Algebraic data types

Overview

Algebraic data types (ADTs) are types formed by combining other types using two operations: sum (OR) and product (AND). This algebraic approach enables precise modeling of domain concepts and makes illegal states unrepresentable.

Sum types (discriminated unions)

Model "OR" relationships - a value is one of several possible variants.

Pattern: Enum types for closed value sets

Use PostgreSQL ENUMs when the set of valid values is fixed and known.

sql

-- PostgreSQL: Enum for order status
CREATE TYPE order_status AS ENUM (
  'pending',
  'confirmed',
  'shipped',
  'delivered',
  'cancelled'
);

CREATE TABLE orders (
  id UUID PRIMARY KEY,
  status order_status NOT NULL DEFAULT 'pending',
  metadata JSONB NOT NULL DEFAULT '{}'
);

-- Enforce: Different states have different data requirements
ALTER TABLE orders ADD CONSTRAINT shipped_has_tracking CHECK (
  status != 'shipped' OR (metadata ? 'tracking_number')
);

ALTER TABLE orders ADD CONSTRAINT cancelled_has_reason CHECK (
  status != 'cancelled' OR (metadata ? 'cancellation_reason')
);

DuckDB compatibility:

DuckDB doesn't support ENUM types - use VARCHAR with CHECK constraints:

sql

-- DuckDB equivalent (SQLMesh transpiles automatically)
CREATE TABLE orders (
  status VARCHAR NOT NULL,
  CHECK (status IN ('pending', 'confirmed', 'shipped', 'delivered', 'cancelled'))
);

ENUM evolution:

sql

-- Safe: Add new value (non-breaking)
ALTER TYPE order_status ADD VALUE 'refunded';

-- Breaking: Cannot remove values
-- Must recreate type or mark as deprecated in application

Pattern: CHECK constraints for inline sum types

When ENUM feels heavyweight or values may evolve:

sql

CREATE TABLE events (
  id UUID PRIMARY KEY,
  event_type VARCHAR NOT NULL,
  payload JSONB NOT NULL,
  CHECK (event_type IN ('UserCreated', 'UserUpdated', 'UserDeleted'))
);

Pattern: Tagged unions with JSONB

For sum types where each variant has a different shape:

sql

CREATE TABLE events (
  id UUID PRIMARY KEY,
  event_type VARCHAR NOT NULL,
  payload JSONB NOT NULL,
  created_at TIMESTAMPTZ NOT NULL DEFAULT now()
);

-- Enforce shape constraints per event type
ALTER TABLE events ADD CONSTRAINT user_created_shape CHECK (
  event_type != 'UserCreated' OR (
    jsonb_typeof(payload) = 'object' AND
    (payload ? 'user_id') AND
    (payload ? 'email') AND
    (payload ? 'name')
  )
);

ALTER TABLE events ADD CONSTRAINT user_updated_shape CHECK (
  event_type != 'UserUpdated' OR (
    jsonb_typeof(payload) = 'object' AND
    (payload ? 'user_id') AND
    (payload ? 'changes')
  )
);

ALTER TABLE events ADD CONSTRAINT user_deleted_shape CHECK (
  event_type != 'UserDeleted' OR (
    jsonb_typeof(payload) = 'object' AND
    (payload ? 'user_id')
  )
);

-- Enforce: event_type must be one of known types
ALTER TABLE events ADD CONSTRAINT valid_event_type CHECK (
  event_type IN ('UserCreated', 'UserUpdated', 'UserDeleted')
);

Cross-language sum types

Python: Discriminated unions with Pydantic

python

from typing import Union, Literal
from pydantic import BaseModel
from uuid import UUID

# Each variant is a separate class with discriminator field
class Pending(BaseModel):
    type: Literal["pending"]

class Confirmed(BaseModel):
    type: Literal["confirmed"]
    confirmed_at: datetime

class Shipped(BaseModel):
    type: Literal["shipped"]
    tracking_number: str
    shipped_at: datetime

class Delivered(BaseModel):
    type: Literal["delivered"]
    delivered_at: datetime

class Cancelled(BaseModel):
    type: Literal["cancelled"]
    reason: str
    cancelled_at: datetime

# Discriminated union - Pydantic uses 'type' field to discriminate
OrderStatus = Union[Pending, Confirmed, Shipped, Delivered, Cancelled]

class Order(BaseModel):
    id: UUID
    status: OrderStatus  # Must be one of the variants

