[Ar-maan05]

When a large bytes write is queued behind pending data, every partial send re-queued the leftover as b[nbytes:], copying the rest of the object each time, so the write cost O(n**2). One write of 48 MiB behind a 2 MiB prefix on a socketpair (Linux, sendmsg path):

before: 1633 ms
after:    14 ms

The fix wraps the remainder in a memoryview in _adjust_leftover_buffer() and _write_send(), the same way write() already does on its first-send path. A bytearray passed to write() now stays exported while its remainder is pending on these paths too, which the first-send path already does today.

Tested by extending the existing partial-write tests to bytes input and checking that the leftover is a view on the original object, test_asyncio passing, and a differential harness that drives the old and new methods with a scripted socket (partial sends, BlockingIOError, send and sendmsg, bytes/bytearray/memoryview input): identical output over 1200 runs.

• Issue: asyncio: large bytes write is quadratic when queued behind pending data #159503

[maurycy]

cc @cmaloney

[Ar-maan05]

Thanks! Pushed: direct slice for memoryviews, bytearray added to the tests, NEWS trimmed, and a line in the write() docs about not modifying the data until it's sent.

[cmaloney]

When making changes look for ways to simplify and what existing docs say. For instance the "must not be mutated" is common in other docs around this same sort of thing.

[Ar-maan05]

Thanks both, applied all three!

[read-the-docs-community]

Documentation build overview

📚 cpython-previews | 🛠️ Build #35498753 | 📁 Comparing 792a686 against main (7e5227f)

  🔍 Preview build  

2 files changed

± library/asyncio-protocol.html
± whatsnew/changelog.html