9 November 2026 · 6 MIN READ
Docker image optimization: multi-stage builds, caching, and .dockerignore
Why images get bloated, and the four habits that make builds smaller, faster, and cheaper to ship.
Smaller images are not vanity
Image size is operational cost you pay forever: slower pulls on every deploy, slower pushes from CI, more attack surface, and registry storage bills that grow quietly.
The fix is rarely one clever trick. It is four habits, applied every time a Dockerfile is touched.
Multi-stage builds: leave the toolchain behind
Compilers, package managers, and build caches have no business inside a running container. Build in one stage, copy the artifact into a minimal runtime stage, and ship only that.
# Stage 1: build
FROM node:22 AS build
WORKDIR /app
COPY package*.json ./
RUN npm ci
COPY . .
RUN npm run build
# Stage 2: run
FROM node:22-slim
WORKDIR /app
COPY --from=build /app/dist ./
CMD ["node", "server.js"]Order layers for cache hits
Docker caches layer by layer. Copying dependency manifests and installing packages before copying application source means an everyday code change rebuilds one fast layer instead of re-downloading everything.
The same reasoning applies in CI: build cache mounts and registry cache imports keep the pipeline from paying full price on every push.
# Cache-hostile
COPY . .
RUN npm ci
# Cache-friendly
COPY package*.json ./
RUN npm ci
COPY . ..dockerignore is a security file too
Most people learn .dockerignore as a size feature. It is also a safety feature: without it, .git history, .env files, and local node_modules ride along into the build context, and anything COPY . . picks up can end up in the image.
A minimal ignore file excludes all three, plus documentation and test fixtures that never belong in a runtime image.
.git
.env*
node_modules
README*
tests/Tag images the way you will need to find them
Every image gets the commit SHA as a tag, so the running container traces back to its source. Release versions get an explicit tag too. The mutable “latest” tag is a last-resort fallback, never the thing a deployment points at.
A tag you cannot trace is a rollback you cannot perform — and rolling back at 2am is the moment this habit earns its keep.