Deploying with Docker
Building a Docker (opens in a new tab) image is a common way to deploy all sorts of applications. However, doing so from a monorepo has several challenges.
The problem
TL;DR: In a monorepo, unrelated changes can make Docker do unnecessary work when deploying your app.
Let's imagine you have a monorepo that looks like this:
├── apps
│ ├── docs
│ │ ├── server.js
│ │ └── package.json
│ └── web
│ └── package.json
├── package.json
└── package-lock.json
You want to deploy apps/docs
using Docker, so you create a Dockerfile:
FROM node:16
WORKDIR /usr/src/app
# Copy root package.json and lockfile
COPY package.json ./
COPY package-lock.json ./
# Copy the docs package.json
COPY apps/docs/package.json ./apps/docs/package.json
RUN npm install
# Copy app source
COPY . .
EXPOSE 8080
CMD [ "node", "apps/docs/server.js" ]
This will copy the root package.json
and the root lockfile to the docker image. Then, it'll install dependencies, copy the app source and start the app.
You should also create a .dockerignore
file to prevent node_modules from being copied in with the app's source.
node_modules
npm-debug.log
The lockfile changes too often
Docker is pretty smart about how it deploys your apps. Just like Turbo, it tries to do as little work as possible (opens in a new tab).
In our Dockerfile's case, it will only run npm install
if the files it has in its image are different from the previous run. If not, it'll restore the node_modules
directory it had before.
This means that whenever package.json
, apps/docs/package.json
or package-lock.json
change, the docker image will run npm install
.
This sounds great - until we realise something. The package-lock.json
is global for the monorepo. That means that if we install a new package inside apps/web
, we'll cause apps/docs
to redeploy.
In a large monorepo, this can result in a huge amount of lost time, as any change to a monorepo's lockfile cascades into tens or hundreds of deploys.
The solution
The solution is to prune the inputs to the Dockerfile to only what is strictly necessary. Turborepo provides a simple solution - turbo prune
.
turbo prune docs --docker
Running this command creates a pruned version of your monorepo inside an ./out
directory. It only includes workspaces which docs
depends on.
Crucially, it also prunes the lockfile so that only the relevant node_modules
will be downloaded.
The --docker
flag
By default, turbo prune
puts all relevant files inside ./out
. But to optimize caching with Docker, we ideally want to copy the files over in two stages.
First, we want to copy over only what we need to install the packages. When running --docker
, you'll find this inside ./out/json
.
out
├── json
│ ├── apps
│ │ └── docs
│ │ └── package.json
│ └── package.json
├── full
│ ├── apps
│ │ └── docs
│ │ ├── server.js
│ │ └── package.json
│ ├── package.json
│ └── turbo.json
└── package-lock.json
Afterwards, you can copy the files in ./out/full
to add the source files.
Splitting up dependencies and source files in this way lets us only run npm install
when dependencies change - giving us a much larger speedup.
Without --docker
, all pruned files are placed inside ./out
.
Example
Our detailed with-docker
example (opens in a new tab) goes into depth on how to utilise prune
to its full potential. Here's the Dockerfile, copied over for convenience.
This Dockerfile is written for a Next.js (opens in a new tab) app that is using the standalone
output mode (opens in a new tab).
FROM node:18-alpine AS base
FROM base AS builder
RUN apk add --no-cache libc6-compat
RUN apk update
# Set working directory
WORKDIR /app
RUN yarn global add turbo
COPY . .
RUN turbo prune web --docker
# Add lockfile and package.json's of isolated subworkspace
FROM base AS installer
RUN apk add --no-cache libc6-compat
RUN apk update
WORKDIR /app
# First install the dependencies (as they change less often)
COPY .gitignore .gitignore
COPY --from=builder /app/out/json/ .
COPY --from=builder /app/out/yarn.lock ./yarn.lock
RUN yarn install
# Build the project
COPY --from=builder /app/out/full/ .
RUN yarn turbo run build --filter=web...
FROM base AS runner
WORKDIR /app
# Don't run production as root
RUN addgroup --system --gid 1001 nodejs
RUN adduser --system --uid 1001 nextjs
USER nextjs
COPY --from=installer /app/apps/web/next.config.js .
COPY --from=installer /app/apps/web/package.json .
# Automatically leverage output traces to reduce image size
# https://nextjs.org/docs/advanced-features/output-file-tracing
COPY --from=installer --chown=nextjs:nodejs /app/apps/web/.next/standalone ./
COPY --from=installer --chown=nextjs:nodejs /app/apps/web/.next/static ./apps/web/.next/static
COPY --from=installer --chown=nextjs:nodejs /app/apps/web/public ./apps/web/public
CMD node apps/web/server.js
Remote caching
To take advantage of remote caches during Docker builds, you will need to make sure your build container has credentials to access your Remote Cache.
There are many ways to take care of secrets in a Docker image. We will use a simple strategy here with multi-stage builds using secrets as build arguments that will get hidden for the final image.
Assuming you are using a Dockerfile similar to the one above, we will bring in some environment variables from build arguments right before turbo build
:
ARG TURBO_TEAM
ENV TURBO_TEAM=$TURBO_TEAM
ARG TURBO_TOKEN
ENV TURBO_TOKEN=$TURBO_TOKEN
RUN yarn turbo run build --filter=web...
turbo
will now be able to hit your remote cache. To see a Turborepo cache hit for a non-cached Docker build image, run a command like this one from your project root:
docker build -f apps/web/Dockerfile . --build-arg TURBO_TEAM=“your-team-name” --build-arg TURBO_TOKEN=“your-token“ --no-cache