opensource/sim
3
0
Fork 0
mirror of https://github.com/simstudioai/sim.git synced 2026-09-24 15:45:35 +08:00
Code Issues Packages Projects Releases Wiki Activity
Files
7c05e36049ae837afe45b830dd1ccdcaae56e68b
sim/docker/app.Dockerfile
T
Waleed 9ce3f51946 perf(ci): build the app image on 16 vCPU and keep the toolchain out of it (#6500) 
Three independent changes to the app image build.

The app image builds on 16 vCPU instead of 8. Its dominant steps -- bun install
(~300-465s) and next build (~262s) -- both scale with cores, and this build gates
every deploy because nothing ships until the image is pushed. The same next build
already runs on 16 vCPU in the Build App verification job, which gates nothing.
The isolated-vm rebuild will not benefit: JOBS=4 caps it deliberately to avoid an
OOM. ARM64 stays at 8 -- that job is off the deploy path, and the workflow warns
an unprovisioned runner label hangs a release in queued.

node-gyp comes from the lockfile rather than npx. npx resolved it from the
registry during every production image build, which pulled 13.x over the pinned
12.4.0 and bypassed the minimumReleaseAge gate in bunfig.toml. It has to be an
apps/sim devDependency because the only other copy is transitive through
@electron/rebuild, which turbo prune sim strips.

base splits into base and build-base. The compiler toolchain exists only to build
isolated-vm against Node and the runner copies the finished binary, so shipping
it inflated every ECS task pull: measured 1.21 GB for base against 1.6 GB for
build-base, so ~390 MB stays out of the final image. ffmpeg and python3 stay --
fluent-ffmpeg is a serverExternalPackage, and node-gyp needs the interpreter.

Verified against real image builds rather than by inspection: g++/make/pip3 are
absent from the runtime image while node/bun/python3/ffmpeg remain, the
isolated-vm native module loads and evaluates, lib0/yjs/y-protocols including the
lib0/logging.js subpath are intact, sharp and libvips load and encode a PNG so
#6499's fix still works against the slimmed image, the container boots to Next.js
Ready, and the full suite passes 21250 tests through the dependency hoisting
reshuffle that promoting node-gyp causes (root which@6.0.1, fluent-ffmpeg keeps
its own 1.3.1).

Not included: an earlier attempt added a cache-key input to
useblacksmith/setup-docker-builder, on the theory that the deps layer re-runs
every build (286s/465s observed) because all five Dockerfiles share one layer
cache. That input does not exist at the pinned SHA or on main, and the key is not
configurable -- setup_builder.ts uses GITHUB_REPO_NAME. CI logs also show the
sticky disk being obtained with a parent snapshot rather than falling back, so
the cache is present and that cause remains unidentified.
2026-08-10 12:42:13 -07:00

