Sqlalchemy

是什么

SQLAlchemy 是 Python 生态里最成熟的 ORM 与 SQL 工具包，既能写对象映射也能写原生 SQL。 用它的效果是：业务代码读起来像在操作对象，运行起来却能拿到接近手写 SQL 的可控性。

怎么用

1. 先用声明式（Declarative）方式建模实体，让表结构与领域模型在同一个文件里成对出现。
2. 通过 Session 管理事务边界，让一次业务操作的所有写入要么全部成功要么全部回滚。
3. 在复杂查询场景下用 Core 层写显式 SQL 表达式，让性能与可读性都不妥协。
4. 用 Alembic 管理迁移脚本，让 schema 演进与代码版本一一对应。
5. 通过 eager loading（饥饿加载）解决 N+1 问题，让接口响应时间稳定可预期。

架构图

flowchart LR
  领域模型 --> 声明映射
  声明映射 --> 会话事务
  会话事务 --> SQL生成
  SQL生成 --> 数据库
  数据库 --> 对象返回

SQLAlchemy 2.0 — Async Patterns

Modern SQLAlchemy 2.0 with full async support (asyncpg) and Alembic migrations.

When to Activate

• Defining ORM models with Mapped / mapped_column
• Writing async queries (select, join, filter, order_by)
• Managing async sessions (AsyncSession, async_sessionmaker)
• Handling relationships and loading strategies (selectin, joined, lazy)
• Running database transactions or bulk operations
• Writing or debugging Alembic migrations
• Converting between ORM models and domain entities

Model Definition (SQLAlchemy 2.0 style)

from datetime import datetime
from uuid import UUID, uuid4
from sqlalchemy import String, ForeignKey, Text, TIMESTAMP, func
from sqlalchemy.orm import DeclarativeBase, Mapped, mapped_column, relationship
from sqlalchemy.dialects.postgresql import UUID as PGUUID, JSONB


class Base(DeclarativeBase):
    pass


class UserORM(Base):
    __tablename__ = "users"

    id: Mapped[UUID] = mapped_column(PGUUID(as_uuid=True), primary_key=True, default=uuid4)
    email: Mapped[str] = mapped_column(String(255), unique=True, nullable=False, index=True)
    name: Mapped[str] = mapped_column(String(255), nullable=False)
    role: Mapped[str] = mapped_column(String(50), nullable=False, default="user")
    metadata_: Mapped[dict] = mapped_column("metadata", JSONB, nullable=False, default=dict)
    created_at: Mapped[datetime] = mapped_column(
        TIMESTAMP(timezone=True), server_default=func.now(), nullable=False
    )
    updated_at: Mapped[datetime] = mapped_column(
        TIMESTAMP(timezone=True), server_default=func.now(),
    )

    # Relationship — loads orders when accessed
    orders: Mapped[list["OrderORM"]] = relationship("OrderORM", back_populates="user")


class OrderORM(Base):
    __tablename__ = "orders"

    id: Mapped[UUID] = mapped_column(PGUUID(as_uuid=True), primary_key=True, default=uuid4)
    user_id: Mapped[UUID] = mapped_column(
        PGUUID(as_uuid=True), ForeignKey("users.id", nullable=False, index=True
    )
    status: Mapped[str] = mapped_column(String(50), nullable=False, default="pending")
    total: Mapped[float] = mapped_column(nullable=False)
    created_at: Mapped[datetime] = mapped_column(TIMESTAMP(timezone=True), server_default=func.now())

    user: Mapped["UserORM"] = relationship("UserORM", back_populates="orders")

Key rules:

• Mapped[T] declares the Python type; mapped_column() declares the column config
• nullable=False is explicit — Mapped[str] without it is still nullable in older versions
• Use PGUUID(as_uuid=True) so SQLAlchemy returns Python UUID objects, not strings
• index=True on FK columns — always

Engine and Session Factory

# config/database.py
from sqlalchemy.ext.asyncio import create_async_engine, AsyncSession, async_sessionmaker

# asyncpg driver — fastest PostgreSQL async driver
engine = create_async_engine(
    "postgresql+asyncpg://user:pass@localhost:5432/mydb",
    pool_size=10,           # max persistent connections
    max_overflow=20,        # extra connections above pool_size under load
    pool_pre_ping=True,     # test connections before use (handles dropped connections)
    echo=False,             # set True to log all SQL (dev only)
)

# Session factory — reuse this, don't recreate per request
AsyncSessionLocal = async_sessionmaker(
    engine,
    class_=AsyncSession,
    expire_on_commit=False,   # keep attributes accessible after commit
    autoflush=False,
)

Dependency (FastAPI)

# api/dependencies.py
from sqlalchemy.ext.asyncio import AsyncSession
from config.database import AsyncSessionLocal

async def get_db() -> AsyncSession:
    async with AsyncSessionLocal() as session:
        try:
            yield session
            await session.commit()
        except Exception:
            await session.rollback()
            raise

FastAPI caches this dependency within a request — one session per request.

