After host reboot kubelet and crio goes into a loop and no container is started.
storage_driver in crio.conf overrides system defaults in etc/containers/storage.conf
/etc/containers/storage.conf is installed by package containers-common dependency
installed from cri-o (centos7) and contains "overlay".
Hosts already configured with overlay2 should be reconfigured and the
/var/lib/containers content removed.
* Add comment from roles/kubespray-defaults/defaults/main.yaml clarifying network allocation and sizes
Signed-off-by: Mikael Johansson <mik.json@gmail.com>
* Rewrite of the comment and added new examples
Signed-off-by: Mikael Johansson <mik.json@gmail.com>
* remove podman cni plugin
* configure networkamanger global dns
* allow installation of python3-libselinux by disabling update repo temporary
* remove ipv4 section because it is not a valid configuration
Removes these startup warnings:
Warning: For remote container runtime, --pod-infra-container-image is ignored in kubelet, which should be set in that remote runtime instead
Using "/var/run/crio/crio.sock" as endpoint is deprecated, please consider using full url format "unix:///var/run/crio/crio.sock".
* add snapshot-controller and v1beta1 snapshot api
* fix typo
* udpate manifest to v1beta1
* update
* update manifests
* fix spelling
* wait until crd is applied
* fix missing info in kube module
* revert snapshotclass
* add snapshot crds before applying the csi driver
* add crds, missed them in last commit
* use pull policy from kubespray
* Update CustomResourceDefinition for kubecontrollersconfigurations.crd.projectcalico.org to v1
* Align ClusterRole for kube-controllers with upstream (calico)
* Add support for openstack application credentials
* Add some lines for readability
* Update external_openstack_tenant_id check
Do not check external_openstack_tenant_id when application credentials are defined
* Add check for external_openstack_domain_id
* Fix typo
By default do not allow "unqualified" (without a registry) images
because it is considered unsecure and subject to mitm attacks.
To enable insecure pull configure for example:
crio_registries:
- "docker.io"
- "quay.io"
* Use proper openssl command to differentiate between host and ip in current certificate check
* fixup! Use proper openssl command to differentiate between host and ip in current certificate check
The bootstrap-os role uses a bootstrap script to provision a
python interpreter on flatcar and container os hosts. As the
pypy project switched to another hoster, the download url changed.
If applied this will use the new proper pypy download url in bootstrap script
* Update the cilium svc proxy test to HA mode
Signed-off-by: Arthur Outhenin-Chalandre <arthur@cri.epita.fr>
* Fix cilium strict kube-proxy in HA
Signed-off-by: Arthur Outhenin-Chalandre <arthur@cri.epita.fr>
* Add a single global endpoint variable
Signed-off-by: Arthur Outhenin-Chalandre <arthur@cri.epita.fr>
* Add cilium docs about kube-proxy replacement
Signed-off-by: Arthur Outhenin-Chalandre <arthur@cri.epita.fr>
* Fix issues in docs
Signed-off-by: Arthur Outhenin-Chalandre <arthur@cri.epita.fr>
* Option for MetalLB to talk BGP
* Check for BGP peers when metallb_protocol is bgp
* README clarification
* Commented values as documentation only in the sample inventory
* layer 2 or BGP, not both
* log level by default increased to 'info'
* cgroup manager by default set to 'systemd'
* stream port (used by kubelet) bound to 127.0.0.1 for security reasons
* metrics can be enabled and port specified
Trying to layer this package on Fedora 32 causes the install to crash
and furthermore it looks like the original bug linked to in the comment
has been resolved for Fedora 31
* Update calico_veth_mtu to FELIX_IPINIP variable
calico_veth_mtu is specified in the configuration, but since it only works for wireguard, modify it to work for IP-in-IP users.
* Update template with more cleaner expression
CI job 624031102 failed with:
fatal: [ubuntu1804]: FAILED! => {"changed": false, "msg": "Failed to download key at https://download.opensuse.org/repositories/devel:kubic:libcontainers:stable/xUbuntu_18.04/Release.key: Request failed: <urlopen error [Errno -3] Temporary failure in name resolution>"}
Assuming its a temporary problem it should get more robust with a
couple of retries like in other roles.
