Add create_datagram_endpoint api

[tarasko]

No description provided.

[codspeed-hq]

Merging this PR will not alter performance

✅ 40 untouched benchmarks

---

_{Comparing create_datagram_endpoint (bdc00fc) with master (389e0f9)}

[tarasko]

@codspeedbot explain why this is faster

[codspeed-hq]

@codspeedbot explain why this is faster

Why test_benchmark_sendfile[tcp-large] is ~10.5% faster

I compared the flame graphs of the two runs behind the report (HEAD 11566a2 vs master ebd3b00). Here's what's actually driving the 3.1 ms → 2.8 ms change.

1. The write / sendfile path was refactored in this PR

The improvement is concentrated in the sendfile submission path, wrapped by __Pyx_PyObject_FastCallDict.constprop.1 (called from SocketTransport._try_sendfile):

Metric	BASE (ebd3b00)	HEAD (11566a2)
Self time	175.1 µs	104.9 µs
Total time	313.7 µs	188.5 µs
Self-time cache misses	30.7 µs	18.6 µs

This lines up with the transport rewrite in the diff: SocketTransport was split into a SocketTransportBase, the write/watermark logic was reorganized (_maybe_pause_protocol / _ensure_writer / _drop_writer inlined helpers), and notably the large inline aiofn_iovec _iovecs[256] array was removed from the transport struct. Shrinking that per-transport structure improves data-cache locality, which shows up as the drop in cache-miss time above.

2. The event loop did slightly less work

The uv__io_poll → __on_uvpoll_event path dropped from 1.4 ms (45%) to 1.1 ms (40%), with one fewer poll event (4 → 3) and one fewer protocol callback cycle (14 → 15 vs the base's counts). So part of the win is simply a marginally different number of loop iterations to move the same data.

3. Not everything changed — and an important caveat

The dominant cost, __memset_avx2_unaligned_erms (~13% self time, from protocol/loop __init__), is essentially flat (390 µs → 387 µs), as are the pure-Python protocol callbacks buffer_updated and get_buffer. So the source changes explain the improvement in the write path specifically, not a wholesale speedup.

⚠️ CodSpeed also flagged that the base and head ran in different runtime environments — the transport, ssl_transport, and utils shared libraries have different build IDs. In Simulation mode the cache model is derived from the runner's CPU, so a different toolchain/runner can shift the estimate on identical source. This means some fraction of the reported +10.51% is attributable to that environment difference rather than your code, and the exact percentage should be read with that in mind.

Bottom line: the real, code-attributable win is the leaner transport struct (better cache locality in the sendfile path) plus slightly fewer event-loop iterations; the headline number is somewhat inflated by a cross-environment comparison.

Summary

• Warming up the performance wizard...
• Handling comment — Comment handling finished successfully

_{Was this response helpful? React with 👍 or 👎 to help us improve CodSpeedBot.}