feat(appview): Notifications, Follower-/Following-Listen, Thread-Route

Bisher erfuhr ein Nutzer nie, dass jemand anderes mit ihm interagiert hat:
Like, Repost, Follow und Reply hinterließen keine Spur, an der der Client
hätte pollen können. Der Tray-/Notification-Pfad im Desktop-Client (Phase 7)
hing damit in der Luft.

Migration 0008:
* notifications(recipient, author, kind, subject_uri, created_at,
  indexed_at, read_at) mit Keyset-Index (recipient, indexed_at DESC, id DESC)
  und Partial-Index auf ungelesene Zeilen für den Badge-Poll.
* Dedupe-Unique-Index über COALESCE(subject_uri, '') — plain NULLs
  kollidieren nicht, sonst gäbe es pro Follow beliebig viele Zeilen.
  Folge: Unlike-Relike erzeugt keine zweite Notification, Toggle-Spam ist
  damit ausgeschlossen.
* Bewusst kein CHECK (recipient <> author): ein Ausrutscher dort würde die
  umgebende Like-Transaktion abbrechen, also das Like wegen eines
  Notification-Bugs verlieren. Gefiltert wird in Rust und im INSERT.

Indexer: record_notification() hängt an upsert_like/-repost (in derselben
Transaktion wie die Counter) sowie upsert_follow/-post. Selbst-Interaktionen
sind still. Empfänger muss uns bekannt sein (profiles- oder posts-Zeile),
sonst würden wir für den gesamten öffentlichen Firehose Zeilen anlegen —
als ein INSERT ... SELECT ... WHERE EXISTS, also ohne TOCTOU-Fenster.
Reply-Notifications tragen die URI der *Antwort* als subject_uri, weil die
Liste den Text zeigt, den der Empfänger noch nicht kennt.

Endpoints: GET /api/notifications, /api/notifications/count,
POST /api/notifications/seen (seenAt als Wasserzeichen),
GET /api/followers, /api/following, GET /api/thread (beide Schreibweisen).
Cursor-Codec, Limit-Clamping und Fehlerform sind die der bestehenden
Endpoints.

/api/post/*uri bleibt wire-kompatibel und teilt sich jetzt
load_thread_context() mit /api/thread — mit max_parents = 1, weil es nur
den direkten Parent serialisiert; die volle Ahnenkette wären bis zu 20
sequenzielle Queries für Zeilen, die danach verworfen werden.

Nebenbei ein Darstellungsfehler: der synthetische Platzhalter-Handle für
Actors ohne bekannten Handle trug ein führendes '@', während jeder Consumer
selbst '@{handle}' rendert — im Feed kam '@@did:plc:abcd…' heraus. Der
Platzhalter ist jetzt durchgängig sigil-frei.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013HC9HLrUU1LNwkzp8nkDLX
This commit is contained in:
tomdebone
2026-09-09 21:36:32 +02:00
co-authored by Claude Opus 5
parent 9d009bfcba
commit c4ca218d97
6 changed files with 2519 additions and 123 deletions
+557
View File
@@ -348,6 +348,31 @@ pub async fn upsert_post(db: &PgPool, row: &mut PostRow) -> Result<()> {
.bind(&row.avatar_cid)
.execute(db)
.await?;
// Reply notification. A post with a `parent_uri` is a reply, so the
// parent's author gets a "someone replied to you" row.
//
// `subject_uri` is the REPLY's own URI, not the parent's: the
// notification list hydrates `subject_uri`'s text, and what the
// recipient wants to read is what the replier wrote — they already
// know the content of their own post. It also makes the row a
// direct link target for "open this reply in the thread view".
//
// The dedupe index keys on the reply URI, so re-indexing (or an
// edit that re-runs the upsert) can't produce a second row.
if let Some(parent_uri) = row.parent_uri.as_deref() {
if let Some(recipient) = post_author_did(db, parent_uri).await? {
record_notification(
db,
&recipient,
&row.did,
NOTIF_REPLY,
Some(&row.uri),
row.created_at,
)
.await?;
}
}
Ok(())
}
@@ -414,6 +439,23 @@ pub async fn upsert_like(
.bind(&post_uri)
.execute(&mut *tx)
.await?;
// Notify the post's author. Runs inside the same transaction as
// the counter bump so a crash can't leave a like counted but
// un-notified (or vice versa). `post_author_did` returns None
// for a post we haven't indexed — then there's nobody to
// notify and we quietly skip.