Co-authored-by: Hans Feldt <hafe@users.noreply.github.com>
* Add additional metadata configuration option to external Openstack CCM (kubernetes-sigs#6338)
* Set the variable external_openstack_metadata_search_order undefined by default
* Fix kubelet cgroup driver detection for crio
Remove fact standalone_kubelet since it is not used
* Fix yamllint complaints of roles/kubernetes/node/tasks/facts.yml
Co-authored-by: Hans Feldt <hafe@users.noreply.github.com>
This changes MetalLB contrib to one of addons for deploying MetalLB with
Kubernetes cluster deployment. By the default, Kubespray doesn't deploy
MetalLB addon.
Support for Ambassador OSS as an Ingress Controller when
settings `ingress_ambassador_enabled: true`.
Signed-off-by: Alvaro Saurin <alvaro.saurin@gmail.com>
* Install Kata Containers as additional container runtime
* Create RuntimeClasses for Kata Containers
* Updated Vagrant to optionally run without Docker as container manager
* Updated Vagrant to optionally use Libvirt nested virtualization
* Add Kata Containers documentation
* Fix lint errors
* Add kata_containers_enabled to kubespray-defaults
* Fixed typo error
* Fixed typo error
We need to specify either external_openstack_tenant_name or
external_openstack_tenant_id. Those values were checked by seeing they
are defined or they have actual values separately.
However those values are always defined because of the following code
of openstack/defaults/main.yml:
external_openstack_tenant_id: "{{ lookup('env','OS_TENANT_ID')| default(lookup('env','OS_PROJECT_ID'),true) }}"
external_openstack_tenant_name: "{{ lookup('env','OS_TENANT_NAME')| default(lookup('env','OS_PROJECT_NAME'),true) }}"
So even if not specifying both values, those checks could not detect
the misconfiguration. This fixes this to detect the misconfiguration.
* MINOR: Check kernel version before enable modprobe nf_conntrack
* CLEANUP: no more need to ignore error of this task
* MINOR: Fixing yaml and ansible lint error - remove trailling-space
If the special parameter "$@" is not quoted, the following command will not work:
./kubectl.sh patch storageclass my-storage-class -p '{"metadata": {"annotations":{"storageclass.kubernetes.io/is-default-class":"true"}}}'
* Add oraclelinux8 and disable firewalld
Add oraclelinux8 image and disable firewalld on oraclelinux VMs
* Fix Oracle Linux repositories
As documented in: http://yum.oracle.com/getting-started.html#installing-software-from-oracle-linux-yum-server
public-yum-ol7.repo was deprecated on release 7.6. Some repos were integrated into oracle-linux-ol7.repo (i.e.: ol7_latest, ol7_addons) and other are available as packages (epel). This also adds support for oraclelinux8
* Fix to use ansible_distribution_version
Instead of ansible_distribution_major_version
* Update README.md
On OpenStack history, we used to call "tenant" for separeted namespace.
However we use "project" now instead.
Then we have replaced "tenant" with "project". Then all "TENANT" variables
also are renamed to "PROJECT".
This makes Kubespray search "PROJECT" variable also for newer OpenStack
clouds.
with the Python ruamel.yml library
- Change True/False to true/false in a few places so file can
be more easily round-tripped with the Python ruamel.yml library
flannel, ovn and multus network plugins did not support all taint keys. This
update changes the tolerations to support them all.
According to the documentation:
```
There are two special cases: An empty key with operator Exists matches all keys,
values and effects which means this will tolerate everything. An empty effect matches
all effects with key key.
```
Usage of the empty `key` and `effect` ensures the network plugin daemonset will
be deployed on every nodes (ex: in case of custom taints, or NoExecute effect)
Since weave 2.5.1, `NoExecute` taint effect is no more supported,
this changes the daemonset tolerations to change this behavior.
Also remove the toleration key `CriticalAddonsOnly` not required anymore.
