A number of changes here. I recommend viewing the diff with the <a href="?w=1">whitespace flag enabled</a>.
- OpenTelemetry is replaced with a custom and lightweight tracing solution.
- Three trace targets are currently supported: console, Zipkin, and NextJS.
- Tracing is now governed by environment variables rather than `--require instrument.js`.
+ `TRACE_TARGET`: one of `CONSOLE`, `ZIPKIN`, or `TELEMETRY`; defaults to `TELEMETRY` if unset or invalid.
+ `TRACE_ID`: an 8-byte hex-encoded value used as the Zipkin trace ID; if not provided, this value will be randomly generated and passed down to subprocesses.
Other sundry:
- I'm missing something, probably a setup step, with the Zipkin target. Traces are captured successfully, but you have to manually enter the Trace ID in order to view the trace - it doesn't show up in queries.
- I'm generally unhappy with [this commit](235cedcb3e). It is... untidy to provide a telemetry object via `setGlobal`, but I don't have a ready alternative. Is `distDir` strictly required when creating a new Telemetry object? I didn't dig too deep here.
As noted, there are a lot of changes, so it'd be great if a reviewer could:
- [ ] pull down the branch and try to break it
- [ ] check the Zipkin traces and identify possible regressions in the functionality
Closes#22570Fixes#22574
This updates to output server chunks to a nested folder to prevent bundling the entire folder when tracing. This also fixes the webpack 5 tests not actually using webpack 5 since https://github.com/vercel/next.js/pull/22583 since the webpack 5 enabling check didn't account for the test environment variable used to enable webpack 5. This also clears up some deprecation warnings from webpack 5 in the mini-css-extract-plugin.
Fixes: https://github.com/vercel/next.js/issues/21297
@timneutkens I think this is ready for a review.
I've made some changes to the original design that _seem_ to have paid off. The parenting relationships for traces of normal builds are applied more uniformly, resulting in more intelligible traces:
<img width="900" alt="Screen Shot 2021-01-29 at 12 53 47 AM" src="https://user-images.githubusercontent.com/5016978/106253732-ba321880-61cc-11eb-98fd-d45af5078273.png">
Hot-reloading is surfaced now, too. I will note, however, that we will want to dig in deeper and find out where the large portion of time at the beginning of hot-reload is spent. Example:
<img width="894" alt="Screen Shot 2021-01-29 at 12 53 28 AM" src="https://user-images.githubusercontent.com/5016978/106253828-e057b880-61cc-11eb-967d-46eaff31ecef.png">
Where did those 180 ms go? At the least, we can now track how long a hot-reload takes, and have a place to start with further investigation.
This picks up on the inlining work in https://github.com/vercel/next.js/pull/20598 to also include webpack loader inlining optimizations.
This includes:
* The dependencies of sass-loader
* resolve-url-loader
And for added benefit:
* babel-plugin-transform-define
* babel-plugin-transform-react-remove-prop-types
style-loader and css-loader didn't inline easily. Perhaps we can come back to these ones.
There's currently two bugs with the font optimization, but we'd really like to ship a stable version.
To unblock the stable release, we're **temporarily** reflagging this. It'll be unflagged on canary again!
Solves the following warning:
> (node:1484) [DEP_WEBPACK_MAIN_TEMPLATE_REQUIRE_FN] DeprecationWarning: MainTemplate.requireFn is deprecated (use "__webpack_require__")
This PR removes the modern mode experiment because:
- It does not yield meaningful bundle size wins when compared to other initiatives we've taken
- It's not compatible with webpack 5 (which we're upgrading to)
- It's currently broken and causes most apps to malfunction
- There's no champion currently owning the experiment
We can re-introduce this in the future when we'd like to make it a default for all Next.js apps.
Note: **Next.js still supports Differential Loading (`nomodule`) and does it by default.** This PR strictly removes the experimental modern _syntax_, and does not disable our existing modern/legacy polyfilling.
