structural-design-principles

Structural Design Principles

These principles originated in object-oriented design but apply to any programming paradigm . They're about code structure, not paradigm.

Paradigm Translations

In functional programming (Elixir), they manifest as:

• Composition Over Inheritance → Function composition, module composition, pipe operators
• Law of Demeter → Minimize coupling between data structures, delegate to owning modules
• Tell, Don't Ask → Push logic to the module owning the data type
• Encapsulation → Module boundaries, immutability, pattern matching, opaque types

In object-oriented programming (TypeScript/React): Apply traditional OO interpretations with classes, interfaces, and encapsulation.

The underlying principle is the same across paradigms: manage dependencies, reduce coupling, and maintain clear boundaries.

---

Four Core Principles

1. Composition Over Inheritance

Favor composition (combining simple behaviors) over inheritance (extending base classes).

Why: Inheritance creates tight coupling and fragile hierarchies. Composition provides flexibility.

Elixir Approach (No Inheritance)

Elixir doesn't have inheritance - it uses composition naturally through:

• Module imports (use, import, alias)
• Function composition (|> pipe operator)
• Struct embedding
• Behaviour protocols

# GOOD - Composition with pipes
def process_payment(order) do
  order
  |> validate_items()
  |> calculate_total()
  |> apply_discounts()
  |> charge_payment()
  |> send_receipt()
end

# GOOD - Compose behaviors
defmodule User do
  use YourApp.Model
  use Ecto.Schema
  import Ecto.Changeset

  # Composes functionality from multiple modules
end

# GOOD - Struct embedding
defmodule Address do
  embedded_schema do
    field :street, :string
    field :city, :string
  end
end

defmodule User do
  schema "users" do
    embeds_one :address, Address
  end
end

TypeScript Examples

// BAD - Inheritance hierarchy
class Animal {
  move() { }
}

class FlyingAnimal extends Animal {
  fly() { }
}

class SwimmingAnimal extends Animal {
  swim() { }
}

class Duck extends FlyingAnimal {
  // Problem: Can't also inherit from SwimmingAnimal
  // Forced to duplicate swim() logic
}

// GOOD - Composition with interfaces
interface Movable {
  move(): void;
}

interface Flyable {
  fly(): void;
}

interface Swimmable {
  swim(): void;
}

class Duck implements Movable, Flyable, Swimmable {
  move() { this.walk(); }
  fly() { /* flying logic */ }
  swim() { /* swimming logic */ }
  private walk() { /* walking logic */ }
}

// GOOD - React composition (pattern)
// Instead of class inheritance, compose components

// Base behaviors as hooks
function useTaskData(gigId: string) {
  // Data fetching logic
}

function useTaskActions(gig: Task) {
  // Action handlers
}

function useTaskValidation(gig: Task) {
  // Validation logic
}

// Compose in component
function TaskDetails({ gigId }: Props) {
  const gig = useTaskData(gigId);
  const actions = useTaskActions(gig);
  const validation = useTaskValidation(gig);

  // Combines all behaviors through composition
  return <View>{/* render */}</View>;
}

Composition Guidelines

• Elixir: Use pipes, protocols, and behaviors instead of inheritance trees
• TypeScript: Use interfaces, hooks, and function composition instead of class hierarchies
• Build complex behavior from simple, reusable parts
• Favor "has-a" over "is-a" relationships
• Keep hierarchies shallow (max 2-3 levels if unavoidable)

2. Law of Demeter (Principle of Least Knowledge)

A module should only talk to its immediate friends, not strangers.

The Rule: Only call methods on:

1. The object itself
2. Objects passed as parameters
3. Objects it creates
4. Its direct properties/fields

DON'T chain through multiple objects (train wrecks)

Why (Elixir Context)

• Reduces coupling: When you reach through multiple data structures (engagement.worker.address.city), you're coupled to the entire chain. If any intermediate structure changes, your code breaks.
• Improves testability: Code that only calls functions on its immediate collaborators is easier to test - you don't need to construct deep object graphs.
• Enables refactoring: You can change internal structure without breaking callers if they only interact with top-level functions.
• Follows functional boundaries: In Elixir, each module should be responsible for its own data type. The Law of Demeter enforces this by pushing you to delegate to the owning module.