CRUD Patterns

from sqlalchemy import select, update, delete
from sqlalchemy.ext.asyncio import AsyncSession

class UserCRUD:
    def __init__(self, session: AsyncSession):
        self.session = session

    async def get(self, user_id: UUID) -> UserORM | None:
        result = await self.session.execute(
            select(UserORM).where(UserORM.id == user_id)
        )
        return result.scalar_one_or_none()

    async def get_by_email(self, email: str) -> UserORM | None:
        result = await self.session.execute(
            select(UserORM).where(UserORM.email == email)
        )
        return result.scalar_one_or_none()

    async def list(self, skip: int = 0, limit: int = 100) -> list[UserORM]:
        result = await self.session.execute(
            select(UserORM).order_by(UserORM.created_at.desc()).offset(skip).limit(limit)
        )
        return list(result.scalars().all())

    async def create(self, email: str, name: str, role: str = "user") -> UserORM:
        user = UserORM(email=email, name=name, role=role)
        self.session.add(user)
        await self.session.flush()   # assigns ID without committing
        await self.session.refresh(user)
        return user

    async def update(self, user_id: UUID, **kwargs) -> UserORM | None:
        await self.session.execute(
            update(UserORM).where(UserORM.id == user_id).values(**kwargs)
        )
        return await self.get(user_id)

    async def delete(self, user_id: UUID) -> None:
        await self.session.execute(
            delete(UserORM).where(UserORM.id == user_id)
        )

    async def count(self, role: str | None = None) -> int:
        from sqlalchemy import func
        q = select(func.count()).select_from(UserORM)
        if role:
            q = q.where(UserORM.role == role)
        result = await self.session.execute(q)
        return result.scalar_one()

Joins and Complex Queries

from sqlalchemy import select, and_, or_, func
from sqlalchemy.orm import selectinload, joinedload

# JOIN — users with their order count
result = await session.execute(
    select(UserORM, func.count(OrderORM.id).label("order_count"))
    .outerjoin(OrderORM, UserORM.id == OrderORM.user_id)
    .group_by(UserORM.id)
    .order_by(func.count(OrderORM.id).desc())
)
rows = result.all()   # list of (UserORM, order_count) tuples

# Relationship loading — selectinload avoids N+1
result = await session.execute(
    select(UserORM)
    .options(selectinload(UserORM.orders))   # one extra query for all orders
    .where(UserORM.role == "admin")
)
users = result.scalars().all()
for user in users:
    print(user.orders)   # no extra query

# joinedload — single JOIN query (good for to-one relationships)
result = await session.execute(
    select(OrderORM)
    .options(joinedload(OrderORM.user))
    .where(OrderORM.status == "pending")
)
orders = result.unique().scalars().all()

# Filtering with operators
result = await session.execute(
    select(UserORM).where(
        and_(
            UserORM.role.in_(["admin", "manager"]),
            UserORM.created_at > datetime(2024, 1, 1),
            or_(
                UserORM.name.ilike("%alice%"),
                UserORM.email.ilike("%alice%"),
            ),
        )
    )
)

# JSONB filtering
result = await session.execute(
    select(UserORM).where(
        UserORM.metadata_["plan"].astext == "pro"
    )
)

Loading Strategies

# Set default loading per model
class OrderORM(Base):
    user: Mapped["UserORM"] = relationship(
        "UserORM",
        back_populates="orders",
        lazy="raise",    # must explicitly use joinedload/selectinload in queries
    )

Transactions

# Session auto-handles transaction — commit/rollback in dependency

# Explicit savepoint (nested transaction)
async with session.begin_nested():
    session.add(obj)
    # rolls back to savepoint on exception, not the whole transaction

# Bulk insert (much faster than add() in a loop)
await session.execute(
    UserORM.__table__.insert(),
    [{"email": f"user{i}@test.com", "name": f"User {i}"} for i in range(1000)],
)

