Rolldown CI Failures

[sheremet-va]

Describe the bug

Rolldown fails on CI for some time. For the longest time it was an issue with source maps, but at one point something else broke too. This PR documents all the known issues.

---

Upstream issue #8754 (comment)

vitest/test/config/fixtures/fs-cached-check/basic.test.ts

Line 11 in c4485c6

const mod = await import("./dist/generated.js")

Rolldown does not resolve the id /dist/generated.js into a full file path and fails in loadAndTransform function. Looks like, it is stored with an incorrect file property in the module graph (perhaps, during the import analysis?) and then never updated. Note that this works correctly in Vite 7.

---

vitest/test/config/test/failures.test.ts

Line 361 in 2827eab

expect(stderr).toMatch('ERROR: Coverage for functions (0%) does not meet global threshold (100%)')

Looks like SSR transform behaves differently because the coverage has no functions:

functions: { total: 0, covered: 0, skipped: 0, pct: 100 },

The module looks like this:

__vite_ssr_exportName__("uncovered", () => { try { return uncovered } catch {} });
function uncovered(condition) {
  return condition ? 1 : 0;
}

New issues will be added

Reproduction

https://github.com/vitest-dev/vitest

System Info

Any

Used Package Manager

pnpm

Validations

• Follow our Code of Conduct
• Read the Contributing Guidelines.
• Read the docs.
• Check that there isn't already an issue that reports the same bug to avoid creating a duplicate.
• Check that this is a concrete bug. For Q&A open a GitHub Discussion or join our Discord Chat Server.
• The provided reproduction is a minimal reproducible example of the bug.

[sapphi-red]

vitest/test/config/fixtures/fs-cached-check/basic.test.ts

Line 11 in c4485c6

const mod = await import("./dist/generated.js")

Rolldown does not resolve the id /dist/generated.js into a full file path and fails in loadAndTransform function. Looks like, it is stored with an incorrect file property in the module graph (perhaps, during the import analysis?) and then never updated. Note that this works correctly in Vite 7.

oxc-resolver caches the fs access by default; this is essentially a revert of #17807, which causes #17760 (though there was one bug vitejs/rolldown-vite#463). This commit makes the test pass (sapphi-red@0a8fefd). The other approachs I can think of are:

• add an option to disable the cache
• add a way to clear the cache so that Vitest can call that before resolving the dynamic import (I guess this is the only place the cache is a problem)

[sapphi-red]

vitest/test/config/test/failures.test.ts

Line 361 in 2827eab

expect(stderr).toMatch('ERROR: Coverage for functions (0%) does not meet global threshold (100%)')

Looks like SSR transform behaves differently because the coverage has no functions:

functions: { total: 0, covered: 0, skipped: 0, pct: 100 },

The module looks like this:

__vite_ssr_exportName__("uncovered", () => { try { return uncovered } catch {} });
function uncovered(condition) {
  return condition ? 1 : 0;
}

It seems this is coming from the source map difference. Vite has a mapping for the closing } of the function (sourcemap visualizer), but rolldown-vite does not have a mapping for it (sourcemap visualizer).
@AriPerkkio Is this mapping needed? or is it possible to handle this on the coverage-v8 side?

[sheremet-va]

This commit makes the test pass (sapphi-red@0a8fefd).

I don't think having watch mode be enabled in run is a good idea. The test should also not be touched, if we ask users to way, we can ask them instead to disable the option in the config.

add an option to disable the cache

I think that could work.

add a way to clear the cache so that Vitest can call that before resolving the dynamic import (I guess this is the only place the cache is a problem)

I guess we could reset the cache before any dynamic import, true. Although if there are any other tests running in parallel, the performance impact might be the same as disabling it, unless this use case is common in someone's code base.

[AriPerkkio]

It seems this is coming from the source map difference. Vite has a mapping for the closing } of the function (sourcemap visualizer), but rolldown-vite does not have a mapping for it (sourcemap visualizer). @AriPerkkio Is this mapping needed? or is it possible to handle this on the coverage-v8 side?