# Pattern matching
def processOrder(order: Order) -> str:
    match order.status:
        case Pending():
            return "Processing payment..."
        case Confirmed(confirmed_at=dt):
            return f"Confirmed at {dt}"
        case Shipped(tracking_number=num):
            return f"Tracking: {num}"
        case Delivered(delivered_at=dt):
            return f"Delivered at {dt}"
        case Cancelled(reason=r):
            return f"Cancelled: {r}"

TypeScript: Discriminated unions

typescript

// Each variant is an object type with discriminator
type OrderStatus =
  | { type: "pending" }
  | { type: "confirmed"; confirmedAt: Date }
  | { type: "shipped"; trackingNumber: string; shippedAt: Date }
  | { type: "delivered"; deliveredAt: Date }
  | { type: "cancelled"; reason: string; cancelledAt: Date };

interface Order {
  id: string;
  status: OrderStatus;
}

// Exhaustive type checking - TypeScript errors if we miss a case
function processOrder(order: Order): string {
  switch (order.status.type) {
    case "pending":
      return "Processing payment...";
    case "confirmed":
      return `Confirmed at ${order.status.confirmedAt}`;
    case "shipped":
      return `Tracking: ${order.status.trackingNumber}`;
    case "delivered":
      return `Delivered at ${order.status.deliveredAt}`;
    case "cancelled":
      return `Cancelled: ${order.status.reason}`;
  }
}

Go: Sealed interfaces

// Sum type via sealed interface
type OrderStatus interface {
    isOrderStatus()  // Marker method - makes interface "sealed"
}

// Each variant implements the interface
type Pending struct{}

type Confirmed struct {
    ConfirmedAt time.Time
}

type Shipped struct {
    TrackingNumber string
    ShippedAt      time.Time
}

type Delivered struct {
    DeliveredAt time.Time
}

type Cancelled struct {
    Reason      string
    CancelledAt time.Time
}

// Implement marker method
func (Pending) isOrderStatus()   {}
func (Confirmed) isOrderStatus() {}
func (Shipped) isOrderStatus()   {}
func (Delivered) isOrderStatus() {}
func (Cancelled) isOrderStatus() {}

// Pattern matching via type switch
func processOrder(status OrderStatus) string {
    switch s := status.(type) {
    case Pending:
        return "Processing payment..."
    case Confirmed:
        return fmt.Sprintf("Confirmed at %v", s.ConfirmedAt)
    case Shipped:
        return fmt.Sprintf("Tracking: %s", s.TrackingNumber)
    case Delivered:
        return fmt.Sprintf("Delivered at %v", s.DeliveredAt)
    case Cancelled:
        return fmt.Sprintf("Cancelled: %s", s.Reason)
    default:
        return "Unknown status"
    }
}

Product types

Model "AND" relationships - a value contains all fields together.

Product types are already implicit in your schema design: tables and records.

python

# Product type: User has id AND email AND created_at
class User(BaseModel):
    id: UUID          # Has id
    email: str        # AND email
    created_at: datetime  # AND created_at

sql

-- Table is a product type
CREATE TABLE users (
  id UUID,           -- Has id
  email VARCHAR,     -- AND email
  created_at TIMESTAMPTZ  -- AND created_at
);

Pattern: Composite types for temporary values

sql

-- PostgreSQL composite type (product)
CREATE TYPE coordinates AS (
  x INTEGER,
  y INTEGER
);

CREATE TYPE address AS (
  street VARCHAR,
  city VARCHAR,
  state VARCHAR,
  zip VARCHAR
);

python

# Python tuple (product)
Coordinates = tuple[int, int]

# Named tuple (product with names)
from typing import NamedTuple

class Coordinates(NamedTuple):
    x: int
    y: int

Algebraic properties

The name "algebraic data types" reflects that sum and product types satisfy algebraic laws analogous to arithmetic. Understanding these properties enables reasoning about type equivalences and refactoring.

Sum types as addition

Sum types correspond to addition (coproduct in category theory). The number of inhabitants of A | B equals the sum of inhabitants of A and B.

Laws:

•Commutativity: A | B ≅ B | A (order doesn't matter)
•Associativity: (A | B) | C ≅ A | (B | C) (grouping doesn't matter)
•Identity: A | Never ≅ A (Never/Void is the zero element)

Product types as multiplication

Product types correspond to multiplication (categorical product). The number of inhabitants of (A, B) equals the product of inhabitants of A and B.

