perf(feed): a shuffle walks a stored random key instead of sorting the whole scope

Every video carries a random number drawn once (videos.shuffle_key, migration 0118, indexed with
id). A seed picks a point on that ring: the feed's Shuffle sort pages along it and wraps, and the
Home surprise reads a 500-video window of it and shuffles that window by the seed, so no video is
favoured for sitting after a wide gap. The md5(id || seed) order read and sorted the whole feed
scope on every call: ~350 ms on prod per surprise, 381-767 ms per Shuffle page on localdev, now
14 ms and 4 ms.

backend/.basedpyright/baseline.json

@@ -185,118 +185,6 @@
                     "lineCount": 1
                 }
             },
-            {
-                "code": "reportArgumentType",
-                "range": {
-                    "startColumn": 50,
-                    "endColumn": 56,
-                    "lineCount": 1
-                }
-            },
-            {
-                "code": "reportArgumentType",
-                "range": {
-                    "startColumn": 50,
-                    "endColumn": 56,
-                    "lineCount": 1
-                }
-            },
-            {
-                "code": "reportArgumentType",
-                "range": {
-                    "startColumn": 50,
-                    "endColumn": 56,
-                    "lineCount": 1
-                }
-            },
-            {
-                "code": "reportArgumentType",
-                "range": {
-                    "startColumn": 50,
-                    "endColumn": 56,
-                    "lineCount": 1
-                }
-            },
-            {
-                "code": "reportArgumentType",
-                "range": {
-                    "startColumn": 50,
-                    "endColumn": 56,
-                    "lineCount": 1
-                }
-            },
-            {
-                "code": "reportArgumentType",
-                "range": {
-                    "startColumn": 50,
-                    "endColumn": 56,
-                    "lineCount": 1
-                }
-            },
-            {
-                "code": "reportArgumentType",
-                "range": {
-                    "startColumn": 50,
-                    "endColumn": 56,
-                    "lineCount": 1
-                }
-            },
-            {
-                "code": "reportArgumentType",
-                "range": {
-                    "startColumn": 50,
-                    "endColumn": 56,
-                    "lineCount": 1
-                }
-            },
-            {
-                "code": "reportArgumentType",
-                "range": {
-                    "startColumn": 50,
-                    "endColumn": 56,
-                    "lineCount": 1
-                }
-            },
-            {
-                "code": "reportArgumentType",
-                "range": {
-                    "startColumn": 50,
-                    "endColumn": 56,
-                    "lineCount": 1
-                }
-            },
-            {
-                "code": "reportArgumentType",
-                "range": {
-                    "startColumn": 50,
-                    "endColumn": 56,
-                    "lineCount": 1
-                }
-            },
-            {
-                "code": "reportArgumentType",
-                "range": {
-                    "startColumn": 50,
-                    "endColumn": 56,
-                    "lineCount": 1
-                }
-            },
-            {
-                "code": "reportArgumentType",
-                "range": {
-                    "startColumn": 50,
-                    "endColumn": 56,
-                    "lineCount": 1
-                }
-            },
-            {
-                "code": "reportArgumentType",
-                "range": {
-                    "startColumn": 50,
-                    "endColumn": 56,
-                    "lineCount": 1
-                }
-            },
             {
                 "code": "reportArgumentType",
                 "range": {

