* Add ClusterRoleBinding for bootstrap token subjects
(`system:bootstrappers`) to have the csr selfnodeclient ClusterRole
* Enable NodeRestriction admission controller to limit the
scope of Node or Pod objects a Kubelet can modify to those of
the node itself
* Ability for a Kubelet to delete its Node object is retained
as preemptible nodes or those in auto-scaling instance groups
need to be able to remove themselves on shutdown. This need
continues to have precedence over any risk of a node deleting
itself maliciously
Security notes:
1. Issued Kubelet certificates authenticate as user `system:node:NAME`
and group `system:nodes` and are limited in their authorization
to perform API operations by Node authorization and NodeRestriction
admission. Previously, a Kubelet's authorization was broader. This
is the primary security motivation.
2. The bootstrap kubeconfig credential has the same sensitivity
as the previous generated TLS client-certificate kubeconfig.
It must be distributed securely to nodes. Its compromise still
allows an attacker to obtain a Kubelet kubeconfig
3. Bootstrapping Kubelet kubeconfig's with a limited lifetime offers
a slight security improvement.
* An attacker who obtains the kubeconfig can likely obtain the
bootstrap kubeconfig as well, to obtain the ability to renew
their access
* A compromised bootstrap kubeconfig could plausibly be handled
by replacing the bootstrap token Secret, distributing the token
to new nodes, and expiration. Whereas a compromised TLS-client
certificate kubeconfig can't be revoked (no CRL). However,
replacing a bootstrap token can be impractical in real cluster
environments, so the limited lifetime is mostly a theoretical
benefit.
* Cluster CSR objects are visible via kubectl which is nice
4. Bootstrapping node-unique Kubelet kubeconfigs means Kubelet
clients have more identity information, which can improve the
utility of audits and future features
Rel: https://kubernetes.io/docs/reference/command-line-tools-reference/kubelet-tls-bootstrapping/
Rel: https://github.com/poseidon/terraform-render-bootstrap/pull/185
* Race: During initial bootstrap, static control plane pods
could hang with Permission denied to bootstrap secrets. A
manual fix involved restarting Kubelet, which relabeled mounts
The race had no effect on subsequent reboots.
* bootstrap.service runs podman with a private unshared mount
of /etc/kubernetes/bootstrap-secrets which uses an SELinux MCS
label with a category pair. However, bootstrap-secrets should
be shared as its mounted by Docker pods kube-apiserver,
kube-scheduler, and kube-controller-manager. Restarting Kubelet
was a manual fix because Kubelet relabels all /etc/kubernetes
* Fix bootstrap Pod to use the shared volume label, which leaves
bootstrap-secrets files with SELinux level s0 without MCS
* Also allow failed bootstrap.service to be re-applied. This was
missing on bare-metal and AWS
* Initial support for Flatcar Linux on Azure used the Flatcar
Linux Azure Marketplace images (e.g. `flatcar-stable`) in
https://github.com/poseidon/typhoon/pull/664
* Flatcar Linux Azure Marketplace images have some unresolved
items https://github.com/poseidon/typhoon/issues/703
* Until the Marketplace items are resolved, revert to requiring
Flatcar Linux's images be manually uploaded (like GCP and
DigitalOcean)
* In v1.18.0, kubectl apply would fail to apply manifests if any
single manifest was unable to validate. For example, if a CRD and
CR were defined in the same directory, apply would fail since the
CR would be invalid as the CRD wouldn't exist
* Typhoon temporary workaround was to separate CNI CRD manifests
and explicitly apply them first. No longer needed in v1.18.1+
* Kubernetes v1.18.1 restored the prior behavior where kubectl apply
applies as many valid manifests as it can. In the example above, the
CRD would be applied and the CR could be applied if the kubectl apply
was re-run (allowing for apply loops).
* Upstream fix: https://github.com/kubernetes/kubernetes/pull/89864
* No change to Fedora CoreOS modules
* For Container Linx AWS and Azure, change the `os_image` default
from coreos-stable to flatcar-stable
* For Container Linux GCP and DigitalOcean, change `os_image` to
be required since users should upload a Flatcar Linux image and
set the variable
* For Container Linux bare-metal, recommend users change the
`os_channel` to Flatcar Linux. No actual module change.
* Add "lb" outbound rule for worker TCP _and_ UDP traffic
* Fix Azure worker nodes clock synchronization being inactive
due to timeouts reaching the CoreOS / Flatcar NTP pool
* Fix Azure worker nodes not providing outbount UDP connectivity
Background:
Azure provides VMs outbound connectivity either by having a public
IP or via an SNAT masquerade feature bundled with their virtual
load balancing abstraction (in contrast with, say, a NAT gateway).
Azure worker nodes have only a private IP, but are associated with
the cluster load balancer's backend pool and ingress frontend IP.
