Files
API/docs/calendar-auth-migration.md
T
Paddy 13a0c07d1b
Jenkins Production Deployment
Read calendar event creators from the admin module, and archive the old ones (#14)
Reviewed-on: #14
Co-authored-by: Patrick Müller <mail@pmueller.me>
Co-committed-by: Patrick Müller <mail@pmueller.me>
2026-09-06 21:12:16 +00:00

266 lines
17 KiB
Markdown

# Migrating the Calendar domain onto the admin identity module
Status: **steps 1-4 implemented 2026-09-06, not yet merged or deployed.** Step 2 dropped by
decision, part of step 5 brought forward. Only step 5, the removal of the legacy path, is
left to write.
> Read the deploy checklist under step 4 before applying anything. "Done" below means the
> code exists on a branch, **not** that production has it - and in particular production has
> none of the three migrations. Step 5 is scoped but deliberately unstarted: it must not be
> built on top of a step 4 that has not been deployed and watched.
Written 2026-09-05 alongside the admin module (step 2 of `docs/plan-admin-auth.md` in the
nachklang-admin repo), which deliberately left the calendar alone. Steps 1-4 of that plan
are now live, so the calendar is the last module still on the legacy query-parameter
sessions.
## Why the calendar was left out
The admin module replaced authentication for feedback and tickets by swapping one
middleware. The calendar cannot be done that way, because its user identity is woven into
its data:
- `users`/`sessions` live in the **calendar** database and are the same tables the
feedback and tickets admin areas used to authenticate against.
- `events.created_by_id` is an **INT** foreign key into `users.user_id`. The admin module's
user ids are **VARCHAR(36)** strings. Migrating identity means migrating that column and
every query that joins it.
- The Angular frontend passes `sessionId`/`sessionKey` as **query parameters**
(`DEFERRED_SECURITY.md` item 1). Cookie sessions remove the parameters entirely, so
every calendar route signature and the frontend's HTTP layer change together.
- `credentials.service.ts` implements a second, parallel authorisation model: the
`MEMBER_CREDENTIAL` / `CHOIR_CREDENTIAL` / `MANAGEMENT_CREDENTIAL` shared secrets that
let non-users read specific calendars. That has no equivalent in the admin module and is
not a per-user permission at all.
What already exists today: `calendar` is a value in the `user_app_permissions.app` enum, so
permissions can be granted before anything else moves.
## What is in place to build on
- Cookie sessions across `*.nachklang.art`, and `requireAppAccess('calendar')` in
`src/models/admin/admin.middleware.ts` - usable the moment a calendar route wants it.
- `res.locals.admin` is `{id, email, displayName, apps}`; `id` is the string user id.
- Invitations, disable/enable and session revocation already cover calendar users, because
they are properties of the account rather than of an app.
## Suggested sequence
Each step is meant to leave production working on its own.
1. **Add a bridging column.** ~~`ALTER TABLE events ADD COLUMN created_by_user_id
VARCHAR(36) NULL`, indexed. Nothing reads it yet.~~ **Done 2026-09-06**, as
`sql/calendar/001_add_admin_user_bridge.sql` - the first migration this repo owns for the
calendar schema, mirrored into `docker/init/01-calendar-schema-dev.sql`. It covers both
`events.created_by_user_id` and `event_versions.version_created_by_user_id`, and carries
no foreign key (see "What the code actually looks like" below). The dev seed leaves two
events on the legacy path and gives one an admin id, so step 3's dual-read has both cases
to exercise. Verified by applying the pre-migration schema and then the migration to a
throwaway MariaDB 11 container, and diffing `SHOW CREATE TABLE` against a fresh dev
schema: identical. Applied to the running dev database on the same day; a dev container
created before then needs it applied, or recreating.
2. ~~**Map the accounts.**~~ **Dropped 2026-09-06.** There is no backfill: since the
creator is only ever a display name (see below), old events keep resolving through the
legacy join until step 5 and then simply lose the name. Re-inviting the people who
actually still need calendar access remains an operational task, but it is no longer a
migration step and nothing is blocked on it.
