Aditya Reddy

web perf, react, opentelemetry. finding out why your click took 400ms. here for the dev news and the ai drama too. made react-inp-blame

[Aditya Reddy @adityareddydev]

the OpenTelemetry demo's recommendation service is on Spring Boot 4 now, my rewrite merged today, docs and the helm chart catch up next https://t.co/V7gcoslrq4

[Aditya Reddy @adityareddydev]

otel js 3.0 is out, three trace packages became one and provider.register() is gone, so that tracing.ts you wrote once and never opened again needs a touch. npx otel-js-upgrade 3 shows the diff before it writes anything https://t.co/5zfgDTXSRq

[Aditya Reddy @adityareddydev]

@tunetheweb yeah anything you wanna count has to go out on `visibilitychange` to hidden, cause thats the last event you can actually trust. and bfcache restores mean the same "session" comes back after you already sent its end lol

[WebPerformance Report @WebPerfReport]

⚡Featured posts this week:
✨ How Swedish sites LCP changes with soft navs: @gernberg
✨ Experimental soft navs in CrUX: @anniesullie
✨ Unload event deprecated: @tunetheweb Fergal Daly
✨ JPEG XL vs AVIF for web: @jaffathecake
🙏Special thanks to the authors!
#webperf

[Aditya Reddy @adityareddydev]

@calvinluttrell @ollama median first token and a tool success rate still hide which call the chain waited on

[Aditya Reddy @adityareddydev]

@BoboLinNotes yeah 7k tok/s is the box. the slow request only felt time to first token, and that wait disappears inside the average

[Aditya Reddy @adityareddydev]

@ApogeeIS yeah redirectStart stays 0 on that hop too. the report can name the hole and still can't split the shortener out of waiting

[Aditya Reddy @adityareddydev]

@WebPerfReport @gernberg @anniesullie @tunetheweb @jaffathecake soft navs still marked experimental in crux is the whole week. lab score is the first load, the click after that is the one they actually waited on

[Aditya Reddy @adityareddydev]

@ahmadparizaad yeah the gateway is the right place for the cap. it still won't say if the wait was the first token or the rest of the answer

[Aditya Reddy @adityareddydev]

@mharshald healthy gpu and no oom is the annoying part. the short chats stutter cause one 128k prefill is sitting on the scheduler, look at queue time versus prefill time first

[Aditya Reddy @adityareddydev]

@AbhishekMadhavn gzip only shrinks the upload. the model still reads the unpacked history, so the sooner first token is a shorter wait on the wire in front of the same prefill

[Aditya Reddy @adityareddydev]

@shrey_sancheti yeah a faster kernel still starts late. the 1744 was the decode worker posting those rdma descriptors inline

[Aditya Reddy @adityareddydev]

@akashsairaja the timeout has to land in the ui. if the page never swaps the spinner for a token or a try again, the user still sits through the slow minute

[Aditya Reddy @adityareddydev]

@ajrgd those three are already the field set. a screenshot after the bot hides the modal is the bot's wait, not the visitor who had to dismiss it

[Aditya Reddy @adityareddydev]

@xspeedcache a content-encoding header is not a measurement. 50 of those responses got bigger and the header still said br

[Aditya Reddy @adityareddydev]

@krudhrah the rum line is the post. ssr and critical css can be perfect and the tap still waits on one long task

[Aditya Reddy @adityareddydev]

@akashsairaja yield only helps if the heavy work is actually after it. if that long task is still inside the click, inp just writes down the whole wait

[Aditya Reddy @adityareddydev]

@unjica213 a dashboard that still shows fid is measuring the old delay. inp waits for the paint, and the page can feel stuck while that old number stays green