192 lines
9.1 KiB
Docker
Raw Blame History

# ========================================
# Base Stage: runtime-only dependencies (inherited by the final image)
# ========================================
FROM oven/bun:1.3.14-slim AS base
# Install Node.js 24 (Active LTS) and the runtime dependencies once in base.
# Node runs only the isolated-vm sandbox worker (the app itself runs under Bun);
# the version is kept in lockstep with the `isolated-vm` pin in
# apps/sim/package.json — Node 24 (ABI 137) requires isolated-vm 6.x.
#
# Only what the running container needs belongs here. ffmpeg backs the
# `fluent-ffmpeg` serverExternalPackage; python3 is the node-gyp interpreter and
# is kept because build-base inherits from this stage. The compiler toolchain
# lives in build-base so the runner does not ship it.
RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
--mount=type=cache,target=/var/lib/apt,sharing=locked \
apt-get update && apt-get install -y --no-install-recommends \
python3 curl ca-certificates bash ffmpeg \
&& curl -fsSL https://deb.nodesource.com/setup_24.x | bash - \
&& apt-get install -y nodejs
# ========================================
# Build Base: adds the native toolchain the isolated-vm rebuild needs
# ========================================
FROM base AS build-base
# The compiler toolchain, needed only to build isolated-vm against Node. The
# runner copies the finished binary from deps, so shipping these would inflate
# every ECS task pull for nothing: measured 1.21 GB for base against 1.6 GB for
# build-base, so ~390 MB stays out of the final image.
RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
--mount=type=cache,target=/var/lib/apt,sharing=locked \
apt-get update && apt-get install -y --no-install-recommends \
python3-pip python3-venv make g++
# ========================================
# Pruner Stage: Emit a minimal monorepo subset that sim depends on
# ========================================
FROM build-base AS pruner
WORKDIR /app
RUN bun install -g turbo@2.9.6
COPY . .
RUN turbo prune sim --docker
# ========================================
# Dependencies Stage: Install Dependencies
# ========================================
FROM build-base AS deps
WORKDIR /app
# Pruned manifests from the pruner stage. This layer only invalidates when
# package.json/bun.lock content changes — not on source edits.
COPY --from=pruner /app/out/json/ ./
# Use the full bun.lock (not the pruned out/bun.lock). turbo prune emits a
# bun.lock that bun 1.3.x rejects with "Failed to resolve prod dependency",
# forcing a slow fresh resolve. The full lockfile parses cleanly and bun
# only installs what the pruned package.jsons reference.
COPY --from=pruner /app/bun.lock ./bun.lock
# Install all dependencies (including devDependencies — tailwindcss/postcss are
# devDeps but required at build time). Then rebuild isolated-vm against Node.js.
# JOBS=4 caps node-gyp parallelism — higher values OOM isolated-vm (laverdet/isolated-vm#428).
#
# node-gyp comes from the lockfile, not `npx`. It is a devDependency of apps/sim
# purely so `turbo prune sim` keeps it: the only other copy is transitive through
# `@electron/rebuild`, which belongs to apps/desktop and is pruned away. `npx`
# resolved it from the registry at build time, which pulled a different major
# (13.x vs the pinned 12.4.0) and bypassed the `minimumReleaseAge` supply-chain
# gate in bunfig.toml on every production image build.
RUN --mount=type=cache,id=bun-cache,target=/root/.bun/install/cache \
--mount=type=cache,id=npm-cache,target=/root/.npm \
HUSKY=0 bun install --ignore-scripts --linker=hoisted && \
cd node_modules/isolated-vm && JOBS=4 /app/node_modules/.bin/node-gyp rebuild --release
# ========================================
# Builder Stage: Build the Application
# ========================================
FROM build-base AS builder
ARG TARGETPLATFORM
WORKDIR /app
# Copy node_modules from deps stage (cached if dependencies don't change)
COPY --from=deps /app/node_modules ./node_modules
# Copy pruned source tree (apps/sim + workspace packages it depends on)
COPY --from=pruner /app/out/full/ ./
# Next.js 16 / Turbopack workspace-root detection looks for a lockfile next to
# the workspace package.json. Without it, `next build` fails with
# "couldn't find next/package.json from /app/apps/sim". turbo also warns
# "Lockfile not found at /app/bun.lock" without it.
COPY --from=pruner /app/bun.lock ./bun.lock
ENV NEXT_TELEMETRY_DISABLED=1 \
VERCEL_TELEMETRY_DISABLED=1 \
DOCKER_BUILD=1
# Dummy values so next build can evaluate modules. Override at runtime.
ARG DATABASE_URL="postgresql://user:pass@localhost:5432/dummy"
