aio-libs/aiohttp - CodSpeed

aiohttp

Blog Docs Changelog

Performance History

Latest Results

Document the multipart_reader_cls constructor contract

multipart-max-size

4 minutes ago

Trim test comments to one line each

ws_tail_bound

23 minutes ago

Use the shared resume helper in resume_reading() It still inlined the transport-resume try/except that _resume_transport_reading() now owns, eight lines below it, along with a _reading_paused = False that could only run when the flag was already False. The helper rechecks the flag itself, so the guard reduces to the buffer hook. Also stop THREAT_MODEL 3.16 claiming the pre-parser pause is always lifted; only set_parser() and set_response_params() drain that buffer.

ws_tail_bound

1 hour ago

Assert only after the synchronisation point

ws_tail_bound

2 hours ago

Document what the multipart drain skips release() and the automatic drain of a skipped part bypass next(), so neither the part cap nor subclass overrides of next() apply there.

multipart-max-size

2 hours ago

Make the functional test exercise the held tail pause

ws_tail_bound

2 hours ago

Resume reading once the parser has failed A peer can push past read_bufsize before set_parser() installs the reader and then send a bad frame. The tail pause taken for that burst was never lifted, because the resume guard could not tell "no parser yet" from "parser failed". Reading stayed off, so the discard branch never ran and the peer's FIN never arrived: the fd leak this was meant to avoid. Only stay paused while the parser has yet to be installed. After a failure the discard branch keeps the tail empty, so there is nothing to refill it and reading is safe to resume.

ws_tail_bound

3 hours ago

Discard data after a WebSocket protocol error instead of closing Closing the transport broke the Autobahn client run: a valid frame that arrives in the same read as the bad one is still delivered, and an echo client answering it hit "Cannot write to closing transport". Pausing was no better, since a paused transport never sees the peer's FIN and the socket leaks. Drop what arrives after the error instead. Nothing can parse it, memory stays flat, the transport stays writable so queued messages can still be answered, and a close from the peer is still delivered. close_code keeps reporting the protocol code, now pinned by test_recv_protocol_error.

ws_tail_bound

3 hours ago

Latest Branches

0%

Cap parts yielded by request.multipart() with client_max_fields#13742

16 minutes ago

a1bcd5e

multipart-max-size

0%

Bound websocket client buffers a peer can fill#13743

36 minutes ago

ff92e64

ws_tail_bound

0%

[PR #13738/8b010057 backport][3.15] Add client_max_fields option to Application#13740

6 hours ago

b1bd57c

patchback/backports/3.15/8b0100572936df28977b9de8a6cc7e464c0b521e/pr-13738

© 2026 CodSpeed Technology

Home Terms Privacy Docs