* fix(kubelet): exec notify restart kubelet service when kube-config.yml changed
* Revert "refactor(kubelet handler): change task name("reload kubelet") this is misleading"
This reverts commit 8f5d29560802c7c997293adb1ce9f84d3b20b6cb.
* fix(handlers,kubelet): setting right notify task name
* Add additional network configuration options to external Openstack CCM (#6083)
* Change the default version of external openstack cloud controller image to v1.18.1 since there was an issue in v1.18.0 where some IPs of the private network were ignored
* Change Network section in external-openstack-cloud-config.j2 to Networking
* Add networking customization information in the openstack documentation
The 98e7a07fba commit udpates the
dashboard version to 2.0.0 but it enable skip login flag wasn't
updated. This change updates its identation to avoid issues when
dashboard_skip_login is enabled.
* bump to dashboard 2.0 rc6 with metrics scrapper
* fix missing yaml seperator making Replicaset complaining about missing ServiceAccount
* unwanted legay gross hack forgot to remove before
* no need namespace on CrBinding
* bump to 2.0.0 release
* remove dashboard_metrics_scrapper_enabled
* replace removed repo with kubic repository for centos 7
* add crio configuration for centos8
* add crio configurations for debian
* use correct crio version for fedora
* simplify calulation of required crio version
- gives possibility to overwrite
* change default path for runc
* change default for seccomp path
* change default for conmon
* declare kubic repo for ubuntu
* do not install crictl twice
* move fedora repo modular tasks to crio_repo file
* move centos repo tasks to crio_repo
* declare crio version matrix for ubuntu
* update documentation crio support for ubuntu
* Add proxy support to CRI-O service
The crio.service requires proxy environment variables when it's
deployed behind a corporated network. This change creates a systemd
configuration file when the proxy variables are defined.
* Remove unnecesary crio's tasks
The playbook that bootstrap openSUSE servers assumes that the
/etc/sysconfig/proxy file exists but the execution fails when
these file is not present. This change guarantees its existence.
* assembly fallback_ips and no_proxy var only one time on localhost and populate result on all hosts
* add tag always, fix ansible lint errors
* workaround to mitogen issue dw/mitogen#663
* do not gather fact before install python on coreos like distros
* try to pass docker molecule test
* fix upgrade of crio on fcos
- update documents
* install conntrack required by kube-proxy
- like commit 48c41bcbe7
* enable fedora modular repo for crio
* allow to override crio configuration
- set cgroup manager same to kubelet_cgroup_driver if defined
- path of seccomp_profile depends on distribution
* allow to override crio configuration
- fix path for ubuntu
* allow to override crio configuration
- fix cni path for fcos
* added required permissions for querying endpointslice resources
* copy-pasted role permissions from cilium install manifests
* bumped cilium version to v1.7.2
* Fix proxy and module_hotfixes
On CentOS 8 with proxy ansible render inline `proxy` and `module_hotfixes` options.
For example:
`proxy=http://127.0.0.1:3128module_hotfixes=True`
But expected result:
```
proxy=http://127.0.0.1:3128
module_hotfixes=True
```
* Use ini_file module for work with ini files
* Prevent duplicates proxy= option in /etc/yum.conf
Module `lineinfile` is weak, use most powerful module `ini_file` and add or remove `proxy=` when `http_proxy` is defined or not.
* etcd: etcd-events doesn't depend on etcd_cluster_setup
Signed-off-by: Etienne Champetier <champetier.etienne@gmail.com>
* etcd: remove condition already present on include_tasks
Signed-off-by: Etienne Champetier <champetier.etienne@gmail.com>
* etcd: fix scaling up
Signed-off-by: Etienne Champetier <champetier.etienne@gmail.com>
* etcd: use *access_addresses, do not delegate to etcd[0]
We want to wait for the full cluster to be healthy,
so use all the cluster addresses
Also we should be able to run the playbook when etcd[0] is down
(not tested), so do not delegate to etcd[0]
Signed-off-by: Etienne Champetier <champetier.etienne@gmail.com>
* etcd: use failed_when for health check
unhealthy cluster is expected on first run, so use failed_when
instead of ignore_errors to remove scary red messages
Also use run_once
Signed-off-by: Etienne Champetier <champetier.etienne@gmail.com>
* kubernetes/preinstall: ensure ansible_fqdn is up to date after changing /etc/hosts
Signed-off-by: Etienne Champetier <champetier.etienne@gmail.com>
* kubernetes/master: regenerate apiserver cert if needed
Signed-off-by: Etienne Champetier <champetier.etienne@gmail.com>
* Fix chicken and egg problem with proxy_env not defined on the first envinronment usage.
