- Exclude kubelet CPU/RAM (kube-reserved) from cgroup. It decreases a
chance of overcommitment
- Add a possibility to modify Kubelet node-status-update-frequency
- Add a posibility to configure node-monitor-grace-period,
node-monitor-period, pod-eviction-timeout for Kubernetes controller
manager
- Add Kubernetes Relaibility Documentation with recomendations for
various scenarios.
Signed-off-by: Sergii Golovatiuk <sgolovatiuk@mirantis.com>
kubelet lost the ability to load kernel modules. This
puts that back by adding the lib/modules mount to kubelet.
The new variable kubelet_load_modules can be set to true
to enable this item. It is OFF by default.
Netchecker is rewritten in Go lang with some new args instead of
env variables. Also netchecker-server no longer requires kubectl
container. Updating playbooks accordingly.
- Docker 1.12 and further don't need nsenter hack. This patch removes
it. Also, it bumps the minimal version to 1.12.
Closes#776
Signed-off-by: Sergii Golovatiuk <sgolovatiuk@mirantis.com>
* Drop linux capabilities for unprivileged containerized
worlkoads Kargo configures for deployments.
* Configure required securityContext/user/group/groups for kube
components' static manifests, etcd, calico-rr and k8s apps,
like dnsmasq daemonset.
* Rework cloud-init (etcd) users creation for CoreOS.
* Fix nologin paths, adjust defaults for addusers role and ensure
supplementary groups membership added for users.
* Add netplug user for network plugins (yet unused by privileged
networking containers though).
* Grant the kube and netplug users read access for etcd certs via
the etcd certs group.
* Grant group read access to kube certs via the kube cert group.
* Remove priveleged mode for calico-rr and run it under its uid/gid
and supplementary etcd_cert group.
* Adjust docs.
* Align cpu/memory limits and dropped caps with added rkt support
for control plane.
Signed-off-by: Bogdan Dobrelya <bogdando@mail.ru>
be run by limit on each node without regard for order.
The changes make sure that all of the directories needed to do
certificate management are on the master[0] or etcd[0] node regardless
of when the playbook gets run on each node. This allows for separate
ansible playbook runs in parallel that don't have to be synchronized.
the openssl tools will fail to create signing requests because
the CN is too long. This is mainly a problem when FQDNs are used
in the inventory file.
THis will truncate the hostname for the CN field only at the
first dot. This should handle the issue for most cases.
Also remove the check for != "RedHat" when removing the dhclient hook,
as this had also to be done on other distros. Instead, check if the
dhclienthookfile is defined.
the tasks fail because selinux prevents ip-forwarding setting.
Moving the tasks around addresses two issues. Makes sure that
the correct python tools are in place before adjusting of selinux
and makes sure that ipforwarding is toggled after selinux adjustments.
* Add restart for weave service unit
* Reuse docker_bin_dir everythere
* Limit systemd managed docker containers by CPU/RAM. Do not configure native
systemd limits due to the lack of consensus in the kernel community
requires out-of-tree kernel patches.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
Current design expects users to define at least one
nameserver in the nameservers var to backup host OS DNS config
when the K8s cluster DNS service IP is not available and hosts
still have to resolve external or intranet FQDNs.
Fix undefined nameservers to fallback to the default_resolver.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
Add BGP route reflectors support in order to optimize BGP topology
for deployments with Calico network plugin.
Also bump version of calico/ctl for some bug fixes.
Also place in global vars and do not repeat the kube_*_config_dir
and kube_namespace vars for better code maintainability and UX.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
Do not repeat options and nameservers in the dhclient hooks.
Do not prepend nameservers for dhclient but supersede and fail back
to the upstream_dns_resolvers then default_resolver. Fixes order of
nameservers placement, which is cluster DNS ip goes always first.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
* For Debian/RedHat OS families (with NetworkManager/dhclient/resolvconf
optionally enabled) prepend /etc/resolv.conf with required nameservers,
options, and supersede domain and search domains via the dhclient/resolvconf
hooks.
* Drop (z)nodnsupdate dhclient hook and re-implement it to complement the
resolvconf -u command, which is distro/cloud provider specific.
Update docs as well.
* Enable network restart to apply and persist changes and simplify handlers
to rely on network restart only. This fixes DNS resolve for hostnet K8s
pods for Red Hat OS family. Skip network restart for canal/calico plugins,
unless https://github.com/projectcalico/felix/issues/1185 fixed.
