The full suite caught it: docs/INDEX.md routes by this metadata, so a page
without it is invisible to the router that exists to make the corpus findable.
Failed on gh-main and would have failed CI.
(cherry picked from commit 0ded9207ab)
Enabling a plugin could not make its routes live. Flask refuses
register_blueprint once the app has served a request, and each gunicorn worker
holds its own app object, so registering in the worker that served the toggle
would leave the others answering 404 for the same URL.
Worse, the window before the restart was a window in which workers disagreed
about reality. Four defects, all of which answered 200 while going wrong:
- Enabling a not-yet-installed plugin ALWAYS 500'd on a running app:
install_plugin ran the migrations, updated the registry, and only then called
register_blueprint and hit Flask's assertion. The toggle half-applied and
reported failure.
- Enabling one plugin could hang the entire site. @jwt_required does SELECT
users, so the request's transaction sits open holding a shared metadata lock;
the plugin's migration runs on its own Alembic connection and needs an
EXCLUSIVE lock on users for a foreign key. Each waits on the other, and MySQL
grants metadata locks in request order, so every later login queued behind the
pending exclusive. Observed twice as a total outage from one plugin toggle.
- Each worker rewrote the WHOLE registry file from a snapshot taken at boot, so
concurrent enables erased each other: seven enables persisted three, with
HTTP 200 for all seven.
- Dependency checks read that same stale snapshot, refusing a plugin whose
dependency another worker had just enabled.
Routes still require the process to restart; what changes is who initiates it.
Each worker fingerprints instance/plugins.json at boot and exits gracefully when
it changes, and the supervisor replaces it. That is the primitive gunicorn's own
--reload uses. Recycles are staggered per worker so the pool rolls rather than
dropping at once, and the exit is scoped to gunicorn: on Windows sys.exit() in a
daemon thread ends the thread rather than the process, so IIS keeps the manual
restart until that path is proven on the Windows VM.
The registry now merges under an exclusive lock and replaces atomically, saves
only the entry being changed, and re-reads before any decision is made from it.
ADR-020 records the decision; ADR-013 is amended where it said the UI must tell
the operator to restart. The restart requirement and the honesty requirement
both stand - only the manual step is removed, and only where it is proven.
(cherry picked from commit 9d6342b3ab)