backend/alembic/versions/0118_videos_shuffle_key.py

@@ -0,0 +1,40 @@
+"""videos: a random key per video, so a seeded shuffle walks an index instead of sorting the scope
+
+Revision ID: 0118_videos_shuffle_key
+Revises: 0117_videos_length_sort_indexes
+Create Date: 2026-11-07
+
+The feed's Shuffle sort and the Home surprise ordered the whole feed scope by md5(id || seed) on every
+call: ~350 ms on prod for a scope of ~213k videos, and a Shuffle page deeper in the list cost more
+(767 ms on page 7 on localdev, against 4 ms for Newest). Each video now carries a random number drawn
+once (`Video.shuffle_key`); a seed picks a point on that ring and the draw reads the index from there
+(`feedquery.shuffle_walk`).
+
+One ADD COLUMN with a volatile default rewrites the table and fills every row with its own value. On
+localdev (278k videos) that took 28 s, against 84 s for an UPDATE of a nullable column, so the rewrite
+it is: the table is locked for that time, and a deploy has the app down while it migrates anyway.
+The index is (shuffle_key, id): the walk's keyset cursor compares the pair.
+"""
+
+from alembic import op
+
+revision: str = "0118_videos_shuffle_key"
+down_revision: str | None = "0117_videos_length_sort_indexes"
+branch_labels: str | None = None
+depends_on: str | None = None
+
+
+def upgrade() -> None:
+    op.execute(
+        "ALTER TABLE videos ADD COLUMN IF NOT EXISTS shuffle_key double precision "
+        "NOT NULL DEFAULT random()"
+    )
+    with op.get_context().autocommit_block():
+        op.execute("DROP INDEX CONCURRENTLY IF EXISTS ix_videos_shuffle_key")
+        op.execute("CREATE INDEX CONCURRENTLY ix_videos_shuffle_key ON videos (shuffle_key, id)")
+
+
+def downgrade() -> None:
+    with op.get_context().autocommit_block():
+        op.execute("DROP INDEX CONCURRENTLY IF EXISTS ix_videos_shuffle_key")
+    op.execute("ALTER TABLE videos DROP COLUMN IF EXISTS shuffle_key")

backend/app/catalog/models.py

@@ -10,6 +10,7 @@ from sqlalchemy import (
     Boolean,
     Computed,
     DateTime,
+    Float,
     ForeignKey,
     Integer,
     JSON,
@@ -217,6 +218,9 @@ class Video(Base, TimestampMixin):
     last_checked_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
     unavailable_since: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
     unavailable_reason: Mapped[str | None] = mapped_column(String(24))
+    # A random number in [0, 1), drawn once when the row is written: the ring a seeded shuffle walks
+    # (`feedquery.shuffle_walk`, index ix_videos_shuffle_key on (shuffle_key, id), migration 0118).
+    shuffle_key: Mapped[float] = mapped_column(Float, nullable=False, server_default=func.random())
 
     @hybrid_property
     def length_seconds(self) -> int | None:

backend/app/discover/queries.py

@@ -10,29 +10,22 @@ explore-visible and in-progress for free, and cannot drift from them. The Plex h
 narrows EVERY rail by the Plex sidebar's `LibraryFilters`, which is right for a surface that HAS a
 sidebar and wrong for a landing page that does not.
 
-**One seed drives both halves.** The feed orders a shuffle by `md5(id || seed)` so a keyset cursor
-stays valid across pages; the Plex surprise rail uses `ORDER BY random()`, which is fine for a single
-un-paged draw but cannot be reproduced. A cross-source reroll needs both halves to move together —
-otherwise "Mást kérek" re-draws the Plex picks while the YouTube ones stand still — so the Plex side
-here takes the feed's seeded-md5 idiom instead of `random()`.
+**One seed drives both halves.** The YouTube half starts on the ring of stored random keys at the
+seed's point and shuffles that window by `md5(id || seed)`; the Plex surprise rail uses
+`ORDER BY random()`, which is fine for a single un-paged draw but cannot be reproduced. A
+cross-source reroll needs both halves to move together — otherwise "Mást kérek" re-draws the Plex
+picks while the YouTube ones stand still — so the Plex side here orders by the same seeded md5
+instead of `random()`.
 """
 
-import json
 from datetime import datetime, timezone
 
 from sqlalchemy import (
-    Float,
-    Select,
     and_,
-    cast,
     func,
-    literal,
-    literal_column,
     or_,
     select,
-    union_all,
 )
-from sqlalchemy.dialects.postgresql import JSONB
 from sqlalchemy.orm import Session, aliased
 
 from app import draw
@@ -43,6 +36,7 @@ from app.feedquery import (
     not_blocked_cond,
     order_by,
     serialize,
+    shuffle_walk,
     sort_keys,
 )
 from app.models import User
@@ -128,66 +122,50 @@ def yt_surprise(db: Session, user: User, budget: str, seed: int, limit: int) ->
     Shorts are opened up only for the shortest band: "I have ten minutes" is the one answer a Short is
     actually a good response to, and the feed's own default keeps them out otherwise."""
     lo, hi = band(budget)
-    params = dict(_FEED_BASE, min_duration=lo, max_duration=hi, include_shorts=budget == "short")
-    query, rank_expr = filtered_query(db, user, **params)
-    keys = sort_keys("shuffle", seed, rank_expr)
-
-    # Over-fetch, then drop anything already started. The feed's `unwatched` scope keeps watched and
-    # hidden out but deliberately keeps in-progress rows in (they belong in the feed). Here they must
-    # go: the resume tier has already had its chance at them, and it may have passed precisely because
-    # the REMAINING time did not fit the budget — offering the same item back as a surprise, whose
-    # runtime does fit, would label a half-finished thing "Meglepetés" and contradict the question.
-    take = limit * 4
-
-    def unstarted(rows) -> list[dict]:
-        return [r for r in map(serialize, rows) if not r["position_seconds"]]
+    params: dict = {
+        **_FEED_BASE,
+        "min_duration": lo,
+        "max_duration": hi,
+        "include_shorts": budget == "short",
+    }
+    query, _rank = filtered_query(db, user, **params)
 
