Repository navigation
fix(graph-db): log the drifting corruption verdict at error level instead of staying silent - #4574
Conversation
A drifting fault (the second open reports different bytes than the first) is terminal and silent: nothing is deleted, nothing is rebuilt, and every later open repeats the verdict. Three containers sat in that state for 4-10 days on one machine, distinguishable from healthy only by a leftover lock file. The refusal now logs one error-level event naming the container, both faults, and the operator action. Half of ScriptedAlchemy#4571; surfacing the state in status/doctor stays tracked there.
|
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
The drifting-fault refusal now has a log-coverage assertion: the operator event fires exactly once per decision. The comment states the general invariant without the incident anecdote. Adds the runtime-core test-helpers dev-dependency for the tracing capture.
|
Both findings addressed in 1ecd052: the drifting-fault test now asserts the |
The drifting-corruption test links tracedecay-runtime-core as a dev-dependency; the Bazel test targets did not carry the label, so the unit_test binary failed with E0433. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
0d7c88d
into
ScriptedAlchemy:master
Logging half of #4571.
Problem
recover_deterministically_corrupt_container_withrefuses deletion when the verification open reports a different fault than the first verdict ("hardware-shaped", the module's words). That refusal is terminal for the container: nothing is deleted, nothing is rebuilt, and every later open repeats the verdict — with no operator-visible event. The store looks healthy from the outside; only a leftover.grafeo.corruption-lockhints at it. On one machine, three containers sat in this state for 4–10 days while callers retried forever.Change
The drifting-fault refusal now emits one
error-level event before returning:No behavior change otherwise — deletion still refuses, the typed
Corruptverdict is unchanged, and the existing corrupt-store tests (7/7) stay green. Surfacing this state intracedecay status/doctor stays tracked in #4571.