mirror of
https://github.com/paperclipai/paperclip.git
synced 2026-10-26 02:00:52 +01:00
Priya RamanandPaperclip
fa14631ba2
The duplex gateway charged an unknown-length request body one time, but readBody retains a chunk list and the Buffer.concat result together for one instant, so its true retained peak is up to two times the body cap. Name that factor as DUPLEX_GATEWAY_UNKNOWN_LENGTH_BODY_RESERVE_FACTOR and charge 2 * maxBodyBytes plus the fixed frame allowance for an unknown-length read. A declared-length read still charges the declared length plus the same fixed allowance, and a bodyless request still charges the zero-length reserve. Update the retained-peak comment to state the true unknown-length peak. The file gateway reserve is unchanged: FILE_GATEWAY_BODY_RESERVE_FACTOR stays 4. Add a regression test that sends four concurrent cap-sized chunked requests with no Content-Length header, holds the host forward so the reserves overlap, and sets the bound to admit exactly three reserves of 2 * maxBodyBytes + the frame allowance. It asserts the fourth request gets a 503 bridge_busy, then asserts a later request is admitted once a terminal releases a reserve. Proven against the pre-fix commit first: the test fails there (all four requests admit, none shed) and passes after this fix. Co-authored-by: Paperclip <noreply@paperclip.ing>