3. **Dual-read.** ~~Change `events.service.ts` to prefer `created_by_user_id` and fall back
to `created_by_id`. Writes fill both.~~ **Done 2026-09-06.** `events.service.ts` now reads
both columns and prefers the admin one, resolving the name through a single
`findDisplayNames` lookup against the admin database per result set (added to
`users.admin.service.ts` for this). Four copies of the same SELECT and four copies of the
row mapper were collapsed into one of each first - the dual read would otherwise have had
to be written four times.
A name now has three possible sources, tried weakest first: the legacy join, then the
`created_by_name` snapshot from migration 002, then the live admin lookup - which wins
because it is the only one that follows an account being renamed. An admin id that no
longer resolves falls back rather than blanking, and a failure to reach the admin database
is caught and logged rather than propagated, so an anonymous read of the public calendar
never depends on the admin database being up. Covered by
`test/calendar/events.service.test.ts`.
**Writes are not dual-written**, contrary to the original plan: before the cutover the
request only ever carries a legacy session, so there is no admin id available to write.
Writes start filling `created_by_user_id` (and stop filling `created_by_id`) in step 4.
4. **Switch the routes.** ~~Replace the query-parameter session checks in `events.router.ts`
and `users.router.ts` with `requireAppAccess('calendar')`, and change the Angular frontend
to `withCredentials: true`.~~ **Done 2026-09-06.** `DEFERRED_SECURITY.md` item 1 is closed:
no route reads `sessionId`/`sessionKey` any more.
How it came out, route by route:
- The four write routes sit behind `requireAppAccess('calendar')` as middleware. They
answer 401 when signed out and 403 without the permission, where they used to answer 403
for both.
- The three read routes cannot use middleware - the same URL serves an anonymous visitor,
an iCal subscription holding a shared password, and a signed-in editor who should see
drafts. They call `resolveAccess` optionally instead (`signedInEditor` in the router),
and a signed-in user *without* the calendar permission is treated as anonymous rather
than refused, so they keep their access to the public calendar.
- `credentials.service.ts` lost its session half entirely and is now just the password
table. `hasAccess(calendar, password)`.
- `/calendar/users/*` was left alone. Nothing calls it and a session it mints opens
nothing, but they are live password-accepting endpoints - step 5 removes them.
Also: `calendar.nachklang.art` joined `DEFAULT_APP_ORIGINS` (better-auth `trustedOrigins`,
without which sign-out from the calendar fails while everything else works), and
`localhost:4200` joined the dev origins for the same reason.
Two things this step had to carry that the original sequence put in step 5:
- **`sql/calendar/003_allow_null_legacy_creator.sql` makes `events.created_by_id` nullable**
(`MODIFY created_by_id INT NULL`). It is `NOT NULL` today, so the first event created after the
cutover would otherwise fail to insert - there is no legacy int id to write any more.
`event_versions.version_created_by_id` is already nullable. The foreign key can stay
until step 5; it permits NULL. It also re-runs 002's idempotent name backfill, to catch
anything created between the two migrations. Applying it early is safe - widening a
column to accept NULL cannot break the running pre-cutover build.
- **The public calendar stays anonymous.** `hasAccess('public')` returns true before any
credential check, and nachklang.art reads `/calendar/events/public/json` and
`/public/json/next` with no session at all. Pinned at both levels - the password table in
`test/calendar/credentials.service.test.ts`, the routes themselves in
`test/calendar/events.router.test.ts` - so this cannot regress quietly.
### Deploy checklist
Production has **none** of the three migrations: 001 and 002 were only ever applied to the
dev database. The API build below selects `created_by_user_id` and `created_by_name` on
every read, so deploying it against a database missing them fails every calendar request
including the anonymous public feed the website uses. In order:
1. **Apply `sql/calendar/001`, `002`, `003`, in that order**, against `CALENDAR_DB`. All
three are re-runnable, so applying one that is already applied is a no-op. Verify
before continuing:
`SHOW COLUMNS FROM events LIKE '%by_user%'; SHOW COLUMNS FROM events LIKE '%by_name%';`
- four rows across the two tables, and `created_by_id` nullable.
