Merge pull request #387 from d3vyce/386-pagination-over-a-to-many-join-duplicates-drops-and-truncates-results

fix: pagination over a to-many join duplicates, drops and truncates results

src/fastapi_toolsets/crud/factory.py

@@ -14,12 +14,25 @@ from typing import Any, ClassVar, Generic, Literal, Self, TypeAlias, cast, overl
 
 from fastapi import Query
 from pydantic import BaseModel
-from sqlalchemy import Date, DateTime, Float, Integer, Numeric, Uuid, and_, func, select
+from sqlalchemy import (
+    Date,
+    DateTime,
+    Float,
+    Integer,
+    Numeric,
+    Uuid,
+    and_,
+    func,
+    select,
+    tuple_,
+)
 from sqlalchemy.dialects.postgresql import insert
 from sqlalchemy.exc import NoResultFound
 from sqlalchemy.ext.asyncio import AsyncSession
 from sqlalchemy.orm import DeclarativeBase, QueryableAttribute, selectinload
+from sqlalchemy.sql import operators
 from sqlalchemy.sql.base import ExecutableOption
+from sqlalchemy.sql.elements import UnaryExpression
 from sqlalchemy.sql.roles import WhereHavingRole
 
 from ..db import transaction
@@ -128,6 +141,32 @@ def _apply_joins(q: Any, joins: JoinType | None, outer_join: bool) -> Any:
     return q
 
 
+def _fans_out(
+    search_joins: Sequence[Any] | None, order_joins: Sequence[Any] | None
+) -> bool:
+    """True if any relationship join yields a collection."""
+    return any(
+        rel.property.uselist for rel in (*(search_joins or ()), *(order_joins or ()))
+    )
+
+
+def _grouped_order(clause: Any, table: Any) -> Any:
+    """Recast an order clause for a query grouped by the entity's key."""
+    inner = clause.element if isinstance(clause, UnaryExpression) else clause
+    expr = inner.__clause_element__() if hasattr(inner, "__clause_element__") else inner
+    tables = {
+        t
+        for c in getattr(expr, "base_columns", ())
+        if (t := getattr(c, "table", None)) is not None
+    }
+    if tables and tables <= {table}:
+        return clause
+    agg = func.min(inner)
+    if isinstance(clause, UnaryExpression) and clause.modifier is operators.desc_op:
+        return agg.desc()
+    return agg.asc()
+
+
 class AsyncCrud(Generic[ModelType]):
     """Generic async CRUD operations for SQLAlchemy models.
 
@@ -163,6 +202,64 @@ class AsyncCrud(Generic[ModelType]):
             ):
                 cls.searchable_fields = [pk_col, *raw_fields]
 
