Canary releases are pretty great! ClassDojo uses them as part of our continuous delivery pipeline: having a subset of real users use & validate our app before continuing with deploys allows us to safely & automatically deploy many times a day.

Our canary releases are conceptually simple:

1. we start canary containers with a new container image
2. we then route some production traffic to these containers
3. we monitor them: if a container sees a problem, we stop our pipeline. If they don't see problems, we start a full production deploy

Simple enough, right? There are a few details that go into setting up a system like this, and I'd like to take you through how ClassDojo does it. Our pipeline works well for our company's needs, and I think it's a good example of what this kind of canary-gated deploy can look like.

The key pieces of our system:

1. We have a logging taxonomy that lets us accurately detect server-errors that we want to fix. ("Errors" that we don't want to fix aren't actually errors!)
2. HAProxy, Consul, and Nomad let us route a subset of production traffic to a group of canary containers running new code
3. Our canary containers expose a route with the count of seen errors and the count of total requests that a monitoring script in our jenkins pipeline can hit
4. The monitoring script will stop our deployment if it sees a single error. If it sees 75,000 successful production requests, it will let the deploy go to production. (75,000 is an arbitrary number that gives us a 99.9% chance of catching errors that happen 1/10^4 requests. )

Starting canary containers

ClassDojo uses Nomad for our container orchestration, so once we've built a docker image and tagged it with our updated_image_id, we can deploy it by running nomad run api-canary.nomad.

1// api-canary.nomad
2job "api-canary" {
3  group "api-canary-group" {
4    count = 8
5    task "api-canary-task" {
6      driver = "docker"
7      config {
8        image = "updated_image_id"
9
10      }
11      service {
12        name = "api-canary"
13        port = "webserver_http"
14       // this registers this port on these containers with consul as eligible for “canary” traffic
15      }
16      resources {
17        cpu = 5000 # MHz
18        memory = 1600
19
20        network {
21          port "webserver_http"{}
22        }
23      }
24    }
25  }
26}

Nomad takes care of running these 8 (count = 8) canary containers on our nomad clients. At this point, we have running containers, but they're not serving any traffic.

Routing traffic to our canary containers

Remember that nomad job file we looked at above? Part of what it was doing was registering a service in consul. We tell consul that the webserver_http port can provide the api-canary service.

1service {
2  name = "api-canary"
3  port = "webserver_http"
4}

We use HAProxy for load-balancing, and we use consul-template to generate updated haproxy configs every 30 seconds based on the service information that consul knows about.

1backend api
2  mode http
3  # I'm omitting a *ton* of detail here!
4  # See https://engineering.classdojo.com/2021/07/13/haproxy-graceful-server-shutdowns talks about how we do graceful deploys with HAProxy
5
6{{ range service "api-canary" }}
7  server canary_{{ .Address }}:{{ .Port }} {{ .Address }}:{{ .Port }}
8{{ end }}
9
10# as far as HAProxy is concerned, the canary containers above should be treated the same as our regularly deployed containers. It will round robin traffic to all of them
11{{ range service "api" }}
12  server api_{{ .Address }}:{{ .Port }} {{ .Address }}:{{ .Port }}
13{{end}}

Monitoring canary

Whenever we see an error, we increment a local counter saying that we saw the error. What counts as an error? For us, an error is something we need to fix (most often 500s or timeouts): if something can't be fixed, it's part of the system, and we need to design around it. If you're curious about our approach to categorizing errors, Creating An Actionable Logging Taxonomy digs into the details. Having an easy way of identifying real problems that should stop a canary deploy is the key piece that makes this system work.

1let errorCount: number = 0;
2export const getErrorCount = () => errorCount;
3export function logServerError(errorDetails: ErrorDetails) {
4  errorCount++;
5  metrics.increment("serverError");
6  winstonLogger.log("error", errorDetails);
7}

Similarly, whenever we finish with a request, we increment another counter saying we saw the request. We can then expose both of these counts on our status route. There are probably better ways of publishing this information to our monitoring script rather than via our main server, but it works well enough for our needs.

