Skip to content

Releases

Mru Field 0.3.0: flood limits

Choose a tag to compare

@wbnns wbnns released this 16 Nov 15:46
05c1655

Flood limits (docs/spec.md, section 11). Field does not run a network and cannot stop a flood on one. Version 0.3.0 makes sure a node does not start one, whatever build of the task it runs, and keeps working when its own source is flooded or its own sink is jammed.

  • Task network lock. task_network = "deny", the default: a seccomp filter in each task process before it starts; it can open AF_UNIX sockets and nothing else. Linux on x86_64, ARM64 and 32-bit ARM.
  • Input backlog. input_backlog = "newest" keeps the newest complete record under a flood and logs the backlog once when it begins and once when it clears. The default, queue, is the 0.2.0 behavior.
  • Fixed: a TCP output or heartbeat write had no deadline, so a peer that stopped reading held the supervisor without end. Writes now end at 5 s.
  • Fixed: a serial line with no newline grew memory without limit.

On the bench (Linux in Docker on an Apple M4, scripts/flood.py, docs/results/2026-11-16-linux-aarch64/flood):

  • A build with right answers that sends 100 datagrams per request: 450,300 out with the lock off, and a staged update accepted it; 0 out with the lock on. 0 wrong either way.
  • A source at 10,000 records a second into 100 cycles a second: the input was 12.0 s old by the last cycle under queue, 0.8 ms at the median under newest.
  • An output peer that stops reading: 0.2.0 hung and was killed at 60 s; 0.3.0 ended at 7.7 s with an error and its checkpoint.

The decision rule is the quorum crate from mruspace/flight v0.4.0, unchanged. A 0.2.0 checkpoint resumes under 0.3.0. Bench only: not on a board in a system, and not qualification evidence. See CHANGELOG.md.

Mru Field 0.2.0: staged updates

Choose a tag to compare

@wbnns wbnns released this 14 Nov 13:28
772fd96

Staged updates (docs/spec.md, section 10). A new build of the task comes in one slot at a time, and the old build checks every delivered output until the last switch. Until then, one fault of a slot on the new build takes it out of every slot and starts the old build there again.

On the bench (Apple M4, 5,000 cycles per run, scripts/staged.sh, docs/results/2026-11-14-darwin-arm64/staged):

  • All at once, a build that crashes or hangs on cycle 500 halts the run on cycle 500. A build that is wrong from cycle 500 delivers 4,501 wrong outputs, and nothing flags them.
  • By staged update, every bad build was rejected on the cycle its bug first showed, with 0 wrong outputs and 3 replicas alive.
  • Limit: a bug that first shows after the last switch gets through. The new build must give the same outputs as the old one.

The decision rule is the quorum crate from mruspace/flight v0.4.0, unchanged. Bench only: not on a board in a system, and not qualification evidence. See CHANGELOG.md.