# Upsert (PostgreSQL ON CONFLICT)
from sqlalchemy.dialects.postgresql import insert

stmt = insert(UserORM).values(email="alice@example.com", name="Alice")
stmt = stmt.on_conflict_do_update(
    index_elements=["email"],
    set_={"name": stmt.excluded.name, "updated_at": func.now()},
)
await session.execute(stmt)

Alembic Migrations

Setup

# In agentex/:
alembic init alembic           # creates alembic/ dir + alembic.ini
# or with make:
make migration NAME="add_users_table"
make apply-migrations

alembic/env.py — connect to async engine and point at your models:

from sqlalchemy.ext.asyncio import async_engine_from_config
from adapters.orm import Base    # import your Base so models are registered

target_metadata = Base.metadata

def run_migrations_online():
    connectable = async_engine_from_config(config.get_section(config.config_ini_section))
    # ... standard async alembic boilerplate

Migration file patterns

# Auto-generated migration — review before applying
def upgrade() -> None:
    op.create_table(
        "users",
        sa.Column("id", pg.UUID(as_uuid=True), primary_key=True),
        sa.Column("email", sa.String(255), nullable=False),
        sa.Column("created_at", sa.TIMESTAMP(timezone=True), server_default=sa.text("now()")),
    )
    op.create_index("ix_users_email", "users", ["email"], unique=True)

def downgrade() -> None:
    op.drop_table("users")

# Add column safely (large tables)
def upgrade() -> None:
    # Step 1: add nullable first (no table lock)
    op.add_column("orders", sa.Column("shipped_at", sa.TIMESTAMP(timezone=True), nullable=True))
    # Step 2: backfill (do in batches in a separate migration or via cron)
    # Step 3: add NOT NULL constraint after backfill

# Create index concurrently (no table lock)
def upgrade() -> None:
    op.execute("CREATE INDEX CONCURRENTLY idx_orders_user_id ON orders(user_id)")

def downgrade() -> None:
    op.execute("DROP INDEX CONCURRENTLY idx_orders_user_id")

# Data migration
def upgrade() -> None:
    op.execute("UPDATE users SET role = 'member' WHERE role = 'user'")

Entity Conversion

Keep ORM models separate from domain entities. Convert at the adapter boundary:

# adapters/crud_store/users.py
from domain.entities.user import User
from adapters.orm import UserORM

def convert_user_to_entity(orm: UserORM) -> User:
    return User(
        id=orm.id,
        email=orm.email,
        name=orm.name,
        role=orm.role,
        created_at=orm.created_at,
    )

class UserRepository:
    async def get(self, user_id: UUID) -> User | None:
        orm = await UserCRUD(self.session).get(user_id)
        return convert_user_to_entity(orm) if orm else None

Red Flags

• Old-style Column() declarations — Column(String, nullable=False) without Mapped[T] loses the Python type information that mypy and editors rely on; use Mapped[str] = mapped_column(String(255), nullable=False) in all new SQLAlchemy 2.0 code
• Accessing relationships without explicit loading — accessing user.orders in an async context without selectinload or joinedload raises MissingGreenlet or emits implicit lazy SQL that blocks the event loop; always declare the loading strategy in the query
• Creating a new AsyncSession per query — instantiating a session for each database call bypasses connection pooling and transaction batching; create one session per request via the FastAPI dependency
• Missing pool_pre_ping=True — without it, connections dropped by the database (idle timeout, network reset) are handed to the application as stale; the first query fails with a connection error rather than transparently reconnecting
• expire_on_commit=False missing — by default SQLAlchemy expires all attributes after commit; accessing them in an async context after the session commits triggers lazy loads that fail; set expire_on_commit=False in async_sessionmaker
• Missing index=True on foreign key columns — SQLAlchemy does not auto-index FK columns; every JOIN or filter on a FK without an index is a sequential scan
• ORM models imported in the domain layer — importing UserORM in use cases or domain entities couples the business logic to the database schema; all ORM ↔ entity conversion belongs in the adapter layer

Checklist

• Mapped[T] + mapped_column() used (not old Column() style)
• FK columns have index=True
• async_sessionmaker with expire_on_commit=False for async
• pool_pre_ping=True on engine to handle dropped connections
• selectinload / joinedload explicit in every query that accesses a relationship
• flush() used after add() to get DB-generated ID without committing
• Alembic migrations add nullable columns first, then backfill, then add NOT NULL
• CREATE INDEX CONCURRENTLY used for large tables
• ORM models never imported in domain layer — conversion happens in adapter