13 KiB
Deployment Details
Table of Contents
- Introduction
- Installation
- Essential Workflow
- Creating a Network
- Creating a Virtual Machine
- Kubernetes
- Creating a Domain
- Conclusion
Introduction
We present here noteworthy details concerning different types of deployments that are possible with the ThreeFold Pulumi plugin.
Please note that the Pulumi plugin for ThreeFold Grid is not yet officially published. We look forward to your feedback on this project.
Installation
If this isn't already done, install Pulumi on your machine.
Essential Workflow
State
We will be creating a state directory and informing pulumi we want to use that local directory to manage the state, no need to use a cloud backend managed by pulumi or other providers (for the sake of testing).
mkdir ${current_dir}/state
pulumi login --cloud-url file://${current_dir}/state
Creating an Empty Stack
pulumi stack init test
Bringing up the Infrastructure
pulumi up --yes
Here we create an empty stack using stack init
and we give it the name test
then to bring up the infrastructure we execute pulumi up --yes
.
The
pulumi up
command shows the plan before agreeing to execute it
Destroy the Infrastructure
pulumi destroy --yes
pulumi stack rm --yes
pulumi logout
Pulumi Makefile
In every example directory, you will find a project file Pulumi.yaml
and a Makefile
to reduce the amount of typing:
current_dir = $(shell pwd)
run:
rm -rf ${current_dir}/state
mkdir ${current_dir}/state
pulumi login --cloud-url file://${current_dir}/state
pulumi stack init test
pulumi up --yes
destroy:
pulumi destroy --yes
pulumi stack rm --yes
pulumi logout
This means that, to execute, you just need to type make run
and to destroy, you need to type make destroy
.
Creating a Network
We address here how to create a network.
Pulumi File
You can find the original file here.
name: pulumi-provider-grid
runtime: yaml
plugins:
providers:
- name: grid
path: ../..
resources:
provider:
type: pulumi:providers:grid
properties:
mnemonic:
scheduler:
type: grid:internal:Scheduler
options:
provider: ${provider}
properties:
farm_ids: [1]
network:
type: grid:internal:Network
options:
provider: ${provider}
dependsOn:
- ${scheduler}
properties:
name: testing
description: test network
nodes:
- ${scheduler.nodes[0]}
ip_range: 10.1.0.0/16
outputs:
node_deployment_id: ${network.node_deployment_id}
nodes_ip_range: ${network.nodes_ip_range}
We will now go through this file section by section to properly understand what is happening.
name: pulumi-provider-grid
runtime: yaml
- name is for the project name (can be anything)
- runtime: the runtime we are using can be code in yaml, python, go, etc.
plugins:
providers:
- name: grid
path: ../..
Here, we define the plugins we are using within our project and their locations. Note that we use ../..
due to the repository hierarchy.
resources:
provider:
type: pulumi:providers:grid
properties:
mnemonic:
We then start by initializing the resources. The provider which we loaded in the plugins section is also a resource that has properties (the main one now is just the mnemonic of TCHhain).
scheduler:
type: grid:internal:Scheduler
options:
provider: ${provider}
properties:
farm_ids: [1]
Then, we create a scheduler grid:internal:Scheduler
, that does the planning for us. Instead of being too specific about node IDs, we just give it some generic information. For example, "I want to work against these data centers (farms)". As long as the necessary criteria are provided, the scheduler can be more specific in the planning and select the appropriate resources available on the TFGrid.
network:
type: grid:internal:Network
options:
provider: ${provider}
dependsOn:
- ${scheduler}
properties:
name: testing
description: test network
nodes:
- ${scheduler.nodes[0]}
ip_range: 10.1.0.0/16
Now, that we created the scheduler, we can go ahead and create the network resource grid:internal:Network
. Please note that the network depends on the scheduler's existence. If we remove it, the scheduler and the network will be created in parallel, that's why we have the dependsOn
section. We then proceed to specify the network resource properties, e.g. the name, the description, which nodes to deploy our network on, the IP range of the network. In our case, we only choose one node.
To access information related to our deployment, we set the section outputs. This will display results that we can use, or reuse, while we develop our infrastructure further.
outputs:
node_deployment_id: ${network.node_deployment_id}
nodes_ip_range: ${network.nodes_ip_range}
Creating a Virtual Machine
Now, we will check an example on how to create a virtual machine.
Just like we've seen above, we will have two files Makefile
and Pulumi.yaml
where we describe the infrastructure.
name: pulumi-provider-grid
runtime: yaml
plugins:
providers:
- name: grid
path: ../..