Laws:

•Commutativity: (A, B) ≅ (B, A)
•Associativity: ((A, B), C) ≅ (A, (B, C))
•Identity: (A, ()) ≅ A (unit type is the identity)

Distributivity

Products distribute over sums, just like multiplication over addition.

code

A × (B + C) ≅ (A × B) + (A × C)

This equivalence enables refactoring between representations:

rust

// Before: product containing sum
struct Tagged<A, B> {
    tag: String,
    value: Either<A, B>,
}

// After: sum of products (often clearer)
enum Tagged<A, B> {
    Left { tag: String, value: A },
    Right { tag: String, value: B },
}

Exponentials as functions

Function types correspond to exponentiation. The type A -> B has |B|^|A| inhabitants.

This completes the arithmetic correspondence:

•0 = Never/Void (uninhabited)
•1 = () / Unit (single inhabitant)
•+ = Sum types
•× = Product types
•^ = Function types

See theoretical-foundations.md#algebraic-data-types-as-initial-algebras for the categorical perspective.

Recursive types and initial algebras

Recursive ADTs like lists and trees are initial algebras of their defining functor. The fold operation (catamorphism) is the unique homomorphism from this initial algebra.

haskell

-- List as initial algebra of ListF
data ListF a r = Nil | Cons a r

-- Natural numbers as initial algebra of NatF
data NatF r = Zero | Succ r

For the full categorical treatment of ADTs as initial algebras, including catamorphisms and anamorphisms, see theoretical-foundations.md#f-algebras-and-catamorphisms.

Newtype pattern

Wrap primitive types to prevent mixing semantically different values.

Pattern: DOMAIN types in PostgreSQL

sql

-- Domain types add semantic meaning to primitives
CREATE DOMAIN email_address AS VARCHAR
  CHECK (VALUE ~ '^[^@]+@[^@]+\.[^@]+$');

CREATE DOMAIN positive_int AS INTEGER
  CHECK (VALUE > 0);

CREATE DOMAIN url AS TEXT
  CHECK (VALUE ~ '^https?://');

CREATE DOMAIN percentage AS NUMERIC(5,2)
  CHECK (VALUE >= 0 AND VALUE <= 100);

-- Use in tables
CREATE TABLE users (
  id UUID PRIMARY KEY,
  email email_address NOT NULL,
  age positive_int,
  website url,
  completion_rate percentage
);

DuckDB compatibility:

DuckDB doesn't support DOMAIN - use inline CHECK constraints:

sql

-- DuckDB equivalent
CREATE TABLE users (
  id UUID PRIMARY KEY,
  email VARCHAR NOT NULL CHECK (email ~ '^[^@]+@[^@]+\.[^@]+$'),
  age INTEGER CHECK (age > 0),
  website TEXT CHECK (website ~ '^https?://'),
  completion_rate NUMERIC(5,2) CHECK (completion_rate BETWEEN 0 AND 100)
);

Pattern: Wrapper types in application code

Python: Single-field Pydantic models

python

from pydantic import BaseModel, validator
from uuid import UUID

# Prevent mixing up different ID types
class UserId(BaseModel):
    value: UUID

class OrderId(BaseModel):
    value: UUID

class ProductId(BaseModel):
    value: UUID

# Validated string types
class EmailAddress(BaseModel):
    value: str

    @validator('value')
    def must_be_valid(cls, v):
        if '@' not in v or '.' not in v.split('@')[1]:
            raise ValueError('invalid email format')
        return v.lower().strip()

class PhoneNumber(BaseModel):
    value: str

    @validator('value')
    def must_be_valid(cls, v):
        # Remove formatting
        digits = ''.join(c for c in v if c.isdigit())
        if len(digits) != 10:
            raise ValueError('must be 10 digits')
        return digits

# Quantity types with units
class Dollars(BaseModel):
    value: Decimal

    @validator('value')
    def must_be_non_negative(cls, v):
        if v < 0:
            raise ValueError('cannot be negative')
        return v

# Type system prevents mistakes
def getUser(id: UserId) -> User:  # Type error if passed OrderId!
    ...

def chargeCard(amount: Dollars) -> PaymentResult:  # Type error if passed raw Decimal!
    ...