if let Some(recipient) = post_author_did(&mut *tx, &post_uri).await? {
record_notification(
&mut *tx,
&recipient,
did,
NOTIF_LIKE,
Some(&post_uri),
created_at,
)
.await?;
}
}
tx.commit().await?;
Ok(())
@@ -489,6 +531,20 @@ pub async fn upsert_repost(
.bind(&post_uri)
.execute(&mut *tx)
.await?;
// Same transactional notification write as the like path — see
// `upsert_like` for the rationale.
if let Some(recipient) = post_author_did(&mut *tx, &post_uri).await? {
record_notification(
&mut *tx,
&recipient,
did,
NOTIF_REPOST,
Some(&post_uri),
created_at,
)
.await?;
}
}
tx.commit().await?;
Ok(())
@@ -517,6 +573,124 @@ pub async fn delete_repost(db: &PgPool, did: &str, rkey: &str) -> Result<()> {
Ok(())
}
// -- notifications ---------------------------------------------------------
/// The four notification kinds the AppView produces. Kept as `&str`
/// constants rather than an enum because the value is a plain `TEXT`
/// column guarded by a CHECK constraint (migration 0008) and every
/// call site is a literal — an enum would only add a `to_string()`.
pub const NOTIF_LIKE: &str = "like";
pub const NOTIF_REPOST: &str = "repost";
pub const NOTIF_FOLLOW: &str = "follow";
pub const NOTIF_REPLY: &str = "reply";
/// Should an interaction by `author_did` aimed at `recipient_did`
/// produce a notification row?
///
/// Pure so it can be unit-tested without a database. Two rules:
///
/// 1. **No self-interactions.** Liking your own post, reposting
/// yourself, self-following or replying to yourself must stay
/// silent — the user already knows they did it, and a "you liked
/// your own post" row is pure noise.
/// 2. **No empty DIDs.** An empty recipient means we failed to resolve
/// the target (e.g. a like against a post that isn't in our index),
/// and an empty author means the event was malformed. Either way
/// the row would be unattributable in the UI.
///
/// The same two rules are ALSO enforced in SQL inside
/// [`record_notification`], so a caller that forgets this guard still
/// can't write a bad row — this function exists to skip the round trip
/// in the common self-interaction case and to make the rule testable.
pub fn should_notify(recipient_did: &str, author_did: &str) -> bool {
!recipient_did.is_empty() && !author_did.is_empty() && recipient_did != author_did
}
/// Insert one notification row, if and only if it is warranted.
///
/// Returns `Ok(true)` when a row was actually written, `Ok(false)` when
/// the write was skipped — either because the interaction failed
/// [`should_notify`], because the recipient isn't a user this AppView
/// serves, or because the same notification already exists.
///
/// **"Local user"**: the AppView deliberately has no `users` table —
/// the PDS owns account state. The closest thing we have is "a DID this
/// AppView knows about", i.e. one with a row in the `profiles` cache
/// (written by the PDS profile push / the Jetstream `app.bsky.actor.profile`
/// arm) or at least one indexed post. That is exactly the set of users
/// a client could ever poll notifications for, so restricting writes to
/// it keeps us from materialising a notification row for every like on
/// the entire public firehose while never dropping one a local user
/// would actually see.
///
/// **Idempotency**: the `ON CONFLICT` target is the expression index
/// `notifications_dedupe_idx` from migration 0008, keyed on
/// `(recipient_did, author_did, kind, COALESCE(subject_uri, ''))`. The
/// `COALESCE` is load-bearing: a plain unique index would let two
/// `follow` rows (whose `subject_uri` is NULL) coexist, because SQL
/// NULLs never collide. Re-indexing the same Jetstream event — on a
/// reconnect replay, or via the PDS `/internal/ingest-commit` push that
/// races the firehose — is therefore a no-op.
///
/// The executor is generic so this can run either on the pool (the
/// reply path) or inside the caller's transaction (the like / repost
/// paths, where the notification must commit atomically with the
/// counter bump).
pub async fn record_notification<'e, E>(
exec: E,
recipient_did: &str,
author_did: &str,
kind: &str,
subject_uri: Option<&str>,
created_at: chrono::DateTime<chrono::Utc>,
) -> Result<bool>
where
E: sqlx::PgExecutor<'e>,
{
if !should_notify(recipient_did, author_did) {
return Ok(false);
}
// One statement, so there's no TOCTOU window between "is the
// recipient local?" and "insert". Every parameter is explicitly
// cast because `INSERT ... SELECT $1, $2, ...` gives Postgres no
// column context to infer the placeholder types from.