+    @classmethod
+    def _pk_attrs(cls: type[Self]) -> list[QueryableAttribute[Any]]:
+        """The model's primary key columns as instrumented attributes."""
+        return [
+            getattr(cls.model, cast(str, col.key))
+            for col in cls.model.__mapper__.primary_key
+        ]
+
+    @classmethod
+    async def _page_entities(
+        cls: type[Self],
+        session: AsyncSession,
+        q: Any,
+        *,
+        order_clauses: Sequence[Any],
+        limit: int,
+        offset: int | None = None,
+        load_options: Sequence[ExecutableOption] | None = None,
+        with_for_update: _ForUpdateMode = False,
+    ) -> list[ModelType]:
+        """Return up to *limit* entities, paging over distinct primary keys."""
+        pk_attrs = cls._pk_attrs()
+        table = cls.model.__table__
+        # Fall back to the primary key so the page boundary is deterministic.
+        grouped = [_grouped_order(c, table) for c in order_clauses] or [pk_attrs[0]]
+        id_q = (
+            q.order_by(None)
+            .with_only_columns(*pk_attrs)
+            .group_by(*pk_attrs)
+            .order_by(*grouped)
+            .limit(limit)
+        )
+        if offset:
+            id_q = id_q.offset(offset)
+        rows = (await session.execute(id_q)).all()
+        ids = [row[0] if len(pk_attrs) == 1 else tuple(row) for row in rows]
+        if not ids:
+            return []
+
+        where = (
+            pk_attrs[0].in_(ids) if len(pk_attrs) == 1 else tuple_(*pk_attrs).in_(ids)
+        )
+        item_q = select(cls.model).where(where)
+        if resolved := cls._resolve_load_options(load_options):
+            item_q = item_q.options(*resolved)
+        item_q = _apply_for_update(item_q, with_for_update)
+        found = (await session.execute(item_q)).unique().scalars().all()
+
+        rank = {pk: n for n, pk in enumerate(ids)}
+
+        def _key(obj: Any) -> Any:
+            values = tuple(getattr(obj, a.key) for a in pk_attrs)
+            return values[0] if len(pk_attrs) == 1 else values
+
+        return cast(
+            list[ModelType], sorted(found, key=lambda o: rank.get(_key(o), len(ids)))
+        )
+
     @classmethod
     def _resolve_load_options(
         cls, load_options: Sequence[ExecutableOption] | None
@@ -1023,9 +1120,21 @@ class AsyncCrud(Generic[ModelType]):
             q = q.where(and_(*filters))
         if resolved := cls._resolve_load_options(load_options):
             q = q.options(*resolved)
-        q = _apply_for_update(q, with_for_update)
         if order_by is not None:
             q = q.order_by(order_by)
+
+        if limit is not None and joins:
+            return await cls._page_entities(
+                session,
+                q,
+                order_clauses=[] if order_by is None else [order_by],
+                limit=limit,
+                offset=offset,
+                load_options=load_options,
+                with_for_update=with_for_update,
+            )
+
+        q = _apply_for_update(q, with_for_update)
         if offset is not None:
             q = q.offset(offset)
         if limit is not None:
@@ -1374,17 +1483,31 @@ class AsyncCrud(Generic[ModelType]):
             q = q.where(and_(*filters))
         if resolved := cls._resolve_load_options(load_options):
             q = q.options(*resolved)
-        if order_by is not None:
-            q = q.order_by(order_by)
+        order_clauses: list[Any] = [] if order_by is None else [order_by]
+        q = q.order_by(*order_clauses)
+
+        fetch_limit = items_per_page if include_total else items_per_page + 1
+        total_count: int | None = None
+        # A to-many join repeats each entity, so LIMIT would slice joined rows
+        # and `.unique()` would shrink the page after the fact.
+        if _fans_out(search_joins, order_joins):
+            raw_items = await cls._page_entities(
+                session,
+                q,
+                order_clauses=order_clauses,
+                limit=fetch_limit,
+                offset=offset,
+                load_options=load_options,
+            )
+        else:
+            result = await session.execute(q.offset(offset).limit(fetch_limit))
+            raw_items = cast(list[ModelType], result.unique().scalars().all())
+        fetched = len(raw_items)
+        raw_items = raw_items[:items_per_page]
 
         if include_total:
-            q = q.offset(offset).limit(items_per_page)
-            result = await session.execute(q)
-            raw_items = cast(list[ModelType], result.unique().scalars().all())
-
             # Count query (with same joins and filters)
-            pk_col = cls.model.__mapper__.primary_key[0]
-            count_q = select(func.count(func.distinct(getattr(cls.model, pk_col.name))))
+            count_q = select(func.count(func.distinct(cls._pk_attrs()[0])))
             count_q = count_q.select_from(cls.model)
 
             # Apply explicit joins to count query
@@ -1397,16 +1520,11 @@ class AsyncCrud(Generic[ModelType]):
                 count_q = count_q.where(and_(*filters))
 
             count_result = await session.execute(count_q)
-            total_count: int = count_result.scalar_one()
+            total_count = count_result.scalar_one()
             has_more = page * items_per_page < total_count
         else:
-            # Fetch one extra row to detect if a next page exists without COUNT
-            q = q.offset(offset).limit(items_per_page + 1)
-            result = await session.execute(q)
-            raw_items = cast(list[ModelType], result.unique().scalars().all())
-            has_more = len(raw_items) > items_per_page
-            raw_items = raw_items[:items_per_page]
-            total_count = None
+            # One extra row was fetched to detect a next page without COUNT
+            has_more = fetched > items_per_page
 
         items: list[Any] = [schema.model_validate(item) for item in raw_items]
 
@@ -1543,18 +1661,30 @@ class AsyncCrud(Generic[ModelType]):
             q = q.options(*resolved)
 
         # Cursor column is always the primary sort; reverse direction for prev traversal
-        if direction is _CursorDirection.PREV:
-            q = q.order_by(cursor_column.desc())
-        else:
-            q = q.order_by(cursor_column)
+        cursor_clause = (
+            cursor_column.desc()
+            if direction is _CursorDirection.PREV
+            else cursor_column
+        )
+        order_clauses: list[Any] = [cursor_clause]
         if order_by is not None:
-            q = q.order_by(order_by)
-
-        # Fetch one extra to detect whether another page exists in this direction
-        q = q.limit(items_per_page + 1)
-        result = await session.execute(q)
-        raw_items = cast(list[ModelType], result.unique().scalars().all())
+            order_clauses.append(order_by)
+        q = q.order_by(*order_clauses)
 
