We have inconsistent sets of options passed to the playbooks during our
CI runs.
Don't run ansible-playbook directly, instead factorize the execution in
a bash function using all the common flags.
Also remove various ENABLE_* variables and instead directly test for the
relevant conditions at execution time, as this makes it more obvious and
does not force one to go back and forth in the script.
With CentOS, kubespray currently produces the following warning:
[WARNING]: TASK: bootstrap-os : Enable Oracle Linux repo: The loop variable
'item' is already in use. You should set the `loop_var` value in the
`loop_control` option for the task to something else to avoid variable
collisions and unexpected behavior.
This could bites us in nasty ways, so fix it.
c58497cde (Refactor bootstrap-os (#10983), 2024-03-27) refactored the
boostrap-os include but didn't adapt the amazon linux tasks to the
actual ID of amazon linux ('amzn')
Re-enable the CI so we can avoid that kind of breakage.
Some packages requirements depends on inventory variables
(`kube_proxy_mode` in that case but it could apply to others).
As the case seems pretty rare, instead of adding complexity to pkgs, we
add an escape hatch to use jinja conditions.
That should be revisited if we find ourselves shoehorning lots of logic
in this later on.
Uses the logic introduced in the previous patch to convert all
kubernetes/preinstall/vars/* os specific files to the `pkgs`
dictionary.
Some niceties for devs:
- always validate the `pkgs` variable to catch mistakes in CI.
- ensure that `pkgs` is always sorted. This makes it easier to find the
packages you're looking for.
Adds infrastructure to install OS packages depending not only on OS
(family, versions, etc) but on groups.
All the informations related to a particular package should reside in
the `pkgs` dictionnary, which takes inspiration from the `downloads`
dictionary structure.
Since the structure we're setting in place for installing packages has
some complexity, add a JSON schema to avoid frustrating errors when
modifying the informations (adding/removing packages install).
This reverts commit 4b0a134bc9.
The mentionned PR break scale.yml. This goes back to the status quo
until a proper fix can be provided, at which point we'll reapply the
PR.
* Move fedora ansible python install to bootstrap-os
* /bin/dir is set in bootstrap-os
* Removing ansible_os_family workarounds
Support for these distributions was merged in Ansible, no need to
override it ourselves now.
https://github.com/ansible/ansible/pull/69324 openEuler
https://github.com/ansible/ansible/pull/77275/ UnionTech OS Server 20
https://github.com/ansible/ansible/pull/78232/ Kylin
* Don't unconditionnaly set VARIANT_ID=coreos in os-release
WTF, this is so wrong.
Furthermore, is_fedora_coreos is already handled in boostrap-os
* Handle Clearlinux generically
Followup of 4eec302e86 (since we're using
package module anyway, let's get rid of the custom task)
* Remove leftover files for Coreos
Coreos was replaced by flatcar in 058438a25 but the file was copied
instead of moved.
* Remove workarounds for resolved ansible issues
* boostrap: Use first_found to include per distro
Using directly ID and VARIANT_ID with first_found allow for less manual
includes.
Distro "families" are simply handled by symlinks.
* boostrap: don't set ansible_python_interpreter
- Allows users to override the chosen python_interpreter with group_vars
easily (group_vars have lesser precedence than facts)
- Allows us to use vars at the task scope to use a virtual env
Ansible python discovery has improved, so those workarounds should not
be necessary anymore.
Special workaround for Flatcar, due to upstream ansible not willing to
support it.
Under the original code, leader election failed for ingress controllers
as a result of mismatch between election-id in the controller config,
and the resourceName in the relevant rule of role 'ingress-nginx'.
This appeared in the controller logs.
To fix the issue, a command-line option was added to container
execution (--election-id=...).
Now, the election-id agrees with the resourceName provided in
the role-ingress-nginx.yml file. A comment in that file was
changed to reflect the new logic.
Co-authored-by: Vasilis Samoladas <vsam@softnet.tuc.gr>
Co-authored-by: Mohamed Omar Zaian <mohamedzaian@gmail.com>
This should avoid permissions problems when the user creating the
directory and the user creating the content are different (when
containers images are saved by root for instances, because the user
can't use the container runtime).
* Refactor of kubeadm images listing
Instead of setting multiples facts, we directly create the dict we need from
kubeadm output.
* Remove useless 'default' filters in roles/download
* Only download kubeadm images where needed
Handle all old dns domains:
- for nodelocaldns: in the same server block as the current dns_domain
- for coredns: uffix rewrite of each of the old dns domains to the
current one
* ci: redefine multinode to node-etcd-client
This should allow to catch several class of problem rather than just
one -> from network plugin such as calico or cilium talking directly to
the etcd.
* Dynamically define etcd host range
This has two benefits:
- We don't play the etcd role twice for no reason
- We have access to the whole cluster (if needed) to use things like
group_by.
* Convert the bug-report template to issue form
* Convert the enchancement issue template to form
* Convert "Failing Test" template to issue form
* github: Remove support request template, direct to slack instead
* Remove checks for docs using exact tags
Instead use a more generic documentation for installing kubespray as a
collection from git.
* Check that we upgraded galaxy.yml to next version
This is only intented to check for human error. The version in galaxy
should be the next (which does not mean the same if we're on master or a
release branch).