Elixir Examples

# BAD - Violates Law of Demeter (train wreck)
def get_worker_city(engagement) do
  engagement.worker.address.city
  # Knows too much about internal structure
end

# What if worker doesn't have address?
# What if address structure changes?
# Tightly coupled to implementation

# GOOD - Delegate to the module that owns the data type
defmodule Assignment do
  def worker_city(%{worker: worker}) do
    User.city(worker)
  end
end

defmodule User do
  def city(%{address: address}) do
    Address.city(address)
  end

  def city(_), do: nil
end

# Now can call: Assignment.worker_city(engagement)
# Each module is responsible for its own data
# Coupling minimized to immediate collaborators

# BAD - Reaching through associations
def total_gig_hours(user_id) do
  user = Repo.get!(User, user_id)
  assignments = user.assignments
  Enum.reduce(assignments, 0, fn eng, acc ->
    acc + eng.shift.hours  # Reaching through
  end)
end

# GOOD - Delegate to the domain
def total_gig_hours(user_id) do
  user = Repo.get!(User, user_id)
  User.total_hours(user)
end

defmodule User do
  def total_hours(%{assignments: assignments}) do
    Enum.reduce(assignments, 0, fn eng, acc ->
      acc + Assignment.hours(eng)
    end)
  end
end

defmodule Assignment do
  def hours(%{shift: shift}), do: WorkPeriod.hours(shift)
end

TypeScript Examples: Law of Demeter

// BAD - Chain of doom
function displayUserLocation(engagement: Assignment) {
  const location = engagement.worker.profile.address.city;
  // Knows about 4 levels of object structure!
  return `Location: ${location}`;
}

// GOOD - Each object provides what you need
function displayUserLocation(engagement: Assignment) {
  const location = engagement.getUserCity();
  return `Location: ${location}`;
}

class Assignment {
  getUserCity(): string {
    return this.worker.getCity();
  }
}

class User {
  getCity(): string {
    return this.address.city;
  }
}

// BAD - GraphQL fragments violating Law of Demeter
const fragment = graphql`
  fragment TaskCard_gig on Task {
    id
    requester {
      organization {
        billing {
          paymentMethod {
            last4
          }
        }
      }
    }
  }
`;
// TaskCard shouldn't know about payment details!

// GOOD - Only query what you need
const fragment = graphql`
  fragment TaskCard_gig on Task {
    id
    title
    payRate
    location {
      city
      state
    }
  }
`;
// TaskCard only knows about gig display data

Law of Demeter Guidelines

• One dot (method call) is okay: object.method()
• Multiple dots is a code smell: object.property.property.method()
• Create wrapper methods instead of chaining
• Each module should only know about its direct collaborators
• Particularly important in GraphQL - don't query deep nested data you don't need

Exception: Fluent interfaces designed for chaining

In Elixir, the pipe operator and certain builder patterns (like Ecto) are designed for chaining:

# This is okay - designed for chaining
User.changeset(%{})
|> cast(attrs, [:email])
|> validate_required([:email])
|> unique_constraint(:email)

# This is okay - Ecto.Query builder pattern
from(u in User)
|> where([u], u.active == true)
|> join(:inner, [u], p in assoc(u, :profile))
|> select([u, p], {u, p})

# These patterns are explicitly designed for method chaining
# Each function returns a chainable structure

The key difference: fluent interfaces are designed for chaining as their primary API, whereas reaching through data structures (.worker.address.city) is accidental coupling.

3. Tell, Don't Ask

Tell objects what to do, don't ask for their data and do it yourself.

Why (Functional Context)

In Elixir, "Tell, Don't Ask" means delegating to the module that owns the data type . Instead of pulling data out of a structure and making decisions based on it, you pass the structure to the owning module and let it handle the logic.

This principle:

• Encapsulates business rules with the data they operate on
• Reduces coupling - callers don't need to know internal state or structure
• Improves cohesion - related logic lives together in the owning module
• Enables polymorphism - different implementations can handle the same "tell" differently

Think of it as: "Don't ask a struct for its fields and decide what to do - tell the module to do it."

