[MDA2AV]

#257 raised TcpOptions.RecvQueueEntries from 64 to 4096, in the same change that made an overflow an error instead of a clean end. This keeps the error and puts the default back to 64.

The queue is sized up front for every connection, pooled ones included, so 4096 costs 96 KiB a connection against 1.5 KiB. In shared mode it also lets one connection whose handler is not reading hold the reactor's whole default ring (RecvSlots 4096) before it overflows. At 64 that connection is closed long before, and since #257 its readers see the error.

Measured

HttpArena's ioxide entry, every published version side by side. The setup was 8 reactors pinned to 8 P-cores of an i9-14900K, gcannon on the E-cores, 4096 connections and the baseline GET/POST mix. The server was saturated at ~2.45M req/s. Versions were measured round-robin; each figure is the median of 6 runs.

	vs 0.7.211
0.15.251	-0.8%
0.16.268, the first release with 4096	-1.9%
0.16.268, RecvQueueEntries = 64	-0.2%
0.17.339	-1.4%
0.17.339, RecvQueueEntries = 64	-0.9%

In the same runs, versions with no change on this path moved by up to ~1%. So the 0.16.268 gap is real, and the 0.17.339 gain is inside the noise. Memory moves for certain: at 16,384 connections the entry held 732 MiB, against 470 MiB at 64.

A handler that lets its peer get more than 64 deliveries ahead now gets #257's IOException (TcpConnection.RecvOverflowed) sooner. It can raise the option.

The samples and docs follow:

• the samples that set it;
• the defaults table in docs/learn/architecture.html;
• the snippets in multiport.html and tls.html;
• the index panes, regenerated by scripts/gen-docs-panes.py and scripts/gen-docs-proxy-panes.py. Both reported the page in sync on main before the change.

Release

The second commit bumps all twelve csproj Version elements to 0.17.340, as #339 did for 0.17.339, so merging publishes the new default.

Tests

No new tests, since this only changes a default. Most fixtures already set RecvQueueEntries = 64 or less themselves, so the suites have been running at 64 all along.

Run locally, with no failures: Unit 78 (1 pending), E2E 285 (6 pending), Http 83 (4 pending), Tls 174 (8 pending, kTLS cases included), File 5, Chaos 79 (1 pending). They ran on the first commit. The release commit only changes Version, and the final head builds clean. I did not run Pg and Redis, since there are no sidecars here. Neither reaches this: only an accepted connection reads RecvQueueEntries (Reactor.Tcp.cs:240).

Overlap

#362, #363 and #365 also edit TcpOptions.cs, and git merge-tree against each open PR is clean. #362 adds SendWaitAll directly above this comment, so the comment's first lines stay as they were to keep the two hunks apart. #307 already conflicts with main, in a file this does not touch.

🤖 Generated with Claude Code