TypeScript: Branded types

typescript

// Nominal typing via branding
type Brand<K, T> = K & { __brand: T };

type UserId = Brand<string, "UserId">;
type OrderId = Brand<string, "OrderId">;
type EmailAddress = Brand<string, "EmailAddress">;
type Dollars = Brand<number, "Dollars">;

// Smart constructors that validate
function makeUserId(s: string): UserId | Error {
  // Validate UUID format
  if (!/^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i.test(s)) {
    return new Error("invalid UUID");
  }
  return s as UserId;
}

function makeEmailAddress(s: string): EmailAddress | Error {
  if (!s.includes("@") || !s.split("@")[1].includes(".")) {
    return new Error("invalid email");
  }
  return s.toLowerCase().trim() as EmailAddress;
}

function makeDollars(n: number): Dollars | Error {
  if (n < 0) {
    return new Error("cannot be negative");
  }
  return n as Dollars;
}

// Type safety at compile time
function getUser(id: UserId): User {
  // ...
}

// This is a compile error:
// getUser(makeOrderId("..."));  // Type error!
// getUser("raw-string");         // Type error!

Go: Struct wrappers

// Newtype via struct wrapper
type UserId struct {
    value uuid.UUID
}

func NewUserId(id uuid.UUID) UserId {
    return UserId{value: id}
}

func (u UserId) String() string {
    return u.value.String()
}

type EmailAddress struct {
    value string
}

func NewEmailAddress(s string) (EmailAddress, error) {
    if !strings.Contains(s, "@") || !strings.Contains(strings.Split(s, "@")[1], ".") {
        return EmailAddress{}, errors.New("invalid email format")
    }
    normalized := strings.ToLower(strings.TrimSpace(s))
    return EmailAddress{value: normalized}, nil
}

func (e EmailAddress) String() string {
    return e.value
}

type Dollars struct {
    value decimal.Decimal
}

func NewDollars(amount decimal.Decimal) (Dollars, error) {
    if amount.LessThan(decimal.Zero) {
        return Dollars{}, errors.New("cannot be negative")
    }
    return Dollars{value: amount}, nil
}

// Type safety
func GetUser(id UserId) (*User, error) {
    // ...
}

// This won't compile:
// GetUser(NewOrderId(...))  // Type error!
// GetUser("raw-string")     // Type error!

When to use newtypes

Always wrap primitives when:

•IDs that should never be mixed: UserId vs OrderId vs ProductId
•Validated strings: EmailAddress, PhoneNumber, URL, SSN
•Quantities with units: Meters, Dollars, Seconds, Bytes
•Opaque tokens: ApiKey, SessionToken, PasswordHash
•Constrained values: PositiveInt, NonEmptyString, Percentage

Making illegal states unrepresentable

Design types so invalid combinations cannot be constructed.

Anti-pattern: Boolean flags with dependent fields

sql

-- Bad: Many illegal states possible
CREATE TABLE reservations (
  id UUID PRIMARY KEY,
  confirmed BOOLEAN NOT NULL,
  cancelled BOOLEAN NOT NULL,
  confirmation_date TIMESTAMPTZ,
  cancellation_reason TEXT
);

-- Illegal states:
-- confirmed=true, cancelled=true (both!)
-- confirmed=true, confirmation_date=NULL (no date!)
-- cancelled=false, cancellation_reason='...' (reason without cancel!)

Pattern: Sum types for mutually exclusive states

sql

-- Good: Only valid states are possible
CREATE TYPE reservation_status AS ENUM ('pending', 'confirmed', 'cancelled');

CREATE TABLE reservations (
  id UUID PRIMARY KEY,
  status reservation_status NOT NULL,
  metadata JSONB NOT NULL DEFAULT '{}'
);

-- Enforce: confirmed requires confirmation_date
ALTER TABLE reservations ADD CONSTRAINT confirmed_has_date CHECK (
  status != 'confirmed' OR (
    jsonb_typeof(metadata->'confirmation_date') = 'string'
  )
);

-- Enforce: cancelled requires cancellation_reason
ALTER TABLE reservations ADD CONSTRAINT cancelled_has_reason CHECK (
  status != 'cancelled' OR (
    jsonb_typeof(metadata->'cancellation_reason') = 'string'
  )
);

-- Illegal states now impossible at schema level!