+    # A window of the ring, not the whole scope: the seed's SURPRISE_WINDOW videos read along the
+    # stored random key (`shuffle_walk`, a millisecond), then shuffled by the seed. Dealing straight
+    # from the ring would favour for good every video that sits after a wide gap of it (the first
+    # video past a random point is the one whose gap the point fell into); shuffling the window makes
+    # each video's chance the window's share, even to within a few percent. A scope no larger than
+    # the window is the whole scope, in exactly the md5 order the draw always had.
+    #
+    # Anything already started is dropped. The feed's `unwatched` scope keeps watched and hidden out
+    # but deliberately keeps in-progress rows in (they belong in the feed). Here they must go: the
+    # resume tier has already had its chance at them, and it may have passed precisely because the
+    # REMAINING time did not fit the budget — offering the same item back as a surprise, whose runtime
+    # does fit, would label a half-finished thing "Meglepetés" and contradict the question.
+    window = [
+        r
+        for r in map(serialize, shuffle_walk(db, query, seed, SURPRISE_WINDOW))
+        if not r["position_seconds"]
+    ]
+    blind = draw.shuffled(window, seed, key=_video_id)
     channels = _channel_affinity(db, user)
     if not channels:
-        return unstarted(db.execute(query.order_by(*order_by(keys)).limit(take)).all())[:limit]
+        return blind[:limit]
 
-    # Both orders from ONE pass over the scope. The pool is the whole feed scope (~200k videos on a
-    # prod-sized catalog, ~390 ms a pass), so a second ordered query for the weighted draw doubled
-    # the cost; a materialized pool of (id, shuffle key, weighted key) is read twice instead.
-    weight = literal_column("1.0") + CHANNEL_WEIGHT * func.coalesce(
-        cast(cast(literal(json.dumps(channels)), JSONB).op("->>")(Video.channel_id), Float), 0.0
-    )
-    shuffle = func.md5(func.concat(Video.id, str(seed)))
-    pool = (
-        query.with_only_columns(
-            Video.id.label("vid"),
-            shuffle.label("plain"),
-            draw.weighted_key(draw.seeded_u(Video.id, seed), weight).label("leaning"),
-        )
-        .cte("surprise_pool")
-        .prefix_with("MATERIALIZED")
-    )
+    def weight(row: dict) -> float:
+        return 1.0 + CHANNEL_WEIGHT * channels.get(row["channel_id"], 0.0)
 
-    def top(side: int, *order) -> Select:
-        part = (
-            select(pool.c.vid, pool.c.plain, pool.c.leaning, literal(side).label("side"))
-            .order_by(*order, pool.c.vid)
-            .limit(take)
-            .subquery()
-        )
-        return select(part)
+    leaning = draw.leaning(window, seed, key=_video_id, weight=weight)
+    return draw.blend(leaning, blind, seed, limit, key=_video_id)
 
-    # One statement, so the pool is built once; each side's order is restored from its keys.
-    got = db.execute(union_all(top(0, pool.c.plain), top(1, pool.c.leaning, pool.c.plain))).all()
-    plain_ids = [
-        r.vid for r in sorted((r for r in got if r.side == 0), key=lambda r: (r.plain, r.vid))
-    ]
-    leaning_ids = [
-        r.vid
-        for r in sorted((r for r in got if r.side == 1), key=lambda r: (r.leaning, r.plain, r.vid))
-    ]
-    rows = {
-        r["id"]: r
-        for r in unstarted(db.execute(query.where(Video.id.in_({*plain_ids, *leaning_ids}))).all())
-    }
-    blind = [rows[i] for i in plain_ids if i in rows]
-    leaning = [rows[i] for i in leaning_ids if i in rows]
-    return draw.blend(leaning, blind, seed, limit, key=lambda r: r["id"])
+
+# How many videos of the ring a surprise draws from: wide enough that a channel the user watches has
+# a few in it, small enough to read in a millisecond or two.
+SURPRISE_WINDOW = 500
+
+
+def _video_id(row: dict) -> str:
+    return row["id"]
 
