perf: middleware to short-circuit on 304 #15586

patak-dev · 2024-01-12T10:24:41Z

Description

I'm checking the work needed in the server to answer on 304 once it is warmed up. We return 304 when:

On a forced reload from the user
On HMR when there is a full reload for all code except for the modified modules
On second startup (both warm and cold), half or more of the modules could already be warmed up on large apps and they will be also hitting 304.

There isn't much work we do to reach the current condition, but we still do a few things.

We add CORS headers. If they are the same, we could skip this, but probably not worth the complexity to take into account the headers when computing the etag. See https://httpwg.org/specs/rfc7234.html#freshening.responses, the CORS headers will be refreshed on 304 if changed.
We have a few middlewares: proxy, open-in-editor, hmr-ping, public. The public one in particular has the most expensive check
Then in the transform middleware, we run a few regexes (CSS, JS, etc) and then need to run a getModuleByUrl that also adds up.

This PR implements a fast path for 304, with a new middleware after CORS that kicks in when there is an etag, and uses a etagToModuleMap.

I'm seeing a ~3-4% improvement when forcing a reload on my M1. I would say the extra map is ok here. It would be interesting to see the effect on slower machines too.

What is the purpose of this pull request?

Bug fix
New Feature
Documentation update
Other

stackblitz · 2024-01-12T10:24:45Z

Run & review this pull request in StackBlitz Codeflow.

bluwy · 2024-01-16T06:08:54Z

packages/vite/src/node/server/middlewares/transform.ts

+        // For direct CSS requests, if the same CSS file is imported in a module,
+        // the browser sends the request for the direct CSS request with the etag
+        // from the imported CSS module. We ignore the etag in this case.
+        const mixedEtag =
+          !req.headers.accept?.includes('text/css') &&
+          isDirectRequest(moduleByEtag.url)


Trying to understand this part. Do you mean that the browser would send the same etag for text/css requests and *.css imports? That sounds strange. Is it a bug that also happens before?

I think it would still be nice to make the request work instead of bailing here.

Without this guard, this test was failing. The issue is that in that test we are both adding global.css as a link in the html and then importing it. I tried to find a reason why we do so and I couldn't. But I think we should leave it as is because it actually helped to spot this bug.

I think our current system for .css isn't ideal. What we do is adding a ?direct flag when we detect that global.css was requested as a link. But that ?direct flag is internal, for the browser, it sees two global.css requests with the same URL. We return different etags for each. When the browser requests the imported global.css module the second time, it uses the etag of the linked one. I thought it could be a bug in chrome but I tested in Firefox and Safari and in all the browsers is the same.

This playground was still working because we did a getModuleByUrl before this PR, and it was done after adding ?direct when it was a link. If the etag was mixed for imported css modules, then it will not match with the etag of the module and a regular 200 will be issued instead of a 304. So this PR is doing the same for this case, I don't know how we could avoid bailing out here. I think it is ok as the setup is an edge case.

An alternative would be to switch from ?direct for linked CSS to using ?import for imported CSS modules instead like we do with other assets (rewriting during import analysis). I think this would actually be cleaner. If we do so, then a bare .css request will always correspond to a link request, and a .css?import will always be a module and these URLs are seen by the browser (and not like ?direct that we add internally) so the requests are for different URLs. But I don't know if we can do this change given that it may have a big effect on the ecosystem, and it is hard to justify it when it is only an issue if the same .css is linked and imported at the same time.

Ah thanks. Sounds like a weird browser bug, but maybe for good reasons. Yeah in that case if it's rare, then I think we can simply bail then. Maybe we could add a debug message to detect this kind of requests if it later turns out to be more common than expected.

The ?import idea does make more sense to me, but yeah maybe we can leave that until we need to make a bigger change.

cc @sapphi-red, this may be interesting to you as I think you mentioned we could remove ?import at one point, no? If we do so, we may hit this same etag issues for that assets.

I guess Vary header (Vary: Accept) is what we should use.

Oh... this looks great, thanks @sapphi-red! I tried it out and it seems it fixes safari but I still see the same issue in Chrome and Firefox. So maybe there are browsers bugs here? I think we should keep the guard for now and aim to use Vary header in the future (for when we'll try to remove ?import)

It seems I misunderstood the Vary header. Vary header doesn't affect conditional requests.
Also browsers seems to only use a single E-Tag in If-None-Match header. I guess there isn't a way to tell browsers to send different If-None-Match header for a same URL and a different header.

sapphi-red · 2024-01-18T16:06:20Z

packages/vite/src/node/server/middlewares/transform.ts

+    // check if we can return 304 early
+    const ifNoneMatch = req.headers['if-none-match']
+    if (ifNoneMatch) {
+      const moduleByEtag = server.moduleGraph.getModuleByEtag(ifNoneMatch)


If a file is added in public directory, I think we need to delete a entry from server.moduleGraph.etagToModuleMap.
For example, if /src/foo.ts existed and later on /public/src/foo.ts is added, then /public/src/foo.ts should be returned instead of the transpileed /src/foo.ts.

Good catch!

More context for others. We are serving public files using sirv, and I see that the way it computes its etag is:

if (isEtag) headers['ETag'] = `W/"${stats.size}-${stats.mtime.getTime()}"`;

There shouldn't be any overlap between public etags and processed etags.

But as this two URLs are the same, the browser may request the public file using the etag of processed module because of the same issue discussed here.

But if the source file existed but was removed, the etag is no longer going to be there. The only case where this is an issue is when there are both a public and a source file at the same time (and the public file was added later). If we find the module by id and remove its etag, then the browser may still request the url with the wrong etag but it should still pick up the public file.

