Repository navigation
Conversation
|
hey @danielroe can you resolve conflicts? |
|
@danielroe Thank you so much for the PR. Just testing a few things of this locally and against my other in the works test and will get back to you |
|
@danielroe Ok, I have gone through the PR and would love to add this; however, I am leaning towards keeping both the Lighthouse INP and this new version. I think having both results to compare would be beneficial. The Lighthouse default INP breakdown is used in lots of places, so I think we should keep it. The next question is do we extend the two lighthouse schemas to include this new metric: Or add this in as its own benchmark for example Let me know what you think. |
812797d to
ad09f3b
Compare
|
done! ✅ it was a hefty rebase, so I pretty much reimplemented this instead, as an alternative measure made from the same trace/click event (so they are directly comparable). some of the data is just therefore exactly the same between the two INP measures, but I thought that was better than having one implicitly depending on the other. I've also chosen to rank by this new metric, but feel free to revert that if you prefer, of course 🙏 |
resolves #573
Note
nuxt is already quite high up on this measurement (in second place), so this isn't motivated by a feeling that we should be in a different position
while running some benchmarks locally, I noticed that most of the latency that frameworks are showing in the current numbers is the wait for the next frame tick after processing, which makes it very dependent on where the click lands in the 16.7 ms frame - rather than on the specific framework.
so this is an alternative proposal: score each click as work to commit
insteadthe idea is to measure input delay + processing duration, plus the renderer main thread's busy time from the end of processing until the Event Timing
commitFinishTimeit skips idle time before the frame, and everything after commit (raster, draw, present, browser UI)
👉 update: this now keeps the Lighthouse INP breakdown, and records work to commit as an additional metric. it's not totally separate, in that we use the same trace/click event to produce both the traditional INP numbers and this rescore. we also now sort by the new metric, as I think it's more accurate, but 🤷
a chart rendered from the numbers in the CI run for the latest commit
Note
Astro's full-document navigation mostly stops at unload under both metrics, so it isn't directly comparable.
Checklist
starter-*package, ensure that the deps, and file content matches the output of the metaframeworks recommended starter/default project