amayer1983/docksentry v2.18.0-beta.36

v2.18.0-beta.36 — which channel hears what, and a switch that now switches

on GitHub

Everything here is about which channel hears what. It started with one
observation on my own server — an update run reported every finished
container to Discord at once and told Telegram nothing until the end —
and a fair question: we pulled the core together so each connection
only renders, so how does that still happen?

Because the update engine never had the seam. It held the notifier half
and nothing else, so the difference was not a decision anybody made. A
survey of all 25 places that say something unprompted found three going
through Broadcast, twelve hand-written fan-outs, and ten reaching
only part of the channels.

Fixed

• A switch that did not switch. Turning the Discord bot channel off
  on the Connections page did not stop it receiving: announce never
  read the setting. And the retry queue delivered messages up to fifteen
  minutes after the switch was thrown, because what it holds is the
  raw transport — it has to be, or a resend would re-enter the code that
  queued it. Both reproduced before either was touched. The queue asks
  again at delivery time now, and every plugin that queues passes the
  question along.

  Telegram had the same hole in one place: send_document never read
  channel_telegram_enabled, so a mass-stop log file and every
  /backup still landed in a chat somebody had switched off.

  Answers to commands keep flowing either way. Silencing the running
  commentary of a /selfupdate somebody just typed turns "off" into
  "half off" — the slash reply arrives, then nothing.

• Quiet hours silenced the wrong half. Broadcast.tell is the door
  for what has to arrive — a rollback, a storage alert, a host that
  stopped answering — and its own docstring says so. The Telegram leg
  passed the gate; the plugins returned early. Between 23:00 and 07:00
  a rollback was therefore Telegram-only. The urgent class now passes on
  every channel, maintenance windows included. Routine traffic is still
  silenced, exactly as before.

• A message that promised what it could not deliver. The mass-stop
  digest told all nine channels "each container's log tail is attached"
  while the attachment is only ever offered to Telegram. Two footers
  now, and Telegram's is chosen by whether the file actually exists
  rather than promising one and finding out afterwards.

• Four notices nobody but Telegram heard. A container held back by a
  closed maintenance window, by the update policy, or because the image
  is younger than you allow; and a major update waiting for
  confirmation. The first of those never entered the "updates available"
  list either, so off Telegram nobody learned the update existed at all.
  They go to every channel now — and once per state rather than once per
  check run, which through the seam would otherwise be ten containers
  times three plugins times an hourly cron.

• A self-update that failed from the Web UI button reached the
  plugins and not Telegram. It also said so in hardcoded English, in the
  one place where every other self-update message is translated.

• /audit ended in silence when everything was fine, leaving you to
  work out that no news was good news (@NotRetarded). It says
  "nothing to do: all N Compose files are reachable from in here".

• /audit printed "Mount to reach it" with a hole where the path
  belongs, for every stack under a prefix it does not recognise.

• Discord labelled local rows @local where Telegram leaves them
  bare on documented purpose, so the same list read two ways. Headings
  that belong to one host still name it, including the local one.

• An idle Telegram long poll that gets reset was logged as an error
  every time it happened. Quiet is right for that; silent is not, so a
  poll that stays dead for three minutes now says so once.

Changed

• The update engine reports through the seam. Same cadence as before —
  a card per container on the plugins, the whole batch in one Telegram
  message — but that is now written down in one place instead of being
  whatever object happened to be reachable.

Built on top of v2.18.0, which @NotRetarded ran for five days before signing it off.

Worth testing: turn a channel off on the Connections page and check it really goes quiet — then check that a /selfupdate you type yourself still answers. And /audit with no container name, which now says so when there is nothing to do.