perf(clips): explore po podzapytaniu okna, kandydaci bez ORM i bez tagow, max(uuid) przez tekst

Strona feedu spadla z 310-470 ms do 50-130 ms (srednio 85 ms na 30 stronach,
0 powtorek). Postgres nie ma max(uuid): koniec okna liczony po rzutowaniu na
tekst (ten sam porzadek dla malych hex w stalym formacie).

app/api/clips.py

@@ -23,7 +23,7 @@ from datetime import UTC, datetime
 
 from fastapi import APIRouter, Depends, HTTPException, Query, Response
 from pydantic import BaseModel, Field
-from sqlalchemy import func, select, text
+from sqlalchemy import String, cast, func, select, text
 from sqlalchemy.dialects.postgresql import insert as pg_insert
 from sqlalchemy.orm import Session
 
@@ -247,39 +247,47 @@ def _visible_clauses(device_id: str, bl: dict[str, set[str]]) -> list:
     return clauses
 
 
-def _candidates(session: Session, ids: list[uuid.UUID], clauses: list) -> list[cf.Candidate]:
-    if not ids:
-        return []
-    rows = session.execute(select(Clip).where(Clip.id.in_(ids), *clauses)).scalars().all()
-    tags = _tags_for(session, [r.id for r in rows])
+_CAND_COLS = (Clip.id, Clip.creator, Clip.likes, Clip.published_at, Clip.niches)
+
+
+def _rows_to_candidates(session: Session, rows, *, with_tags: bool) -> list[cf.Candidate]:
+    """Wiersze `_CAND_COLS` → Candidate. Tagi tylko na życzenie: pula explore punktuje
+    wyłącznie lajkami, więc ciągnięcie tagów dla 300 kandydatów było czystym kosztem."""
+    tags = _tags_for(session, [r[0] for r in rows]) if with_tags else {}
     return [
-        cf.Candidate(
-            id=r.id, creator=r.creator, likes=r.likes, published_at=r.published_at,
-            tags=tags.get(r.id, []), niches=list(r.niches or []),
-        )
+        cf.Candidate(id=r[0], creator=r[1], likes=r[2], published_at=r[3], tags=tags.get(r[0], []), niches=list(r[4] or []))
         for r in rows
     ]
 
 
-def _explore_window(session: Session, pk: uuid.UUID) -> tuple[list[uuid.UUID], bool]:
-    """Kolejne `_EXPLORE_WINDOW` id po `pk`; gdy okno się kończy, zawija do zera."""
-    ids = [
-        r[0]
-        for r in session.execute(
-            select(Clip.id).where(Clip.id > pk, Clip.dead_at.is_(None)).order_by(Clip.id).limit(_EXPLORE_WINDOW)
-        )
-    ]
-    wrapped = False
-    if len(ids) < _EXPLORE_WINDOW and pk != _UUID_ZERO:
-        wrapped = True
-        more = [
-            r[0]
-            for r in session.execute(
-                select(Clip.id).where(Clip.dead_at.is_(None)).order_by(Clip.id).limit(_EXPLORE_WINDOW - len(ids))
-            )
-        ]
-        ids.extend(more)
-    return ids, wrapped
+def _candidates(session: Session, ids: list[uuid.UUID], clauses: list) -> list[cf.Candidate]:
+    if not ids:
+        return []
+    rows = session.execute(select(*_CAND_COLS).where(Clip.id.in_(ids), *clauses)).all()
+    return _rows_to_candidates(session, rows, with_tags=True)
+
+
+def _explore_candidates(
+    session: Session, pk: uuid.UUID, clauses: list, limit: int
+) -> tuple[list[cf.Candidate], uuid.UUID, bool]:
+    """Widoczni kandydaci z okna `_EXPLORE_WINDOW` id po `pk` (podzapytanie, bez
+    przenoszenia 2000 uuid przez SQL). Zwraca (kandydaci, koniec okna, czy okno było
+    ostatnie). Ostatnie okno = katalog wyczerpany od `pk`, następna strona zawija do zera."""
+    window = (
+        select(Clip.id).where(Clip.id > pk, Clip.dead_at.is_(None)).order_by(Clip.id).limit(_EXPLORE_WINDOW)
+    ).subquery()
+    # Postgres nie ma max(uuid); tekstowo porządek jest ten sam (małe hex, stały format).
+    end_txt, n = session.execute(
+        select(func.max(cast(window.c.id, String)), func.count()).select_from(window)
+    ).one()
+    if not n:
+        return [], _UUID_ZERO, True
+    end_id = uuid.UUID(end_txt)
+    rows = session.execute(
+        select(*_CAND_COLS).where(Clip.id.in_(select(window.c.id)), *clauses).order_by(Clip.id).limit(limit)
+    ).all()
+    last = n < _EXPLORE_WINDOW
+    return _rows_to_candidates(session, rows, with_tags=False), (end_id if not last else _UUID_ZERO), last
 
 
 def _affinity_candidate_ids(session: Session, aff: cf.Affinity, rng) -> set[uuid.UUID]:
@@ -344,19 +352,19 @@ def clip_feed(
     taken = {c.id for c in picked_aff}
 
     # Explore: okno po PK; gdy w oknie brak widocznych, przesuwamy sie dalej (max 3 okna).
+    # Braki z afinicznosci (za malo kandydatow po filtrach) dopelnia explore.
     picked_explore: list[cf.Candidate] = []
     wrapped = False
     next_pk = pk
-    want = n_explore + max(0, n_aff - len(picked_aff))  # braki z afinicznosci dopelnia explore
+    want = n_explore + max(0, n_aff - len(picked_aff))
     for _ in range(3):
-        ids, w = _explore_window(session, next_pk)
-        wrapped = wrapped or w
-        if not ids:
-            break
-        next_pk = ids[-1] if not w else _UUID_ZERO
-        pool = [c for c in _candidates(session, ids, clauses) if c.id not in taken]
+        pool, next_pk, last = _explore_candidates(session, next_pk, clauses, limit=max(60, want * 15))
+        pool = [c for c in pool if c.id not in taken]
         picked_explore += cf.pick_explore(pool, want - len(picked_explore), rng)
-        if len(picked_explore) >= want or len(ids) < _EXPLORE_WINDOW:
+        taken.update(c.id for c in picked_explore)
+        if last:
+            wrapped = True
+        if len(picked_explore) >= want or last:
             break
 
     page_items = cf.compose_page(picked_aff, picked_explore, rng)