Elixir Examples: Tell Don't Ask

# BAD - Asking for data and making decisions
def process_engagement(engagement) do
  if engagement.status == "pending" and engagement.worker_id != nil do
    attrs = %{status: "confirmed", confirmed_at: DateTime.utc_now()}
    Repo.update(Assignment.changeset(engagement, attrs))
  end
end
# We're asking about the engagement's state and deciding what to do

# GOOD - Delegate to the module that owns the Assignment struct
def process_engagement(engagement) do
  Assignment.confirm(engagement)
end

defmodule Assignment do
  def confirm(%{status: "pending", worker_id: worker_id} = engagement)
      when not is_nil(worker_id) do
    changeset = change(engagement, %{
      status: "confirmed",
      confirmed_at: DateTime.utc_now()
    })
    Repo.update(changeset)
  end

  def confirm(engagement), do: {:error, :invalid_state}
  # Assignment module knows its own business rules
  # Callers just "tell" it to confirm, don't "ask" about status
end

# BAD - Asking and deciding
def charge_gig(gig) do
  if gig.payment_type == "per_hour" do
    rate = gig.hourly_rate
    hours = gig.total_hours
    Money.multiply(rate, hours)
  else
    gig.fixed_amount
  end
end

# GOOD - Delegate to the module that owns the Task struct
def charge_gig(gig) do
  Task.total_charge(gig)
end

defmodule Task do
  def total_charge(%{payment_type: "per_hour", hourly_rate: rate,
total_hours: hours}) do
    Money.multiply(rate, hours)
  end

  def total_charge(%{payment_type: "fixed", fixed_amount: amount}) do
    amount
  end
  # Business logic lives with the data in the owning module
end

TypeScript Examples: Tell Don't Ask

// BAD - Asking for data
function renderTaskStatus(gig: Task) {
  let statusText: string;
  let statusColor: string;

  if (gig.status === 'active' && gig.workerCount > 0) {
    statusText = 'In Progress';
    statusColor = 'green';
  } else if (gig.status === 'active') {
    statusText = 'Waiting for Users';
    statusColor = 'yellow';
  } else {
    statusText = 'Completed';
    statusColor = 'gray';
  }

  return <Badge text={statusText} color={statusColor} />;
}

// GOOD - Tell the gig to provide display info
function renderTaskStatus(gig: Task) {
  const { text, color } = gig.getStatusDisplay();
  return <Badge text={text} color={color} />;
}

class Task {
  getStatusDisplay(): { text: string; color: string } {
    if (this.status === 'active' && this.workerCount > 0) {
      return { text: 'In Progress', color: 'green' };
    } else if (this.status === 'active') {
      return { text: 'Waiting for Users', color: 'yellow' };
    }
    return { text: 'Completed', color: 'gray' };
  }
}

Tell, Don't Ask Guidelines

• Push behavior into the module that owns the data type (Elixir) or object (TypeScript)
• Commands over queries (when possible)
• Modules/objects should protect their own invariants
• Reduces coupling - callers don't need to know internal state
• Particularly important in Command handlers - they tell, don't ask

Pattern

Command handlers follow "Tell, Don't Ask":

# Command tells the system what to do
%CreateTask{requester_id: id, title: "Landscaping"}
|> CreateTaskHandler.handle()
# Handler tells domain objects to execute
# Not: Handler asks domain for data and decides

4. Encapsulation

Hide internal implementation details. Expose minimal, stable interfaces.

Why: Encapsulation (Elixir Context)

• Enables change: You can change internal implementation without breaking callers
• Enforces invariants: Internal state can only change through controlled functions
• Improves testability: Test the public interface, not implementation details
• Reduces cognitive load: Callers only need to understand the public API
• Supports modularity: Clear boundaries between modules

In Elixir, encapsulation is achieved through:

• Module boundaries (private functions with defp)
• Opaque types (@opaque)
• Pattern matching guards
• Changesets for validation at boundaries
• Minimizing public API surface

Elixir Examples: Encapsulation

# BAD - Exposing internals
defmodule PaymentProcessor do
  defstruct [:stripe_client, :api_key, :retry_count]

  def process(processor, amount) do
    # Callers can access processor.stripe_client directly
    # Breaks if we change internal implementation
  end
end

# GOOD - Encapsulate internals
defmodule PaymentProcessor do
  @type t :: %__MODULE__{
    stripe_client: term(),
    api_key: String.t(),
    retry_count: integer()
  }

  @enforce_keys [:stripe_client, :api_key]
  defstruct [:stripe_client, :api_key, retry_count: 3]

  # Public API
  def new(api_key), do: %__MODULE__{
    stripe_client: Stripe.Client.new(api_key),
    api_key: api_key
  }

  def process(%__MODULE__{} = processor, amount) do
    # Internal implementation hidden
    do_process(processor, amount)
  end

  # Private implementation
  defp do_process(processor, amount) do
    # Can change internals without affecting callers
  end
end

# BAD - Ecto schema with map fields (no structure)
defmodule Task do
  schema "tasks" do
    field :data, :map  # Anything goes!
  end
end