* Disable fact gathering for the first proxy_env evaluation.
* Move proxy_env var set up from the role defaults to the root playbooks as fact.
* kubernetes-sigs-kubespray #5824
Added support nodes which are part of Virtual Machine Scale Sets(VMSS)
* kubernetes-sigs-kubespray #5824
* kubernetes-sigs-kubespray #5824
Added comments and updatetd azure docs.
* kubernetes-sigs-kubespray #5824
Added supported values comments for "azure_vmtype" in azure.yml
The variable is defined in `kubernetes/preinstall` role and used in several roles. Since `kubernetes/preinstall` is not always included when `ansible-playbook` is run with tag selectors (see #5734 for reason), they will fail, or individual roles must copy the same fact definitions (as in #3846). Moving the definition to the always-included `kubespray-defaults` role will resolve the dependency problem.
- This solves issue #5721 & #5713 (dupes)
- Provide a cleaner default usage pattern for the download role
around etcd that supports 'host' and 'docker' properly
- Extract the 'etcdctl' as a separate task install piece and reuse it where
appropriate
- Update the kubeadm-etcd task to reflect the above change
* fedora coreos support
- bootstrap and new fact for
* fedora coreos support
- fix bootstrap condition
* fedora coreos support
- allow customize packages for fedora coreos bootstrap
* fedora coreos support
- prevent install ptyhon3 and epel via dnf for fedora coreos
* fedora coreos support
- handle all ostree like os in same way
* fedora coreos support
- handle all ostree like os in same way for crio
* fedora coreos support
- add fcos documentations
* Support configuring the insert mode
Defaults to the upstream default https://docs.projectcalico.org/v3.9/reference/felix/configuration
so nothing should change for existing deployments.
This allows coexistence with other firewall management technologies.
* Add a note to the sample config
* Add docker-ce 19.03 packages for Debian & Ubuntu
K8s has updated the recommended Docker version to 19.03. More
specifically it should be 19.03.4, but since we used 18.06.7 instead of
.2, I'm assuming the latest patch version should be used here as well.
* Add docker 19.03 for redhat
The 'regexp' parameter matches last occurrence of a line starting with 'proxy=' and replaces it with the one defined in 'line' parameter. If no match - it works same way as before. This fixes resuming cluster deployments failed after that task (if there was no more than one line starting with 'proxy' in the yum.conf file - this condition should also be reassured with the change introduced here) eg. if they were initiated with Terraform.
refs #5277
As the issue describes, when no external or local load-balanced is used,
kube-proxy won't be able to contact apiserver at 127.0.0.1. So the
config map should be left as is.
* Upgrade etcd to 3.3.18
* Try with etcd 3.3.15 (kubeadm 1.16.7 default)
* Back to square one
* Try with 3.3.11
* Upgrade etcd to 3.3.18 (take 2)
* Try with 3.3.12
* download file
* download containers
* fix push image to nodes
* pull if none image on host
* fix
* improve docker image tag checks.
do not pull already cached images
* rebase fix merge conflict
* add support download_run_once when upgrade and scale cluster
add some test with download_run_once
* set default values to temp flag for every download cycle
* add save,load abilty for containerd and crio when download_run_once=true
* return redefine image save/load command to set_docker_image_facts.yml
* move set command to set_container_facts
* ctr in containerd_bin_dir
* fix order of ctr image export arguments
* temporary disable download_run_once for containerd and crio
due https://github.com/containerd/containerd/issues/4075
* remove unused files
* fix strict yaml linter warning and errors
* refactor logical conditions to pull and cache container images
* remove comment due lint check
* document role
* remove image_load_on_localhost, because cached images are always loaded to docker on remote sites
* remove XXX from debug output
* Run 'container-engine' after drain.