resources:
provider:
type: pulumi:providers:grid
properties:
mnemonic: <to be filled>
scheduler:
type: grid:internal:Scheduler
options:
provider: ${provider}
properties:
mru: 256
sru: 2048
farm_ids: [1]
network:
type: grid:internal:Network
options:
provider: ${provider}
dependsOn:
- ${scheduler}
properties:
name: test
description: test network
nodes:
- ${scheduler.nodes[0]}
ip_range: 10.1.0.0/16
deployment:
type: grid:internal:Deployment
options:
provider: ${provider}
dependsOn:
- ${network}
properties:
node_id: ${scheduler.nodes[0]}
name: deployment
network_name: test
vms:
- name: vm
flist: https://hub.grid.tf/tf-official-apps/base:latest.flist
entrypoint: "/sbin/zinit init"
network_name: test
cpu: 2
memory: 256
planetary: true
mounts:
- disk_name: data
mount_point: /app
env_vars:
SSH_KEY:
disks:
- name: data
size: 2
outputs:
node_deployment_id: ${deployment.node_deployment_id}
planetary_ip: ${deployment.vms_computed[0].planetary_ip}
We have a scheduler, and a network just like before. But now, we also have a deployment grid:internal:Deployment
object that can have one or more disks and virtual machines.
deployment:
type: grid:internal:Deployment
options:
provider: ${provider}
dependsOn:
- ${network}
properties:
node_id: ${scheduler.nodes[0]}
name: deployment
network_name: test
vms:
- name: vm
flist: https://hub.grid.tf/tf-official-apps/base:latest.flist
entrypoint: "/sbin/zinit init"
network_name: test
cpu: 2
memory: 256
planetary: true
mounts:
- disk_name: data
mount_point: /app
env_vars:
SSH_KEY: <to be filled>
disks:
- name: data
size: 2
The deployment can be linked to a network using network_name
and can have virtual machines in the vms
section, and disks in the disks
section. The disk can be linked and mounted in the VM if disk_name
is used in the mounts
section of the VM.
We also specify a couple of essential properties, like how many virtual cores, how much memory, what FList to use, and the environment variables in the env_vars
section.
That's it! You can now execute make run
to bring the infrastructure up.
Kubernetes
We now see how to deploy a Kubernetes cluster using Pulumi.
content was removed for brevity
kubernetes:
type: grid:internal:Kubernetes
options:
provider: ${provider}
dependsOn:
- ${network}
properties:
master:
name: kubernetes
node: ${scheduler.nodes[0]}
disk_size: 2
planetary: true
cpu: 2
memory: 2048
workers:
- name: worker1
node: ${scheduler.nodes[0]}
disk_size: 2
cpu: 2
memory: 2048
- name: worker2
node: ${scheduler.nodes[0]}
disk_size: 2
cpu: 2
memory: 2048
token: t123456789
network_name: test
ssh_key: <to be filled>
outputs:
node_deployment_id: ${kubernetes.node_deployment_id}
planetary_ip: ${kubernetes.master_computed.planetary_ip}
Now, we define the Kubernetes resource grid:internal:Kubernetes
that has master and workers slice. You define almost everything like a normal VM except for the FLiist. Also note that the token is the cluster token
. This will ensure that the workers and the master communicate properly.
Creating a Domain
The ThreeFold Pulumi repository also covers examples on how to work with TFGrid gateways.
The basic idea is that you have a virtual machine workload on a specific IP, e.g. public IPv4, IPv6, or Planetary Network, and you want to access it using domains.
There are two versions to achieve this, a simple and a fully controlled version.
- Simple domain version:
- subdomain.gent01.dev.grid.tf
- This is a generous service from ThreeFold to reserve a subdomain on a set of defined gateway domains like gent01.dev.grid.tf.
- Fully controlled domain version:
- e.g.
mydomain.com
where you manage the domain with the name provider.
- e.g.
Example of a Simple Domain Prefix
We present here the file for a simple domain prefix.
content was removed for brevity
scheduler:
type: grid:internal:Scheduler
options:
provider: ${provider}
properties:
mru: 256
farm_ids: [1]
ipv4: true
free_ips: 1
gatewayName:
type: grid:internal:GatewayName
options:
provider: ${provider}
dependsOn:
- ${scheduler}
properties:
name: pulumi
node_id: ${scheduler.nodes[0]}
backends:
- "http://69.164.223.208"
outputs:
node_deployment_id: ${gatewayName.node_deployment_id}
fqdn: ${gatewayName.fqdn}
In this example, we create a gateway name resource grid:internal:GatewayName
for the name pulumi.gent01.dev.grid.tf
.
Some things to note:
- pulumi is the prefix we want to reserve.
- It's assuming that the gateway domain we received by scheduler was the one managed by freefarm
gent01.dev.grid.tf
. - backends: defines a list of IPs to load balance against when a request for
pulumi.gent01.dev.grid.tf
is received on the gateway.
Example of a Fully Controlled Domain
Here's an example of a more complicated, but fully controlled domain.
code removed for brevity
gatewayFQDN:
type: grid:internal:GatewayFQDN
options:
provider: ${provider}
dependsOn:
- ${deployment}
properties:
name: testing
node_id: 14
fqdn: mydomain.com
backends:
- http://[${deployment.vms_computed[0].planetary_ip}]:9000
Here, we informed the gateway that any request coming for the domain mydomain.com
needs to be balanced through the backends.
Note: You need to create an A record for your domain (here
mydomain.com
) pointing to the gateway IP.
Conclusion
We covered in this guide some basic details concerning the use of the ThreeFold Pulumi plugin.
If you have any questions, you can ask the ThreeFold community for help on the ThreeFold Forum or on the ThreeFold Grid Tester Community on Telegram.