Progressive Canary With Armory Continuous Deployment-as-a-Service

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:

• Send 25% of incoming traffic to the new version of my software
• Run a metric analysis (e.g., verify that no more than 1% of HTTP requests return a 5xx status code during the analysis period)
• Run integration tests
• If the preceding two steps succeed, bump the traffic weight to 100% and tear down the old version of my software

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:

• Traffic granularity: in a pod-ratio canary, it’s not always possible to achieve highly granular traffic shifts. If your application typically runs as a cluster of four pods, it’s not possible to send a tiny percentage of traffic — e.g., 1% — to the new version of your software. If you’re trying to maintain an ambitious SLA, coarse-grained traffic weighting may not be acceptable to you.
• Traffic shift and rollback time: since the pod-ratio strategy scales pods up and down to shift traffic, traffic shifts and rollbacks can be slow if your application boots slowly. This can be scary if you need to roll back quickly.

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.

Service Meshes and Service Mesh Interface

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.

Progressive Canary with SMI Traffic Split

Armory CD-as-a-Service uses SMI’s TrafficSplit resource to implement progressive canary deployments.

A 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:

• You can choose any integer traffic weight for the new version of your software, regardless of the size of your deployment. In the future we may support fractional traffic weights (e.g., 0.1%)
• Traffic shifts and rollbacks are nearly instantaneous and not dependent on scale up or scale downtime.

These features are now available for use in Armory CD-as-a-Service. Refer to our docs to get started.