There are several considerations that you need to make before deciding how and where to deploy your cluster. This section of the documentation will discuss different common practices and considerations that are important to creating a cluster that is able to handle the demands of a production environment.
The first things you need to consider before creating your cluster is the specifications around networking. The following are common and important questions that you must answer before you begin building any production clusters.
A cluster, like most of the resources on the Catalyst Cloud, must sit on a network inside your project. However, it is up to you whether you want to have your cluster sitting on its own network isolated from your other resources, or if you want to have your cluster sitting on the same network as other objects; so that they are able to interact with one another. This decision is dependant on what resources your cluster is going to be using or interacting with and how you want to construct your system.
Another consideration that you need to make in regards to networking is, whether you want the cluster to be publicly or privately accessible. The majority of the time, production clusters are deployed in an internal network and do not have access to the outside world; this is for security purposes and are talked about in detail later on. However, there may be times when you want to take advantage of the features offered by the kubernetes platform for a service that you want to provide public access to. It is important to know before creating your cluster what the intended purpose is, so you can take the time to learn the security practices necessary to run a public cluster.
The final thing to consider about networking is the actual address of the cluster itself. Before creating a cluster it is important to consider which subnet you are going to use for the network that the kubernetes cluster creates; both for the cluster address itself and the internal supernet that the cluster creates. The internal supernet uses the address 10.100.0.0/16 and shouldn’t conflict with any networks that you normally use but you do need to be mindful of this in the event that it does clash. If you maintain a table of the subnets that your company has in use, it is recommended that you update this list to include the new subnet space that is created with your cluster.
There are two important CIDR ranges that are defined in the cluster template.
The first is the
fixed_subnet_cidr which controls the address range that
is used by the cluster nodes. The default value for this is 10.0.0.0/24.
The second is
calico_ipv4pool which controls the address range used for
the Pod IP address pool. The default for this 10.100.0.0./16.
It is possible to modify either of these two address space by supplying a new label value a the time the cluster is created. For example, if we wished to use the range 172.16.0.0/24 for our pod IP addresses we would change the label to the following:
For the specifics on how to change label values in a cluster template when creating your cluster please see here.
The security of a cluster is mainly affected by how many access points there are to the cluster. As such, the following mainly speaks on important options for access to your cluster that you should consider.
The first issue which follows on from your networking considerations is, where am I going to be able to access the cluster? Because you have the option to make the cluster publicly accessible, you could create it so that you and those who need access to the cluster can do so wherever they are. This does come with the same risks as exposing anything to the public internet however.
Alternatively, if you are creating a private cluster, you can refine the location from where you are able to access the cluster. This customization goes beyond just your internal network. You could limit the access to the API’s from only a fixed ip address range, whether this is for your entire company’s subnet or you may only want the APIs visible from a management subnet? Or an office specific subnet? Regardless of where you may want the cluster exposed, the options are there for you to decide.
If you have already opted to go with a private cluster then this consideration is of less importance to you. If, however, you have deployed a publicly accessible cluster you can minimise your exposure to risk by applying the following.
To restrict access to the cluster API we can supply a comma separated list of
CIDR values to the
master_lb_allowed_cidrs label when we create the cluster.
This limits which IP addresses the load balancer will accept external requests
The default value is “” which means access is open to 0.0.0.0/0.
This will only work when the cluster has been deployed with a loadbalancer in front of the Kubernetes API as is the case for all of the Catalyst Cloud production templates.
As an example of what the create command could look like, let’s assume we wish to create a cluster based on the following conditions:
It is based on a production template
It is publicly accessible via the internet
That access will be restricted to to a single IP address
The resulting command would look like this.
$ openstack coe cluster create k8s-cluster \ --cluster-template kubernetes-v1.18.2-prod-20200630 \ --labels master_lb_floating_ip_enabled=true,master_lb_allowed_cidrs=18.104.22.168/32 \ --merge-labels \ --keypair glyndavies \ --node-count 2 \ --master-count 3
Your capacity needs will vary wildly depending on what you need to utilize a cluster for. Therefore when we talk about capacity considerations for a production cluster it is difficult to be specific, as each users needs will differ. However, there are some key factors and options that are available that you should know in regards to the size of your clusters and the scale of how many nodes you need.
For the size of your individual nodes. The templates that we provide for your clusters have a default flavor set that should be sufficient for most uses. Generally, we do not use a large flavor size as tasks performed in a kubernetes cluster are more reliant on scaling horizontally than on each individual node requiring a large amount of resources.
In the case of scaling, this is entirely dependant on what action you are trying to perform using your cluster. For any individual cluster we recommended that your master node count is at minimum three nodes, but always an odd number if you can help it. This is to ensure that your cluster always remains highly available as the fault tolerance for your system will scale as the cluster does. For scaling to meet the demands of your system, we recommended using the autoscaling feature, as this will allow your cluster to perform optimally no matter the amount of work it needs to complete.
Both of these capacity considerations are reliant on having a sufficient quota for your project. If you are utilizing autoscaling but are working with a quota that is smaller than your demand requires than you will run in to errors constantly. That is why it is also important that you increase your quota size based on demand.
An important part of running a production cluster is making sure that it is healthy and that you can track what actions have taken place on your clusters. You can monitor the status of your cluster at any time using the container infra or the orchestration tabs via the Dashboard.
Additionally, if you need to review or set up logging for your cluster, you can find more information on this topic under the logging tab of this documentation.