 
 # How much a channel the user watches most outweighs one they never watch (1 + this), and how far back

backend/app/draw.py

@@ -11,8 +11,13 @@ often. Two things keep it a surprise rather than a ranking:
   seeded (the Plex Discover rail) takes u from random().
 - **One slot of three is blind** (`blend`): dealt from the unweighted order, so the same few
   favourites cannot fill the rail.
+
+The YouTube surprise draws over a window of the feed scope rather than all of it (`yt_surprise`):
+the window is read in SQL, and `shuffled` / `leaning` order it here with the same u.
 """
 
+import hashlib
+import math
 from collections.abc import Callable, Hashable
 from typing import Any
 
@@ -44,6 +49,27 @@ def random_u() -> ColumnElement[float]:
     return func.random()
 
 
+def u_of(key: str, seed: int) -> float:
+    """`seeded_u` in Python: the same u for the same key and seed."""
+    return int(hashlib.md5(f"{key}{seed}".encode()).hexdigest()[:8], 16) / _SPAN
+
+
+def shuffled(items: list, seed: int, key: Callable[[Any], str]) -> list:
+    """`items` in the seed's md5 order — the order `ORDER BY md5(key || seed)` gives."""
+    return sorted(items, key=lambda it: hashlib.md5(f"{key(it)}{seed}".encode()).hexdigest())
+
+
+def leaning(
+    items: list, seed: int, key: Callable[[Any], str], weight: Callable[[Any], float]
+) -> list:
+    """`items` in the weighted draw's order (`weighted_key` in Python), ties in the md5 order."""
+    order = {key(it): i for i, it in enumerate(shuffled(items, seed, key))}
+    return sorted(
+        items,
+        key=lambda it: (-math.log(1.0 - u_of(key(it), seed)) / weight(it), order[key(it)]),
+    )
+
+
 def blend(
     weighted: list, blind: list, offset: int, n: int, key: Callable[[Any], Hashable] = lambda x: x
 ) -> list:

backend/app/feedquery.py

@@ -13,7 +13,19 @@ import binascii
 import json
 from datetime import date, datetime, timedelta, timezone
 
-from sqlalchemy import BigInteger, Select, and_, case, cast, false, func, or_, select
+from sqlalchemy import (
+    BigInteger,
+    Select,
+    and_,
+    case,
+    cast,
+    false,
+    func,
+    literal,
+    or_,
+    select,
+    tuple_,
+)
 from sqlalchemy.orm import Session, aliased
 
 from app.collation import collated
@@ -536,15 +548,17 @@ def sort_keys(sort: str, seed: int, rank_expr=None, lang: str | None = None) ->
             },
             pub,
         ],
-        # Deterministic per-seed shuffle: the same seed yields the same total order, so
-        # the cursor stays valid across pages of one shuffled run.
+        # The ring of stored random keys from the seed's point (`shuffle_walk`): the keys at or after
+        # it, then the ones before. The same seed yields the same total order, so the cursor stays
+        # valid across pages; `feed_page` reads it as two index ranges, never as this sort.
         "shuffle": [
             {
-                "expr": func.md5(func.concat(Video.id, str(seed))),
+                "expr": case((Video.shuffle_key < shuffle_start(seed), 1), else_=0),
                 "asc": True,
                 "nullable": False,
-                "kind": "str",
-            }
+                "kind": "num",
+            },
+            {"expr": Video.shuffle_key, "asc": True, "nullable": False, "kind": "num"},
         ],
     }
     # Relevance (search-result ranking) only exists when there's a query to rank against;
@@ -700,6 +714,40 @@ class BadCursor(ValueError):
     """A `cursor` this module did not write (or one from another sort): the caller's bad input."""
 