2. **Check `APP_ORIGINS` on the API vhost.** `calendar.nachklang.art` is in the code's
default list, but the environment variable *replaces* that list rather than adding to
it - so if it is set at all (the tickets/feedback cutover may have set it), append
`https://calendar.nachklang.art` or the calendar's sign-out will 403 while everything
else works. That is the failure mode the comment in `admin.config.ts` warns about.
3. **Deploy the API.**
4. **Deploy the calendar frontend immediately after.** Do not leave a gap - see below.
5. **Re-run 002's two `UPDATE` statements.** Between step 1 and step 3 the old API was
still writing `created_by_id` with no snapshot; those few rows would otherwise lose
their author at step 5.
6. **Rebuild the admin app** if `NEXT_PUBLIC_ALLOWED_REDIRECT_ORIGINS` does not already
contain `https://calendar.nachklang.art`. It is a **build-time** value, so a restart
does nothing.
**The window between steps 3 and 4 does not look broken, which is the danger.** The old
Angular bundle starts by calling `POST /calendar/users/checkSessionValid`, and those
legacy routes are untouched - so it still succeeds and the page renders as signed in. What
the user then sees is an empty event table and saves that silently do nothing. It looks
like the calendar lost its data, not like a deploy in progress. Keep the gap to minutes,
or take the frontend offline for it.
**One-way door:** any iCal subscription whose URL carries `?sessionId=&sessionKey=` rather
than `?password=` stops working permanently. The shared-password URLs are unaffected.
5. **Drop the legacy path.** Not started - and deliberately not started until step 4 has been
deployed and watched, because it removes the fallback step 4 still leans on. Scoped and
decided 2026-09-06; what follows is the agreed shape, not a suggestion.
**Prerequisite: step 4 live in production and behaving.** Until then the legacy join is
what renders the author of every pre-cutover event, and the legacy routes are what an old
cached bundle talks to. Doing this first turns a recoverable deploy into an unrecoverable
one.
Code, in one branch:
- **Delete `src/models/calendar/users/` entirely** - `users.router.ts`, `users.service.ts`,
`session.interface.ts`, `user.interface.ts` - and the `calendarRouter.use('/users', ...)`
line in `Calendar.router.ts`. *(Decided: delete outright rather than unmount.)* This
removes the last unauthenticated account-creation and mail-sending endpoint in the API.
A survey on 2026-09-06 confirmed nothing outside that directory imports it, and nothing
outside it touches the `users`/`sessions` tables except the two joins below.
- **Drop the legacy half of the read** in `events.service.ts`: the two
`LEFT OUTER JOIN users` clauses, the `legacy_*` aliases, and `created_by_id` /
`version_created_by_id` from the SELECT and the row mapper. The snapshot fallback stays -
it is what makes this safe. Remove `createdById` / `lastModifiedById` from
`event.interface.ts` and their (already deprecated) swagger properties.
- **Remove `X-Session-Id` / `X-Session-Key`** from the CORS `allowedHeaders` in
`src/app.factory.ts`. Nothing has sent them since the tickets and feedback frontends were
redeployed.
- **Drop the obsolete test mocks**: `test/feedback/feedback.auth.test.ts`,
`test/tickets/tickets.auth.test.ts` and `test/admin/auth-binding.ts` each mock
`calendar/users/users.service.js` and assert `checkSession` is never called. That
tripwire is meaningless once the module does not exist; remove the mock and the
assertion, keep the rest.
Database, as `sql/calendar/004_*.sql`:
- Drop the foreign keys `events_users_user_id_fk` and `event_versions_users_user_id_fk`,
then the `created_by_id` and `version_created_by_id` columns.
