[PR #24928] feat: configurable aiohttp read_bufsize via AIOHTTP_READ_BUFSIZE env #131597

New Issue

GiteaMirror · 2026-05-21T17:17:45-05:00

GiteaMirror commented

2026-05-21 17:17:45 -05:00

📋 Pull Request Information

Original PR: https://github.com/open-webui/open-webui/pull/24928
Author: @shoemoney
Created: 5/19/2026
Status: 🔄 Open

Base: dev ← Head: feat/aiohttp-read-bufsize-env-v2

📝 Commits (1)

d1a8722 feat: configurable aiohttp read_bufsize via AIOHTTP_READ_BUFSIZE env

📊 Changes

2 files changed (+23 additions, -1 deletions)

View changed files

📝 backend/open_webui/env.py (+14 -0)
📝 backend/open_webui/utils/session_pool.py (+9 -1)

📄 Description

Problem

The shared aiohttp.ClientSession in utils/session_pool.py is created with the library default read_bufsize (64 KiB). Some upstream providers emit a single SSE chunk that exceeds this ceiling, causing the stream to fail with:

400, message: Got more than 131072 bytes when reading: b'data: {"id":"gen-...","object":"chat.completion.chunk",...

(The 131072 = 2 × 64 KiB comes from aiohttp internally doubling the buffer once before raising.)

Triggers seen in the wild

OpenRouter routing reasoning models (Kimi K2.6, R1-style, etc.) through providers that batch the entire reasoning trace into a single chunk — observed with Novita, WandB, SiliconFlow.
Image-generation responses returned as a single large SSE message — see discussion #24543, which reports the same error signature and is currently unresolved.

There is no current env or config knob to override this; the only workaround is patching the source.

Fix

Expose AIOHTTP_READ_BUFSIZE (default 1 MiB) and pass it through to ClientSession(read_bufsize=...).

Default of 1 MiB is conservative — handles every real-world case I've reproduced while still rejecting truly runaway payloads.
Falls back gracefully on invalid/empty/negative values.
Logged at session creation alongside the other pool params.

Files

backend/open_webui/env.py — new env constant with validation.
backend/open_webui/utils/session_pool.py — import it, pass to ClientSession, update docstring + log line.

Validated against

Reproduced the failure with moonshotai/kimi-k2.6 via OpenRouter (Novita provider) on v0.9.5. With the default 64 KiB buffer the stream dies mid-reasoning; with AIOHTTP_READ_BUFSIZE=1048576 (or the new default) the stream completes cleanly through data: [DONE].

Related: #24543

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/open-webui/open-webui/pull/24928 **Author:** [@shoemoney](https://github.com/shoemoney) **Created:** 5/19/2026 **Status:** 🔄 Open **Base:** `dev` ← **Head:** `feat/aiohttp-read-bufsize-env-v2` --- ### 📝 Commits (1) - [`d1a8722`](https://github.com/open-webui/open-webui/commit/d1a87225454f43f86645b4c91d419f7d1b522292) feat: configurable aiohttp read_bufsize via AIOHTTP_READ_BUFSIZE env ### 📊 Changes **2 files changed** (+23 additions, -1 deletions) <details> <summary>View changed files</summary> 📝 `backend/open_webui/env.py` (+14 -0) 📝 `backend/open_webui/utils/session_pool.py` (+9 -1) </details> ### 📄 Description ## Problem The shared `aiohttp.ClientSession` in `utils/session_pool.py` is created with the library default `read_bufsize` (64 KiB). Some upstream providers emit a **single** SSE chunk that exceeds this ceiling, causing the stream to fail with: ``` 400, message: Got more than 131072 bytes when reading: b'data: {"id":"gen-...","object":"chat.completion.chunk",... ``` (The 131072 = 2 × 64 KiB comes from aiohttp internally doubling the buffer once before raising.) ### Triggers seen in the wild - **OpenRouter routing reasoning models** (Kimi K2.6, R1-style, etc.) through providers that batch the entire reasoning trace into a single chunk — observed with Novita, WandB, SiliconFlow. - **Image-generation responses** returned as a single large SSE message — see discussion #24543, which reports the same error signature and is currently unresolved. There is no current env or config knob to override this; the only workaround is patching the source. ## Fix Expose `AIOHTTP_READ_BUFSIZE` (default **1 MiB**) and pass it through to `ClientSession(read_bufsize=...)`. - Default of 1 MiB is conservative — handles every real-world case I've reproduced while still rejecting truly runaway payloads. - Falls back gracefully on invalid/empty/negative values. - Logged at session creation alongside the other pool params. ## Files - `backend/open_webui/env.py` — new env constant with validation. - `backend/open_webui/utils/session_pool.py` — import it, pass to `ClientSession`, update docstring + log line. ## Validated against Reproduced the failure with `moonshotai/kimi-k2.6` via OpenRouter (Novita provider) on `v0.9.5`. With the default 64 KiB buffer the stream dies mid-reasoning; with `AIOHTTP_READ_BUFSIZE=1048576` (or the new default) the stream completes cleanly through `data: [DONE]`. Related: #24543 --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

GiteaMirror added the pull-request label 2026-05-21 17:17:45 -05:00

Sign in to join this conversation.

Branches Tags

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: github-starred/open-webui#131597