* Replace linefiles line plus with_items to block mode as it's faster.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
Co-authored-by: Matthew Mosesohn <mmosesohn@mirantis.com>
In order to enable offline/intranet installation cases:
* Move DNS/resolvconf configuration to preinstall role. Remove
skip_dnsmasq_k8s var as not needed anymore.
* Preconfigure DNS stack early, which may be the case when downloading
artifacts from intranet repositories. Do not configure
K8s DNS resolvers for hosts /etc/resolv.conf yet early (as they may be
not existing).
* Reconfigure K8s DNS resolvers for hosts only after kubedns/dnsmasq
was set up and before K8s apps to be created.
* Move docker install task to early stage as well and unbind it from the
etcd role's specific install path. Fix external flannel dependency on
docker role handlers. Also fix the docker restart handlers' steps
ordering to match the expected sequence (the socket then the service).
* Add default resolver fact, which is
the cloud provider specific and remove hardcoded GCE resolver.
* Reduce default ndots for hosts /etc/resolv.conf to 2. Multiple search
domains combined with high ndots values lead to poor performance of
DNS stack and make ansible workers to fail very often with the
"Timeout (12s) waiting for privilege escalation prompt:" error.
* Update docs.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
The variale etcd_access_addresses is used to determine
how to address communication from other roles to
the etcd cluster.
It was set to the address that ansible uses to
connect to instance ({{ item }})s and not the
the variable:
ip_access
which had already been created and could already
be overridden through the access_ip variable.
This change allows ansible to connect to a machine using
a different address than the one used to access etcd.
Override GCE sysctl in /etc/sysctl.d/99-sysctl.conf instead of
the /etc/sysctl.d/11-gce-network-security.conf. It is recreated
by GCE, f.e. if gcloud CLI invokes some security related changes,
thus losing customizations we want to be persistent.
Update cloud providers firewall requirements in calico docs.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
Fixes#655.
This is a teporary solution for long-polling idle connections to
apiserver. It will make Nginx not cut them for the duration of expected
timeout. It will also make Nginx extremely slow in realizing that there
is some issue with connectivity to apiserver as well, so it might not be
perfect permanent solution.
* Add an option to deploy K8s app to test e2e network connectivity
and cluster DNS resolve via Kubedns for nethost/simple pods
(defaults to false).
* Parametrize existing k8s apps templates with kube_namespace and
kube_config_dir instead of hardcode.
* For CoreOS, ensure nameservers from inventory to be put in the
first place to allow hostnet pods connectivity via short names
or FQDN and hostnet agents to pass as well, if netchecker
deployed.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
* Add dns_replicas, dns_memory/cpu_limit/requests vars for
dns related apps.
* When kube_log_level=4, log dnsmasq queries as well.
* Add log level control for skydns (part of kubedns app).
* Add limits/requests vars for dnsmasq (part of kubedns app) and
dnsmasq daemon set.
* Drop string defaults for kube_log_level as it is int and
is defined in the global vars as well.
* Add docs
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
W/o this patch, the "Download containers" task may be skipped
when running on the delegate node due to wrong "when" confition.
Then it fails to upload nginx image to the nodes as well.
Fix download nginx dependency so it always can be pushed to
nodes when download_run_once is enabled.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
According to http://kubernetes.io/docs/user-guide/images/ :
By default, the kubelet will try to pull each image from the
specified registry. However, if the imagePullPolicy property
of the container is set to IfNotPresent or Never, then a local\
image is used (preferentially or exclusively, respectively).
Use IfNotPresent value to allow images prepared by the download
role dependencies to be effectively used by kubelet without pull
errors resulting apps to stay blocked in PullBackOff/Error state
even when there are images on the localhost exist.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
Pre download all required container images as roles' deps.
Drop unused flannel-server-helper images pre download.
Improve pods creation post-install test pre downloaded busybox.
Improve logs collection script with kubectl describe, fix sudo/etcd/weave
commands.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
Fix unreliable waiting for the apiserver to become ready.