Outbound traffic uses SNAT with this frontend IP. A subtle detail
with Azure SNAT seems to be that since both inbound lb_rule's are
TCP only, outbound UDP traffic isn't SNAT'd (highlights the reasons
Azure shouldn't have conflated inbound load balancing with outbound
SNAT concepts). However, adding a separate outbound rule and
disabling outbound SNAT on our ingress lb_rule's we can tell Azure
to continue load balancing as before, and support outbound SNAT for
worker traffic of both the TCP and UDP protocol.
Fixes clock synchronization timeouts:
```
systemd-timesyncd[786]: Timed out waiting for reply from
45.79.36.123:123 (3.flatcar.pool.ntp.org)
```
Azure controller nodes have their own public IP, so controllers (and
etcd) nodes have not had clock synchronization or outbound UDP issues
* Fix bootstrap error for missing `manifests-networking/crd*yaml`
when `networking = "flannel"`
* Cleanup manifest-networking directory left during bootstrap
* Regressed in v1.18.0 changes for Calico https://github.com/poseidon/typhoon/pull/675
* Change kube-proxy, flannel, and calico-node DaemonSet
tolerations to tolerate `node.kubernetes.io/not-ready`
and `node-role.kubernetes.io/master` (i.e. controllers)
explicitly, rather than tolerating all taints
* kube-system DaemonSets will no longer tolerate custom
node taints by default. Instead, custom node taints must
be enumerated to opt-in to scheduling/executing the
kube-system DaemonSets
* Consider setting the daemonset_tolerations variable
of terraform-render-bootstrap at a later date
Background: Tolerating all taints ruled out use-cases
where certain nodes might legitimately need to keep
kube-proxy or CNI networking disabled
Related: https://github.com/poseidon/terraform-render-bootstrap/pull/179
* Fix delete-node service that runs on worker (cloud-only)
shutdown to delete a Kubernetes node. Regressed in #669
(unreleased)
* Use rkt `--exec` to invoke kubectl binary in the kubelet
image
* Use podman `--entrypoint` to invoke the kubectl binary in
the kubelet image
* Kubernetes plans to stop releasing the hyperkube container image
* Upstream will continue to publish `kube-apiserver`, `kube-controller-manager`,
`kube-scheduler`, and `kube-proxy` container images to `k8s.gcr.io`
* Upstream will publish Kubelet only as a binary for distros to package,
either as a DEB/RPM on traditional distros or a container image on
container-optimized operating systems
* Typhoon will package the upstream Kubelet (checksummed) and its
dependencies as a container image for use on CoreOS Container Linux,
Flatcar Linux, and Fedora CoreOS
* Update the Typhoon container image security policy to list
`quay.io/poseidon/kubelet`as an official distributed artifact
Hyperkube: https://github.com/kubernetes/kubernetes/pull/88676
Kubelet Container Image: https://github.com/poseidon/kubelet
Kubelet Quay Repo: https://quay.io/repository/poseidon/kubelet
* Accept `os_image` "flatcar-stable" and "flatcar-beta" to
use Kinvolk's Flatcar Linux images from the Azure Marketplace
Note: Flatcar Linux Azure Marketplace images require terms be
accepted before use
* Add support for `terraform-provider-azurerm` v2.0+. Require
`terraform-provider-azurerm` v2.0+ and drop v1.x support since
the Azure provider major release is not backwards compatible
* Use Azure's new Linux VM and Linux VM Scale Set resources
* Change controller's Azure disk caching to None
* Associate subnets (in addition to NICs) with security groups
(aesthetic)
* If set, change `worker_priority` from `Low` to `Spot` (action required)
Related:
* https://www.terraform.io/docs/providers/azurerm/guides/2.0-upgrade-guide.html
* Quay has historically generated ACI signatures for images to
facilitate rkt's notions of verification (it allowed authors to
actually sign images, though `--trust-keys-from-https` is in use
since etcd and most authors don't sign images). OCI standardization
didn't adopt verification ideas and checking signatures has fallen
out of favor.
