9.5 KiB
Maintenance
Best Practices
- Run multiple Kubernetes clusters. Run across platforms. Plan for regional and cloud outages.
- Require applications be platform agnostic. Moving an application between a Kubernetes AWS cluster and a Kubernetes bare-metal cluster should be normal.
- Strive to make single-cluster outages tolerable. Practice performing failovers.
- Strive to make single-cluster outages a non-event. Load balance applications between multiple clusters, automate failover behaviors, and adjust alerting behaviors.
Versioning
Typhoon provides tagged releases to allow clusters to be versioned using ordinary Terraform configs.
module "yavin" {
source = "git::https://github.com/poseidon/typhoon//google-cloud/fedora-coreos/kubernetes?ref=v1.18.4"
...
}
module "mercury" {
source = "git::https://github.com/poseidon/typhoon//bare-metal/container-linux/kubernetes?ref=v1.18.4"
...
}
Master is updated regularly, so it is recommended to pin modules to a release tag or commit hash. Pinning ensures terraform get --update only fetches the desired version.
Upgrades
Typhoon recommends upgrading clusters using a blue-green replacement strategy and migrating workloads.
- Launch new (candidate) clusters from tagged releases
- Apply workloads from existing cluster(s)
- Evaluate application health and performance
- Migrate application traffic to the new cluster
- Compare metrics and delete old cluster when ready
Blue-green replacement reduces risk for clusters running critical applications. Candidate clusters allow baseline properties of clusters to be assessed (e.g. pod-to-pod bandwidth). Applying application workloads allows health to be assessed before being subjected to traffic (e.g. detect any changes in Kubernetes behavior between versions). Migration to the new cluster can be controlled according to requirements. Migration may mean updating DNS records to resolve the new cluster's ingress or may involve a load balancer gradually shifting traffic to the new cluster "backend". Retain the old cluster for a time to compare metrics or for fallback if issues arise.
Blue-green replacement provides some subtler benefits as well:
- Encourages investment in tooling for traffic migration and failovers. When a cluster incident arises, shifting applications to a healthy cluster will be second nature.
- Discourages reliance on in-place opaque state. Retain confidence in your ability to create infrastructure from scratch.
- Allows Typhoon to make architecture changes between releases and eases the burden on Typhoon maintainers. By contrast, distros promising in-place upgrades get stuck with their mistakes or require complex and error-prone migrations.
Bare-Metal
Typhoon bare-metal clusters are provisioned by a PXE-enabled network boot environment and a Matchbox service. To upgrade, re-provision machines into a new cluster.
Failover application workloads to another cluster (varies).
kubectl config use-context other-context
kubectl apply -f mercury -R
# DNS or load balancer changes
Power off bare-metal machines and set their next boot device to PXE.
ipmitool -H node1.example.com -U USER -P PASS power off
ipmitool -H node1.example.com -U USER -P PASS chassis bootdev pxe
Delete or comment the Terraform config for the cluster.
- module "mercury" {
- source = "git::https://github.com/poseidon/typhoon//bare-metal/container-linux/kubernetes"
- ...
-}
Apply to delete old provisioning configs from Matchbox.
$ terraform apply
Apply complete! Resources: 0 added, 0 changed, 55 destroyed.
Re-provision a new cluster by following the bare-metal tutorial.
Cloud
Create a new cluster following the tutorials. Failover application workloads to the new cluster (varies).
kubectl config use-context other-context
kubectl apply -f mercury -R
# DNS or load balancer changes
Once you're confident in the new cluster, delete the Terraform config for the old cluster.
- module "yavin" {
- source = "git::https://github.com/poseidon/typhoon//google-cloud/container-linux/kubernetes"
- ...
-}
Apply to delete the cluster.
$ terraform apply
Apply complete! Resources: 0 added, 0 changed, 55 destroyed.
Alternatives
In-place Edits
Typhoon uses a static pod Kubernetes control plane which allows certain manifest upgrades to be performed in-place. Components like kube-apiserver, kube-controller-manager, and kube-scheduler are run as static pods. Components flannel/calico, coredns, and kube-proxy are scheduled on Kubernetes and can be edited via kubectl.