let res = sqlx::query(
r#"INSERT INTO notifications
(recipient_did, author_did, kind, subject_uri, created_at)
SELECT $1::text, $2::text, $3::text, $4::text, $5::timestamptz
WHERE $1::text <> $2::text
AND $1::text <> ''
AND $2::text <> ''
AND (EXISTS (SELECT 1 FROM profiles WHERE did = $1::text)
OR EXISTS (SELECT 1 FROM posts WHERE did = $1::text))
ON CONFLICT (recipient_did, author_did, kind, COALESCE(subject_uri, ''))
DO NOTHING"#,
)
.bind(recipient_did)
.bind(author_did)
.bind(kind)
.bind(subject_uri)
.bind(created_at)
.execute(exec)
.await?;
Ok(res.rows_affected() > 0)
}
/// Look up the author DID of an indexed post. `None` when the post
/// isn't in our index — which is the normal case for a like/reply
/// aimed at a post hosted somewhere we don't follow. The caller then
/// simply skips the notification rather than guessing a recipient.
async fn post_author_did<'e, E>(exec: E, uri: &str) -> Result<Option<String>>
where
E: sqlx::PgExecutor<'e>,
{
let did: Option<String> = sqlx::query_scalar("SELECT did FROM posts WHERE uri = $1")
.bind(uri)
.fetch_optional(exec)
.await?;
Ok(did)
}
// -- follows ---------------------------------------------------------------
pub async fn upsert_follow(
@@ -540,6 +714,22 @@ pub async fn upsert_follow(
.bind(created_at)
.execute(db)
.await?;
// Notify the followed user. `subject_uri` is NULL — a follow isn't
// about a post — which is exactly the case the dedupe index's
// `COALESCE(subject_uri, '')` exists for. Not wrapped in a
// transaction with the follow upsert: the follow row is the source
// of truth and a missed notification is recoverable noise, whereas
// taking a transaction here would serialise every follow write.
record_notification(
db,
subject_did,
follower_did,
NOTIF_FOLLOW,
None,
created_at,
)
.await?;
Ok(())
}
@@ -1181,6 +1371,373 @@ mod tests {
}
}
#[cfg(test)]
mod notification_tests {
use super::*;
use serde_json::json;
fn did(tag: &str) -> String {
format!("did:plc:notif_{}_{}", tag, uuid::Uuid::new_v4().simple())
}
/// Seed one post so `did` counts as a user this AppView knows
/// about (see [`record_notification`]'s "local user" note) and so
/// there's something to like / reply to.
///
/// The `handle` is deliberately non-empty. `handle_sync`'s tests
/// assert exact counts over a *global* scan of empty-handle rows,
/// so a fixture that left the column blank would silently break
/// them whenever both suites share a database.
async fn seed_post(db: &PgPool, author: &str, rkey: &str) -> String {
let uri = format!("at://{author}/app.twi.post/{rkey}");
sqlx::query(
r#"INSERT INTO posts
(uri, did, handle, rkey, collection, text, cid,
parent_uri, root_uri, langs, created_at)
VALUES ($1,$2,'notif-fixture.test','x','app.twi.post','seed','bafy',
NULL,NULL,NULL, now())
ON CONFLICT (uri) DO NOTHING"#,
)
.bind(&uri)
.bind(author)
.execute(db)
.await
.unwrap();
uri
}
async fn count_notifications(db: &PgPool, recipient: &str, kind: &str) -> i64 {
sqlx::query_scalar(
"SELECT COUNT(*)::BIGINT FROM notifications \
WHERE recipient_did = $1 AND kind = $2",
)
.bind(recipient)
.bind(kind)
.fetch_one(db)
.await
.unwrap()
}
async fn cleanup(db: &PgPool, dids: &[&str]) {
for d in dids {
let _ = sqlx::query(
"DELETE FROM notifications WHERE recipient_did = $1 OR author_did = $1",
)
.bind(d)
.execute(db)
.await;
let _ = sqlx::query("DELETE FROM likes WHERE did = $1")
.bind(d)
.execute(db)
.await;
let _ = sqlx::query("DELETE FROM reposts WHERE did = $1")
.bind(d)
.execute(db)
.await;
let _ = sqlx::query("DELETE FROM follows WHERE follower_did = $1 OR subject_did = $1")
.bind(d)
.execute(db)
.await;
let _ = sqlx::query("DELETE FROM posts WHERE did = $1")
.bind(d)
.execute(db)
.await;
}
}
#[test]
fn should_notify_rejects_self_and_empty() {
assert!(should_notify("did:plc:a", "did:plc:b"));
// Self-interaction: liking / replying to / following yourself.
assert!(!should_notify("did:plc:a", "did:plc:a"));
// Unresolvable ends of the edge.
assert!(!should_notify("", "did:plc:b"));
assert!(!should_notify("did:plc:a", ""));
assert!(!should_notify("", ""));
// Case matters — DIDs are compared verbatim, never folded.
assert!(should_notify("did:plc:A", "did:plc:a"));
}
/// A like by someone else must produce exactly one notification,
/// and re-indexing the same event (Jetstream replay racing the PDS
/// push) must not produce a second.