Remove logfile mount to align with the rest of static pods
and because containers shall write logs to stdout only.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
- Move CNI configuration from `kubernetes/node` role to
`network_plugin/canal`
- Create SSL dir for Canal and symlink etcd SSL files
- Add needed options to `canal-config` configmap
- Run flannel and calico-node containers with proper configuration
- Drop debugs from collect-info playbook
- Drop sudo from collect-info step and add target dir var (required for travis jobs)
- Label all k8s apps, including static manifests
- Add logs for K8s apps to be collected as well
- Fix upload to GCS as a public-read tarball
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
Related bug: https://github.com/ansible/ansible/issues/15405
Uses tar and register because synchronize module cannot sudo on the
remote side correctly and copy is too slow.
This patch dramatically cuts down the number of tasks to process
for cert synchronization.
* Don't push containers if not changed
* Do preinstall role only once and redistribute defaults to
corresponding roles
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
Change the kubelet --hostname-override flag to use the ansible_hostname variable which should be more consistent with the value required by cloud providers
Add ansible_hostname alias to /etc/hosts when it is different from inventory_hostname to overcome node name limitations see https://github.com/kubernetes/kubernetes/issues/22770
Signed-off-by: Chad Swenson <chadswen@gmail.com>
The requirements for network policy feature are described here [1]. In
order to enable it, appropriate configuration must be provided to the CNI
plug in and Calico policy controller must be set up. Beside that
corresponding extensions needed to be enabled in k8s API.
Now to turn on the feature user can define `enable_network_policy`
customization variable for Ansible.
[1] http://kubernetes.io/docs/user-guide/networkpolicies/
Also adds all masters by hostname and localhost/127.0.0.1 to
apiserver SSL certificate.
Includes documentation update on how localhost loadbalancer works.
* Add a var for ndots (default 5) and put it hosts' /etc/resolv.conf.
* Poke kube dns container image to v1.7
* In order to apply changes to kubelet, notify it to
be restarted on changes made to /etc/resolv.conf. Ignore errors as the kubelet
may yet to be present up to the moment of the notification being processed.
* Remove unnecessary kubelet restart for master role as the node role ensures
it is up and running. Notify master static pods waiters for apiserver,
scheduler, controller-manager instead.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
Change additional dnsmasq opts:
- Adjust caching size and TTL
- Disable resolve conf to not create loops
- Change dnsPolicy to default (similarly to kubedns's dnsmasq). The
ClusterFirst should not be used to not create loops
- Disable negative NXDOMAIN replies to be cached
- Make its very installation as optional step (enabled by default).
If you don't want more than 3 DNS servers, including 1 for K8s, disable
it.
- Add docs and a drawing to clarify DNS setup.
- Fix stdout logs for dnsmasq/kubedns app configs
- Add missed notifies to resolvconf -u handler
- Fix idempotency of resolvconf head file changes
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
* Add the retry_stagger var to tweak push and retry time strategies.
* Add large deployments related docs.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
Move version/repo vars to download role.
Add container to download params, which overrides url/source_url,
if enabled.
Fix networking plugins download depending on kube_network_plugin.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
Shorten deployment time with:
- Remove redundand roles if duplicated by a dependency and vice versa
- When a member of k8s-cluster, always install docker as a dependency
of the etcd role and drop the docker role from cluster.yaml.
- Drop etcd and node role dependencies from master role as they are
covered by the node role in k8s-cluster group as well. Copy defaults
for master from node role.
- Decouple master, node, secrets roles handlers and vars to be used w/o
cross references.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
Creating the unit using default settings early on
and then changing it during network_plugin section
leads to too many docker restarts and duplicated code.
Reversed Wants= dependence on docker.service so it does not
restart docker when reloading systemd
Consolidated all docker restart handlers.
* Add for docker system units:
ExecReload=/bin/kill -s HUP $MAINPID
Delegate=yes
KillMode=process.
* Add missed DOCKER_OPTIONS for calico/weave docker systemd unit.
* Change Requires= to a less strict and non-faily Wants=, add missing
Wants= for After=.
* Align wants/after in a wat if Wants=foo, After= has foo as well.
* Make wants/after docker.service to ask for the docker.socket as well.
* Move "docker rm -f" commands from ExecStartPre= to ExecStopPost=.
hooks to ensure non-destructive start attempts issued by Wants=.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
etcd facts are generated in kubernetes/preinstall, so etcd nodes need
to be evaluated first before the rest of the deployment.
Moved several directory facts from kubernetes/node to
kubernetes/preinstall because they are not backward dependent.
