The transcription endpoint was async but called the synchronous transcribe() function directly, blocking the single-threaded uvicorn event loop for the entire duration of inference. This caused all HTTP and WebSocket connections to stall for every user on the instance during STT processing.
- Add asyncio import
- Use async UploadFile.read() instead of synchronous file.file.read()
- Offload the blocking transcribe() call via asyncio.to_thread()
Closes#24169
The chat table has no computed columns (no DEFAULT, SERIAL/IDENTITY,
or TRIGGER that populate server-side values on UPDATE), and every
column modified by update_chat_by_id is set explicitly from Python
values earlier in the function. db.refresh therefore issues a SELECT
that replaces those just-written Python values with the round-tripped
database representation of the same values, which is a no-op for
functional purposes but pulls the entire chat.chat JSON blob back over
the network and through the driver's JSON decoder.
On large, active chats where chat.chat can reach tens of megabytes,
skipping the refresh measurably reduces latency and eliminates one
~JSON-sized transient allocation per write.
- Return 302 to /static/favicon.png instead of streaming the same PNG per
model id so browsers can cache one asset for default avatars.
- Validate stored /static/ paths with decode, normpath, and /static
prefix checks; invalid paths fall back to favicon.
Made-with: Cursor
