AWS Spots Booster

A controller for Kubernetes that increase AWS ASGs capacity on RebalanceRecommendation events, and drain cordoned nodes in a controlled way

Use this controller ONLY if your cluster has Multi-AZ enabled. If not, when a new instance is created on a RebalanceRecommendation event, AWS will provision it from datacenter's unbalanced pool, so after some seconds, will mark it again for termination. This ends with an endless loop of scaling up nodes until ASG limit is reached and datacenter is rebalanced.

Motivation

Spot instances are good for saving costs but reliability is pretty important on production environments. Spots are spots, so AWS can reclaim them on any moment, emitting RebalanceRecommentation and SpotInterruption events.

The SpotInterruption event gives you only 2 minutes to react and move the load. However, a RebalanceRecommendation event is emitted some minutes before.

On Kubernetes, this unreliability issues can be driven using AWS Node Termination Handler to cordon and drain the nodes on SpotInterruption and RebalanceRecommendation notices, together with Cluster Autoscaler to restore the capacity, creating new nodes when pods become Pending after draining

There is a gap between the time some capacity is removed and new one is joined to Kubernetes, causing outages and making spots unusable on real production scenarios with heavy loads.

This controller is the missing piece in the middle, just to boost your ASGs containing Spot instances on RebalanceRecommendarion and drain cordoned nodes on a controlled way, keeping your capacity over the time.

Requirements on your cluster

Cluster Autoscaler: This controller relies on Cluster Autoscaler's capacity calculations as a starting point to estimate how big is the boost needed for the ASGs, so flag --write-status-configmap must be set to true
AWS Node Termination Handler: This controller relies NTH to create Kubernetes events on RebalanceRecommendation. This way less permissions are needed, because no need to read them from the AWS.

For doing it, you need to set enableRebalanceMonitoring to true on its Helm chart, and be sure that enableRebalanceDraining is disabled (don't worry, this is the default)

Architecture

This toy is composed by two main processes:

A process to do the calculations and set the numbers into the cloud provider's ASGs
A process to drain batches of nodes in a controlled way

There are a lot of goroutines running in the background just to have the pools (EventPool, ASGPool and NodePool) always up-to-date and use them as a single point of truth. For better understanding, please, dig deeper into the source code.

Main processes using those pools are executed asynchronously and their behaviour is described in the following images:

Permissions

AWS Spots Booster require some permissions on the provider side to be able to terminate instances on drain process or read tags from ASGs. Several ways to do this can be used, but we recommend to use OIDC on EKS clusters

Using OIDC Federated Authentication

OIDC federated authentication allows your service to assume an IAM role and interact with AWS services without having to store credentials as environment variables. For an example of how to use AWS IAM OIDC with the AWS Spots Booster, please see here.

Using AWS Credentials

NOTE The following is not recommended for Kubernetes clusters running on AWS. If you are using Amazon EKS, consider using IAM roles for Service Accounts instead.

For on-premise clusters, you may create an IAM user subject to the above policy and provide the IAM credentials as environment variables in the AWS Spots Booster deployment manifest. AWS Spots Booster will use these credentials to authenticate and authorize itself.

apiVersion: v1
kind: Secret
metadata:
  name: aws-secret
type: Opaque
data:
  aws_access_key_id: BASE64_OF_YOUR_AWS_ACCESS_KEY_ID
  aws_secret_access_key: BASE64_OF_YOUR_AWS_SECRET_ACCESS_KEY

Please refer to the relevant Kubernetes documentation for creating a secret manually.

env:
  - name: AWS_ACCESS_KEY_ID
    valueFrom:
      secretKeyRef:
        name: aws-secret
        key: aws_access_key_id
  - name: AWS_SECRET_ACCESS_KEY
    valueFrom:
      secretKeyRef:
        name: aws-secret
        key: aws_secret_access_key
  - name: AWS_REGION
    value: YOUR_AWS_REGION

Shared config is enabled by default, this means that it's possible to provide a whole config file using AWS_SDK_LOAD_CONFIG environment variable with several profiles inside. One of them can be selected just setting its name on AWS_PROFILE. Even when this is possible, it's not recommended to have the credentials of several profiles inside all your environments, so our recommendation is to use this feature only when testing the controller from local without deploying it into Kubernetes.

How to deploy

We are working on a public Helm Chart which will be hosted on this repository. Until then, the simplest deployment example from the documentation can be used instead.

Flags

There are several flags that can be configured to change the behaviour of the application. They are described in the following table:

Name	Description	Default	Example
`--connection-mode`	Connect from inside or outside Kubernetes	`kubectl`	`--connection-mode incluster`
`--kubeconfig`	Path to the kubeconfig file	`~/.kube/config`	`--kubeconfig "~/.kube/config"`
`--dry-run`	Skip actual changes	`false`	`--dry-run true`
`--ca-status-namespace`	Namespace where to look for Cluster Autoscaler's status configmap	`kube-system`	`--ca-status-namespace "default"`
`--ca-status-name`	Name of Cluster Autoscaler's status configmap	`cluster-autoscaler-status`	`--ca-status-name "another-cm"`
`--ignored-autoscaling-groups`	Comma-separated list of autoscaling-group names to ignore on ASGs boosting	-	`--ignored-autoscaling-groups "eks-one,eks-two"`
`--extra-nodes-over-calculation`	Extra nodes to add to ASGs over calculated ones	`0`	`--extra-nodes-over-calculation 3`
`--disable-drain`	Disable drain-and-destroy process for nodes under risk (not recommended)	`false`	`--disable-drain true`
`--drain-timeout`	Duration to consider a drain as done when not finished	`120s`	`--drain-timeout 2m`
`--max-concurrent-drains`	Nodes to drain at once	`5`	`--max-concurrent-drains 7`
`--time-between-drains`	Duration between scheduling a drainages batch and the following (when new nodes are ready)	`60s`	`--time-between-drains "1m"`
`--ignore-pods-grace-period`	Ignore waiting for pod's grace period on termination when draining	`false`	`--ignore-pods-grace-period true`
`--max-time-consider-new-node`	Max time to consider a node as new after joined to the cluster	`-10m`	`--max-time-consider-new-node -20m`
`--metrics-port`	Port where metrics web-server will run	`2112`	`--metrics-port 8080`
`--metrics-host`	Host where metrics web-server will run	`0.0.0.0`	`--metrics-host 0.1.0.2`
`--help`	Show this help message	-	-

FAQ

Why not using Kubebuilder?

This is the first iteration of this project, and we needed to validate the idea. Coding a solution that several developers can review was faster for us. If you are reading this, that means the idea worked pretty well, and a refactor is already on our roadmap using kubebuilder.

How to release

Each release of this container is done following several steps carefully in order not to break the things for anyone.

Test the changes on the code:
```
make test
```
A release is not done if this stage fails
Define the package information
```
export VERSION="0.0.1"
```
Generate and push the Docker image (published on Docker Hub).
```
make docker-buildx
```

How to collaborate

We are open to external collaborations for this project: improvements, bugfixes, whatever.

For doing it, open an issue to discuss the need of the changes, then:

Open an issue, to discuss what is needed and the reasons
Fork the repository
Make your changes to the code
Open a PR and wait for review

License

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

DocPlanner / aws-spots-booster