#[tokio::test]
async fn like_notifies_author_once() {
let Some(db) = try_test_db().await else {
eprintln!("appview DB unavailable; skipping");
return;
};
let author = did("author");
let liker = did("liker");
// The liker also needs to exist for nothing in particular —
// only the *recipient* is checked — but seeding both keeps the
// fixture symmetric with reality.
let post_uri = seed_post(&db, &author, "p1").await;
seed_post(&db, &liker, "p1").await;
let record = json!({
"subject": { "uri": post_uri, "cid": "bafysubject" },
"createdAt": "2026-07-01T12:00:00Z"
});
upsert_like(&db, &liker, "lrk1", Some("bafylike"), Some(&record))
.await
.unwrap();
assert_eq!(count_notifications(&db, &author, NOTIF_LIKE).await, 1);
// Replay the exact same event.
upsert_like(&db, &liker, "lrk1", Some("bafylike"), Some(&record))
.await
.unwrap();
assert_eq!(
count_notifications(&db, &author, NOTIF_LIKE).await,
1,
"replayed like must not duplicate the notification"
);
// Unlike + re-like under a NEW rkey. The like row is recreated
// but the notification tuple is unchanged, so the dedupe index
// suppresses it — see migration 0008's "Idempotency" note.
delete_like(&db, &liker, "lrk1").await.unwrap();
upsert_like(&db, &liker, "lrk2", Some("bafylike2"), Some(&record))
.await
.unwrap();
assert_eq!(
count_notifications(&db, &author, NOTIF_LIKE).await,
1,
"toggling a like must not be a notification-spam vector"
);
// The stored row must point at the liked post and name the
// liker as the author.
let row: (String, Option<String>) = sqlx::query_as(
"SELECT author_did, subject_uri FROM notifications \
WHERE recipient_did = $1 AND kind = $2",
)
.bind(&author)
.bind(NOTIF_LIKE)
.fetch_one(&db)
.await
.unwrap();
assert_eq!(row.0, liker);
assert_eq!(row.1.as_deref(), Some(post_uri.as_str()));
// Unread by default — that's what /api/notifications/count sees.
let unread: i64 = sqlx::query_scalar(
"SELECT COUNT(*)::BIGINT FROM notifications \
WHERE recipient_did = $1 AND read_at IS NULL",
)
.bind(&author)
.fetch_one(&db)
.await
.unwrap();
assert_eq!(unread, 1);
cleanup(&db, &[&author, &liker]).await;
}
/// Liking / reposting your own post is silent.
#[tokio::test]
async fn self_interaction_writes_no_notification() {
let Some(db) = try_test_db().await else {
eprintln!("appview DB unavailable; skipping");
return;
};
let author = did("selfie");
let post_uri = seed_post(&db, &author, "p1").await;
let record = json!({
"subject": { "uri": post_uri, "cid": "bafysubject" },
"createdAt": "2026-07-01T12:00:00Z"
});
upsert_like(&db, &author, "lrk1", None, Some(&record))
.await
.unwrap();
upsert_repost(&db, &author, "rrk1", None, Some(&record))
.await
.unwrap();
// Self-follow is legal in the protocol; it must stay silent too.
upsert_follow(&db, &author, &author, None).await.unwrap();
let total: i64 = sqlx::query_scalar(
"SELECT COUNT(*)::BIGINT FROM notifications WHERE recipient_did = $1",
)
.bind(&author)
.fetch_one(&db)
.await
.unwrap();
assert_eq!(total, 0, "self-interactions must not notify");
// The like/repost themselves still landed — the notification
// suppression must not swallow the interaction.
let likes: i64 = sqlx::query_scalar("SELECT COUNT(*)::BIGINT FROM likes WHERE did = $1")
.bind(&author)
.fetch_one(&db)
.await
.unwrap();
assert_eq!(likes, 1);
cleanup(&db, &[&author]).await;
}
/// A repost notifies, and a reply notifies the *parent's* author
/// with the reply's own URI as the subject.