Move possibly disruptive role 'container-engine' to run after the node
is drained.
As that role have to be run on non-cluster nodes as well (etcd and
calico-rr), and those nodes are not drained, add play for that case.
* Check if api is up before upgrade.
If container engine is restarted in previous role, api controller can
take some time to start. This check ensures api is up before upgrade.
When kube-router is used as cni, rules might be added to the mangle table
to support external IPs. Therefore, mangle table should be flushed during
reset as well.
This 38688a4486 change replaces the
value for dockerproject_.+_repo_.+ docker variables but their new
value was previously defined in other variables. This change removes
the dockerproject_.+_repo_.+ docker variables in favor of the older
ones.
* Fix incorrect assertion comparison for kube_network_node_prefix
* Ignore assertion comparison for kube_network_node_prefix when using calico
* Adding more var docs description for kube_network_node_prefix
* Fixing trailing whitespaces
* External OpenStack Cloud Controller Manager implementation
* Adding controller image tag
* Minor fixes
* Restructuring the external cloud controller to work with KubeADM
* Added in code to allow control over pull policy for local path provisioner
* change to imagePullPolicy to use globally used variable k8s_image_pull_policy
* removed unusued variable from defaults
* updated contiv-etcd and cinder-csi-controllerplugin to use k8s_image_pull_policy variable
* Introduce kubelet_config_extra_args and kubelet_node_config_extra_args to pass params to kubelet via YAML config
* kubelet_config_extra_args is not the alternative
* Fix recover-control-plane to work with etcd 3.3.x and add CI
* Set default values for testcase
* Add actual test jobs
* Attempt to satisty gitlab ci linter
* Fix ansible targets
* Set etcd_member_name as stated in the docs...
* Recovering from 0 masters is not supported yet
* Add other master to broken_kube-master group as well
* Increase number of retries to see if etcd needs more time to heal
* Make number of retries for ETCD loops configurable, increase it for recovery CI and document it
* containerd: add proxy support
Signed-off-by: Etienne Champetier <champetier.etienne@gmail.com>
* kubespray-defaults: add kube_service_addresses / kube_pods_subnet to no_proxy
CIDR notation in no_proxy is supported by a lot of programs/languages,
including go: https://github.com/golang/go/issues/16704
Without that containerd cannot talk the the API server (kube_apiserver_ip),
but it should not go through an external proxy for the nodes/pods/services
Signed-off-by: Etienne Champetier <champetier.etienne@gmail.com>
* Change dockerproject.org to download.docker.com
dockerproject.org was deprecated in 2017 and has gone down.
* Restore yum repo for containerd
Change-Id: I883bb512a2164a85865b1bd4fb569af0358c8c2b
Co-authored-by: Craig Rodrigues <rodrigc@crodrigues.org>
When running with serial != 100%, like upgrade_cluster.yml, we need to apply this fixup each time
Problem was introduced in 05dc2b3a09
Signed-off-by: Etienne Champetier <champetier.etienne@gmail.com>
Raises limit from 100 to 300 because the default is far too low
and the pod can handle 300 with the given resources.
Change-Id: Ib1eec10da3d09d198933fcfe87291587e58d7cdb
I've tested this update by deploying a containerd / etcd cluster on top CentOS7,
MetalLB + NGINX Ingress. Upgrade using upgrade-cluster.yml
Signed-off-by: Etienne Champetier <champetier.etienne@gmail.com>
Resolves issue where kubectl cache of <v1.16 api schema
interferes with interacting with daemonsets and deployments.
Change-Id: I63b7046958f2008eb144b6da0004c598f945e0ae
* Fix crictl
* Reload systemd daemon before enabling service
* Typo
* Add crictl template
* Remove seccomp.json for ubuntu
* Set runtime path of runc for ubuntu
* Change path to conmon
There is no cri-tools package in CentOS/EPEL/Red Hat.
Additionally, cri-tools is provided into the installation via
roles/download/defaults/main.yml:104:crictl_download_url.