# GOOD - Explicit fields (encapsulation via type system)
defmodule Task do
  schema "tasks" do
    field :title, :string
    field :description, :string
    field :pay_rate, Money.Ecto.Composite.Type
    field :status, Ecto.Enum, values: [:draft, :published, :active, :completed]
  end

  # Changesets enforce valid transitions
  def publish_changeset(gig) do
    gig
    |> change(%{status: :published})
    |> validate_required([:title, :description, :pay_rate])
  end
  # Can't publish without required fields (encapsulated business rule)
end

TypeScript Examples: Encapsulation

// BAD - Public mutable state
class ShoppingCart {
  public items: Item[] = [];  // Anyone can modify directly!

  public total(): number {
    return this.items.reduce((sum, item) => sum + item.price, 0);
  }
}

const cart = new ShoppingCart();
cart.items.push(invalidItem);  // Bypasses validation!

// GOOD - Encapsulated state
class ShoppingCart {
  private items: Item[] = [];  // Hidden implementation

  public addItem(item: Item): void {
    if (this.isValid(item)) {
      this.items.push(item);
    } else {
      throw new Error('Invalid item');
    }
  }

  public removeItem(itemId: string): void {
    this.items = this.items.filter(item => item.id !== itemId);
  }

  public getTotal(): number {
    return this.items.reduce((sum, item) => sum + item.price, 0);
  }

  public getItemCount(): number {
    return this.items.length;
  }

  private isValid(item: Item): boolean {
    return item.price > 0 && item.quantity > 0;
  }
  // All state changes go through controlled methods
}

// GOOD - React component encapsulation
function TaskCard({ gigRef }: Props) {
  // Encapsulate internal state
  const [expanded, setExpanded] = useState(false);
  const gig = useFragment(fragment, gigRef);

  // Private helpers
  const handleToggle = () => setExpanded(!expanded);

  // Public interface is just the props
  return (
    <Pressable onPress={handleToggle}>
      {/* Internal implementation */}
    </Pressable>
  );
}
// Parent components don't know about 'expanded' state
// Clean interface: pass gig data, get rendered card

Encapsulation Guidelines

• Make fields/functions private by default, public only when needed
• Use TypeScript private, protected modifiers
• Use Elixir module attributes (@) for internal data
• Elixir: prefix private functions with defp
• Validate inputs at boundaries (changesets, prop validation)
• Don't expose internal data structures
• Provide focused, minimal public APIs
• Change detection should be internal (don't expose dirty flags)

Patterns

• Ecto Changesets: Encapsulate validation and constraints
• GraphQL Types: Only expose fields needed by frontend
• Command Handlers: Encapsulate business rules
• Relay Fragments: Encapsulate data requirements in component

Application Checklist

Before implementing

• Can I compose instead of inherit? (Composition > Inheritance)
• Am I reaching through multiple objects? (Law of Demeter)
• Should this object do the work instead of me? (Tell, Don't Ask)
• Are internals hidden? (Encapsulation)

During implementation

• Use composition (pipes, hooks, interfaces)
• Call methods on direct collaborators only
• Push behavior to data owners
• Make fields/functions private by default
• Validate at boundaries

During code review

• Any deep inheritance hierarchies? (max 2-3 levels)
• Train wreck chains? (a.b.c.d())
• Logic operating on other objects' data?
• Public mutable state?

Red Flags

Composition Over Inheritance

• Deep class hierarchies (>3 levels)
• Can't extend because already inheriting
• Duplicating code because can't multi-inherit

Law of Demeter

• Multiple dots: user.profile.address.city
• GraphQL queries 5+ levels deep
• Functions taking many parameters to avoid chaining

Tell, Don't Ask

• Lots of getters used in if statements
• Business logic outside the entity
• Type checking with if/else (if type == 'x')

Encapsulation

• Public mutable fields
• Map/any types instead of structured data
• Callers modifying internal state directly
• No validation at boundaries

Integration with Existing Skills

Works with

• solid-principles: Particularly Single Responsibility and Dependency Inversion
• ecto-patterns: Changesets encapsulate validation
• cqrs-pattern: Commands tell, don't ask
• atomic-design-pattern: Components composed from atoms/molecules
• simplicity-principles: Simple composition over complex inheritance

Remember

Favor object composition over class inheritance (Gang of Four)

• Compose simple behaviors into complex ones
• Delegate to direct collaborators only
• Tell objects what to do, don't interrogate them
• Hide implementation details behind clean interfaces

Good design is about managing dependencies and protecting invariants - regardless of paradigm.