Generated code enforces at type level:

python

class Pending(BaseModel):
    type: Literal["pending"]
    # No extra fields

class Confirmed(BaseModel):
    type: Literal["confirmed"]
    confirmation_date: datetime  # Required!

class Cancelled(BaseModel):
    type: Literal["cancelled"]
    cancellation_reason: str  # Required!

ReservationStatus = Union[Pending, Confirmed, Cancelled]

class Reservation(BaseModel):
    id: UUID
    status: ReservationStatus

# Impossible to construct invalid state:
# - Can't be confirmed without date
# - Can't be cancelled without reason
# - Can't be both confirmed and cancelled

Anti-pattern: Optional fields that should be mutually exclusive

sql

-- Bad: Can have both payment_method types
CREATE TABLE payments (
  id UUID PRIMARY KEY,
  credit_card_token VARCHAR,
  bank_account_number VARCHAR,
  amount NUMERIC NOT NULL
);

-- Illegal: both credit card and bank account
-- Illegal: neither payment method

Pattern: Sum type for payment methods

sql

-- Good: Exactly one payment method
CREATE TYPE payment_method_type AS ENUM ('credit_card', 'bank_account', 'paypal');

CREATE TABLE payments (
  id UUID PRIMARY KEY,
  payment_method_type payment_method_type NOT NULL,
  payment_details JSONB NOT NULL,
  amount NUMERIC NOT NULL,

  CHECK (
    (payment_method_type = 'credit_card' AND payment_details ? 'card_token') OR
    (payment_method_type = 'bank_account' AND payment_details ? 'account_number') OR
    (payment_method_type = 'paypal' AND payment_details ? 'paypal_email')
  )
);

Generated code:

python

class CreditCardPayment(BaseModel):
    type: Literal["credit_card"]
    card_token: str

class BankAccountPayment(BaseModel):
    type: Literal["bank_account"]
    account_number: str
    routing_number: str

class PaypalPayment(BaseModel):
    type: Literal["paypal"]
    paypal_email: str

PaymentMethod = Union[CreditCardPayment, BankAccountPayment, PaypalPayment]

class Payment(BaseModel):
    id: UUID
    payment_method: PaymentMethod  # Exactly one!
    amount: Decimal

State machines with sum types

Sum types naturally model state machines where entities transition through discrete states.

Pattern: Shopping cart states

python

from typing import Literal, Union
from pydantic import BaseModel
from datetime import datetime
from decimal import Decimal

class EmptyCart(BaseModel):
    type: Literal["empty"]
    session_id: str

class ActiveCart(BaseModel):
    type: Literal["active"]
    session_id: str
    items: list[CartItem]
    subtotal: Decimal

class PaidCart(BaseModel):
    type: Literal["paid"]
    session_id: str
    items: list[CartItem]
    total: Decimal
    payment_id: str
    paid_at: datetime

CartState = EmptyCart | ActiveCart | PaidCart

# State transitions
def add_item(cart: EmptyCart | ActiveCart, item: CartItem) -> ActiveCart:
    """Adding item always results in ActiveCart."""
    items = cart.items if isinstance(cart, ActiveCart) else []
    new_items = items + [item]
    subtotal = sum(i.price * i.quantity for i in new_items)
    return ActiveCart(
        type="active",
        session_id=cart.session_id,
        items=new_items,
        subtotal=subtotal
    )

def checkout(cart: ActiveCart, payment_id: str) -> PaidCart:
    """Can only checkout ActiveCart, produces PaidCart."""
    return PaidCart(
        type="paid",
        session_id=cart.session_id,
        items=cart.items,
        total=cart.subtotal,
        payment_id=payment_id,
        paid_at=datetime.now()
    )

# Type system prevents invalid transitions:
# - Can't checkout EmptyCart (type error)
# - Can't add items to PaidCart (type error)
# - PaidCart has payment_id that ActiveCart lacks

Pattern: Email verification workflow

python

class UnverifiedEmail(BaseModel):
    type: Literal["unverified"]
    email: str
    verification_token: str
    sent_at: datetime

class VerifiedEmail(BaseModel):
    type: Literal["verified"]
    email: str
    verified_at: datetime