I haven't tested this myself yet, but missing end } should work fine.

https://github.com/AriPerkkio/ast-v8-to-istanbul/blob/adb80ab089073dde2b46b20e2dbfd230cc8efa42/src/location.ts#L87-L91

[sapphi-red]

It seems this is coming from the source map difference. Vite has a mapping for the closing } of the function (sourcemap visualizer), but rolldown-vite does not have a mapping for it (sourcemap visualizer). @AriPerkkio Is this mapping needed? or is it possible to handle this on the coverage-v8 side?

I haven't tested this myself yet, but missing end } should work fine.

https://github.com/AriPerkkio/ast-v8-to-istanbul/blob/adb80ab089073dde2b46b20e2dbfd230cc8efa42/src/location.ts#L87-L91

It worked when I add the mapping for } manually. So probably there's a bug around that. Perhaps it doesn't work if it's the last mapping?

[AriPerkkio]

Before debugging V8 coverage it would be good to have basic Istanbul coverage working. For example the test/coverage-test/test/results-snapshot.test.ts is pretty good smoke-test for verifying source maps.

https://evanw.github.io/source-map-visualization/#NjAyNgBfX3ZpdGVfc3NyX2V4cG9ydE5hbWVfXygic3VtIiwgKCkgPT4geyB0cn

 FAIL   istanbul  test/results-snapshot.test.ts > coverage results matches snapshot
Error: Snapshot `coverage results matches snapshot 1` mismatched

- Expected
+ Received

  {
    "<process-cwd>/fixtures/src/even.ts": {
      "branches": "0/0 (100%)",
-     "functions": "1/2 (50%)",
+     "functions": "0/0 (100%)",
      "lines": "1/2 (50%)",
      "statements": "1/2 (50%)",
    },
    "<process-cwd>/fixtures/src/math.ts": {
      "branches": "0/0 (100%)",
-     "functions": "1/4 (25%)",
+     "functions": "0/0 (100%)",
      "lines": "1/4 (25%)",
      "statements": "1/4 (25%)",
    },
    "<process-cwd>/fixtures/src/untested-file.ts": {
-     "branches": "0/4 (0%)",
+     "branches": "0/0 (100%)",
-     "functions": "0/4 (0%)",
+     "functions": "0/0 (100%)",
-     "lines": "0/8 (0%)",
+     "lines": "0/6 (0%)",
-     "statements": "0/8 (0%)",
+     "statements": "0/6 (0%)",
    },
  }

[sapphi-red]

Oh, I didn't notice that failure. Probably because ecosystem-ci fails with @vitest/test-core and @vitest/test-coverage and after doesn't get run. I'll check that later.

[sapphi-red]

@AriPerkkio I've checked what's happening there. It seems @vitest/coverage-instanbul is not returning the expected source map (#9001). I'm not sure why it worked with Vite though.

[sapphi-red]

When I override the internal combined sourcemap with the one returned from instrumenter.lastSourceMap, I get:

 FAIL   istanbul  test/results-snapshot.test.ts > coverage results matches snapshot
Error: Snapshot `coverage results matches snapshot 1` mismatched

- Expected
+ Received

  {
    "<process-cwd>/fixtures/src/even.ts": {
      "branches": "0/0 (100%)",
-     "functions": "1/2 (50%)",
+     "functions": "0/0 (100%)",
      "lines": "1/2 (50%)",
      "statements": "1/2 (50%)",
    },
    "<process-cwd>/fixtures/src/math.ts": {
      "branches": "0/0 (100%)",
-     "functions": "1/4 (25%)",
+     "functions": "0/0 (100%)",
      "lines": "1/4 (25%)",
      "statements": "1/4 (25%)",
    },
-   "<process-cwd>/fixtures/src/untested-file.ts": {
-     "branches": "0/4 (0%)",
-     "functions": "0/4 (0%)",
-     "lines": "0/8 (0%)",
-     "statements": "0/8 (0%)",
-   },
  }

which looks similar to the one happening with v8 coverage.