- **`RENAME TABLE users TO users_legacy_archive`**, same for `sessions`. *(Decided: rename
rather than drop.)* The reasoning: the display names are already snapshotted so nothing
visible depends on these rows, but they still hold the old e-mail addresses and password
hashes, and a rename makes the tables unreachable without destroying anything. Dropping
them later is one statement, at a moment when nobody is mid-deploy.
- Mirror all of it in `docker/init/01-calendar-schema-dev.sql` (the archive tables need no
mirror - a fresh dev database has nothing to archive).
Documentation: `DEFERRED_SECURITY.md` items **3** (activation token has no expiry) and
**4** (password reset token has no expiry) close outright - both describe code that ceases
to exist. Item 2 (no event ownership check) stays open.
Two consequences to accept explicitly rather than discover:
- Any activation or password-reset e-mail already sent points at
`api.nachklang.art/calendar/users/activate` and becomes a 404. Those links were only ever
valid for legacy accounts, which no longer open anything.
- `Event.createdById` disappears from the API response. The Angular frontend never read it
(its `Event` model has only `createdBy`, the name), so this is not a breaking change for
the only known consumer - but it is a wire-format removal, so check anything else that
reads `/calendar/events/*/json` first.
## What the code actually looks like (surveyed 2026-09-06)
Four things found while doing step 1 that change how the later steps should be built:
- **`created_by_id` is display-only.** Nothing authorises on it. `events.router.ts` gates
PUT, POST, DELETE and `/move` on `user?.isActive` alone - there is no "only the creator may
edit" rule anywhere - and the column is read back solely to render `created_by_name` and
`last_modified_by_name`. That de-risks steps 2, 3 and 5 considerably: an event whose
creator never gets re-invited loses a name in the UI, it does not become uneditable or
invisible. It also means the step 2 backfill is best-effort, not a precondition.
- **The two schemas are separate databases.** `nachklang_calendar` and `nachklang_admin`
have their own connection pools (`Calendar.db.ts` vs the admin module's Kysely instance).
So the bridging columns get no foreign key, and - the part the original sequence missed -
**the `LEFT OUTER JOIN users` that produces the creator's name cannot simply be repointed**.
It would have to become a cross-schema join, which hardcodes the admin database name into
calendar SQL and ties the two schemas together exactly as an FK would. Recommendation for
step 3: drop the join for the new path and resolve names in the service layer instead -
collect the distinct ids from the result set and do one lookup against the admin users
service. One extra query per listing, no coupling, and it keeps working if the admin
database ever moves.
- **`events.created_by_id` is `NOT NULL`.** Step 5 cannot simply stop writing it; that step
has to drop the column (and its FK to `users`) in the same migration that stops the writes,
or make it nullable first.
- **Every calendar read already hits the session table.** `/:calendar/json` calls
`UserService.checkSession` before falling back to `credentials.service.ts`, so the shared
credentials are the *fallback*, not the primary path. Step 4 replaces the first half of
that with `requireAppAccess('calendar')` and has to decide what happens to the second half
- which is the first open question below.
## Open questions to settle before starting
**Settled 2026-09-06:**
- **The shared calendar credentials keep working, but only for iCal.** The web app goes
cookie-only at step 4; `MEMBER_CREDENTIAL` and friends survive on
`GET /calendar/events/{calendar}/ical`, which is the one case where the client genuinely
cannot send a cookie. Everything else in `credentials.service.ts` goes with step 5.
`public` stays anonymous everywhere - see the note under step 4.
- **The iCal export keeps its own scheme.** Same reasoning; it is the reason the shared
credentials survive at all rather than an exception to their removal.
- **No account backfill.** See step 2 above.
- **Pre-cutover authorship is archived, not discarded.** `events.created_by_name` and
`event_versions.version_created_by_name`, backfilled once by migration 002 and never
written again. This was originally listed as a step 5 question; it was brought forward so
the data is safe well before the table that holds it is dropped.
**Nothing is open.** The last one - ~~`event_versions.version_created_by_id`~~, the same INT
reference on the version rows - was handled in passing: step 1 gave it a sibling bridging
column, step 2 a sibling snapshot, and step 3 reads it exactly like `events`.