EmailState = UnverifiedEmail | VerifiedEmail

# Transitions
def send_verification(email: str) -> UnverifiedEmail:
    """Create unverified email with token."""
    return UnverifiedEmail(
        type="unverified",
        email=email,
        verification_token=generate_token(),
        sent_at=datetime.now()
    )

def verify(unverified: UnverifiedEmail, token: str) -> Result[VerifiedEmail, str]:
    """Verify email if token matches."""
    if token != unverified.verification_token:
        return Error("Invalid verification token")

    return Ok(VerifiedEmail(
        type="verified",
        email=unverified.email,
        verified_at=datetime.now()
    ))

# Can only send password reset to verified emails
def send_password_reset(email: VerifiedEmail) -> None:
    """Only accepts VerifiedEmail - type system enforces."""
    send_email(email.email, "Password reset...")

# Type error if try to pass UnverifiedEmail:
# send_password_reset(unverified_email)  # ⊘ Type error!

Pattern: Document approval workflow

typescript

// Document lifecycle states
interface Draft {
  type: "draft";
  content: string;
  author: string;
  lastModified: Date;
}

interface Submitted {
  type: "submitted";
  content: string;
  author: string;
  submittedAt: Date;
  reviewers: string[];
}

interface Approved {
  type: "approved";
  content: string;
  author: string;
  approvedBy: string;
  approvedAt: Date;
}

interface Rejected {
  type: "rejected";
  content: string;
  author: string;
  rejectedBy: string;
  rejectedAt: Date;
  reason: string;
}

type DocumentState = Draft | Submitted | Approved | Rejected;

// Transitions
function submit(draft: Draft, reviewers: string[]): Submitted {
  return {
    type: "submitted",
    content: draft.content,
    author: draft.author,
    submittedAt: new Date(),
    reviewers,
  };
}

function approve(submitted: Submitted, approver: string): Approved {
  if (!submitted.reviewers.includes(approver)) {
    throw new Error("Approver must be in reviewers list");
  }

  return {
    type: "approved",
    content: submitted.content,
    author: submitted.author,
    approvedBy: approver,
    approvedAt: new Date(),
  };
}

// Type system enforces:
// - Can't approve a Draft (only Submitted)
// - Can't reject an Approved document
// - Rejected documents have required reason

See also: domain-modeling.md#pattern-3-state-machines-for-entity-lifecycles for workflows and domain-modeling patterns

Type organization and dependencies

How you organize types affects compilation order and maintainability.

Principle: Dependencies should form a DAG

Types should depend on each other in a directed acyclic graph (no circular dependencies).

code

Low-level types (primitives, newtypes)
       ↓
Value objects (email, money, quantity)
       ↓
Simple domain types (product, customer)
       ↓
Aggregates (order with order lines)
       ↓
Workflows (place order, fulfill order)

Pattern: Define types in dependency order

python

# 1. Primitive types first (no dependencies)
class EmailAddress(BaseModel):
    value: str
    @field_validator('value')
    @classmethod
    def must_be_valid(cls, v: str) -> str:
        if '@' not in v:
            raise ValueError('invalid email')
        return v

class ProductCode(BaseModel):
    value: str

class Quantity(BaseModel):
    value: int
    @field_validator('value')
    @classmethod
    def must_be_positive(cls, v: int) -> int:
        if v <= 0:
            raise ValueError('quantity must be positive')
        return v

# 2. Simple domain types (depend on primitives)
class Product(BaseModel):
    code: ProductCode
    name: str
    price: Decimal

class Customer(BaseModel):
    email: EmailAddress
    name: str

# 3. Aggregates (depend on simple types)
class OrderLine(BaseModel):
    product_code: ProductCode
    quantity: Quantity
    price: Decimal

class Order(BaseModel):
    customer_email: EmailAddress
    lines: list[OrderLine]
    total: Decimal

# 4. Workflows (depend on aggregates)
def place_order(
    customer: Customer,
    lines: list[OrderLine]
) -> Result[Order, OrderError]:
    ...

