adapter-cloudflare 8: concurrent Vite servers each start a platform proxy, so Vitest projects collide on the same SQLite (SQLITE_BUSY_RECOVERY) #17344

Activity

1. svelte-triage-bot

   added

   triage-bot

   and removed

   triage-bot

   on Oct 4, 2026
2. teemingc

   added

   pkg:adapter-cloudflare

   on Oct 5, 2026
3. glw907 commented on Oct 6, 2026

   

   glw907

   on Oct 6, 2026

   More actions

   The same global also fails in the opposite direction: a second Vite server that never created the proxy still disposes it when it closes.

   In virtual_workers_module() (packages/adapter-cloudflare/index.js at 2ec690e), setup() returns early when globalThis.__sveltekit_cloudflare_platform is already set (line 226). The dispose that #17238 added (lines 233-237) runs from closeServer for every server (lines 242-245), whether or not that server's setup created the proxy. A plugin that starts and closes a short-lived second server during vite dev therefore disposes the proxy the main dev server still uses, and clears the global. A route that reads cloudflare:workers then fails.

   To reproduce, with @sveltejs/kit 3.0.1, @sveltejs/adapter-cloudflare 8.0.0, Vite 8.3.3, and wrangler 4.147.0:

   1. Scaffold an app with sv create app --template minimal --types jsdoc --no-add-ons, and pass adapter: adapter() to sveltekit() in vite.config.js.

   2. Add a wrangler.jsonc with main and assets set as the adapter docs describe, and "vars": { "FOO": "bar" }.

   3. Add src/routes/+page.server.js:

      import { env } from 'cloudflare:workers';
      export const load = () => ({ foo: env.FOO });

   4. Add this plugin after sveltekit():

      {
      	name: 'nested-server',
      	apply: 'serve',
      	async buildStart() {
      		if (process.env.NESTED) return;
      		process.env.NESTED = '1';
      		const { createServer } = await import('vite');
      		const server = await createServer({ server: { middlewareMode: true } });
      		await server.close();
      	}
      }

   5. Run vite dev and request /.

   The request returns a 500:

   Error: Cannot access cloudflare:workers in a prerenderable route
       at load (src/routes/+page.server.js:2:39)
   

   Without the plugin, / returns 200. With the plugin but without the server.close() call, it also returns 200, so the failure comes from the close.

   Both this case and the concurrent-start case come from the global holding one proxy with no record of which servers use it. I've opened #17368, which shares one pending proxy promise across servers, counts the servers using it, and disposes it when the last one closes.
4. imlargo commented on Oct 7, 2026

   

   imlargo

   on Oct 7, 2026

   More actions

   Same failure on a clean GitHub Actions runner, so journal recovery after an unclean shutdown is not the only trigger. ubuntu-latest, pnpm install --frozen-lockfile, no .wrangler/ directory, no bindings besides ASSETS, the two Vitest projects sv generates (client in Chromium, server in Node), then vitest --run:

   ⎯⎯⎯⎯⎯⎯⎯ Startup Error ⎯⎯⎯⎯⎯⎯⎯⎯
   AggregateError: Failed to initialize projects. There were errors during projects setup. See below for more details.
       MiniflareCoreError [ERR_RUNTIME_FAILURE]: The Workers runtime failed to start. There was likely a problem with the workerd binary or your configuration.
         code: 'ERR_RUNTIME_FAILURE',
   *** Fatal uncaught kj::Exception: workerd/util/sqlite.c++:1733: failed: SQLite failed; dbErrorMessage(err, db) = database is locked: SQLITE_BUSY
   

   Plain SQLITE_BUSY here, not SQLITE_BUSY_RECOVERY. Log: https://github.com/imlargo/svelte-template/actions/runs/37266332359 (Test job). @sveltejs/kit 3.0.0, @sveltejs/adapter-cloudflare 8.0.0, vitest 4.1.11, wrangler 4.147.0.