#[tokio::test]
async fn repost_and_reply_notify() {
let Some(db) = try_test_db().await else {
eprintln!("appview DB unavailable; skipping");
return;
};
let author = did("parent");
let other = did("replier");
let parent_uri = seed_post(&db, &author, "p1").await;
seed_post(&db, &other, "p1").await;
let record = json!({
"subject": { "uri": parent_uri, "cid": "bafysubject" },
"createdAt": "2026-07-01T12:00:00Z"
});
upsert_repost(&db, &other, "rrk1", None, Some(&record))
.await
.unwrap();
assert_eq!(count_notifications(&db, &author, NOTIF_REPOST).await, 1);
// Now a reply from `other` to `author`'s post.
let reply_record = json!({
"text": "nice one",
"createdAt": "2026-07-01T12:05:00Z",
"reply": {
"parent": { "uri": parent_uri, "cid": "bafyparent" },
"root": { "uri": parent_uri, "cid": "bafyparent" }
}
});
let mut row = PostRow::from_record(
&other,
"replykey",
"app.twi.post",
"bafyreply",
&reply_record,
None,
);
let reply_uri = row.uri.clone();
upsert_post(&db, &mut row).await.unwrap();
let subject: Option<String> = sqlx::query_scalar(
"SELECT subject_uri FROM notifications \
WHERE recipient_did = $1 AND kind = $2",
)
.bind(&author)
.bind(NOTIF_REPLY)
.fetch_one(&db)
.await
.unwrap();
assert_eq!(
subject.as_deref(),
Some(reply_uri.as_str()),
"a reply notification's subject is the REPLY, so the list can \
preview what was written"
);
// Re-indexing the reply must not duplicate.
upsert_post(&db, &mut row).await.unwrap();
assert_eq!(count_notifications(&db, &author, NOTIF_REPLY).await, 1);
cleanup(&db, &[&author, &other]).await;
}
/// A follow notifies the followed user with a NULL subject_uri —
/// the case the dedupe index's `COALESCE(subject_uri, '')` exists
/// for, since plain SQL NULLs never collide.
#[tokio::test]
async fn follow_notifies_subject_and_dedupes_on_null_subject() {
let Some(db) = try_test_db().await else {
eprintln!("appview DB unavailable; skipping");
return;
};
let subject = did("followee");
let follower = did("follower");
seed_post(&db, &subject, "p1").await;
let record = json!({ "subject": subject, "createdAt": "2026-01-01T00:00:00Z" });
upsert_follow(&db, &follower, &subject, Some(&record))
.await
.unwrap();
upsert_follow(&db, &follower, &subject, Some(&record))
.await
.unwrap();
assert_eq!(
count_notifications(&db, &subject, NOTIF_FOLLOW).await,
1,
"two NULL-subject follow rows must collide, not coexist"
);
let subject_uri: Option<String> = sqlx::query_scalar(
"SELECT subject_uri FROM notifications \
WHERE recipient_did = $1 AND kind = $2",
)
.bind(&subject)
.bind(NOTIF_FOLLOW)
.fetch_one(&db)
.await
.unwrap();
assert!(subject_uri.is_none(), "a follow is not about a post");
cleanup(&db, &[&subject, &follower]).await;
}
/// A recipient the AppView has never seen (no profile row, no
/// posts) gets nothing — this is what keeps us from materialising
/// a row for every like on the public firehose.
#[tokio::test]
async fn unknown_recipient_is_skipped() {
let Some(db) = try_test_db().await else {
eprintln!("appview DB unavailable; skipping");
return;
};
let stranger = did("stranger");
let author = did("author");
let wrote = record_notification(
&db,
&stranger,
&author,
NOTIF_FOLLOW,
None,
chrono::Utc::now(),
)
.await
.unwrap();
assert!(!wrote, "unknown recipient must not get a notification row");
// Give the recipient a post — now they're a user we serve.
seed_post(&db, &stranger, "p1").await;
let wrote = record_notification(
&db,
&stranger,
&author,
NOTIF_FOLLOW,
None,
chrono::Utc::now(),
)
.await
.unwrap();
assert!(wrote, "known recipient must get the row");
// And the second call is a no-op thanks to ON CONFLICT.
let wrote = record_notification(
&db,
&stranger,
&author,
NOTIF_FOLLOW,
None,
chrono::Utc::now(),
)
.await
.unwrap();
assert!(!wrote, "duplicate must report false, not error");
cleanup(&db, &[&stranger, &author]).await;
}
}
/// Backfill the `posts.handle` column for every row belonging to
/// `did`. Used by the Jetstream `identity` handler when Jetstream
/// tells us a DID's handle has changed — every existing post row