SQLAlchemy 2.0 Patterns (framework-agnostic core)

Detection

Read pyproject.toml or requirements.txt before writing any model code:

grep -E "sqlalchemy" pyproject.toml requirements.txt

• SQLAlchemy 2.0+: use Mapped/mapped_column syntax (shown throughout this skill). This is the assumed baseline.
• SQLAlchemy 1.x: use Column()/relationship() style. Mark a comment in the code noting the legacy version; do not silently mix styles.

Always prefer 2.0 style for new code. When working in an existing project using 1.x style throughout, match the existing style. Never mix 1.x Column() and 2.0 mapped_column() in the same model.

---

Declarative base and mapped classes

Define one Base class per project. All models inherit from it. (Framework plugins may supply the base for you — e.g., Flask-SQLAlchemy's db.Model — but the mapped-column rules below are identical.)

from datetime import datetime
from decimal import Decimal
from typing import Optional

from sqlalchemy import DateTime, ForeignKey, Numeric, String, func
from sqlalchemy.orm import DeclarativeBase, Mapped, mapped_column, relationship


class Base(DeclarativeBase):
    pass


class User(Base):
    __tablename__ = "users"

    id: Mapped[int] = mapped_column(primary_key=True)
    email: Mapped[str] = mapped_column(String(255), unique=True, index=True)
    display_name: Mapped[str] = mapped_column(String(100))
    hashed_password: Mapped[str] = mapped_column(String(255))
    is_active: Mapped[bool] = mapped_column(default=True)
    created_at: Mapped[datetime] = mapped_column(
        DateTime(timezone=True), server_default=func.now()
    )

    orders: Mapped[list["Order"]] = relationship(
        "Order", back_populates="user", lazy="selectin"
    )

    def __repr__(self) -> str:
        return f"<User id={self.id} email={self.email!r}>"

Key rules:

• Mapped[T] without Optional means NOT NULL. Mapped[Optional[T]] means nullable.
• mapped_column() without a SQLAlchemy type uses Python type inference — always provide the type explicitly (e.g., String(255)) so the migration specialist can finalize column lengths, precision, and constraints correctly.
• Use server_default=func.now() for database-side default timestamps, not Python-side default=datetime.utcnow (Python-side defaults are not reflected in the DB schema, and utcnow is deprecated).

---

Column type guidance

Python type	SQLAlchemy column type	Notes
str	String(N)	Always set length; never bare String
Decimal	Numeric(precision, scale)	Never Float for money or precise values
datetime	DateTime(timezone=True)	Always set timezone=True
int	Integer or BigInteger	Use BigInteger for large tables (users, events)
bool	Boolean
UUID	Uuid (SA 2.0+) or String(36)	Uuid stores as native UUID on PostgreSQL
enum	Enum(MyEnum, native_enum=False)	native_enum=False for DB portability
float	Float	Only for non-monetary approximations (lat/lon, scores)

---

Querying patterns (2.0 style)

Always use 2.0-style select() statements executed through the session. The session/execution style (sync db.session vs async AsyncSession) comes from the framework delta skill; the statement construction is identical.

from sqlalchemy import select
from sqlalchemy.orm import selectinload

select(User).where(User.id == user_id)                      # single row by predicate
select(User).where(User.email == email)                     # lookup by unique column
select(User).offset(skip).limit(limit)                      # pagination
select(User).options(selectinload(User.orders)).where(...)  # explicit eager load

Result handling rules:

• Use scalar_one_or_none() for single-row queries.
• Use scalars().all() for multi-row queries.
• Use session.get(Model, pk) for primary-key lookups where the session API supports it.

Transaction/flush discipline:

• Use flush() inside a unit of work to get the generated PK without committing.
• Commit once per request/unit of work at the boundary the framework defines (FastAPI: the get_db() dependency on yield exit; Flask: the view/service layer at the end of the request) — never scattered through helper functions.

