Introduce Async Hooks into the SDK

[AbhiPrasad]

ref: #3660

Let's work toward adding async hooks into the SDK. What we probably want to use is AsyncLocalStorage, which was made stable in v16.4.0, but added in v13.10.0, v12.17.0

Important to note that for Node 8, 10, 12 we should still use domains, and then for Node 14 and above we can use AsyncLocalStorage.

• feat(core): Add async context abstraction #7753
• feat(node): Adds domain implementation of AsyncContextStrategy #7767
• feat(node): Migrate to domains used through AsyncContextStrategy #7779
• fix(node): Fix domain scope inheritance #7799
• feat(node): Add AsyncLocalStorage implementation of AsyncContextStrategy #7800
• feat(node): Auto-select best AsyncContextStrategy for Node.js version #7804
• Test performance differences between ASL and domains
• Expose and document methods for async context management (for manual instrumentation)
• feat(node): Add docs for async context apis sentry-docs#6641

Requirements:

1. There must be only one hub management strategy active at a time (no exceptions)
2. No breaking changes

Additional Details:

We can get this done with the following steps.

// @sentry/node
import { setAsyncContextStrategy } from '@sentry/core';

interface AsyncContextStrategy {
  getCurrentHub(): Hub | undefined;
  runWithAsyncContext<T>(callback: (hub: Hub) => T): T;
}

init() {
  setHubStrategy(strategy)
}

1. Update getCurrentHub to reference some global strategy for grabbing the current hub. When you initialize the SDK in Node 8-12, it uses a domain strategy. For Node 14+ it'll use async local storage strategy.
2. Instead of using domains directly in our code, we should define a runWithHub method that creates a new domain/asynclocalstorage/whatever for that code. Then we can go in and replace domain usage with runWithHub instead.

The global strategy can just just be registered onto the global object. This means there can only be one global strategy at a time.

We have to add this concept of a strategy to @sentry/core, so it can also be used by non-node SDKs (think vercel edge, cloudflare workers etc.).

[AbhiPrasad]

Did a little load testing with our changes with artillery (good tech). This was with production express, that had been warmed up, single route:

const express = require("express");
const app = express();
const Sentry = require("@sentry/node");

Sentry.init({
  dsn: "https://5382661235ae4a2da41a88dd2a4cc141@o19635.ingest.sentry.io/5338414",
  tracesSampleRate: 1.0,
  integrations: [
    // enable HTTP calls tracing
    new Sentry.Integrations.Http({ tracing: true }),
    // enable Express.js middleware tracing
    new Sentry.Integrations.Express({
      // to trace all requests to the default router
      app,
      // alternatively, you can specify the routes you want to trace:
      // router: someRouter,
    }),
  ],
});

// RequestHandler creates a separate execution context using domains, so that every
// transaction/span/breadcrumb is attached to its own Hub instance
app.use(Sentry.Handlers.requestHandler());
// TracingHandler creates a trace for every incoming request
app.use(Sentry.Handlers.tracingHandler());

app.get("/hello/:name", function (req, res) {
  res.send("hello " + req.params.name);
});

app.use(Sentry.Handlers.errorHandler());

app.listen(3000);
console.log("Listening on port 3000");

Domains: 2500 req/s

"http.response_time": {
  "min": 0,
  "max": 106,
  "count": 100000,
  "p50": 7,
  "median": 7,
  "p75": 8.9,
  "p90": 10.9,
  "p95": 13.1,
  "p99": 19.1,
  "p999": 49.9
},

Async Local Storage: 3300 req/s

"http.response_time": {
  "min": 0,
  "max": 85,
  "count": 100000,
  "p50": 5,
  "median": 5,
  "p75": 7,
  "p90": 8.9,
  "p95": 10.1,
  "p99": 13.1,
  "p999": 21.1
},

This represents a 30% improvement in throughput 🚀

Looking at the Sentry data:

first peak is domains - second is async local storage

[timfish]

Thanks for doing the load testing Abhi!

Out of interest, what throughput do you get if you don't set a strategy (ie. it just does the default which calls the callback directly)?

[AbhiPrasad]

No strategy enabled is 3600 rps

"http.response_time": {
  "min": 0,
  "max": 91,
  "count": 100000,
  "p50": 5,
  "median": 5,
  "p75": 6,
  "p90": 7.9,
  "p95": 8.9,
  "p99": 12.1,
  "p999": 16
},

Around a 8% increase when using ALS, and 60% increase in response time when using domain in the p99 case (oooof)

[AbhiPrasad]

Considering we have just the docs left - going to close this issue. Will go back and close the related issues once the release has happened!

[joshperry]

I'm working an update from 6.16 to latest for our app and just want to say I appreciate you guys shipping this. Domains were mentioned in the comments of our code and I was a bit shocked and ended up learning a ton about Sentry today. Also thank you for having this code opensource, very helpful and looks awesome!