Basically, both services foo
and bar
depend on a common
library.
Let's assume that the common
package has already been published to the npm registry.
|
├── packages
| ├── common
| | ├── src
| | ├── package.json
| | ├── tsconfig.build.json
| | ├── tsconfig.json
| ├── foo
| | ├── src
| | ├── Dockerfile
| | ├── package.json
| | ├── tsconfig.build.json
| | ├── tsconfig.json
| ├── bar
| | ├── src
| | ├── Dockerfile
| | ├── package.json
| | ├── tsconfig.build.json
| | ├── tsconfig.json
├── tsconfig.json
├── package.json
├── yarn.lock
├── docker-compose.init.yml
├── docker-compose.yml
├── Dockerfile
├── Dockerfile.init
├── .dockerignore
I've added all devDependencies that are common to all of the packages in the root package.json, as follows:
"scripts": {
"build": "lerna run build --stream",
"setup": "yarn && yarn build",
"docker:bootstrap": "docker-compose --file=docker-compose.init.yml build",
"docker:up": "docker-compose up --build"
},
"devDependencies": {
"@nestjs/cli": "^7.5.1",
"@nestjs/common": "^7.4.4",
"@nestjs/core": "^7.4.4",
"@nestjs/platform-express": "^7.4.4",
"@nestjs/schematics": "^7.1.2",
"@nestjs/testing": "^7.4.4",
"@types/express": "^4.17.8",
"@types/jest": "^26.0.13",
"@types/node": "^14.10.2",
"@types/supertest": "^2.0.10",
"@typescript-eslint/eslint-plugin": "^4.1.1",
"@typescript-eslint/parser": "^4.1.1",
"eslint": "^7.9.0",
"eslint-config-prettier": "^6.11.0",
"eslint-plugin-prettier": "^3.1.4",
"express": "^4.17.1",
"husky": "^4.3.0",
"jest": "^26.4.2",
"lerna": "^3.22.1",
"lint-staged": "^10.4.0",
"prettier": "^2.1.2",
"reflect-metadata": "^0.1.13",
"rimraf": "^3.0.2",
"rxjs": "^6.6.3",
"supertest": "^4.0.2",
"ts-jest": "^26.3.0",
"ts-loader": "^8.0.3",
"ts-node": "^9.0.0",
"typescript": "3.9.5"
}
The foo
package needs to use a relational database, so I've installed the following packages independently.
$ yarn workspace foo add @nestjs/typeorm mysql typeorm
To resolve an error message "<package> has unmet peer dependency <package>"
, I've hit the following command.
$ yarn workspace foo add @nestjs/common @nestjs/core @nestjs/platform-express rxjs
I'm kinda lost here. What's the point of organizing multiple applications in a monorepo manner if I keep repeating myself installing the same packages from one package to another? It makes me even more difficult to write a Dockerfile after all.
My first question is, when a developer working on a monolithic codebase, is this a normal behavior to install libraries to a specific package if necessary?
This is how my Dockerfile looks like:
// docker-compose.init.yml
# This file triggers the initial build
version: "3.8"
services:
pkg_builder:
image: pkg-builder
build:
context: .
dockerfile: Dockerfile.init
First, execute the command below.
$ yarn docker:bootstrap
Dockerfile.init
creates an initial builder image from which the “real” builder image can copy the build directory.
// Dockerfile.init
FROM scratch
# Copy files from the root to build directory
COPY package.json lerna.json yarn.lock tsconfig.json /build/
# This line is required to install dependencies from foo's package.json
COPY ./packages/foo/package.json /build/packages/foo/package.json
From then on, build the images using the command:
$ yarn docker:up
// docker-compose.yml
version: "3.8"
services:
pkg_builder:
image: pkg-builder
build: .
mariadb:
image: mariadb:10.3
ports:
- "3306:3306"
environment:
- MYSQL_USER=root
- MYSQL_ROOT_PASSWORD=root
- MYSQL_DATABASE=tutorial
restart: always
foo:
container_name: foo
build: ./packages/foo
ports:
- "8000:8000"
depends_on:
- mariadb
restart: always
// Dockerfile
FROM node:12-alpine
COPY --from=pkg-builder /build /build
WORKDIR /build
RUN rm -rf node_modules
RUN yarn
CMD ["true"]
The problem is, the size of an image way too big. This is because, all the dev dependencies have been copied from the pkg-builder
.
// foo's Dockerfile
FROM node:12-alpine
WORKDIR /app/current
COPY --from=pkg-builder /build/node_modules /app/current/packages/foo/node_modules
COPY --from=pkg-builder /build/tsconfig.json ./tsconfig.json
WORKDIR /app/current/packages/foo
COPY . .
RUN yarn build
EXPOSE 8000
CMD [ "node", "./dist/main" ]
Lastly, how am I supposed to reduce the image? In this case, I believe that multi-stage build is not the right strategy for the size reduction. What am I missing here?
Every
package.json
file needs to list the complete immediate dependencies of its application. I should be able to check out your source repository, runyarn install
, and have a working application tree. Where your question says "and by the way these other dependencies are installed in the environment and I just assume them", that's a problem for anyone who isn't working on the exact system you are, and it's more specifically a problem for Docker and other automated-build tools.Your library dependencies can have their own library dependencies. These will get listed out in the
yarn.lock
file, but they don't need to be directly listed in thepackage.json
file.Taking the database-access libraries as an example: if your main application uses them, they need to be included in your
dependencies
. But if all of the database access is encapsulated in yourcommon
shared library, your applications only need to refer to that library (infoo/package.json
), and the library needs to include the database dependencies (incommon/package.json
).You should split
dependencies
fromdevDependencies
. The things you need to run your application (express
) need to be listed independencies
; the things you only need to build your application (eslint
) should bedevDependencies
. You discuss image size; this gives you a way to install a much smaller group of packages in the container when you actually run it.(Note that Yarn doesn't actually support not installing
devDependencies
; npm does, though it's otherwise much slower to use.)Then a multi-stage build can produce a smaller image. The idea here is that the first stage installs the entire dev dependencies, and builds the application; the second stage only includes the runtime dependencies and the built code. This more or less looks like:
You shouldn't ever need a "builder container"; the setup you show is basically identical to a multi-stage build, except spread across three separate Dockerfiles. In particular if you have an image that doesn't run a command but just contains files, it's a good candidate to be an early stage in a multi-stage build.