kops incorporates management of some addons; we have to manage some addons which are needed before the kubernetes API is functional.
In addition, kops offers end-user management of addons via the channels
tool (which is still experimental,
but we are working on making it a recommended part of kubernetes addon management). We ship some
curated addons in the addons directory, more information in the addons document.
kops uses the channels
tool for system addon management also. Because kops uses the same tool
for system addon management as it does for user addon management, this means that
addons installed by kops as part of cluster bringup can be managed alongside additional addons.
(Though note that bootstrap addons are much more likely to be replaced during a kops upgrade).
The general kops philosophy is to try to make the set of bootstrap addons minimal, and to make installation of subsequent addons easy.
Thus, kube-dns
and the networking overlay (if any) are the canonical bootstrap addons.
But addons such as the dashboard or the EFK stack are easily installed after kops bootstrap,
with a kubectl apply -f https://...
or with the channels tool.
In future, we may as a convenience make it easy to add optional addons to the kops manifest, though this will just be a convenience wrapper around doing it manually.
The channels tool adds a manifest-of-manifests file, of Kind: Addons
, which allows for a description
of the various manifest versions that are available. In this way kops can manage updates
as new versions of the addon are released. For example,
the dashboard addon
lists multiple versions.
For example, a typical addons declaration might looks like this:
- version: 1.4.0
selector:
k8s-addon: kubernetes-dashboard.addons.k8s.io
manifest: v1.4.0.yaml
- version: 1.5.0
selector:
k8s-addon: kubernetes-dashboard.addons.k8s.io
manifest: v1.5.0.yaml
That declares two versions of an addon, with manifests at v1.4.0.yaml
and at v1.5.0.yaml
.
These are evaluated as relative paths to the Addons file itself. (The channels tool supports
a few more protocols than kubectl
- for example s3://...
for S3 hosted manifests).
The version
field gives meaning to the alternative manifests. This is interpreted as a
semver. The channels tool keeps track of the current version installed (currently by means
of an annotation on the kube-system
namespace), and it will not reapply the same version
of the manifest. This means that a user can edit a deployed addon, and changes will not
be replaced, until a new version of the addon is installed.
The long-term direction here is that addons will mostly be configured through a ConfigMap or Secret object, and that the addon manager will (TODO) not replace the ConfigMap.
The selector
determines the objects which make up the addon. This will be used
to construct a --prune
argument (TODO), so that objects that existed in the
previous but not the new version will be removed as part of an upgrade.
The addon manager now supports a kubernetesVersion
field, which is a semver range specifier
on the kubernetes version. If the targeted version of kubernetes does not match the semver
specified, the addon version will be ignored.
This allows you to have different versions of the manifest for significant changes to the kubernetes API. For example, 1.6 changed the taints & tolerations to a field, and RBAC moved to beta. As such it is easier to have two separate manifests.
For example:
- version: 1.5.0
selector:
k8s-addon: kube-dashboard.addons.k8s.io
manifest: v1.5.0.yaml
kubernetesVersion: "<1.6.0"
id: "pre-k8s-16"
- version: 1.6.0
selector:
k8s-addon: kube-dashboard.addons.k8s.io
manifest: v1.6.0.yaml
kubernetesVersion: ">=1.6.0"
id: "k8s-16"
On kubernetes versions before 1.6, we will install v1.5.0.yaml
, whereas from kubernetes
versions 1.6 on we will install v1.6.0.yaml
.
Note that we remove the pre-release
field of the kubernetes semver, so that 1.6.0-beta.1
will match >=1.6.0
. This matches the way kubernetes does pre-releases.
However, semver is insufficient here with the kubernetes version selection. The problem is arises in the following scenario:
- Install k8s 1.5, 1.5 version of manifest is installed
- Upgrade to k8s 1.6, 1.6 version of manifest is installed
- Downgrade to k8s 1.5; we want the 1.5 version of the manifest to be installed but the 1.6 version will have a semver that is greater than or equal to the 1.5 semver.
We need a way to break the ties between the semvers, and thus we introduce the id
field.
Thus a manifest will actually look like this:
- version: 1.6.0
selector:
k8s-addon: kube-dns.addons.k8s.io
manifest: pre-k8s-16.yaml
kubernetesVersion: "<1.6.0"
id: "pre-k8s-16"
- version: 1.6.0
selector:
k8s-addon: kube-dns.addons.k8s.io
manifest: k8s-16.yaml
kubernetesVersion: ">=1.6.0"
id: "k8s-16"
Note that the two addons have the same version, but a different kubernetesVersion
selector.
But they have different id
values; addons with matching semvers but different id
s will
be upgraded. (We will never downgrade to an older semver though, regardless of id
)
So now in the above scenario after the downgrade to 1.5, although the semver is the same,
the id will not match, and the pre-k8s-16
will be installed. (And when we upgrade back
to 1.6, the k8s-16
version will be installed.
A few tips:
- The
version
can now more closely mirror the upstream version. - The manifest names should probably incorporate the
id
, for maintainability.