* Set collection version to KUBESPRAY_NEXT_VERSION
* Disable control plane allocating podCIDR for nodes when using calico
Calico does not use the .spec.podCIDR field for its IP address
management.
Furthermore, it can false positives from the kube controller manager if
kube_network_node_prefix and calico_pool_blocksize are unaligned, which
is the case with the default shipped by kubespray.
If the subnets obtained from using kube_network_node_prefix are bigger,
this would result at some point in the control plane thinking it does
not have subnets left for a new node, while calico will work without
problems.
Explicitely set a default value of false for calico_ipam_host_local to
facilitate its use in templates.
* Don't default to kube_network_node_prefix for calico_pool_blocksize
They have different semantics: kube_network_node_prefix is intended to
be the size of the subnet for all pods on a node, while there can be
more than on calico block of the specified size (they are allocated on
demand).
Besides, this commit does not actually change anything, because the
current code is buggy: we don't ever default to
kube_network_node_prefix, since the variable is defined in the role
defaults.
We take advantage of group_by to create the list of nodes needing new
certs, instead of manually looping inside a Jinja template.
This should make the role more readable and less susceptible to
white space problems.
* Decouple role kubespray-defaults from download
Avoids doing re-importing the download role on every invocation of
kubespray-defaults (and skipping everything).
This has a measurable effect on playbook performance.
* Update docs refering to moved download defaults
* Mask systemd swap.target do disable swap
This is a more generic way to disable swap, since it pulls .swap units
in systemd distributions; fstab is only one way to generate .swap units.
* Unconditionally disable swap
We only care to disable it (the "swapon" registered variable is not used
anywhere else.
This allows to get rid of the ignore_errors, since this was added
because swapon.stdout does not exist in check_mode (see issue #6642).
* Don't explicitly disable swapOnZram
We're already masking the swap.target, which would pull the zram unit,
hence no need to handle zram-generator specifically.
Allow to fail early (pre-commit time) for jinja error, rather than
waiting until executing the playbook and the invalid template.
I could not find a simple jinja pre-commit hook in the wild.
* Clean up redondant defaulting
drain_{timeout,grace_period}_after_failure don't exist at this point, so
they always default.
* Remove useless facts
The drain_*_after_failure are never used
* Try both conntrack modules instead of checking kernel version
Depending on kernel distributor, the kernel version might not be a
correct indicator of the conntrack module use.
Instead, we check both (and use the first found).
* Use modproble.persistent rather than manual persistence
When installed as an ansible collection, roles in
ansible_play_role_names will be designated by their FQDN (i.e
'kubernetes-sigs.kubespray.<role-name>).
It means we need to check for both when checking for roles in the play.
* Validate systemd unit files
This ensure that we fail early if we have a bad systemd unit file
(syntax error, using a version not available in the local version, etc)
* Hack to check systemd version for service files validation
factory-reset.target was introduced in system 250, same version as the
aliasing feature we need for verifying systemd services with ansible.
So we only actually executes the validation if that target is present.
This is an horrible hack which should be reverted as soon as we drop
support for distributions with systemd<250.
* Use RandomizedDelaySec to spread out control certificates renewal plane
If the number of control plane node is superior to 6, using (index * 10
minutes) will fail (03:60:00 is not a valid timestamp).
Compared to just fixing the jinja expression (to use a modulo for
example), this should avoid having two control planes certificates
update node being triggered at the same time.
* Make k8s-certs-renew.timer Persistent
If the control plane happens to be offline during the scheduled
certificates renewal (node failure or anything like that), we still want
the renewal to happen.
* containerd: refactor handlers to use 'listen'
* cri-dockerd: refactor handlers to use 'listen'
* cri-o: refactor handlers to use 'listen'
* docker: refactor handlers to use 'listen'
* etcd: refactor handlers to use 'listen'
* control-plane: refactor handlers to use 'listen'
* kubeadm: refactor handlers to use 'listen'
* node: refactor handlers to use 'listen'
* preinstall: refactor handlers to use 'listen'
* calico: refactor handlers to use 'listen'
* kube-router: refactor handlers to use 'listen'
* macvlan: refactor handlers to use 'listen'
This make native ansible features (dry-run, changed state) easier to
have, and should have a minimal performance impact, since it only runs
on the etcd members.
The blockSize attribute from Calico IPPool resources cannot be changed
once set [1]. Consequently, we use the one currently defined when
configuring the existing IPPool, avoiding upgrade errors by trying to
change it.
In particular, this can be useful when calico_pool_blocksize default
changes in kubespray, which would otherwise force users to add an
explicit setting to their inventories.
[1]: https://docs.tigera.io/calico/latest/reference/resources/ippool#spec
This allows this task to work with a forks count > 10 and the default
configuration of sshd, which is to limit sessions to 10. (see
MaxSessions in sshd_config).
Since this is a delegate_to task, it connects to the same host (first
etcd) for each node in the cluster, thus easily going above 10.
Raising the ssh connection attempts allow for more robustness, without
decreasing the forks count or serialising the tasks, which could slow
the task (or the playbook as a whole, if decreasing forks).
Max Gautier
