DYODE 1.4.3

Fixed

• A file taking longer than batch_timeout (default 5 minutes) to transfer
  was discarded on arrival. The receiving box closed the batch while the
  file was still arriving, because only finished files reset the clock. The
  intact file was then thrown away as "received a file with no manifest
  pending". With the default FEC capping a gigabit link at about 350–400
  Mbit/s, a 10 GB file takes about that long. Now any data still landing in
  the staging folder keeps open batches alive, so a batch only times out after
  batch_timeout with nothing received at all. Verified in the field: a 10 GB
  transfer that failed this way arrived intact. (On earlier versions: set
  batch_timeout: 3600 in the receiving folder module.)

Changed

• The receive-side writeback ceiling scales with RAM. vm.dirty_bytes in
  /etc/sysctl.d/90-dyode-receive.conf is now 20% of RAM (like the kernel's
  own default), at most 2 GB and at least 128 MB. In 1.4.2 it was a fixed
  1 GB: too tight on large boxes and too much for a small one such as a
  Raspberry Pi. Writeback still starts at 64 MB, or a quarter of the ceiling
  on boxes under ~1.25 GB of RAM. Run sudo ./install.sh --tuning-only to
  update an existing box.

Documentation

• CAVEATS.md:
  • new entry for the batch-timeout problem
  • a section on sustained disk speed, which the writeback settings cannot
    fix: how to measure it and how to set bitrate from it
  • an explanation of what the writeback ceiling is for