1router.get("/api/errorAndRequestCount", () => {
2  return {
3    errorCount: getErrorCount(),
4    requestCount: getRequestsSeenCount(),
5    ...otherInfo,
6  });
7});

Finally, we can use consul-template to re-generate our list of canary hosts & ports, and write a monitoring script to check the /api/errorAndRequestCount route on all of them. If we see an error, we can run nomad job stop api-canary && exit 1, and that will stop our canary containers & our deployment pipeline.

consul-template -template canary.tpl:canary.txt -once

1{{ range service "api-canary" }}
2  {{ .Address }}:{{ .Port }}
3{{end -}}

Our monitoring script watches our canary containers until it sees that they've handled 75,000 requests without an error. (75,000 is a little bit of an arbitrary number: it's large enough that we'll catch relatively rare errors, and small enough that we can serve that traffic on a small number of containers within a few minutes.)

1const fs = require("fs");
2const canaryContainers = fs
3  .readFileSync("./canary.txt")
4  .toString()
5  .split("\n")
6  .map((s) => s.trim())
7  .filter(Boolean);
8const fetch = require("node-fetch");
9const { execSync } = require("child_process");
10const GOAL_REQUEST_COUNT = 75_000;
11
12const delay = (ms) => new Promise((resolve) => setTimeout(resolve, ms));
13
14(async function main() {
15  while (true) {
16    let totalRequestCount = 0;
17    for (const container of canaryContainers) {
18      const { errorCount, requestCount } = await fetch(
19        `${container}/api/errorAndRequestCount`
20      ).then((res) => res.json());
21      totalRequestCount += requestCount;
22      if (errorCount) {
23        // stopping our canary containers is normally handled by the next stage in our pipeline
24        // putting it here for illustration
25        console.error("oh no! canary failed");
26        execSync(`nomad job stop api-canary`);
27        return process.exit(1);
28      }
29    }
30
31    if (totalRequestCount >= GOAL_REQUEST_COUNT) {
32      console.log("yay! canary succeeded");
33      execSync(`nomad job stop api-canary`);
34      return process.exit(0);
35    }
36
37    await delay(1000);
38  }
39})();

Nary an Error with Canary

We've been running this canary setup (with occasional changes) for over eight years now, and it's been a key part of our continuous delivery pipeline, and has let us move quickly and safely. Without it, we would have shipped a lot more errors fully out to production, our overall error rate would likely be higher, and our teams would not be able to move as quickly as they can. Our setup definitely isn't perfect, but it's still hugely valuable, and I hope that sharing our setup will help your team create a better one.

When multiple teams share a resource like a code-base, a database, or an analytics pipeline, that resource can often suffer from the Tragedy of the Commons. Well intentioned engineers can add slow code, not prioritize important improvements because those improvements aren't aligned with team incentives, and just generally not leave the codebase or database better than they found it.

ClassDojo has a monolithic backend web-server API, but it and its backing databases often suffered from hard to diagnose performance problems because engineers would accidentally add code with poorly indexed queries, with database fanouts, or that blocked the event loop, and they wouldn't realize that that code was causing any problems. We intentionally made our Model layer request-agnostic, which helps us create better tests & simpler code, but it meant that we didn't have a way of instrumenting database & performance cost by request until Node 14 started supporting AsyncLocalStorage.

AsyncLocalStorage "creates stores that stay coherent through asynchronous operations." This lets us reference a consistent context over the course of a request to get information about the route, store information about how many database requests that request has triggered, count how many documents we've fetched, and dramatically increase our visibility into what every request is doing. Because every single one of our routes is tied to a product area, and each of our product areas is owned by a team, we were suddenly able to give each team insight into any performance problems they'd inadvertently added to the codebase. None of our engineers and teams ever wanted to cause performance problems, but the lack of pre-AsyncLocalStorage system legibility meant that these issues were difficult to track down, and teams always had higher priority work than tracking down unattributed performance problems.

