refactor(db): move the authentication store into its own schema - #64
Merged
Conversation
The port was generic in name only. Its methods returned DiscordGuildMembership, DiscordMessage and DiscordReactionUser, and every identifier was validated against the Discord snowflake format, so the identifier regex reached callers that have no reason to know what a snowflake is. Replacing the provider would have meant editing the interface and both api_v0 services rather than swapping an adapter. The port now speaks CommunityRole, CommunityMembership, CommunityMessage, CommunityReactionUser and CommunityAccountProfile, treats identifiers as opaque strings, and drops the Discord message length limit. Discord's snowflake format, its 2000-character limit, and the global_name field it returns are refinements applied in discord/schema.ts, which is the only layer that issues those values. The provider-specific field name is mapped to displayName at the adapter boundary, matching the provider_display_name column it is stored in. Behaviour is unchanged: the adapter still rejects a malformed provider response and a guild member response for another user, both of which depend on the snowflake assertion that moved. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The reaction summary returned the complete member record — student ID, student email, emergency contact, insurance and allergy details — inside every reaction's user list as well as in `members`. A member who reacted with three emoji had those fields serialised four times, so the private data on the wire grew with the number of reactions rather than the number of members. The badges only ever rendered names: the client maps that list through getDisplayName and reads nothing else from it. They now carry a ReactionParticipant with the identity and name fields, while `members` keeps the full record the admin table and its CSV export need, including the emergency contact and allergy details an organiser relies on. Adds a regression test asserting no private field appears in the badge payload. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Better Auth's tables move to app_auth, and every reference that crossed between them and the domain is removed, so the authentication store can be lifted into its own database without touching a domain table. Moving the tables alone would not have achieved that: a foreign key cannot span databases, and three of them crossed this boundary, one of which only surfaced when the new assertion ran. The domain now keeps its own account record in public.app_accounts, holding the membership link and the application role. role has to live on this side because the membership trigger authorizes against it, and a trigger cannot read another database. user_id is stored as a value rather than a foreign key; the reviewer and community identity references become snapshots too, matching member_status_history, which already recorded its actor that way. Losing those foreign keys costs the guarantee that a reviewed user cannot be deleted. That is the price of the move, and it is now stated in the schema comments and asserted in the pgTAP suite rather than left implicit. The auth middleware is the only place that reads both sides. It resolves the subject from the authentication store and the account from the domain in two separate queries, so a future split turns the first into a remote call and leaves the second untouched. The admin views that display an account email do the same thing in batch instead of joining. A replay test asserts the tables sit in app_auth, that "user" carries no domain column, and that no foreign key crosses the boundary in either direction. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The RLS suite still required a reviewed account to be undeletable, which was the guarantee the dropped foreign key provided. That key crossed into the authentication store, so the assertion now states what replaced it: the delete succeeds and the review record keeps the reviewer it was written with. Caught by the pgTAP job, which the membership work added and which is the only thing that executes these suites. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Sign-in now requires a public.app_accounts row, which the Better Auth create hook provisions. If that hook ever fails, or a subject is seeded outside the flow, the account exists in the authentication store with nothing on the domain side, and the JWT endpoint and the auth middleware both reject it. The user is then locked out permanently with no path back except editing the database. The sign-in path now inserts the row when it is absent, which costs one statement per sign-in and makes the missing row self-correcting. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
yufoxda
force-pushed
the
refactor/reaction-payload-pii
branch
from
July 27, 2026 07:46
4299e4d to
f2650c9
Compare
yufoxda
force-pushed
the
refactor/split-auth-schema
branch
from
July 27, 2026 07:46
049739a to
2e77823
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
「auth を別スキーマへ。将来は別DBへ」という方針の実装です。
「移すだけ」では成立しなかった理由
テーブルを
app_authへ移すだけでは目的を達成できません。外部キーはDBを跨げないためです。同一DB内なら跨ぐFKはそのまま動いてしまい、別DBへ切り出す当日に初めて全部が壊れます。跨いでいたFKは3本ありました:
4本目(
auth_account_id)は設計時に見落としていたもので、新しく追加した「跨ぐFKが存在しないこと」の表明が検出しました。さらに、ドメインのトリガーが
public."user"を読んで role を検証していました。トリガーは別DBを読めないため、これも解消が必要でした。設計
roleはドメイン側に置きました。メンバーシップのトリガーが role で認可判定しており、トリガーは別DBを読めないためですapp_accountsもドメイン側です。承認処理がmembersと会員リンクを1トランザクションで更新しており、これが別DBに分かれると原子性を失うためですadditionalFieldsを撤去。認証テーブルは完全にライブラリの所有物になりましたdatabaseHooks.user.create.afterでドメイン側の口座行を作ります(トリガーにすると認証スキーマがドメインに依存し直すため)失うもの(正直に)
members.reviewed_by_user_idのon delete restrict(「レビュー履歴がある限りユーザーを削除できない」)が失われます。これが今回の対価です。ただし既存設計と矛盾しません。
member_status_history.changed_by_user_idは元々FKなしのスナップショット方式で、pgTAP にも表明がありました。同じ方針を reviewer と identity にも広げた形です。暗黙にせず、スキーマコメントと pgTAP の表明として明文化しました。境界層
auth とドメインの両方を読むのは認証ミドルウェアだけです。JOINせず2クエリに分けてあるため、別DB化したとき1つ目がリモート呼び出しになるだけで2つ目は無変更で済みます。管理画面が表示する認証メールも同様にバッチ取得へ変更しました。
退行防止
再生テストで以下を表明しています(これが将来の別DB化を守ります):
app_authにありpublicに無いことapp_auth."user"にドメイン列が無いこと検証
tsc --noEmitクリーンpgTAP 3スイートと runbook 3件も新配置に追従させました。
purgerunbook は分離マイグレーションより前に実行される手順のため参照先はpublicのままが正しく、順序をコメントで明記してあります。🤖 Generated with Claude Code