In certain scenarios, in-place edits can be useful for quickly rolling out security patches (e.g. bumping coredns) or prioritizing speed over the safety of a proper cluster re-provision and transition.
!!! note Rarely, we may test certain security in-place edits and mention them as an option in release notes.
!!! warning Typhoon does not support or document in-place edits as an upgrade strategy. They involve inherent risks and we choose not to make recommendations or guarentees about the safety of different in-place upgrades. Its explicitly a non-goal.
Node Replacement
Typhoon supports multi-controller clusters, so it is possible to upgrade a cluster by deleting and replacing nodes one by one.
!!! warning Typhoon does not support or document node replacement as an upgrade strategy. It limits Typhoon's ability to make infrastructure and architectural changes between tagged releases.
Upgrade terraform-provider-ct
The terraform-provider-ct plugin parses, validates, and converts Container Linux Configs into Ignition user-data for provisioning instances. Since Typhoon v1.12.2+, the plugin can be updated in-place so that on apply, only workers will be replaced.
Add the terraform-provider-ct plugin binary for your system to ~/.terraform.d/plugins/, noting the final name.
wget https://github.com/poseidon/terraform-provider-ct/releases/download/v0.5.0/terraform-provider-ct-v0.5.0-linux-amd64.tar.gz
tar xzf terraform-provider-ct-v0.5.0-linux-amd64.tar.gz
mv terraform-provider-ct-v0.5.0-linux-amd64/terraform-provider-ct ~/.terraform.d/plugins/terraform-provider-ct_v0.5.0
Binary names are versioned. This enables the ability to upgrade different plugins and have clusters pin different versions.
$ tree ~/.terraform.d/
/home/user/.terraform.d/
└── plugins
├── terraform-provider-ct_v0.2.1
├── terraform-provider-ct_v0.3.0
├── terraform-provider-ct_v0.5.0
└── terraform-provider-matchbox_v0.3.0
Update the version of the ct plugin in each Terraform working directory. Typhoon clusters managed in the working directory must be v1.12.2 or higher.
# providers.tf
provider "ct" {
version = "0.5.0"
}
Run init and plan to check that no diff is proposed for the controller nodes (a diff would destroy cluster state).
terraform init
terraform plan
Apply the change. Worker nodes' user-data will be changed and workers will be replaced. Rollout happens slightly differently on each platform:
AWS
AWS creates a new worker ASG, then removes the old ASG. New workers join the cluster and old workers disappear. terraform apply will hang during this process.
Azure
Azure edits the worker scale set in-place instantly. Manually terminate workers to create replacement workers using the new user-data.
Bare-Metal
No action is needed. Bare-Metal machines do not re-PXE unless explicitly made to do so.
DigitalOcean
DigitalOcean destroys existing worker nodes and DNS records, then creates new workers and DNS records. DigitalOcean lacks a "managed group" notion. For worker droplets to join the cluster, you must taint the secret copying step to indicate it must be repeated to add the kubeconfig to new workers.
# old workers destroyed, new workers created
terraform apply
# add kubeconfig to new workers
terraform state list | grep null_resource
terraform taint -module digital-ocean-nemo null_resource.copy-worker-secrets[N]
terraform apply
Expect downtime.
Google Cloud
Google Cloud creates a new worker template and edits the worker instance group instantly. Manually terminate workers and replacement workers will use the user-data.
Terraform v0.12.x
Terraform v0.12 introduced major changes to the provider plugin protocol and HCL language (first-class expressions, formal list and map types, nullable variables, variable constraints, and short-circuiting ternary operators).
Typhoon modules have been adapted for Terraform v0.12. Provider plugins requirements now enforce v0.12 compatibility. However, some HCL language changes were breaking (list type hint workarounds in v0.11 now have new meaning). Typhoon cannot offer both v0.11 and v0.12 compatibility in the same release. Upcoming releases require upgrading Terraform to v0.12.
| Typhoon Release | Terraform version |
|---|---|
| v1.15.0 - ? | v0.12.x |
| v1.10.3 - v1.15.0 | v0.11.x |
| v1.9.2 - v1.10.2 | v0.10.4+ or v0.11.x |
| v1.7.3 - v1.9.1 | v0.10.x |
| v1.6.4 - v1.7.2 | v0.9.x |