---
Fixes#19200Fixes#18960Fixes#14707Fixes#14465
We accidentally regressed back in 9.5 and dropped support for inline CSS comments. PostCSS always parses these as pass-through (and not a syntax error), which can cause problems when minifying.
Browsers do a similar thing and ignore the comments.
To ensure we generate valid CSS, this adds support for stripping the CSS comments from the build.
---
Fixes#15589Closes#17130
This upgrades to ncc@0.25.0 and fixes the previous bugs including:
* ncc not referenced correctly in build
* Babel type errors
* node-fetch, etag, chalk and raw-body dependencies not building with ncc - these have been "un-ncc'd" for now. As they are relatively small dependencies, this doesn't seem too much of an issue and we can follow up in the tracking ncc issue at https://github.com/vercel/ncc/issues/612.
* `yarn dev` issues
Took a lot of bisecting, but the overall diff isn't too bad here in the end.
This adds inlining for Babel and the Babel plugins used in next.
This is based to the PR at https://github.com/vercel/next.js/pull/18823.
The approach is to make one large bundle and then separate out the individual packages from that in order to avoid duplications.
In the first attempt the Babel bundle size was 10MB... using "resolutions" in the Yarn workspace to reduce the duplicated packages this was brought down to a 2.8MB bundle for Babel and all the used plugins which is exactly the expected file size here.
This will thus add a 2.8MB download size to the next package, but save downloading any babel dependencies separately, removing a large number of package dependencies from the overall install.
This is a prerequisite to being able to ncc inline the Babel dependencies in next.js.
The removal of preset-modules is based on replacing it with preset-env under `targets: { esmodules: true }`, as per the guidance from the package (https://www.npmjs.com/package/@babel/preset-modules):
> Starting from @babel/preset-env 7.9.0, you can enable the bugfixes: true option to get the same behavior as using @babel/preset-modules, but with support for custom targets. If you need to target browsers with native modules support (like this preset does), you can use targets: { esmodules: true }.
From the above, I'm pretty sure this is entirely a backwards compatible change, apart from the change to the runtime plugin list being visible. Perhaps @developit can confirm this as well.
To prevent FOUC, discussed in #10557 i need to store information about css file dependencies for chunk. Right now current implementation just throws away everything but js.
Can there be more than one css file in chunk? If no - code will be simplified.
closes#10557
This pull request edits the `BuildManifest` that is sent to `/_document` instead of modifying a single input array to decouple its implementation details.
Optimally, we'd eliminate the `files` key all together.
---
Related to #16182
- Using `namedChunks` where possible, this will also allow for faster access to the chunks as we no longer have to look them up like we did before using `find`
- Using the new asset hooks introduced in the latest webpack beta
- Using the new externals function signature
Initial PR to make `next build` work with webpack 5, still needs more work to make sure runtimeChunk and such are shared between pages.
- No longer needs the custom ChunkNamesPlugin as the default behavior was changed
- Dropping AMP First client page bundles is now compatible
Updates the way filenames are generated for browser compilation.
Notably:
- All entry bundles now have hashes in production, this includes pages (previously pages used a buildId in the path)
- The AmpFiles no longer depends on hardcoded bundle names, it uses the buildManifest instead (internals)
- All cases where we match the page name from the chunk/entrypoint name now use the same function `getRouteFromEntrypoint` (internals)
- In development we no longer include the "faked" `buildId` set to `development` for page files, instead we just use the `/_next/static/pages` path (was `/_next/static/development/pages`). This was changed as it caused unneeded complexity and makes generating the bundles easier (internals)
- Updated tons of tests to be more resilient to these changes by relying on the buildManifest instead of hardcoded paths (internals)
Follow up of these PRs:
https://github.com/vercel/next.js/pull/13759https://github.com/vercel/next.js/pull/13870https://github.com/vercel/next.js/pull/13937https://github.com/vercel/next.js/pull/14130https://github.com/vercel/next.js/pull/14176https://github.com/vercel/next.js/pull/14268Fixes#6303Fixes#12087Fixes#1948Fixes#4368Fixes#4255Fixes#2548
Webpack will randomly execute script order if its runtime is not prioritized before chunks execute.
