Improvements for httpx2.Client.sse timeouts

[rafmagns-skepa-dreag]

When using Server-Sent Events, the timeout parameter defaults USE_CLIENT_DEFAULT which defaults to the httpx2 default of DEFAULT_TIMEOUT_CONFIG (Timeout(timeout=5.0) at time of writing). The current implementation seems to reset the timeout any time a real SSE is read (I had trouble finding the actual implementation here including where the "timed out" message gets printed from).

My understanding is that ping events/messages in SSEs are generally implemented as "comments" and are ignored per the spec

A colon as the first character of a line is in essence a comment,

and are ignored due to this block

httpx2/src/httpx2/httpx2/_sse.py

Lines 87 to 88 in 4bb6753

	if line.startswith(":"):
	return None

The effect of this is that clients using default parameters will disconnect from SSE streams any time there's not a real message every 5 seconds.

I'm implementing both an SSE producer (sse-starlette cause I'm fastapi version limited) and SSE consumer (httpx2). Is there a best practice that I'm not aware of to send "real" messages frequently? If not, I'd like to discuss the possibility of handling SSE timeouts differently in httpx2 either by potentially letting comment/ping messages reset the timeout or by setting a higher read timeout by default.

Other than this, SSEs in httpx2 have been really easy to work with and I'm really grateful for the work done on this fork to modernize httpx.

[rksharma-owg]

You are not missing anything obvious here. There are two separate layers involved:

• The SSE parser correctly ignores comment lines that start with :. That is required by the SSE format, so turning ping comments into user-visible events would not be the right fix.
• The read timeout is enforced lower down by the HTTP transport. HTTPX2 documents the default as 5 seconds of network inactivity, not 5 seconds for the whole stream.

So a ping comment should still reset the read timer if the bytes actually arrive before the deadline, even though the parser yields no event for it. If the server can stay completely silent for longer than 5 seconds, configure a longer read timeout or disable just the read timeout for the SSE request, for example:

timeout = httpx2.Timeout(10.0, read=None)

(or use a suitably large read= value if you still want a dead-connection cutoff). I would keep the parser behavior as-is and document the long-lived-stream timeout pattern instead. It may also be worth adding a test that feeds comment-only traffic through the transport and confirms that it prevents a read timeout without producing an SSE message.

[rafmagns-skepa-dreag]

Thanks for the confirmation.

Your proposal is exactly what I ended up doing in my implementation - Timeout(3.0, read=None).

I believe that httpx2 should either have a default read timeout of None for SSEs or that ping messages should reset the timeout

[rksharma-owg]

Thanks for confirming the read=None workaround. I checked the current sync and async paths: SSE comment lines are discarded only after response bytes have passed through the transport. Each socket read applies the read timeout to that individual receive, so a ping comment that actually arrives before the deadline already resets it. If you still see a timeout with pings sent more often than the read interval, buffering by the server or a proxy would be worth checking. Making read=None the SSE default is a separate API choice for the maintainers; your current per-request override keeps the connection and other timeout settings explicit.