- Introduction
- Features
- Install
- Usage
- Workspace
- Deployment
- The "built packages" strategy
- The "internal packages" strategy
- Live code changes from internal packages
- Firebase
There is an accompanying article "My quest for the perfect TS monorepo" that you might want to read for context.
This is a personal quest for the perfect Typescript monorepo setup.
It is the best I could come up with given the tooling that is available, so expect this repository to change over time as the ecosystem around Typescript evolves.
My current projects are based on Node.js, Next.js, and Firebase, so that is what I am focussing on primarily. If you use different a different stack, I believe this can still be a great reference, as the approach itself does not depend on it.
Contributions and suggestions are welcome within the scope of this example, but I doubt there ever will be a one-size-fits-all solution, so this code should be viewed as opinionated.
I ended up basing a lot of things on the Turborepo starter, and I recommend reading their monorepo handbook.
For demonstration purposes, mono-ts uses the "internal packages approach" for
@repo/common
and a traditional built approach for@repo/backend
. Read below for more info.
- Turborepo to orchestrate the build process and dependencies, including the v2 watch task.
- Showcasing a traditional "built package" with multiple entry points, as well as the "internal package" strategy referencing Typescript code directly.
- Multiple isolated Firebase deployments, using firebase-tools-with-isolate
- Firebase emulators with hot reloading
- A web app based on Next.js with ShadCN and Tailwind CSS
- Working IDE go-to-definition and go-to-type-definition using
.d.ts.map
files - ES modules for everything
- Path aliases
- Shared configurations for ESLint
- Simple standardized configuration for TypeScript
- Vitest
In the main branch of this repo, packages are managed with PNPM.
There is also a branch for NPM
Originally, I included branches for Yarn classic (v1), and modern (v4), but I stopped updating them as Yarn is not that commonly used anymore.
I recommended using pnpm
over npm
or yarn
. Apart from being fast and
efficient, I believe PNPM has better support for monorepos.
You can install PNPM with corepack
which is part of modern Node.js versions:
corepack enable
(if you have not used it before)corepack prepare pnpm@latest --activate
Then run pnpm install
from the repository root.
There are 4 commands to run in separate processes:
- From the monorepo root, run
pnpm watch
. This builds all the code (except the web app) using the Turborepo watch task. - From
apps/web
runpnpm dev
. This start the Next.js dev server and builds its pages on request. - From
services/api
runpnpm emulate
. This starts the emulators for the API server. - From
services/fns
runpnpm emulate
. This starts the emulators for the (other) Firebase functions.
The web app should become available on http://localhost:3000 and the emulators UI on http://localhost:4000.
You should now have a working local setup, in which code changes to any package are picked up.
Note that hot-reloading for the firebase packages like API are not as instant as you are used to with front-end tooling like Next.js. When code changes are detected, the isolate process needs to run again to compile new output and the function needs to reload.
Typically in a monorepo, you will never publish the packages to NPM, and because of that, the namespace you use to prefix your package names does not matter. You might as well pick a generic one that you can use in every private codebase.
At first I used @mono
, and later I switched to @repo
when I discovered that
in the Turborepo examples. I like both, because they are equally short and
clear, but I went with @repo
because I expect it will become the standard.
- common Code that can shared across both front-end and back-end environments.
- backend Code that is shared between server environments like cloud functions.
- web A Next.js based web application configured to use Tailwind CSS and ShadCN components.
- fns Various Firebase functions that execute on document writes, pubsub events etc. This package shows how to use [isolate-package] explicitly as part of the predeploy phase.
- api A 2nd gen Firebase function (based on Cloud Run) serving
as an API endpoint, using Express. This package shows how to use
firebase-tools-with-isolate
to have the isolation integrated as part of the
firebase deploy
command. In addition, it illustrates how to use secrets.
I consider deployment a bit out-of-scope for this demo.
For deployment to Firebase, you will have to set up and configure an actual project, but it is not required to run this demo since by default it runs on local emulators. Additional info about the use of isolate-package (used by fns) and firestore-tools-with-isolate (used by api) can be found in the instructions for each package.
In a traditional monorepo setup, each package exposes its code as if it was a
published NPM package. For Typescript this means the code has to be transpiled
and the manifest entry points reference to the build output files. You can use
Typescript tsc
compiler for this, but it is likely you will want to use a
bundler for the reasons explained below.
The services
in this codebase use TSUP as a bundler. It is a modern, simple to
use Rollup-inspired bundler for Typescript.
The advantages of using a bundler are discussed below.
If you use path aliases like ~/*
or @/*
to conveniently reference top-level
folders from deeply nested import statements, these paths are not converted by
the standard Typescript tsc
compiler.
A bundler will typically remove path aliases, because it combines your code into self-contained files that do not import other local files themselves.
If you target platforms without using bundler, you will have to convert them.
You can run something like tsc-alias
after your build step. Note that path
aliases can also end up in d.ts
files.
