Summary

What is this repo?

• A sandbox for testing different methods to make Sentry work within a family of web application architectures collectively referred to as Micro-frontend (MFE).
  • See Sandbox how-to
• A collection of code snippets inside ("methods/") directory which are documented below, in Current Methods
• This documentation.
  • Problem Overview
  • Sentry support
    • Current Methods
    • Fundamental technical challenges
  • Sandbox (for testing)

Problem Overview

Micro-frontend is not a specific technology, but rather a concept (or buzzword). Modeled after the idea of microservices vs monolith backend, it is a design pattern where a single user-facing web application is composed of 2 or more separate frontend components each owned (and in the case of true, remote, MFEs - also deployed and operated) by a separate team. The goal is to allow each component-owner team to build and ship independently. Naturally, the developers want the errors from different components go into their own separate Sentry projects/DSNs. Unfortunately, the naive approach of calling Sentry.init() in each component does not work.

Not everyone uses the term "micro-frontend". Developers may describe their micro-component (or micro) as widget, plugin, library or module. Typically, there is a top-level component responsible for tying everything together into a single user-facing web app, which we refer to as host-application or host. Sometimes a micro is consumed by multiple hosts (as in 3rd party scenario described later).

There are 4 subtypes of this architecture. There is no (fundamental) difference between these 4 types from the browser runtime perspective. However, when it comes to integrating Sentry, each has its own distinct requirements and obstacles:

remote	lib
3premote	3plib

• True (remote) micro-frontends: components are built, deployed and served separately. Often associated with Webpack's module federation.¹
• Library (lib) components are package dependencies included (npm, yarn) during host-application's build process. They are deployed together with the host app and served from the same origin and sometimes even the same bundle file.

On top of that, components may be owned by:

• Same org (remote, lib) host and micro are owned by different teams within the same company.
• 3rd Party (3premote, 3plib) Typically multiple host apps owned by separate 3rd party companies who are micro's customers.

	remote	lib	3premote	3plib
Component-level ownership	yes	yes	yes	yes
micro team controls their code's deployment to prod	yes	no	yes	no
micro team controls their code's minification and URL paths	yes	no	yes	no
Code changes in the host are undesirable	no	no	yes	yes

Sentry support

As of August 2022 Sentry does not officially support MFE use case.

See Fundamental technical challenges to understand why.

This repository documents current workarounds (aka methods).

Current methods

All the code included in this repository is intended as example only and should NOT be adopted for use in production software without first undergoing full review and rigorous testing. This code is provided on an "AS-IS" basis without warranty of any kind, either express or implied, including without limitation any implied warranties of condition, uninterrupted use, merchantability, fitness for a particular purpose, or non-infringement. The details of your application or component, architecture of the host-application, and your target browser support, among many other things, may require you to modify this code. Issues regarding these code examples should be submitted through GitHub.

Feature support and limitations	simple-lib.js	simple-remote.js	WrapALL	flex-micro.js
Supported use cases	lib	remote	lib 3plib remote 3premote	remote 3premote
Recommended for use case	lib	-	3premote remote (2+ MICROs)	remote (1 MICRO)
Sentry SDK support	v7, v6	v7, v6	v7, v6	v7, (v6?)
Separate projects	yes	yes	yes	yes
No errors leak from MICRO into HOST project	yes	yes	yes	yes***
Source mapping MICRO	yes (tricky)*	yes	yes	yes
Works if host doesn't use Sentry	no**	no**	no**	yes
Works if MICRO initialized before HOST Sentry loaded	no**	no**	no**	yes
Works if MICRO and HOST use different versions of Sentry	no, might break	no, might break	no, might break	yes, defers to HOST version
Works if more than 1 MICRO-component	not impl.	not impl.	yes	?
Code change needed in HOST	yes, custom	yes, generic	no	no
Requires broad application code changes	no	no	yes, in micro	no
Separate breadcrumbs, tags, context	possibly, see mlmmn's code	possibly, see mlmmn's code	possibly, see mlmmn's code	possibly, see mlmmn's code
React support	not impl.	not impl.	not impl.	not impl.
Performance (see footnote †)	H,M->Hp 0->Mp	H,M->Hp 0->Mp	H,M->Hp 0->Mp	H,M->Hp 0->Mp

• not impl. = possible, but not implemented yet
• no = not feasible with this approach
• ? = feasibility has not been evaluated yet
• † H,M->Hp 0->Mp means HOST transactions/spans (H) and MICRO transactions/spans (M) go into Host-project, nothing (0) send to Micro-project (Mp)

* Source mapping (lib)

(this is not needed for true, remote, MFEs) A lib-type micro can potentially have multiple host applications consuming it. Each host might use a different minification algorithm, serve micro code at different URL path or even bundle it together with other code into one big .js file. Naturally it is the responsibility of the host team to upload their source mappings during their build process, because micro team simply doesn't possess the information to generate those mappings. Source maps are associated with and uploaded for each individual release, each file can have only one mapping in a given release. This leaves room for a few options:

• Option 1 (recommended - less things can go wrong)
  • Each host build process registers its own releases of micro using Sentry API in the following format v2.0.1.host1. Then micro team can filter issues with a glob in the Sentry UI v2.0.1.*.
  • Look into dist flag: docs/sourcemaps/troubleshooting and docs/cli/releases/--dist
• Option 2
  • Component owners (teams) make a contract to serve micro at a defined URL path, separate from all other code.
  • Additionally (assuming build toolchain supports this) micro ships pre-transpiled/minified.
  • micro team uploads their own source mappings.

