Jun 28, 2022 by Parth Srivastav
We’re very proud to announce that Armory’s new product — Armory Continuous-Deployment-as-a-Service or CD-as-a-Service — now supports progressive canary deployments using Service Mesh Interface (SMI). For those put off by the preceding jumble of buzzwords and acronyms, let me try to give a simple explanation.
Armory CD-as-a-Service allows users (you!) to safely deploy software to Kubernetes using deployment strategies. These strategies have obscure names, and each has its own pros and cons, but they tend to have shared properties: they limit the blast radius of a bad deployment, allow user-defined checks like metric analysis or integration tests, and provide an easy way to roll back to a previous version if something goes wrong.
A progressive canary deployment — not to be confused with canary analysis — incrementally shifts traffic between two versions of your software. You can, for example, define a deployment that looks like this:
A progressive canary deployment is a kind of interface. Armory CD-as-a-Service offers two progressive canary implementations: pod-ratio canary and SMI canary.
In a pod-ratio canary, we shape traffic by controlling the relative number of pods of each version of your software. Pod-ratio canaries are great, and most of the internal services that comprise Armory CD-as-a-Service use it to deploy themselves. There are, however, some limitations:
For those that like to get straight to the point (and would like to skip to the end of the article): our implementation of progressive canary with SMI solves both of these problems.
For those who want to hear more, I’ll get into some of the details of service meshes, SMI, and our implementation below.
There are many guides and introductions to service meshes. The best I’ve found is the Service Mesh Manifesto. For our purposes, I’ll give the shortest explanation possible.
Most guides to the service mesh concept start by describing a control plane and a data plane. Ignore both of these terms. Instead, focus on the proxy, which is typically auto-injected as a sidecar in your application pods.
This proxy intercepts all inbound and outbound network traffic and understands application-level protocols (think HTTP and gRPC). Because it understands application protocols, the proxy can implement application features: HTTP metrics, path-and-header-based routing, access control, or retries. The service mesh, as a whole — the previously ignored data and control plane — is a system to implement, configure, and observe these features. The magic of a mesh is that it’s possible to get consistent HTTP metrics (or retries, access control, etc.) across a set of heterogeneous microservices without the wrangling or even buy-in of the service owners.
The Service Mesh Interface(SMI) is an abstraction of the features common among service mesh implementations; the abstraction takes the form of Kubernetes Custom Resource Definitions (CRDs).
SMI solves a common integration problem: n consumers want to integrate with m producers, each having its own interface, creating a miserable m ⋅ n integration matrix. Because of SMI, tools that want to integrate with service mesh providers — Linkerd, Istio, Consul — don’t need to integrate with each mesh individually. Instead, they can just integrate with SMI. Popular service meshes understand SMI natively or via adapters.
Armory CD-as-a-Service uses SMI’s
TrafficSplit resource to implement progressive canary deployments.
TrafficSplit defines, declaratively, how traffic should be weighted across two or more groups of pods under a single cluster-local domain name. Clients can call a single address and talk, with a defined likelihood, to one of several different versions of your software.
A split is a powerful tool but, on its own, it can’t deploy your software. It needs to be used in concert with other tools: as the SMI spec says, “the resource itself [i.e., SMI
TrafficSplit] is not a complete solution, as there must be some kind of controller managing the traffic shifting over time.”
Armory CD-as-a-Service is this controller — it’s the coordinator that scales up the new version of your software, waits for it to become healthy, safely shifts traffic by updating a
TrafficSplit, and shuts down the old version of your software once you’re certain that your deployment is healthy. Apart from installation, you don’t need any knowledge of SMI or your service mesh to use CD-as-a-Service – we’re responsible for creating all the auxiliary resources and cleaning them up when we’re done.
As promised, a canary with SMI has the following advantages over a pod-ratio canary:
These features are now available for use in Armory CD-as-a-Service. Refer to our docs to get started.
Autoscaling, which helps to optimize utilization and minimize costs, is one of the most important aspects of cloud computing. Autoscaling dynamically allocates resources to your cloud application based on load. It handles bursts of load and traffic during peak hours, ensuring that end user experience is not degraded. Autoscaling also guarantees that you don’t have […]
Read more →
original post For teams that practice CI/CD and ship multiple times a day, waiting weeks to communicate these updates to customers defeats the purpose. After all, if a feature ships and nobody knows about it, did it really ship? You miss out on delivering value to customers sooner, learning which features they’re most excited about, […]
Read more →