This seems to be somehow triggered in #13870 because of slightly different script ordering.
This had actually broke CSS, which is why our tests are failing 50% of the time:
Without this PR:
![image](https://user-images.githubusercontent.com/616428/84221491-57f0a000-aaa3-11ea-9dff-c27c87d29ac5.png)
However, it's still problematic to use `async` in development since we rely on script execution order. So, this PR disables `async` in development.
We're exploring `defer` in the future anyway (over `async`), which will be ordered, so I don't mind diverging between dev and prod in this way.
---
Fixes#13911
Initial work to use chunkhashes instead of buildid for the page files in production. This does not change the calculation of the filename itself initially.
Disambiguate between pages/index.js and pages/index/index.js so that they resolve differently.
It all started with a bug in pagesmanifest that propagated throughout the codebase. After fixing pagesmanifest I was able to remove a few hacks here and there and more logic is shared now. especially the logic that resolves an entrypoint back into a route path. To sum up what happened:
- `getRouteFromEntrypoint` is the inverse operation of `getPageFile` that's under `pages/_document.tsx`
- `denormalizePagePath` is the inverse operation of `normalizePagePath`.
Everything is refactored in terms of these operations, that makes their behavior uniform and easier to update/patch in a central place. Before there were subtle differences between those that made `index/index.js` hard to handle.
Some potential follow up on this PR:
- [`hot-reloader`](https://github.com/vercel/next.js/pull/13699/files#diff-6161346d2c5f4b7abc87059d8768c44bR207) still has one place that does very similar behavior to `getRouteFromEntrypoint`. It can probably be rewritten in terms of `getRouteFromEntrypoint`.
- There are a few places where `denormalizePagePath(normalizePagePath(...))` is happening. This is a sign that `normalizePagePath` is doing some validation that is independent of its rewriting logic. That should probably be factored out in its own function. after that I should probably investigate whether `normalizePagePath` is even still needed at all.
- a lot of code is doing `.replace(/\\/g, '')`. If wanted, that could be replaced with `normalizePathSep`.
- It looks to me like some logic that's spread across the project can be centralized in 4 functions
- `getRouteFromEntrypoint` (part of this PR)
- its inverse `getEntrypointFromRoute` (already exists in `_document.tsx` as `getPageFile`)
- `getRouteFromPageFile`
- its inverse `getPageFileFromRoute` (already exists as `findPageFile ` in `server/lib/find-page-file.ts`)
It could be beneficial to structure the code to keep these fuctionalities close together and name them similarly.
- revise `index.amp` handling in pagesmanifest. I left it alone in this PR to keep it scoped, but it may be broken wrt nested index files as well. It might even make sense to reshape the pagesmanifest altogether to handle html/json/amp/... better
This removes remaining references to `granularChunks` in configs, error messages, and comments.
Also removed the `process.env.__NEXT_GRANULAR_CHUNKS` value.
---
Follow up to: https://github.com/vercel/next.js/pull/13663
Was going through _document and noticed some variable shadowing going on. Added a rule for it to our eslint configuration and went through all warnings with @Timer.
The experimental modern mode runs the type checking plugin twice, which **occasionally** suffers from a race condition that hangs the build.
This PR fixes type checking to only be run once.
While this test cannot 100% reproduce/capture the race condition, I don't feel strongly about the test case:
- We're planning on eliminating this type checking plugin ASAP (for a faster alternative)
- Modern mode implementation as-is will probably go away with webpack 5
Adding a conformance plugin the make sure users don't undo the benefits of the granularChunks config.
The plugin makes sure that minSize, maxInitialRequests values aren't overridden. Also ensures the cacheGroups - vendors, framework, libs, common, shared are maintained.
The warning and error messages do not break the build with this change. They only display a message.
cc - @prateekbh, @atcastle