** No host Sentry

The problem here is how do you know for sure that host is never loading Sentry or just didn't have a chance yet. You can have a timeout but then either delay micro's widgets own initialization or miss on reporting errors that occur while you waiting. In flex-micro.js it is handled by patching temporary queueing handlers but at that point you could as well implement the entire flex-micro.

WrapALL method

This method may be the best choice for 3premote use case, because of its simplicity and the fact that it works without any change to the host-application code. The idea is to simply wrap all of micro's entry points (including event handlers, anonymous function callbacks, etc.) in try-catch statements and then use a separate instance of Sentry client to report the errors to the right DSN/project.

• Code snippet below assume that all of micro's host-applications use Sentry and that it's initialized at the time of micro's initialization. If those assumptions don't always hold it may be necessary to put some if-statements to check and either wait for the host-sentry to be initialized or dynamically load Sentry SDK through injecting a script element.

var sentry_micro_client = new Sentry.BrowserClient({
 dsn: "https://[email protected]/12345" 
 release: "[email protected]",
 transport: ("fetch" in window ? Sentry.makeFetchTransport : Sentry.makeXHRTransport),
 integrations: []
}

var sentry_wrap = function(callback) {
	return () => {
		try {
			callback();
		} catch (e) {
			sentry_micro_client.captureException(e);
			// throw e; // if desired, host-sentry won't report it again
		}
	}
}

/* Component's entry point */
sentry_wrap(micro_widget_init)();

my_element.addEventListener('click', sentry_wrap(my_click_event_handler));

window.setTimeout(sentry_wrap(() => {
	// my code here
	// more code
}), 1000);

var req = new XMLHttpRequest();
req.addEventListener("load", sentry_wrap(() => {
	// my code here
	// more code	
}));
req.open("GET", "http://www.example.org/example.txt");
req.send();

*** Flex: minimal error leakage

In one unlikely circumstance flex-micro.js will leak the very first error into host project/DSN. This will happen if (1) at the time of micro's initialization host-Sentry has not only not been initialized yet but the SDK hasn't even been loaded. In that situation whe can not detect the exact moment Sentry.init() is called and only detect it when 1 error may have already been incorrectly reported to host. (TODO: would it be possible to patch onload events of all <script> elements on the page to detect that?).

Fundamental technical challenges

TLDR: You catch unhandled error from an event handler function. How do you know which FE/widget/micro-component added that event handler, and therefore, which Sentry project to send this error to?

The nature of Javascript/browser environment presents significant obstacles to implementing first class support of MFEs in Sentry SDK.

1. Javascript's concurrency model (single thread event loop) makes it very difficult for Sentry Javascript SDK to maintain 'current hub' state. (e.g. client code async waiting inside Hub.run()).

2. Sentry's reliance on global event handlers for catching errors in UI and async callbacks (error, unhandledrejection) as well as auto-instrumenting certain things, for example XHR breadcrumbs. Most wrapper workarounds (e.g. mlmmn or ScriptedAlchemy) forget that part and consequently only correctly route errors originating in the code that was explicitly wrapped (see WrapALL). That eliminates one of the big benefits of Sentry - not having to manually instrument things.

   2.1. When using frameworks, e.g. React, errors in component callbacks can be correctly captured using the framework facilities, e.g. React error boundaries. However a lot of the event-based asyncronous code will still rely on globabl event handlers and therefore escape the "walls" of the framework (except in Angular/Zone.js).

3. Loss of function/file -> Component mapping and function names during build process (lib architecture).

Sandbox

A tool for manual end-to-end testing of Sentry in micro-frontend environment. Currently offers a basic DIY micro-frontend "framework" with 1 micro and 1 host application.

Sandbox how-to

To try out the sandbox:

git clone [email protected]:realkosty/sentry-micro-frontend.git
cd sentry-micro-frontend
npm install
node index.js

Then rename env.js.example to env.js, fill in your DSNs, releases and project links.

Finally, open http://localhost:8000/

Sandbox tips

Sandbox sets mv tag on all events sent to Sentry which is <module>@<SDK version>, for example:

mv:[email protected]

Footnotes

1. Since this is a new and evolving space we try, whenever possible, to provide solutions based on basic principles that work regardless of the composition technology. ↩