 
+# The golden ratio's fraction: consecutive seeds land far apart on the ring, never next to each other.
+_GOLDEN = 0.6180339887498949
+
+
+def shuffle_start(seed: int) -> float:
+    """The point in [0, 1) where a seed's shuffle starts on the ring of `Video.shuffle_key`."""
+    return (seed * _GOLDEN) % 1.0
+
+
+def shuffle_walk(
+    db: Session, query: Select, seed: int, n: int, after: tuple[int, float, str] | None = None
+) -> list:
+    """Up to `n` rows of `query` in the seed's ring order: shuffle_key from the seed's point upwards,
+    then from 0 up to that point. `after` is a cursor's (phase, key, id): the walk resumes past it.
+
+    Each arc is an ascending range read along ix_videos_shuffle_key with the scope's filters checked
+    row by row, so a draw costs the rows it reads, not a sort of the scope — the md5(id || seed) order
+    it replaced read and sorted every row the filters let through (~350 ms on prod for ~213k)."""
+    start = shuffle_start(seed)
+    phase_from, key, last_id = after if after is not None else (0, None, None)
+    rows: list = []
+    for phase in (0, 1):
+        if phase < phase_from or len(rows) >= n:
+            continue
+        arc = Video.shuffle_key >= start if phase == 0 else Video.shuffle_key < start
+        q = query.where(arc)
+        if key is not None and phase == phase_from:
+            q = q.where(
+                tuple_(Video.shuffle_key, Video.id) > tuple_(literal(key), literal(last_id))
+            )
+        rows += db.execute(q.order_by(Video.shuffle_key, Video.id).limit(n - len(rows))).all()
+    return rows
+
+
 def feed_page(
     db: Session,
     user: User,
@@ -734,6 +782,14 @@ def feed_page(
             k["value"] = v
     if ranked and window is not None:
         rows = windowed_relevance_rows(db, query, keys, window, cursor_id, limit + 1)
+    elif sort == "shuffle":
+        after = None
+        if vals is not None and cursor_id is not None:
+            try:
+                after = (int(vals[0]), float(vals[1]), str(cursor_id))
+            except (TypeError, ValueError) as e:
+                raise BadCursor("Invalid feed cursor") from e
+        rows = shuffle_walk(db, query, seed, limit + 1, after)
     else:
         query = query.order_by(*order_by(keys))
         if cursor_id is not None:

