[shikhar]

Why

The server never sets TCP_NODELAY. axum-server's default acceptor leaves Nagle's algorithm on, and turning it off is opt-in via NoDelayAcceptor.

Under HTTP/2, a fetch response ends with a small separate frame: the trailers frame from the body stream, or an empty END_STREAM data frame. That frame often goes out while the previous small segment is still unacknowledged, so Nagle holds it until the ACK arrives. A Linux client delays that ACK by up to 40 ms, so the client finishes reading the response about 40 ms late.

HTTP/1.1 clients don't hit this. With Content-Length the response ends at its last byte, and no trailing frame is sent.

We saw this after moving a client from HTTP/1.1 to HTTP/2 with prior knowledge. Client-observed fetches in the 40–80 ms range became about 7× more common. The server's own first-chunk and download latencies didn't change.

Repro

A minimal Linux loopback test mirrors the fetch response: a 206 with Content-Length, the body streamed through StreamBody, and a trailing trailers frame. The client is reqwest with http2_prior_knowledge(), or HTTP/1.1 for comparison. The server is axum-server 0.8 with rustls, with and without NoDelayAcceptor. Responses are 2 KiB, 2000 requests per case:

client	concurrency	Nagle (current)	TCP_NODELAY
HTTP/1.1	1	p99 6.7 ms, 0% > 30 ms	p99 6.6 ms, 0% > 30 ms
HTTP/2	1	p50 47 ms, 99.9% > 30 ms	p99 6.8 ms, 0% > 30 ms
HTTP/2	8	p99 50 ms, 27% > 30 ms	p99 6.9 ms, 0% > 30 ms

On a real network the stall happens less often than on loopback, because other traffic on the connection usually carries the ACK sooner. The fix is the same either way.

What

Use NoDelayAcceptor for all three listener modes: TLS with a provided certificate, self-signed TLS, and plain HTTP.

🤖 Generated with Claude Code

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no actionable issues were found.

Summary

Adds NoDelayAcceptor to all three listener modes to avoid delayed HTTP/2 response completion.

• Every server listener disables Nagle on accepted connections.

_{Reviews (1) · Last reviewed commit: "fix(server): set TCP_NODELAY on accepted..." · Reviewed by Greptile}