* Fix python3-libselinux installation for RHEL/CentOS 8
In bootstrap-centos.yml we haven't gathered the facts,
so #5127 couldn't work
Minimum ansible version to run kubespray is 2.7.8,
so ansible_distribution_major_version is defined an there is no need to default it
Signed-off-by: Etienne Champetier <champetier.etienne@gmail.com>
* Restart NetworkManager for RHEL/CentOS 8
network.service doesn't exist anymore
# systemctl status network
Unit network.service could not be found.
Signed-off-by: Etienne Champetier <champetier.etienne@gmail.com>
* Add module_hotfixes=True to docker / containerd yum repo config
https://bugzilla.redhat.com/show_bug.cgi?id=1734081https://bugzilla.redhat.com/show_bug.cgi?id=1756473
Without this setting you end up with the following error:
# yum install docker-ce
Failed to set locale, defaulting to C
Last metadata expiration check: 0:03:21 ago on Thu Sep 26 22:00:05 2019.
Error:
Problem: package docker-ce-3:19.03.2-3.el7.x86_64 requires containerd.io >= 1.2.2-3, but none of the providers can be installed
- cannot install the best candidate for the job
- package containerd.io-1.2.2-3.3.el7.x86_64 is excluded
- package containerd.io-1.2.2-3.el7.x86_64 is excluded
- package containerd.io-1.2.4-3.1.el7.x86_64 is excluded
- package containerd.io-1.2.5-3.1.el7.x86_64 is excluded
- package containerd.io-1.2.6-3.3.el7.x86_64 is excluded
(try to add '--skip-broken' to skip uninstallable packages or '--nobest' to use not only best candidate packages)
Signed-off-by: Etienne Champetier <champetier.etienne@gmail.com>
Initially this was to fix a mis-indented approvers key. However, it turns
out that 'oilbeater' is not a member of kubernetes-sigs nor
kubernetes-incubator (the org this repo was migrated from). Thus this
OWNERS file is failing prow's validation check.
As a workaround I've opted to move them to emeritus_approver, which
isn't valiated and can be used as a hint for other approvers in this
repo
This fixes the scenario where masters are upgraded one at a time
and coredns gets improperly scaled back up to 2 replicas.
Change-Id: I7cc9283f40efcfd61b5813c89a5805c95d901567
Kubespray Pull Request #5084 (https://github.com/kubernetes-sigs/kubespray/pull/5084) caused more problems than it solved due to limitations with the synchronize module. See comments on Kubespray Issues #5059 (https://github.com/kubernetes-sigs/kubespray/issues/5059) and #5116 (https://github.com/kubernetes-sigs/kubespray/issues/5116). Details from Ansible documentation: "Currently, synchronize is limited to elevating permissions via passwordless sudo. This is because rsync itself is connecting to the remote machine and rsync doesn’t give us a way to pass sudo credentials in. ... Currently there are only a few connection types which support synchronize (ssh, paramiko, local, and docker) because a sync strategy has been determined for those connection types. Note that the connection for these must not need a password as rsync itself is making the connection and rsync does not provide us a way to pass a password to the connection. ..." Thus, reverting Pull Request #5084.
* Add support for Kubernetes 1.16.1
* Defaults to 1.16.1
* add 1.16.2 checksums and set new version as default
* correct 1.16.2 checksums and add 1.15.5 checksums
When using cluster.yml or scale.yml to add/scale nodes in the existing
k8s cluster, the `kubeadm init` wouldn't run. As a result, kube-proxy
wouldn't be created, and therefore the kube-proxy deletion task would
fail, e.g. in the case where kube-router is used and "kube_proxy_remove"
is set to true. As a workaround, add ignore_errors to the kube-proxy
deletion task.
The script is not usable unless you are in the '.vagrant/provisioners/ansible/inventory/artifacts' folder.
This update makes this usable from anywhere.
- do not run etcd role when etcd_kubeadm_enabled == true
- remove default value 'systemd' for cgroup driver in containerd role.
this value override autodetect in kubelet_cgroup_driver_detected from docker info