+        # One extra row detects whether another page exists in this direction.
+        # Under a to-many join that extra row may be a duplicate of one already
+        # on the page, which reads as "no next page" and ends traversal early.
+        if _fans_out(search_joins, order_joins):
+            raw_items = await cls._page_entities(
+                session,
+                q,
+                order_clauses=order_clauses,
+                limit=items_per_page + 1,
+                load_options=load_options,
+            )
+        else:
+            result = await session.execute(q.limit(items_per_page + 1))
+            raw_items = cast(list[ModelType], result.unique().scalars().all())
         has_more = len(raw_items) > items_per_page
         items_page = raw_items[:items_per_page]
 

tests/test_crud_search.py

@@ -5,6 +5,7 @@ import uuid
 
 import pytest
 from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.orm import selectinload
 from sqlalchemy.sql.elements import ColumnElement, UnaryExpression
 
 from fastapi_toolsets.crud import (
@@ -16,7 +17,7 @@ from fastapi_toolsets.crud import (
     get_searchable_fields,
 )
 from fastapi_toolsets.exceptions import InvalidOrderFieldError
-from fastapi_toolsets.schemas import OffsetPagination, PaginationType
+from fastapi_toolsets.schemas import OffsetPagination, PaginationType, PydanticBase
 
 from .conftest import (
     Article,
@@ -29,15 +30,19 @@ from .conftest import (
     OrderCrud,
     OrderRead,
     OrderStatus,
+    Post,
+    PostCrud,
     Role,
     RoleCreate,
     RoleCrud,
     RoleCursorCrud,
     RoleRead,
+    Tag,
     User,
     UserCreate,
     UserCrud,
     UserRead,
+    post_tags,
 )
 
 
@@ -397,6 +402,225 @@ class TestBuildSearchFilters:
         assert "CAST" in str(filters[0])
 