When we set up our AsyncLocalStorage per-request instrumentation system, we found some crazy things:

• a route that occasionally made 30,000+ database requests because it was fanning out over a large list of items
• another route that blocked the event-loop for 15-20 seconds a few times a day, and caused timeouts for any other requests that our server was handling at the same time
• a third route that was occasionally fetching 500,000+ items to render information to return to clients
• and plenty of other routes that weren't doing things nearly this egregious, but were still bad enough that we wanted to fix them

Once we had this instrumentation, it was incredibly simple for the right team to find and fix these issues. We'd looked into some of these & similar issues before, but finding & fixing them was incredibly time-consuming. Now, we have a single log that triggers an alert and tells you everything you need to know to track down the problem.

What did we track?

1. Whenever we query one of our databases, we now include a comment with the query with the route name that triggered the request. When we go through our database's slow query logs, this comment shows us where the query actually came from, and lets us accurately attribute our database load. This makes our database optimization work so much simpler because many of our query patterns could be caused by many different routes!
2. We also track:

• the number of database queries made per database and per table-model over the course of a request
• the number of items returned over the course of a request per database and per table-model
• the total duration spent querying per-database and per table-model

Whenever we exceed safe thresholds for any of these metrics over the course of a request, we log out the details and send an alert to the appropriate team. We also send all of this data to our data warehouse, Redshift, so we can run more detailed queries about what each route is doing.

3. We track overall event-loop-lag per container, and whenever event-loop-lag crosses 1 second (which is incredibly high!), we log out the same AsyncLocalStorage based request-cost details for all recent requests. This always points to situations where we've fetched a lot of data, and are then doing something computationally complex with it.

Monoliths need Transparency

It's hard to overestimate how much using AsyncLocalStorage like this has helped easily improve our backend web-server's performance. One of the major drawbacks to a monolith like ours can be that teams aren't able to easily isolate their code's performance from that of the rest of the monolith. Having insight into what is actually happening on a granular level is what's going to allow us to continue having an easy time scaling our monolith for a long time.

When the ClassDojo engineering team started using TypeScript on our backend, I was incredibly frustrated. It felt like I needed to constantly work around the type system, and it felt like the types weren't providing any value at all. And, they weren't providing any value to me: I had fundamental misconceptions about how TypeScript worked, and my attempts to work around the Type system by using lots of any and as created some truly abominable TypeScript code.

Type Inference

Whenever you create a variable using let, const, or var, and don't provide a type, TypeScript will infer the type: this means it will use its best guess for the type. These best guesses are pretty good, but they aren't magic!

If you use let or var, TypeScript will assume that it can re-assign to any similar value. If you write let method = "GET";, TypeScript will infer that method is a string, and will let you later reassign method: method = "KonMari". If instead you use const, const method = "GET", TypeScript will infer that method is of type "GET". This means when you use let, you'll often need to use a type. let method: HTTPMethod = "GET" will allow only type-safe reassignment.

When TypeScript infers types for Objects and Arrays, it will similarly give its best guess. In JavaScript, objects are mutable: if you set up a variable like const request = { method: "GET" }, it'll assume that the type that you want is { method: string } to let you update the method field. A type of { method: string } won't be usable by something that wants { method: HTTPMethod }, so you need to either explicitly tell it the type (const request: { method: HTTPMethod } = { method: "GET" }, or tell TypeScript that it's safe to infer a stricter type (const request = { method: "GET" as const }).

Here's a TypeScript playground that explores type inference in a bit more depth, and the type inference part of the handbook is great if you want a detailed introduction.

Why isn't it narrowing?

One of the things that I found most frustrating about TypeScript was writing a clause that felt like it should have narrowed the type of a variable and not have it actually narrow. I'd often use as to bypass the type system entirely and force something to narrow, but overriding TypeScript like this is a real anti-pattern. When I finally learned about type narrowing, and type guards, TypeScript became so much more pleasant to use.

One of the very first things I learned about JavaScript was that you should never use in with Objects because in doesn't differentiate between an object's properties and those of its prototype. In TypeScript, in is crucial to use for type narrowing. If you have a function that takes a type like Error | (Error & { type: "NotFound", table: string }) | (Error & { type: "NotAllowed", reason: string }), you can't write if (!err.type) return because TypeScript doesn't know whether err has that field or not. Instead, you need to write if (!("type" in err)) return and TypeScript won't error, and will correctly narrow the type.