A bundler will allow you to output ESM-compatible code without having to adhere
to the ESM import rules. ESM requires all relative imports to be explicit,
appending a .js
file extension for importing TS files and appending
/index.js
when importing from the root of a directory.
The reason you need to use .js
and not .ts
is because these imports, like
path aliases are not converted by the Typescript compiler, and so at runtime the
transpiled JS file is what is getting imported.
Because a bundler, by nature, will bundle code into one or more isolated files,
those files do not use relative imports and only contain imports from
node_modules
, which do not require a file extension. For this reason, a
bundled js file that uses import and export keywords is an ES module.
An advantage of writing your code as ESM is that you can import both ES modules and CommonJS without conversion. An application that uses CJS can not import ESM directly, because CJS imports are synchronous and ESM imports are asynchronous.
Not having to use ESM import extensions can be especially valuable if you are trying to convert a large codebase to ESM, because I have yet to find a solution that can convert existing imports. There is this ESLint plugin that you could use it in combination with the --fix flag to inject the extensions, but at the time of writing it does not understand path aliases.
Some bundlers like TSUP are capable of eliminating dead code by tree-shaking the build output, so that less code remains to be deployed. Eliminating dead code is most important for client-side code that is shipped to the user, but for the backend it can also reduce cold-start times for serverless functions, although in most situations, it is probably not going to be noticeable.
The internal packages strategy, as it was coined by Jared Palmer of Turborepo, removes the build step from the internal packages by linking directly to the Typescript source files in the package manifest.
There are some advantages to this approach:
- Code and type changes can be picked up directly, removing the need for a watch task in development mode.
- Removing the build step reduces overall complexity where you might otherwise use a bundler with configuration.
- IDE go-to-definition, in which cmd-clicking on a reference takes you to the
source location instead of the typed exports, works without the need for
Typescript project references or generating
d.ts.map
files.
But, as always, there are also some disadvantages you should be aware of:
- You can not publish the shared packages to NPM, as you do not expose them as Javascript.
- If you use path aliases like
~/
, you will need to make sure every package has its own unique aliases. You might not need aliases, because shared packages typically do not have a deeply nested folder structure anyway. - Since all source code gets compiled by the consuming application, build times can start to suffer when the codebase grows. See caveats for more info.
- The shared package is effectively just a source folder, and as a whole it
needs to be transpiled and bundled into the consuming package. This means that
its dependencies must also be available in the consuming package. Next.js can
do this for you with the
transpilePackage
setting, but this is the reasonservices/api
includesremeda
, as it is used bypackages/common
.
For testing and comparison, mono-ts uses the internal packages approach for
@repo/common
and a traditional built approach for @repo/backend
. Both are
compatible with isolate-package
for deploying to Firebase.
Traditionally in a monorepo, each package is treated similarly to a released NPM package, meaning that the code and types are resolved from the built "dist" output for each module. Adding new types and changing code in shared packages therefore requires you to rebuild these, which can be cumbersome during development.
Turborepo does not (yet) include a watch task, so Turbowatch was created to solve this issue. I haven't tried it but it looks like a neat solution. However, you might want to use the internal packages strategy instead.
Throughout this repository, we use a Firebase demo project called demo-mono-ts
A demo project allows you to run emulators for the different components like
database without creating a Firebase projects with resources. To make this work
you pass the --project
flag when starting the emulator, and you need to use a
name that starts with demo-
.
When passing configuration to initializeApp you can use any non-empty string for the API keys as you can see in apps/web/.env.development.
Deploying code to Firebase that uses shared packages from a monorepo comes with its own set of challenges, because the Firebase deploy pipeline requires you to upload a self-contained package that can be treated similarly to an NPM package, by installing its dependencies and executing the main entry.
This repo includes a solution based on isolate-package I wrote an article explaining what it does and why it is needed.
This demo can be run using only the emulators, but if you would like to see the
deployment to Firebase working you can simply execute
npx firebase deploy --project your-project-name
from any of the service
packages. For services/fns
this will trigger a deploy using isolate-package
and the standard firebase-tools
, and for services/api
this will invoke a
deploy using the
firestore-tools-with-isolate
fork where both are integrated.
You might notice @google-cloud/functions-framework
as a dependency in the
service package even though it is not being used in code imports. It is
currently required for Firebase to be able to deploy a PNPM workspace. Without
it you will get an error asking you to install the dependency. I don't quite
understand how the two are related, but it works.
For Firebase Functions each service (api and fns) start separate emulators on port 5001 and 5002. The backend service (using the firebase-admin api) connects to emulators by setting various environment variables.
I have stored these in .env
files in the respective service packages. Normally
you would want to store them in a file that is not part of the repository like
.env.local
but by placing them in .env
I prevent having to give instructions
for setting them up just for running the demo.
The api service uses a secret for DEMO_API_KEY. To make secrets work with the
emulator you currently have to add the secret to .secret.local
and also a
.env
or .env.local
file. See
this issue for more
info. I have placed it in .env
which is part of the repo, so you don't have to
set anything up, but .env.local is the proper location probably because that
file is not checked into git.