Extreme lagging on Langsmith Playground

[ErzErzHerzog]

Description

• Langsmith playground extreme lagging appeared in the last week or so
• Loading an llm call in plaground (via trace) loads frozed + brings up chrome 'wait for page' toast
• Adding a new message box takes ~60 seconds to respond
• Typing in message box takes 4-5 seconds to render

=> Can someone take a look at this?

Steps to Reproduce

1. Select an LLM call from trace (in this case ChatOpenAI ~40k context....not large)
2. top right hand corner, click on 'Playground'
3. Wait for page to load....1 minute or more
4. Click '+Message' ....wait some more
5. Start typing...wait more
   => FYI: use case here is to open up a live chat and branch into to 'debrief' the agent / get their feedback...this is different from trying parallel calls on the same prompt. Generally as UI changes are made they tend to make the parallel call scenario easier and the 'debrief' scenario harder/extra steps (just FYI)

Expected / Observed Behaviour

No lag expected. Fairly unusable as is.

Notes

System RAM not exceptionally high (20/32gb used).
Windows 11 using Comet Version 145.0.7632.117 (Official Build) (64-bit) Perplexity build number: 5936

Dev tools:
(from keyboard typing).....its something in the background bottlenecking:

Digging further on a single key stoke, it looks like the entire style sheet is being rendered for all 40k x 4-5 = 200k characters on the page on every key press

Perplexity contributes its opinion:

Root cause identified: On every keystroke, a CSS-in-JS style mount is triggered. getRules() takes ~1,400ms to collect CSS rules, then set textContent takes ~3,060ms to write them to a <style> tag. This accounts for 79% of the 5,592ms interaction cost. ProseMirror/updatePlugins itself is only ~83ms and is not the issue. The fix is preventing style remounting on every editor transaction — either by memoising the style sheet, fixing a React re-render that's recreating CSS instances, or guarding the mountStyles call behind a dirty-check.

[jacoblee93]

Hey UPwind GS&O (@upwindgso)! Thank you for reporting - will take a look.

[madams0013]

Hi UPwind GS&O (@upwindgso)! We've merged some perf updates to this view (will be live later this afternoon), but are continuing to investigate and improve. There are multiple factors causing some lag here, but able to reproduce. Thanks for the report!

[ErzErzHerzog]

Thanks Maddy Adams (@madams0013) !
Greatly improved!

The slowest paint (that ~2 second paint) is when I added a new message to a ~25 message chain (fine as its an infrequent op). Typing/editing messages etc are now back to normal/fast.

Thanks for the super quick turnaround!