I think it should be fixed by e9a6011

sapphi-red · 2024-01-19T08:27:00Z

/ecosystem-ci run

vite-ecosystem-ci · 2024-01-19T08:41:02Z

📝 Ran ecosystem CI on e9a6011: Open

suite	result	latest scheduled
analogjs	✅ success	✅ success
astro	✅ success	✅ success
histoire	✅ success	✅ success
ladle	✅ success	✅ success
laravel	❌ failure	❌ failure
marko	✅ success	✅ success
nuxt	❌ failure	✅ success
nx	❌ failure	❌ failure
previewjs	✅ success	✅ success
qwik	❌ failure	❌ failure
rakkas	✅ success	✅ success
remix	❌ failure	❌ failure
sveltekit	✅ success	✅ success
unocss	✅ success	✅ success
vike	❌ failure	❌ failure
vite-plugin-pwa	✅ success	✅ success
vite-plugin-react	✅ success	✅ success
vite-plugin-react-pages	✅ success	✅ success
vite-plugin-react-swc	✅ success	✅ success
vite-plugin-svelte	✅ success	✅ success
vite-plugin-vue	✅ success	✅ success
vite-setup-catalogue	✅ success	✅ success
vitepress	✅ success	✅ success
vitest	✅ success	✅ success

patak-dev · 2024-01-19T08:47:10Z

I saw the same Nuxt fail in other PRs when running ecosystem-ci, I think it isn't related to this PR

sapphi-red · 2024-01-19T09:21:02Z

I noticed that this breaks vite-plugin-static-copy (similar edge case with #15586 (comment)). That plugin injects a middleware just before the serve public middleware. I will add something similar to https://github.com/vitejs/vite/pull/15586/files#diff-abb3345b6e3b2ec6d297c2ebc54cca85ae4487a31bac3cc9e78457f5114adb26R687-R697 on the plugin side as I guess the middlewares.stack is not part of the public API.

patak-dev · 2024-01-19T09:26:32Z

injects a middleware just before the serve public middleware

The names in middlewares.stack are part of the public API, but we have added and removed middlewares in minors in the past. In this case, IIUC, vite-plugin-static-copy will still work and only break when a static file is added with the same name as a source file (and that source file isn't deleted). I think it is an edge case worth solving in the plugin, but we could still move forward with this one in a minor.

bluwy

LGTM 👍 I'm not sure how confident are you with the new module graph API, and whether we should mark them as internal if we ever need to change it, but otherwise we can give this a shot!

patak-dev · 2024-01-19T15:54:14Z

I'm not sure if we should mark it as internal yet, because if not @sapphi-red won't be able to fix the edge case he talked about here 🤔

Let's merge it without marking as internal, but we can revisit this before releasing Vite 5.1.
I think the API will be useful for plugins in general.

patak-dev added 2 commits January 12, 2024 11:00

perf: middleware to short-circuit on 304

f24ab2d

fix: move after cors middleware

1c4ec51

patak-dev added the performance Performance related enhancement label Jan 12, 2024

fix: mixed etag for direct CSS that is also imported

7f189df

bluwy reviewed Jan 16, 2024

View reviewed changes

sapphi-red reviewed Jan 18, 2024

View reviewed changes

patak-dev added 2 commits January 18, 2024 17:51

chore: merge main

c5820eb

fix: adding a public file with the same path as a module

e9a6011

sapphi-red approved these changes Jan 19, 2024

View reviewed changes

bluwy approved these changes Jan 19, 2024

View reviewed changes

patak-dev merged commit 35ae4f8 into main Jan 19, 2024
15 checks passed

patak-dev deleted the perf/fast-304-path branch January 19, 2024 15:54

patak-dev mentioned this pull request Feb 15, 2024

fix: cachedTransformMiddleware for direct css requests #15919

Merged

4 tasks

bluwy mentioned this pull request Sep 1, 2024

[Bug] Dev server resolves wrong module for provided ETag, leading to stale content #17987

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

perf: middleware to short-circuit on 304 #15586

perf: middleware to short-circuit on 304 #15586

patak-dev commented Jan 12, 2024

stackblitz bot commented Jan 12, 2024

bluwy Jan 16, 2024

patak-dev Jan 16, 2024

bluwy Jan 16, 2024

patak-dev Jan 16, 2024

sapphi-red Jan 16, 2024

patak-dev Jan 16, 2024

sapphi-red Jan 18, 2024

sapphi-red Jan 18, 2024

patak-dev Jan 18, 2024

sapphi-red commented Jan 19, 2024

vite-ecosystem-ci bot commented Jan 19, 2024

patak-dev commented Jan 19, 2024

sapphi-red commented Jan 19, 2024 •

edited

Loading

patak-dev commented Jan 19, 2024

bluwy left a comment

patak-dev commented Jan 19, 2024

perf: middleware to short-circuit on 304 #15586

perf: middleware to short-circuit on 304 #15586

Conversation

patak-dev commented Jan 12, 2024

Description

What is the purpose of this pull request?

stackblitz bot commented Jan 12, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sapphi-red commented Jan 19, 2024

vite-ecosystem-ci bot commented Jan 19, 2024

patak-dev commented Jan 19, 2024

sapphi-red commented Jan 19, 2024 • edited Loading

patak-dev commented Jan 19, 2024

bluwy left a comment

Choose a reason for hiding this comment

patak-dev commented Jan 19, 2024

sapphi-red commented Jan 19, 2024 •

edited

Loading