SSE (Server-Sent Events) response body not flushed to downstream on macOS

[janpoem]

Description

When proxying SSE (Server-Sent Events) responses through Pingora on macOS, the response body chunks are received from upstream but never flushed to the downstream client. The same setup works correctly on Linux.

Environment

• macOS: Darwin 24.6.0 (Apple Silicon, aarch64)
• Linux: Ubuntu 22.04 x86_64 (works correctly)
• Pingora version: 0.8.0
• Rust: stable

Steps to Reproduce

1. Set up a simple upstream server that serves SSE responses (Content-Type: text/event-stream, Transfer-Encoding: chunked)
2. Configure Pingora as a reverse proxy forwarding to this upstream
3. Send a GET request to the proxy endpoint from macOS

Expected Behavior

SSE events should be streamed to the client in real-time as they arrive from upstream.

Actual Behavior (macOS only)

• Pingora successfully connects to upstream and receives the response headers
• Upstream body chunks are received (confirmed via response_body_filter logging)
• However, no data reaches the downstream client — curl shows no output
• After curl timeout, the connection is reset with: Fail to proxy: Downstream ConnectionClosed ... Prematurely before response body is complete
• Debug logs show the downstream socket's BufWriter has written: 0 despite data being in the buffer

Debug Log Evidence

[DEBUG] Detected streaming response: text/event-stream
[DEBUG] Streaming chunk: 937 bytes, eof: false     ← data received from upstream
[DEBUG] Streaming chunk: 318 bytes, eof: false     ← more data received
[ERROR] Fail to proxy: Downstream ConnectionClosed
        context: Prematurely before response body is complete
        status: 0                                   ← 0 bytes written downstream

The BufWriter state at connection close:

BufStream { inner: BufReader { reader: BufWriter {
  writer: ..., buffer: 0/1460, written: 0           ← 1460 bytes buffered, 0 written
}}}

Analysis

The issue appears to be in the HTTP/1.1 downstream write path. Pingora's do_write_chunked_body in body.rs does call stream.flush() after each chunk write. However, on macOS, the data appears to remain in the BufWriter's buffer (1460 bytes = typical TCP MSS) without being flushed to the socket.

This suggests a platform-specific difference in how AsyncWrite::flush() behaves on macOS vs Linux TCP sockets, possibly related to:

• TCP_NODELAY not being set on the downstream socket
• macOS TCP stack buffering behavior differences
• BufWriter not propagating flush to the underlying TCP stream

Workarounds Attempted (none worked on macOS)

• session.set_keepalive(None)
• Custom response_body_filter returning Ok(None) (no delay)
• Custom upstream_response_filter setting chunked encoding headers
• Removing all custom filters (bare Pingora proxy) — still fails on macOS

Impact

• Normal HTTP requests (GET, POST with complete responses) work fine on macOS
• Only streaming/long-lived responses (SSE, potentially WebSocket) are affected
• Linux works perfectly — this is macOS-only

Notes

macOS is commonly used for development. While Pingora is primarily designed for Linux production deployments, macOS development support would significantly improve the developer experience.

---

Reported by siyuan — discovered while building gateway-rs