Anti-pattern: Circular dependencies

python

# Bad: A depends on B, B depends on A
class Order(BaseModel):
    customer: Customer  # Order depends on Customer

class Customer(BaseModel):
    orders: list[Order]  # Customer depends on Order ⊘

Fix: Use IDs for references across aggregates

python

# Good: Break cycle with ID reference
class Order(BaseModel):
    customer_id: CustomerId  # Reference by ID

class Customer(BaseModel):
    id: CustomerId
    # To get orders, query separately:
    # orders = get_orders_for_customer(customer.id)

Guideline: Module organization

Organize code to reflect dependency hierarchy:

code

src/
  domain/
    primitives.py      # EmailAddress, ProductCode, Quantity
    value_objects.py   # Money, Address
    entities.py        # Product, Customer (depend on primitives)
    aggregates.py      # Order (depends on entities)
  workflows/
    place_order.py     # Depends on domain types
    fulfill_order.py

This ensures:

•Low-level types defined before high-level types
•No circular imports
•Clear dependency structure
•Easy to test in isolation

See also:

•domain-modeling.md#pattern-1-types-as-domain-vocabulary for domain type patterns
•domain-modeling.md#pattern-5-aggregates-as-consistency-boundaries for aggregate design

Testing ADTs

Property-based testing

Use hypothesis (Python) or similar to test ADT invariants:

python

from hypothesis import given, strategies as st
from hypothesis.strategies import builds

# Generate arbitrary EmailAddress values
email_strategy = builds(
    EmailAddress,
    value=st.emails()
)

@given(email_strategy)
def test_email_roundtrip(email: EmailAddress):
    """Property: parsing serialized email should give original"""
    serialized = email.value
    parsed = EmailAddress(value=serialized)
    assert parsed == email

@given(email_strategy)
def test_email_normalized(email: EmailAddress):
    """Property: emails are always lowercase and trimmed"""
    assert email.value == email.value.lower().strip()

# Generate arbitrary OrderStatus values
order_status_strategy = st.one_of(
    builds(Pending, type=st.just("pending")),
    builds(Shipped,
           type=st.just("shipped"),
           tracking_number=st.text(min_size=5, max_size=20)),
    # ... other variants
)

@given(order_status_strategy)
def test_order_status_serialization(status: OrderStatus):
    """Property: status can roundtrip through JSON"""
    json_data = status.json()
    parsed = parse_order_status(json_data)  # Your parser
    assert parsed == status

Exhaustiveness testing

Ensure pattern matching handles all cases:

python

def test_all_order_statuses_handled():
    """Ensure processOrder handles all status variants"""
    statuses = [
        Pending(type="pending"),
        Confirmed(type="confirmed", confirmed_at=datetime.now()),
        Shipped(type="shipped", tracking_number="ABC123", shipped_at=datetime.now()),
        Delivered(type="delivered", delivered_at=datetime.now()),
        Cancelled(type="cancelled", reason="test", cancelled_at=datetime.now()),
    ]

    for status in statuses:
        order = Order(id=uuid4(), status=status)
        result = processOrder(order)
        assert result is not None  # Ensure no case is missed

ADTs in event-sourced systems

In event-sourced architectures, algebraic data types form the foundation of the Decider pattern—the canonical structure for command handling and event application.

Commands are typically sum types (enums) representing possible operations: PlaceOrder | CancelOrder | UpdateQuantity. Each variant carries the data necessary to execute that specific command.

Events are also sum types representing facts that occurred: OrderPlaced | OrderCancelled | QuantityUpdated. Event variants include all data necessary to apply the state change, including failure events like PaymentFailed or InsufficientInventory.

State is typically a product type (struct) containing all aggregate fields together: Order { id, customer, items, status, total }. State fields are often monoidal (can be combined associatively), enabling efficient event replay and incremental updates.

The Decider pattern composes these ADTs into a cohesive structure:

rust

struct Decider<Command, Event, State> {
    decide: fn(&Command, &State) -> Vec<Event>,
    evolve: fn(&State, &Event) -> State,
    initial_state: State,
}

The decide function validates commands against current state and produces events (possibly including rejection events). The evolve function applies events to state, producing new state. Together they form an algebra where commands flow through validation to produce facts, and facts accumulate into state.

See event-sourcing.md#the-decider-pattern for operational patterns and theoretical-foundations.md#coalgebra-algebra-duality for the categorical foundation.

Integration with schema versioning

See ~/.claude/commands/preferences/schema-versioning.md for:

•How to configure sqlc to generate ADT types
•Cross-database compatibility (PostgreSQL ENUMs → DuckDB CHECK)
•Migration patterns for evolving ADTs

See ~/.claude/commands/preferences/railway-oriented-programming.md for:

•How to use ADTs with Result types for error handling
•Composing operations on sum types with bind/apply