Skip to content
Source
Sign inSign up
TanStack/router

[Start] vite build keeps every environment's server-fn compiler alive until the build ends (memory stacks across client β†’ ssr β†’ server) #8794

Send to my agent
Openkmkwagner-anlargroup(GitHub profile, opens in a new tab) opened this issue 0 comments

Which project does this relate to?

Start

Describe the bug

In a production vite build, startCompilerPlugin keeps one compiler per environment in the compilers Map (packages/start-plugin-core/src/vite/start-compiler-plugin/plugin.ts). In build mode an entry is only removed in buildStart of the same environment, so it is never removed during a normal build: the client environment's compiler stays reachable while ssr and the server/provider environment build, and the ssr compiler stays reachable while the server environment builds.

Each compiler holds:

  • moduleCache: per-module info whose bindings keep Babel AST nodes (init: declarator.init in extractModuleInfoFromAst, @tanstack/router-utils);
  • loadModule / resolveId closures over the transform hook's this. The comment in buildStart already notes this ("The compiler host closes over this hook context for load/resolve"). That context keeps the finished phase's plugin context, and through it its module graph, reachable.

So live heap only grows from one environment to the next. In a large app this is the difference between fitting and not fitting under --max-old-space-size (we hit "JavaScript heap out of memory" in CI at a 6 GB cap).

Suggested fix: also drop the compiler when its environment's build finishes:

buildEnd() {
  if (this.environment.mode === 'build') {
    compilers.delete(this.environment.name)
  }
},

In build mode the compiler is only used from transform (hotUpdate, watchChange and the lookup-capture plugin are serve-only), and Rollup/Rolldown's buildEnd runs after the last transform and this.load, so nothing uses it afterwards. Happy to open a PR with this and a changeset if that helps.

Reproducer

The app where we measured this is private, so I can't share it. The retention follows directly from the source on main: https://github.com/TanStack/router/blob/main/packages/start-plugin-core/src/vite/start-compiler-plugin/plugin.ts (the compilers Map and buildStart) Any Start app with enough modules shows it with node --trace-gc node_modules/vite/bin/vite.js build: the heap after each full GC keeps rising from the client build into ssr and the server build. I'm glad to try a patch or a canary on our build and report back.

Complete minimal reproducer

https://github.com/TanStack/router/blob/main/packages/start-plugin-core/src/vite/start-compiler-plugin/plugin.ts (our app is private: see "Reproducer" at the end of the description)

Steps to Reproduce the Bug

  1. In a large TanStack Start app (ours: ~5,500 client modules, ~5,800 server modules), run node --max-old-space-size=6144 --trace-gc node_modules/vite/bin/vite.js build.
  2. Note the heap after each Mark-Compact at the end of each environment's build.
  3. Apply the buildEnd deletion above to start-compiler-plugin/plugin.js and build again.

Measured (heap after full GC, MB; Linux x64, Node 24.14, 4 vCPU, same machine, back to back):

buildend of clientend of ssrend of server (nitro)max live heappeak heaptime
unpatched1,8092,7185,9305,9305,974191 s
patched1,4562,2234,7094,7105,030163 s
patched, --max-old-space-size=51201,4562,2184,7024,7034,771170 s (completes)

About 1.2 GB less live heap (βˆ’21%). The unpatched build cannot complete under a 5 GB cap. Of the 973 files in .output, 972 were byte-identical between the unpatched and patched builds; the one that differed (server/index.mjs) embeds build timestamps.

Expected behavior

Once an environment's build is done, its server-fn compiler (module cache, AST nodes and the captured plugin context) is released, so peak memory is set by the largest single environment rather than the sum of all of them.

Screenshots or Videos

No response

Platform

  • Router / Start Version: @tanstack/react-start 1.168.60, @tanstack/start-plugin-core 1.171.49 (same code in 1.171.55 and on main)
  • OS: Linux x64
  • Browser: n/a (build-time)
  • Browser Version: n/a
  • Bundler: vite (+ nitro 3.0.260603-beta)
  • Bundler Version: 7.3.6

Additional context

Even with the fix, about 2.2 GB is still live when the server environment starts. A likely second holder, which I haven't measured: the import-protection plugin's per-environment transformResultCache, which is only cleared at that environment's buildStart, so the ssr cache also lives through the server build. It's related to #8614, which is about dev rather than build. Happy to measure that one too.

Sign in to join this conversation.