 
+class _PostTitle(PydanticBase):
+    """Minimal read schema for the to-many pagination tests."""
+
+    id: uuid.UUID
+    title: str
+
+
+class _TagName(PydanticBase):
+    name: str
+
+
+class _PostWithTags(PydanticBase):
+    """Serialising `tags` fails unless the relation was eager-loaded."""
+
+    title: str
+    tags: list[_TagName]
+
+
+PostTagSearchCrud = CrudFactory(
+    Post,
+    searchable_fields=[Post.title, (Post.tags, Tag.name)],
+    cursor_column=Post.id,
+)
+
+_POST_COUNT = 10
+
+
+async def _seed_posts_with_tags(session) -> None:
+    """10 posts, each with 3 tags whose names all match the search term."""
+    author = await UserCrud.create(
+        session, UserCreate(username="fanout", email="fanout@test.com")
+    )
+    for i in range(_POST_COUNT):
+        tags = [Tag(name=f"shared-{i}-{j}") for j in range(3)]
+        session.add_all(tags)
+        session.add(Post(title=f"post{i:02d}", author_id=author.id, tags=tags))
+    await session.flush()
+
+
+class TestPaginateToManyJoin:
+    """Searching a to-many relationship must not truncate or duplicate pages."""
+
+    @pytest.mark.anyio
+    async def test_offset_pages_are_full_and_complete(self, db_session: AsyncSession):
+        """Every page is full and every post is returned exactly once."""
+        await _seed_posts_with_tags(db_session)
+
+        seen: list[str] = []
+        for page in (1, 2):
+            result = await PostTagSearchCrud.offset_paginate(
+                db_session,
+                page=page,
+                items_per_page=5,
+                search="shared",
+                schema=_PostTitle,
+            )
+            assert result.pagination.total_count == _POST_COUNT
+            assert len(result.data) == 5, f"page {page} came back short"
+            seen += [p.title for p in result.data]
+
+        assert len(set(seen)) == _POST_COUNT, "posts duplicated or missing"
+
+    @pytest.mark.anyio
+    async def test_offset_without_total_reports_has_more(
+        self, db_session: AsyncSession
+    ):
+        """``has_more`` counts entities, not joined rows."""
+        await _seed_posts_with_tags(db_session)
+
+        first = await PostTagSearchCrud.offset_paginate(
+            db_session,
+            page=1,
+            items_per_page=5,
+            search="shared",
+            include_total=False,
+            schema=_PostTitle,
+        )
+        assert len(first.data) == 5
+        assert first.pagination.has_more is True
+
+        last = await PostTagSearchCrud.offset_paginate(
+            db_session,
+            page=2,
+            items_per_page=5,
+            search="shared",
+            include_total=False,
+            schema=_PostTitle,
+        )
+        assert len(last.data) == 5
+        assert last.pagination.has_more is False
+
+    @pytest.mark.anyio
+    async def test_cursor_traverses_every_row(self, db_session: AsyncSession):
+        """Cursor traversal must not stop early."""
+        await _seed_posts_with_tags(db_session)
+
+        seen: list[str] = []
+        cursor: str | None = None
+        for _ in range(_POST_COUNT):
+            result = await PostTagSearchCrud.cursor_paginate(
+                db_session,
+                cursor=cursor,
+                items_per_page=5,
+                search="shared",
+                schema=_PostTitle,
+            )
+            seen += [p.title for p in result.data]
+            cursor = result.pagination.next_cursor
+            if cursor is None:
+                break
+
+        assert len(seen) == _POST_COUNT, "traversal stopped early or repeated rows"
+        assert len(set(seen)) == _POST_COUNT
+
+    @pytest.mark.anyio
+    async def test_orders_by_a_to_many_column(self, db_session: AsyncSession):
+        """Ordering by a related column has to be collapsed to an aggregate."""
+        await _seed_posts_with_tags(db_session)
+
+        result = await PostTagSearchCrud.offset_paginate(
+            db_session,
+            page=1,
+            items_per_page=5,
+            search="shared",
+            order_by=Tag.name.asc(),
+            order_joins=[Post.tags],
+            schema=_PostTitle,
+        )
+
+        assert len(result.data) == 5
+        assert result.pagination.total_count == _POST_COUNT
+
+    @pytest.mark.anyio
+    async def test_orders_by_a_bare_to_many_column(self, db_session: AsyncSession):
+        """An order clause with no explicit direction still needs aggregating."""
+        await _seed_posts_with_tags(db_session)
+
+        result = await PostTagSearchCrud.offset_paginate(
+            db_session,
+            page=1,
+            items_per_page=5,
+            search="shared",
+            order_by=Tag.name,
+            order_joins=[Post.tags],
+            schema=_PostTitle,
+        )
+
+        assert len(result.data) == 5
+
+    @pytest.mark.anyio
+    async def test_page_past_the_end_is_empty(self, db_session: AsyncSession):
+        """No keys on the page means no second query and an empty result."""
+        await _seed_posts_with_tags(db_session)
+
+        result = await PostTagSearchCrud.offset_paginate(
+            db_session, page=99, items_per_page=5, search="shared", schema=_PostTitle
+        )
+
+        assert result.data == []
+        assert result.pagination.total_count == _POST_COUNT
+
+    @pytest.mark.anyio
+    async def test_load_options_apply_on_the_fan_out_path(
+        self, db_session: AsyncSession
+    ):
+        """The entity query still honours loader options."""
+        await _seed_posts_with_tags(db_session)
+
+        result = await PostTagSearchCrud.offset_paginate(
+            db_session,
+            page=1,
+            items_per_page=5,
+            search="shared",
+            load_options=[selectinload(Post.tags)],
+            schema=_PostWithTags,
+        )
+
+        assert len(result.data) == 5
+        assert all(len(p.tags) == 3 for p in result.data)
+
+    @pytest.mark.anyio
+    async def test_get_multi_is_not_truncated_by_a_to_many_join(
+        self, db_session: AsyncSession
+    ):
+        """`get_multi` takes a raw join, so cardinality cannot be inspected."""
+        await _seed_posts_with_tags(db_session)
+
+        rows = await PostCrud.get_multi(
+            db_session,
+            joins=[(post_tags, post_tags.c.post_id == Post.id)],
+            outer_join=True,
+            limit=5,
+        )
+
+        assert len(rows) == 5
+        assert len({r.id for r in rows}) == 5
+
+    def test_grouped_order_only_aggregates_foreign_columns(self):
+        """A base-table column is left alone; anything else collapses to min()."""
+        from fastapi_toolsets.crud.factory import _grouped_order
+
+        table = Post.__table__
+
+        assert "min" not in str(_grouped_order(Post.title.desc(), table)).lower()
+
+        asc_on_join = str(_grouped_order(Tag.name.asc(), table))
+        desc_on_join = str(_grouped_order(Tag.name.desc(), table))
+        assert "min" in asc_on_join.lower() and asc_on_join.endswith("ASC")
+        assert "min" in desc_on_join.lower() and desc_on_join.endswith("DESC")
+
+    def test_to_one_join_does_not_take_the_fan_out_path(self):
+        """A to-one join keeps the single-query path."""
+        from fastapi_toolsets.crud.factory import _fans_out
+
+        assert _fans_out([User.role], None) is False
+        assert _fans_out([Post.tags], None) is True
+        assert _fans_out(None, [Post.tags]) is True
+
+
 class TestSearchEnumColumn:
     """Searching an enum column must reach the database, not just build SQL."""