One related confusion was why I couldn't use if statements to narrow a type. I'd try to write code like the following:

Don't do this!

1type NotFound = Error & { NotFound: true; table: string };
2type ServerError = Error & { ServerError: true; message: string };
3type ClientError = Error & { ClientError: true; status: number };
4
5function getResponse(err: Error | NotFound | ServerError | ClientError) {
6  if ((err as ClientError).ClientError) {
7    return { status: (err as ClientError).status, message: "client error" };
8  }
9  if ((err as NotFound).NotFound) {
10    const notFoundError = err as NotFound;
11    return {
12      status: 404,
13      message: `not found in ${notFoundError.table}`;
14    }
15  }
16
17  return {
18    status: 500,
19    message: (err as ServerError).message || "server error",
20  }
21}

TypeScript can automatically discriminate a union type if the members of a union all share a field. This can lead to much simpler, type-safe, and easier to work with code!

1type NotFound = Error & { type: "NotFound", table: string };
2type ServerError = Error & { type: "ServerError", message: string; };
3type ClientError = Error & { type: "ClientError", status: number; };
4
5function getResponse(err: Error | NotFound | ServerError | ClientError) {
6  // doing this first lets TypeScript discriminate using the `type` property
7  if (!("type" in err)) {
8    return {
9      status: 500,
10      message: "server error",
11    }
12  }
13
14  if (err.type === "ClientError") {
15    // err now has type ClientError, so we can use status!
16    return { status: err.status, message: "client error" };
17  }
18
19  if (err.type === "NotFound") {
20    return {
21      status: 404,
22      message: `not found in ${err.table}`
23    }
24  }
25
26  // it even narrows this to ServerError!
27  // although it may be wiser to have an explicit `if` statement and assert that every case is handled
28  return {
29    status: 500,
30    message: err.message,
31  }
32}

Finally, if there are more complex types that you need to narrow, writing a custom type guard function is always an option! This is a function that returns a boolean that lets TypeScript narrow a type whenever you call it!

1// `method isHTTPMethod` tells callers that any string that you've checked with this function is an HTTPMethod
2function isHTTPMethod (method: string): method is HTTPMethod {
3  return ["GET", "PUT", "POST", "DELETE"].includes(method);
4}

Here's a TypeScript playground where you can see how this works.

"Brands" are required for Nominal Typing

TypeScript uses "structural typing", which means that only the shape of something matters, and the name of the type is completely irrelevant. If something looks like a duck, it's a duck as far as TypeScript is concerned. This is surprising if you're coming from a "nominal typing" background where two types with the same shapes can't be assigned to each other.

1// this is valid TypeScript
2type EmailAddress = string;
3type Url = string;
4const emailAddress: EmailAddress = "myEmail@classdojo.com";
5const url: Url = emailAddress;

If you want nominal-style types for a string or number, you need to use "Brands" to create "impossible" types.

1type EmailAddress = string & { __impossible_property: "EmailAddress" };
2type URL = string & { __impossible_property: "Url" };
3const emailAddress = "myEmail@classdojo.com" as EmailAddress; // we need to use `as` here because the EmailAddress type isn't actually possible to create in JS
4const url: URL= emailAddress; // this now errors the way we'd want

(The above isn't how you'd actually want to write this code: I'm writing it in a way that hopefully makes it a bit more clear what's going on. If you'd like to see a more realistic example, take a look at this playground)

TypeScript isn't just "JS with types"

TypeScript is a great language, but just treating it as "JS with types" is a surefire recipe for frustration. It takes time to learn the language, and I wish I'd spent more of that time up-front reading through the docs: it'd have saved a ton of pain. The TypeScript handbook is great, and there are tons of great resources online to help understand TypeScript better. There's still a ton I don't know about TypeScript, but finally learning about inference, type narrowing, and structural typing has made developing with TypeScript so much nicer.