ENV DATABASE_URL=${DATABASE_URL}
ARG NEXT_PUBLIC_APP_URL="http://localhost:3000"
ENV NEXT_PUBLIC_APP_URL=${NEXT_PUBLIC_APP_URL}
# Per-platform cache id keeps arm64/amd64 SWC artifacts isolated.
RUN --mount=type=cache,id=next-cache-${TARGETPLATFORM},target=/app/apps/sim/.next/cache \
--mount=type=cache,id=turbo-cache-${TARGETPLATFORM},target=/app/.turbo \
bun run build
# Bundle the secrets-loading bootstrap into a self-contained entrypoint. It runs
# before (and outside) the Next standalone server, so its dependencies
# (@sim/runtime-secrets, AWS SDK) are inlined here rather than resolved from the
# pruned standalone node_modules. The dynamic import of ./server.js stays a
# runtime import.
RUN bun build apps/sim/bootstrap.ts --target=bun --outfile=apps/sim/bootstrap.js
# ========================================
# Runner Stage: Run the actual app
# ========================================
FROM base AS runner
WORKDIR /app
# Node.js 24, Python, ffmpeg, etc. are already installed in base stage
ENV NODE_ENV=production
# Create non-root user and group
RUN groupadd -g 1001 nodejs && \
useradd -u 1001 -g nodejs nextjs
# Copy application artifacts from builder
COPY --from=builder --chown=nextjs:nodejs /app/apps/sim/public ./apps/sim/public
COPY --from=builder --chown=nextjs:nodejs /app/apps/sim/.next/standalone ./
COPY --from=builder --chown=nextjs:nodejs /app/apps/sim/.next/static ./apps/sim/.next/static
# Self-contained secrets-loading bootstrap (bundled in the builder stage). Runs
# before the standalone server.js to hydrate process.env from the runtime secret.
COPY --from=builder --chown=nextjs:nodejs /app/apps/sim/bootstrap.js ./apps/sim/bootstrap.js
# Copy blog/author content for runtime filesystem reads (not part of the JS bundle)
COPY --from=builder --chown=nextjs:nodejs /app/apps/sim/content ./apps/sim/content
# Copy isolated-vm native module (compiled for Node.js in deps stage)
COPY --from=deps --chown=nextjs:nodejs /app/node_modules/isolated-vm ./node_modules/isolated-vm
# The collab-doc seed/merge/persist routes run the converter (markdown <-> Yjs) server-side. `yjs` is a
# serverExternalPackage, and the Next standalone tracer copies it only partially — it misses ESM subpath
# files that `yjs/dist/yjs.mjs` imports through `lib0`'s exports map (e.g. `lib0/logging`), so the seed
# 500s ("Cannot find module 'lib0/logging'") and every collaborative doc is stuck read-only. Overwrite
# the partial trace with the complete packages from the full install (outputFileTracingIncludes can't:
# its globs resolve against apps/sim, but these deps hoist to the monorepo-root node_modules).
COPY --from=deps --chown=nextjs:nodejs /app/node_modules/lib0 ./node_modules/lib0
COPY --from=deps --chown=nextjs:nodejs /app/node_modules/yjs ./node_modules/yjs
COPY --from=deps --chown=nextjs:nodejs /app/node_modules/y-protocols ./node_modules/y-protocols
# `@img/sharp-<platform>` loads libvips from `@img/sharp-libvips-<platform>` through the dynamic
# linker, not a JS require, so the tracer copies the binding but not the library and sharp dies with
# "ERR_DLOPEN_FAILED: libvips-cpp.so: cannot open shared object file". Same hoisting reason as the Yjs
# stack above. Copying whole directories keeps these arch-agnostic (each build's deps stage holds only
# its own platform's packages) and keeps sharp and its binding on the same install. Must stay below
# the standalone COPY, which ships its own partial node_modules that would otherwise win.
COPY --from=deps --chown=nextjs:nodejs /app/node_modules/sharp ./node_modules/sharp
COPY --from=deps --chown=nextjs:nodejs /app/node_modules/@img ./node_modules/@img
# Copy the isolated-vm worker script
COPY --from=builder --chown=nextjs:nodejs /app/apps/sim/lib/execution/isolated-vm-worker.cjs ./apps/sim/lib/execution/isolated-vm-worker.cjs
# Copy the pre-built sandbox library bundles (pptxgenjs, docx, pdf-lib) that
# run inside the V8 isolate. Committed into the repo; see
# apps/sim/lib/execution/sandbox/bundles/build.ts to regenerate.
COPY --from=builder --chown=nextjs:nodejs /app/apps/sim/lib/execution/sandbox/bundles ./apps/sim/lib/execution/sandbox/bundles
# Guardrails PII runs in a standalone Presidio service (combined analyzer +
# anonymizer, docker/pii.Dockerfile), reached over the network via PII_URL —
# no Python/Presidio in this image.
# Create .next/cache directory with correct ownership
RUN mkdir -p apps/sim/.next/cache && \
chown -R nextjs:nodejs apps/sim/.next/cache
# Switch to non-root user
USER nextjs
EXPOSE 3000
ENV PORT=3000 \
HOSTNAME="0.0.0.0"
CMD ["bun", "apps/sim/bootstrap.js"]
Reference in New Issue View Git Blame Copy Permalink