mirror of
https://github.com/p1neappleXpress/OpenFlux.git
synced 2026-10-02 05:04:39 +08:00
acac97b07d
* phpbox: survive any host's limits and keep streams alive across handovers Found on a local emulation of a free host (Apache + PHP-FPM, 60 s CPU per request, the host's disabled functions, a gateway that gives up on the request): - disabled set_time_limit/ignore_user_abort/getmypid/getenv/ini_set ended the node at once under PHP 8 (a disabled function is undefined): guarded everywhere; ping lists what the host took away - time budget per generation on both clocks (CPU on Linux, wall on Windows/macOS, RLIMIT_CPU), handover early when CPU runs short, and learning: a generation that died early leaves its age in host.json and the next ones plan below it; on a host not yet proven, handovers start at 40 s and grow as generations end well - cups: the carrier slept 18 ms per message inside the loop (a 64 KB frame ~0.4 s), so under a download new connections timed out and acks waited; now a paced outbox in the mux: control frames first, streams in turn, whole packets per message, at most ~72 KB/s (the server drops a member that sends faster) - two generations in one room: each now ignores the other's messages (and its own echoes, before decoding them), and messages carry whole packets, so a packet cut across messages is never mixed with another member's bytes; the Go cupsonline receiver keeps one reassembly buffer per member - a successor takes streams only after its link stayed up 8 s (a Mail.ru document drops new connections right after they join, and streams handed over then died) - a repeated OPEN is answered again instead of redialled; the client asks again after 4 s without an answer - the page restarts a node that ended without Stop while it is open - --mode=stream without --inbound is SOCKS5 on every OS again (macOS took utun) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * phpbox: send to cups only as fast as the server confirms; release rebuilds an existing tag - the cups server shows every member's message back to it; our own coming back is its confirmation. At most 6 unconfirmed messages: past that the server is behind, and the client's OPENs and its pings queued behind our echoes (close 3012 'no pong' under a download). The link's close code and reason go to the log. - release.yml: a moved tag replaces the release's assets and notes, as the apps do. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * phpbox: count the cups server's echoes reliably; learn a host limit only from two deaths - echoes were matched with a regex on the raw line that never matched the server's JSON, so the window waited out its 3 s timeout each time (~6 KB/s under load): a substring test on the member uuid instead; a live room confirms 10 of 10 - one sudden end (a host restart) no longer teaches a limit that shortens every run for a week: two ends at about the same age do Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * phpbox: hand over early and consistently; never reach for a longer run than the host allows The emulated hosts exposed the real cause of the node dying: the 'hold idle to prove a longer run' recorded a lifetime (107s) longer than a host's hidden wall-clock kill (90s), so later generations set their handover past that limit and were killed before they could start a successor - the chain broke with no generation left to carry the lesson. - a generation now hands over at ~2/3 of the known limit, never later than 45s, even on a host that looks generous: a free host's real limit is unknown until a generation dies by it, and that death must happen after a successor exists. Deaths only lower it. - removed the idle hold and the lifetime-growth; a learned wall/CPU limit still lowers the handover, and a wall limit is believed only after two deaths at a similar age - while a successor shares the room, the predecessor caps its reading (OVERLAP_RATE): the document server copies our download to the successor too, and at full speed the client's OPENs to the successor waited behind it Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * CHANGELOG: the node-survival and stream-proxy fixes Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * CHANGELOG: fold the node-survival fixes into 0.3.0 (rebuilt in place) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: p1neappleXpress <a@a.a> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>