Skip to content

Commit

Permalink
doc: Add a design doc for user customization & persistence
Browse files Browse the repository at this point in the history
Initially, user defined configurations such as changing the number of
replicas for a given deployment in a MetalK8s cluster are lost during
an upgrade/downgrade scenario.

This document explains some of the design choices considered while
designing a simplistic tool for MetalK8s that guarantees that user defined
configurations are persisted throughout.

Closes: #2233
  • Loading branch information
Ebaneck committed Feb 18, 2020
1 parent 17cbd14 commit 7c2fc38
Show file tree
Hide file tree
Showing 3 changed files with 205 additions and 0 deletions.
1 change: 1 addition & 0 deletions docs/developer/architecture/index.rst
Original file line number Diff line number Diff line change
Expand Up @@ -7,3 +7,4 @@ Architecture Documents
deployment
monitoring
requirements
user_customization_and_persistence
192 changes: 192 additions & 0 deletions docs/developer/architecture/user_customization_and_persistence.rst
Original file line number Diff line number Diff line change
@@ -0,0 +1,192 @@
User Customization & Persistence
================================

Context
-------

.. todo::

This section will be handled by the requirements PR.


Design Choices
--------------

:term:`ConfigMap` is chosen as a unified data access and storage media for user
editable configurations in a MetalK8s cluster based on the above requirements
for the following reasons:

* Ability to support Update operations on ConfigMap's with CLI and UI easily
using our already existing python kubernetes module.
* Guarantee of adaptability and ease of changing the design and implementation
in cases were customer needs evolve rapidly.
* ConfigMaps are stored in the :term:`etcd` database which is generally being
backed up. This ensures that user settings cannot be lost easily.


How it works
~~~~~~~~~~~~

Before Bootstrap, :file:`defaults.yaml` will be pre-filled with default service
settings. Service pods and deployments will be configured to either
consume configuration data directly from ConfigMaps or fallback to
the default values if the service specific ConfigMap was not found.

**Using Salt states**

Once a ConfigMap is updated by the user (say a user changes the number of
replicas for Prometheus deployments to a new value), then perform the
following actions:

- Apply a Salt states that reads the ConfigMap object, validates the schema
and checks the new values passed.

- Finally, we make sure that the updated values are picked up by their
respective consumers (this might require Pod restarts for changes to take
effect).

Storage format
~~~~~~~~~~~~~~

A sample ConfigMap can be defined with the following fields.

An example of such a ConfigMap:

.. code-block:: yaml
apiVersion: v1
kind: ConfigMap
metadata:
namespace: <namespace>
name: <config-name>
data:
config.yaml: |-
apiVersion: <object-version>
kind: <kind>
spec:
<key>: <values>
**Use case 1:**

Configure and store the number of replicas for service specific Deployments
found in the `metalk8s-monitoring` namespace using the ConfigMap format.

.. code-block:: yaml
apiVersion: v1
kind: ConfigMap
metadata:
namespace: metalk8s-monitoring
name: metalk8s-grafana-userconfig
data:
config.yaml: |-
apiVersion: metalk8s.scality.com/v1alpha1
kind: GrafanaUserConfig
spec:
replicas: 2
Non-goals
~~~~~~~~~

- Persisting newly added Grafana dashboards or new Grafana datasources
especially for modifications added via the Grafana UI cannot be stored in
ConfigMaps.
To handle this particular user case, we need to configure Grafana deployments
to use persistent volume claims and then document how to create volumes that
will handle the persistence of these settings across Pod restarts.

- ConfigMaps as used in the above design for storing and persisting user
settings need periodic backups for recovery purposes. This implies that
tested backup and restore measures for the :term:`etcd` database need to be
put in place to guarantee recovery in cases where there is a full site loss.

Rejected design choices
~~~~~~~~~~~~~~~~~~~~~~~

Consul KV vs ConfigMap
~~~~~~~~~~~~~~~~~~~~~~

This approach offers a full fledge KV store with a /kv endpoint which allows
CRUD operations on all KV data stored in it.
Consul KV also allows access to past versions of objects and has an optimistic
concurrency when manipulating multiple objects.

Note that, Consul KV store was rejected because managing operations such as
performing full backups, system restores for a full fledged KV system
requires time and much more efforts than the ConfigMap approach.

Operator (Custom Controller) vs ConfigMap
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Operators are useful in that, they provide self-healing functionalities on a
reactive basis. When a user changes a given configuration, it is easy to
reconcile and apply these changes to the in-cluster objects.

The Operator approach was rejected because it is much more complex, requires
much more effort to realize and there is no real need for applying changes
using this method because configuration changes are not frequent
(for a typical MetalK8s admin, changing the number of replicas for a given
deployment could happen once in 3 months) as such, having an operator watch for
object changes is not significant and not very useful at this point in time.


Implementation Details
----------------------

Iteration 1
~~~~~~~~~~~

- Define and deploy new ConfigMap stores that will hold user configurations
as listed in the requirements
- Template and render Deployment and Pod manifests that will make use of
this user customizable configurations
- Document how to change user configurations using kubectl
- Document how to add Prometheus alerting rules by editing the Prometheus rule
CRD.

Iteration 2
~~~~~~~~~~~

- Provide a CLI tool for changing any of the user configurations:

- Count of replicas for chosen Deployments (Prometheus)
- Updating a Dex authentication connector (OpenLDAP, AD and
staticUser store)
- Updating the Alertmanager notification configuration

- Provide a UI interface for adding, updating and deleting service specific
configurations for example Dex-LDAP connector integration.
- Provide a UI interface for listing MetalK8s available/supported Dex
authentication Connectors
- Provide a UI interface for enabling or disabling Dex authentication
connectors (LDAP, Active Directory, StaticUser store)
- Provide a UI interface for changing the number of replicas for a chosen set
of MetalK8s deployments (Prometheus, ....)
- Add a UI interface for listing Alertmanager notification systems MetalK8s
will support (Slack, email, hipchat)
- Provide a UI interface for adding, modifying and deleting Alertmanager
configurations from the listing above

Documentation
-------------

In the Operational Guide:

* Document how to customize or change any given service settings using the CLI
tool
* Document how to customize or change any given service settings using the UI
interface

Test Plan
---------

- Add test that ensures that update operations on user configurations are
propagated down to the various services

- Other corner cases that require testing to reduce error prone setups include:

- Checking for invalid values in a user defined configuration (e.g setting
the number of replicas to a string ("two"))
- Checking for invalid formats in a user configuration
- Checking that a user lost a configuration and we can actually revert
to default values within the pillars
12 changes: 12 additions & 0 deletions docs/glossary.rst
Original file line number Diff line number Diff line change
Expand Up @@ -31,6 +31,18 @@ Glossary
and from where the cluster will be deployed to other machines. It also
serves as the entrypoint for upgrades of the cluster.


ConfigMap
A ConfigMap is a Kubernetes object that allows one to store general
configuration information such as environment variables in a key-value
pair format.
ConfigMaps can only be applied to namespaces and once created, they can
be updated automatically without the need of restarting containers that
depend on it.

|see K8s docs|
`ConfigMap <https://kubernetes.io/docs/tasks/configure-pod-container/configure-pod-configmap/#understanding-configmaps-and-pods/>`_.

Controller Manager
``kube-controller-manager``
The Kubernetes controller manager embeds the core control loops shipped
Expand Down

0 comments on commit 7c2fc38

Please sign in to comment.