* Fix an issue where Quay no longer seems to be generating ACI
signatures for new images (e.g. quay.io/coreos/etcd:v.3.4.4)
* Don't be alarmed by rkt `--insecure-options=image`. It refers
to disabling image signature checking (i.e. docker pull doesn't
check signatures either)
* System containers for Kubelet and bootstrap have transitioned
to the docker:// transport, so there is precedent and this brings
all the system containers on Container Linux controllers into
alignment
* Configure kube-proxy --metrics-bind-address=0.0.0.0 (default
127.0.0.1) to serve metrics on 0.0.0.0:10249
* Add firewall rules to allow Prometheus (resides on a worker) to
scrape kube-proxy service endpoints on controllers or workers
* Add a clusterIP: None service for kube-proxy endpoint discovery
* Change kubelet.service on Container Linux nodes to ExecStart Kubelet
inline to replace the use of the host OS kubelet-wrapper script
* Express rkt run flags and volume mounts in a clear, uniform way to
make the Kubelet service easier to audit, manage, and understand
* Eliminate reliance on a Container Linux kubelet-wrapper script
* Typhoon for Fedora CoreOS developed a kubelet.service that similarly
uses an inline ExecStart (except with podman instead of rkt) and a
more minimal set of volume mounts. Adopt the volume improvements:
* Change Kubelet /etc/kubernetes volume to read-only
* Change Kubelet /etc/resolv.conf volume to read-only
* Remove unneeded /var/lib/cni volume mount
Background:
* kubelet-wrapper was added in CoreOS around the time of Kubernetes v1.0
to simplify running a CoreOS-built hyperkube ACI image via rkt-fly. The
script defaults are no longer ideal (e.g. rkt's notion of trust dates
back to quay.io ACI image serving and signing, which informed the OCI
standard images we use today, though they still lack rkt's signing ideas).
* Shipping kubelet-wrapper was regretted at CoreOS, but remains in the
distro for compatibility. The script is not updated to track hyperkube
changes, but it is stable and kubelet.env overrides bridge most gaps
* Typhoon Container Linux nodes have used kubelet-wrapper to rkt/rkt-fly
run the Kubelet via the official k8s.gcr.io hyperkube image using overrides
(new image registry, new image format, restart handling, new mounts, new
entrypoint in v1.17).
* Observation: Most of what it takes to run a Kubelet container is defined
in Typhoon, not in kubelet-wrapper. The wrapper's value is now undermined
by having to workaround its dated defaults. Typhoon may be better served
defining Kubelet.service explicitly
* Typhoon for Fedora CoreOS developed a kubelet.service without the use
of a host OS kubelet-wrapper which is both clearer and eliminated some
volume mounts
* Rename Container Linux Config (CLC) files to *.yaml to align
with Fedora CoreOS Config (FCC) files and for syntax highlighting
* Replace common uses of Terraform `element` (which wraps around)
with `list[index]` syntax to surface index errors
* Original tutorials favored including the platform (e.g.
google-cloud) in modules (e.g. google-cloud-yavin). Prefer
naming conventions where each module / cluster has a simple
name (e.g. yavin) since the platform is usually redundant
* Retain the example cluster naming themes per platform
* Allow generated assets (TLS materials, manifests) to be
securely distributed to controller node(s) via file provisioner
(i.e. ssh-agent) as an assets bundle file, rather than relying
on assets being locally rendered to disk in an asset_dir and
then securely distributed
* Change `asset_dir` from required to optional. Left unset,
asset_dir defaults to "" and no assets will be written to
files on the machine that runs terraform apply
* Enhancement: Managed cluster assets are kept only in Terraform
state, which supports different backends (GCS, S3, etcd, etc) and
optional encryption. terraform apply accesses state, runs in-memory,
and distributes sensitive materials to controllers without making
use of local disk (simplifies use in CI systems)
* Enhancement: Improve asset unpack and layout process to position
etcd certificates and control plane certificates more cleanly,
without unneeded secret materials
Details:
* Terraform file provisioner support for distributing directories of
contents (with unknown structure) has been limited to reading from a
local directory, meaning local writes to asset_dir were required.
https://github.com/poseidon/typhoon/issues/585 discusses the problem
and newer or upcoming Terraform features that might help.
* Observation: Terraform provisioner support for single files works
well, but iteration isn't viable. We're also constrained to Terraform
language features on the apply side (no extra plugins, no shelling out)
and CoreOS / Fedora tools on the receive side.
* Take a map representation of the contents that would have been splayed
out in asset_dir and pack/encode them into a single file format devised
for easy unpacking. Use an awk one-liner on the receive side to unpack.
In pratice, this has worked well and its rather nice that a single
assets file is transferred by file provisioner (all or none)
Rel: https://github.com/poseidon/terraform-render-bootstrap/pull/162
* Set small CPU requests on static pods kube-apiserver,
kube-controller-manager, and kube-scheduler to align with
upstream tooling and for edge cases
* Effectively, a practical case for these requests hasn't been
observed. However, a small static pod CPU request may offer
a slight benefit if a controller became overloaded and the
below mechanisms were insufficient
Existing safeguards:
* Control plane nodes are tainted to isolate them from
ordinary workloads. Even dense workloads can only compress
CPU resources on worker nodes.
* Control plane static pods use the highest priority class, so
contention favors control plane pods (over say node-exporter)
and CPU is compressible too.
See: https://github.com/poseidon/terraform-render-bootstrap/pull/161
Dalton Hubble
Bo Huang
Suraj Deshmukh