backend/tests/test_shuffle_walk.py

@@ -0,0 +1,151 @@
+"""A seeded shuffle walks a stored random key instead of sorting the whole scope. DB lane.
+
+Every video carries a random number drawn once (`Video.shuffle_key`). A seed picks a point on that
+ring; the feed's Shuffle sort pages along the ring from there and wraps, and the Home surprise reads
+a window of it and shuffles that window by the seed (`discover.queries.yt_surprise`). Ordering the
+scope by md5(id || seed) cost ~350 ms per call on prod and grew with every Shuffle page.
+"""
+
+from contextlib import contextmanager
+from datetime import datetime, timedelta, timezone
+
+from sqlalchemy import event
+
+from app import feedquery
+from app.catalog.models import Channel, Subscription, Video
+from app.discover import queries
+from app.discover.queries import _FEED_BASE
+
+NOW = datetime.now(timezone.utc)
+KEYS = {"k05": 0.05, "k20": 0.20, "k40": 0.40, "k60": 0.60, "k80": 0.80, "k95": 0.95}
+
+
+def _seed(db, make_user, keys=KEYS):
+    user = make_user()
+    db.add(Channel(id="UCring", title="Ring"))
+    db.flush()
+    db.add(Subscription(user_id=user.id, channel_id="UCring"))
+    db.add_all(
+        [
+            Video(
+                id=vid,
+                channel_id="UCring",
+                title=vid,
+                duration_seconds=20 * 60,
+                published_at=NOW - timedelta(days=i + 1),
+                shuffle_key=key,
+            )
+            for i, (vid, key) in enumerate(keys.items())
+        ]
+    )
+    db.commit()
+    return user
+
+
+def _ring(seed, keys=KEYS):
+    """The ring order from the seed's point: the keys at or after it ascending, then the rest."""
+    r = feedquery.shuffle_start(seed)
+    return sorted(keys, key=lambda v: (keys[v] < r, keys[v]))
+
+
+def _seed_landing_between(lo, hi):
+    return next(s for s in range(1000) if lo < feedquery.shuffle_start(s) < hi)
+
+
+def test_the_walk_starts_at_the_seeds_point_and_wraps(db, make_user):
+    user = _seed(db, make_user)
+    query, _ = feedquery.filtered_query(db, user, **_FEED_BASE)
+    seed = _seed_landing_between(0.45, 0.55)  # starts at k60, wraps to k05 after k95
+    got = [r.id for r in feedquery.shuffle_walk(db, query, seed, 6)]
+    assert got == ["k60", "k80", "k95", "k05", "k20", "k40"]
+    assert [r.id for r in feedquery.shuffle_walk(db, query, seed, 4)] == got[:4]
+
+
+def test_the_feed_shuffle_pages_along_the_ring(db, make_user, as_user):
+    user = _seed(db, make_user)
+    client = as_user(user)
+    for seed in (_seed_landing_between(0.1, 0.15), _seed_landing_between(0.85, 0.9)):
+        walked, cursor = [], None
+        for _ in range(5):
+            url = f"/api/feed?sort=shuffle&seed={seed}&limit=2" + (
+                f"&cursor={cursor}" if cursor else ""
+            )
+            body = client.get(url).json()
+            walked += [it["id"] for it in body["items"]]
+            cursor = body["next_cursor"]
+            if not body["has_more"]:
+                break
+        assert walked == _ring(seed), seed
+
+
+def test_a_new_seed_opens_the_shuffle_elsewhere(db, make_user, as_user):
+    user = _seed(db, make_user)
+    client = as_user(user)
+    first = {
+        seed: client.get(f"/api/feed?sort=shuffle&seed={seed}&limit=2").json()["items"][0]["id"]
+        for seed in (_seed_landing_between(0.1, 0.15), _seed_landing_between(0.65, 0.7))
+    }
+    assert sorted(first.values()) == ["k20", "k80"]
+
+
+@contextmanager
+def _statements(db):
+    seen: list[str] = []
+
+    def grab(conn, cursor, statement, *args):
+        seen.append(statement)
+
+    engine = db.get_bind()
+    event.listen(engine, "before_cursor_execute", grab)
+    try:
+        yield seen
+    finally:
+        event.remove(engine, "before_cursor_execute", grab)
+
+
+def _draws_of_videos(seen):
+    return [s for s in seen if "FROM videos" in s and "LIMIT" in s]
+
+
+def test_neither_draw_orders_the_scope_by_a_hash(db, make_user):
+    # Each draw reads the videos as an ascending range of the key (what ix_videos_shuffle_key serves),
+    # never ordered by a hash or by the ring's CASE (which no index can serve).
+    user = _seed(db, make_user)
+    with _statements(db) as surprise:
+        queries.yt_surprise(db, user, "any", 7, 3)
+    with _statements(db) as page:
+        feedquery.feed_page(
+            db, user, dict(_FEED_BASE), sort="shuffle", seed=7, limit=2, cursor=None, lang="en"
+        )
+    for seen in (surprise, page):
+        draws = _draws_of_videos(seen)
+        assert draws
+        for s in draws:
+            assert "md5(" not in s.lower()
+            assert "ORDER BY videos.shuffle_key, videos.id" in s.replace("\n", " "), s
+
+
+def test_the_surprise_reads_a_window_of_the_ring(db, make_user, monkeypatch):
+    user = _seed(db, make_user)
+    monkeypatch.setattr(queries, "SURPRISE_WINDOW", 3)
+    for seed in range(20):
+        drawn = {r["id"] for r in queries.yt_surprise(db, user, "any", seed, 3)}
+        assert drawn == set(_ring(seed)[:3]), seed
+
+
+def test_a_wide_gap_on_the_ring_does_not_make_a_favourite(db, make_user):
+    # One video after an empty half of the ring, twenty packed into the other half. Taking the first
+    # video after the seed's point would deal the lonely one about half the time; the window is
+    # shuffled by the seed, so it leads about as often as any other.
+    keys = {"lonely": 0.5} | {f"p{i:02}": 0.51 + i * 0.02 for i in range(20)}
+    user = _seed(db, make_user, keys)
+    leads = [queries.yt_surprise(db, user, "any", s, 1)[0]["id"] for s in range(90)]
+    assert leads.count("lonely") < 15, leads.count("lonely")
+
+
+def test_a_new_video_gets_its_own_key(db, make_user):
+    _seed(db, make_user, {})
+    db.add_all([Video(id=f"n{i}", channel_id="UCring", title="n") for i in range(3)])
+    db.commit()
+    keys = [db.get(Video, f"n{i}").shuffle_key for i in range(3)]
+    assert len(set(keys)) == 3 and all(0 <= k < 1 for k in keys)