* Add HA docs for API server.
* Add auto-evaluated internal endpoints and clarify the loadbalancer_apiserver
vars and usecases.
* Use facts for kube_apiserver to not repeat code and enable LB endpoints use.
* Use /healthz check for the wait-for apiserver.
* Use the single endpoint for kubelet instead of the list of apiservers
* Specify kube_apiserver_count to for HA layout
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
Nearly the last stage of source all components to containers.
Kubectl will be called from hyperkube image.
Remaining tasks:
* Move kube_version variable to kubernetes/preinstall
* Drop placeholder download.nothing requirement
kubelet via docker
kube-apiserver as a static pod
Fixed etcd service start to be more tolerant of slow start.
Workaround for kube_version to stay in download role, but not
download an files by creating a new "nothing" download entry.
* Add auto-evaluated internal endpoints and clarify the loadbalancer_apiserver
vars and usecases.
* Add loadbalancer_apiserver_localhost (default false). If enabled, override
the external LB and expect localhost:443/8080 to be new internal only frontends.
* Add kube_apiserver_multiaccess to ignore loadbalancers, and make clients
to access the apiservers as a comma-separated list of access_ip/ip/ansible ip
(a default mode). When disabled, allow clients to use the given loadbalancers.
* Define connections security mode for kube controllers, schedulers, proxies.
It is insecure be default, which is the current deployment choice.
* Rework the groups['kube-master'][0] hardcode defining the apiserver
endpoints.
* Improve grouping of vars and add facts for kube_apiserver.
* Define kube_apiserver_insecure_bind_address as a fact, add more
facts for ease of use.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
Move set_facts to the preinstall scope, so every role
may see it. For example, network plugins to see the etcd_endpoint.
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
* Enforce a etcd-proxy role to a k8s-cluster group members. This
provides an HA layout for all of the k8s cluster internal clients.
* Proxies to be run on each node in the group as a separate etcd
instances with a readwrite proxy mode and listen the given endpoint,
which is either the access_ip:2379 or the localhost:2379.
* A notion for the 'kube_etcd_multiaccess' is: ignore endpoints and
loadbalancers and use the etcd members IPs as a comma-separated
list. Otherwise, clients shall use the local endpoint provided by a
etcd-proxy instances on each etcd node. A Netwroking plugins always
use that access mode.
* Fix apiserver's etcd servers args to use the etcd_access_endpoint.
* Fix networking plugins flannel/calico to use the etcd_endpoint.
* Fix name env var for non masters to be set as well.
* Fix etcd_client_url was not used anywhere and other etcd_* facts
evaluation was duplicated in a few places.
* Define proxy modes only in the env file, if not a master. Del
an automatic proxy mode decisions for etcd nodes in init/unit scripts.
* Use Wants= instead of Requires= as "This is the recommended way to
hook start-up of one unit to the start-up of another unit"
* Make apiserver/calico Wants= etcd-proxy to keep it always up
Signed-off-by: Bogdan Dobrelya <bdobrelia@mirantis.com>
Co-authored-by: Matthew Mosesohn <mmosesohn@mirantis.com>
This should make things a little more composable,
by making these roles meta roles that perform no
actions by default we allow each role to own its own
resources.
Kubernetes API server has an option:
```
--advertise-address=<nil>: The IP address on which to advertise the apiserver to members of the cluster. This address must be reachable by the rest of the cluster. If blank, the --bind-address will be used. If --bind-address is unspecified, the host's default interface will be used.
```
kargo does not set --bind-address, thus it binds to eth0, in vagrant and similar
environments this causes issues because nodes cannot talk to eachother over eth0.
This sets `--advertise-address` to `ip` if its set, otherwise the default behavior
of is persisted by using `ansible_default_ipv4.address`.
check_certs task "Check_certs | Set 'sync_certs' to true" was failing
due to the dict not existing, this sets defaults that allows the
correct behavior of the conditionals.
This allows you to simply run `vagrant up` to get a 3 node HA cluster.
* Creates a dynamic inventory and uses the inventory/group_vars/all.yml
* commented lines in inventory.example so that ansible doesn't try to use it.
* added requirements.txt to give easy way to install ansible/ipaddr
* added gitignore files to stop attempts to save unwated files
* changed `Check if kube-system exists` to `failed_when: false` instead of
`ignore_errors`
