ADR 0217, to-be 44: no change to a machine takes effect unseen

Two incidents in two days had one shape: a change took effect that nobody saw first. Three guards
where the change is made — a settings layer read before it is replaced, a push that says what it
will change, a running module's data move acknowledged — each silent when nothing is at stake.
This commit is contained in:
2026-10-05 15:14:28 +02:00
parent d222091fe9
commit e7126bc1e6
3 changed files with 141 additions and 0 deletions
@@ -0,0 +1,81 @@
---
topic: the mesh
status: accepted
date: 2026-10-05
deciders: jochen
reconstructed: false
extends: 02-DECISIONS/0030-data-outlives-the-mesh-that-declared-it.md
---
# 217. No change to a machine takes effect unseen
## Context
Two incidents in two days had one shape: **a change took effect that nobody saw before it did.**
- Issue 241: a provisioner read an unreadable grants file as "nobody asks" and withdrew every
consumer; the postgres provider dropped seven databases. Nothing announced the withdrawal before it
happened.
- Issue 246: one placement was set for one module on one machine with `settings set`. The layer is
replaced whole, so the module's other settings on that machine — three directory placements, eight
media accesses, an exposure, four endpoints, the account's ids — were dropped, and the command
answered only "places". The machine's plan then ran the module on an empty configuration
directory. It was caught by reading the plan before a push, and the lost layer was read back from a
database backup a few hours old.
Since issue 241 the host no longer destroys what it did not make (ADR 0030 for directories, issue
241's fix for files) and no provider drops a store. What remains is the class above: the mesh doing
exactly what it was told, when what it was told was not what the person meant, and nothing showing
the difference until a machine had changed.
## Considered Options
1. **Confirm every push.** Rejected: every push would ask, and a confirmation asked every time is
answered without reading. The guard must speak only when something is at stake.
2. **Make `settings set` merge.** Rejected: removing a key would then need a second command, and the
layer stops being a statement of the whole (ADR 0046, ADR 0164). The fault is that the removal was silent,
not that it was possible.
3. **Three guards, each where the change is made, each silent when nothing is at stake.** Adopted.
## Decision
**A change that removes, moves or replaces something a machine is running is shown before it takes
effect; one that only adds is not interrupted.**
1. **A settings layer is read before it is replaced, and what a replacement removes is said.**
`settings show` (and the verb) reads a layer. `settings set` answers with the keys it adds, changes
and removes, and **refuses to remove a key unless told `--replace`**. Every change keeps the layer
it replaced, so the previous value is one command away, not in a backup.
2. **A push says what it will change.** `plan <node> --diff` lists what differs from what the machine
was last sent: resources added, changed and removed, and every container that will be recreated
with the reason. **`push` with no machine named is refused** unless `--all` says so.
3. **Moving a running module's data is acknowledged.** When a push would give a running container a
different host directory at a mount it already has — its data moving, or not following — the push
to that machine is held, naming the module, the mount and both directories; the other machines in
the same push go ahead. It is sent when the operator says so for that module (`--move <module>`).
None of the three adds a step to an ordinary change: a new module, a rebuilt image, a setting that
only adds keys and a push to a named machine behave exactly as before.
## Consequences
- **The controller keeps what it last sent each machine**, not only its digest, so a plan can be
compared with it. Migration 0014 declined to keep it — "a second account of what a machine should
be" — and that reason stands for what it was about: the copy is never read as what a machine
*should* be, only as what it *was told*, which is the one thing a comparison needs and the digest
cannot give.
- Settings carry a history; `settings show` reads it.
- A script that ran `push` with no machine must say `--all`.
- A changed placement for a running module needs one acknowledgement, once.
## How it is checked
Tests in the controller: `settings set` that drops a key without `--replace` is refused and changes
nothing; with it, the answer names the removed key and the history holds the previous layer. `plan
--diff` of an unchanged node is empty. A push changing a running container's mount source is held,
and goes ahead with `--move` for that module. `push` with no machine and no `--all` is refused.
## References
- [04-ISSUES/241](../04-ISSUES/241-one-unreadable-grants-file-dropped-every-database-on-the-control-node/00-report.md), [04-ISSUES/246](../04-ISSUES/246-setting-a-modules-settings-replaces-the-whole-layer-and-nothing-shows-it-first/00-report.md)
- ADR 0030 (data outlives its declaration), ADR 0046 (a module's configuration is its assignments), ADR 0164 (a setting is declared with its default and its cost), ADR 0214 (backups)
+1
View File
@@ -194,6 +194,7 @@ python3 00-META/checks/index.py fail if stale
- **0207** — [A module depends on the node seats that apply its resources](0207-a-module-depends-on-the-node-seats-that-apply-its-resources.md)
- **0210** — [A tool's configuration is its seat holder's, and every other module extends it through the seat](0210-a-tools-configuration-is-its-seat-holders-and-every-other-module-extends-it-through-the-seat.md)
- **0212** — [A seat says what it receives, and the machine's hotkeys are a seat](0212-a-seat-says-what-it-receives-and-the-machines-hotkeys-are-a-seat.md)
- **0217** — [No change to a machine takes effect unseen](0217-no-change-to-a-machine-takes-effect-unseen.md)
### Its tiers, from the bottom up