---

Relationships

Define relationships with explicit lazy and cascade settings. Never rely on defaults.

from sqlalchemy.orm import Mapped, mapped_column, relationship


class User(Base):
    __tablename__ = "users"

    id: Mapped[int] = mapped_column(primary_key=True)

    orders: Mapped[list["Order"]] = relationship(
        "Order",
        back_populates="user",
        lazy="selectin",
        cascade="all, delete-orphan",
    )

    profile: Mapped[Optional["UserProfile"]] = relationship(
        "UserProfile",
        back_populates="user",
        lazy="selectin",
        uselist=False,
        cascade="all, delete-orphan",
    )


class Order(Base):
    __tablename__ = "orders"

    id: Mapped[int] = mapped_column(primary_key=True)
    user_id: Mapped[int] = mapped_column(ForeignKey("users.id", ondelete="CASCADE"))

    user: Mapped["User"] = relationship("User", back_populates="orders", lazy="selectin")
    lines: Mapped[list["OrderLine"]] = relationship(
        "OrderLine",
        back_populates="order",
        lazy="selectin",
        cascade="all, delete-orphan",
    )

Structural rules:

• Always pair both sides with back_populates.
• Use uselist=False for the scalar side of one-to-one relations.
• Set ondelete on the ForeignKey (e.g., "CASCADE") so the constraint matches the ORM cascade.

Lazy loading strategy guide:

• lazy="selectin" — loads the related collection with a separate SELECT IN query. Best for small-to-medium collections that are always needed. Safe in both sync and async contexts.
• lazy="select" — separate SELECT on first attribute access (SQLAlchemy default). Only safe in synchronous contexts; in async it triggers implicit IO and fails.
• lazy="joined" — loads the relation with a JOIN in the same query. Best for one-to-one relations or small, always-needed collections.
• lazy="raise" — raises if accessed without explicit eager loading, forcing selectinload()/joinedload() at query time. Best for large or rarely-needed collections to prevent N+1.
• lazy="subquery" — sync-only; not supported by async drivers.
• lazy="dynamic" — deprecated in SQLAlchemy 2.0. Do not use. Replace with explicit select() queries.

Which strategies are permitted per framework is refined by the framework delta skill (async contexts forbid select/subquery).

To prevent N+1 queries on list endpoints, apply eager-load options at query time:

from sqlalchemy.orm import joinedload, selectinload

select(User).options(selectinload(User.orders))

---

Migration metadata rules (Alembic-agnostic)

Both FastAPI (raw Alembic) and Flask (Flask-Migrate, which wraps Alembic) rely on --autogenerate reading the metadata object (Base.metadata / db.metadata).

• Import all models before target_metadata is set. Any model module not imported is invisible to autogenerate and its tables are silently missed.
• Centralize model imports in one module (e.g., app/db/base.py or app/models/__init__.py) so a single import guarantees complete metadata.
• The metadata object is the source of truth for --autogenerate; explicit column types (lengths, precision, timezone=True) are what make the generated migrations correct.
• The migration specialist agent runs migration commands; the architect agent only defines models.

Framework-specific wiring (async env.py for FastAPI, flask db init scaffolding for Flask) lives in the respective delta skill.

---

Anti-patterns (shared)

Anti-pattern	Problem	Correct approach
String without length	Alembic autogenerate produces VARCHAR with no length; some databases use TEXT or reject it	Always String(N)
Float for monetary values	IEEE 754 rounding errors on financial calculations	Numeric(precision, scale)
lazy="dynamic"	Deprecated in SQLAlchemy 2.0; raises a warning	Use lazy="selectin"/lazy="select" (per framework) or explicit selectinload()
default=datetime.utcnow	Python-side default — not reflected in DB schema; utcnow is deprecated	server_default=func.now() with DateTime(timezone=True)
Importing models only in routers	Migration tooling never sees them; --autogenerate misses tables	Import all models in the central metadata module
Mixing Column() and mapped_column()	Produces inconsistent metadata; confuses tooling	One style per project; prefer mapped_column() for new code
Scattered commit() calls in helpers	Makes unit testing harder; unclear transaction boundaries	Commit once per unit of work at the framework-defined boundary; flush() for PKs