FP in TypeScript -- Functional Software Crafter Skill

Cross-references: fp-principles | fp-domain-modeling | pbt-typescript

When to Choose TypeScript

• Best for: teams already on the Node/browser ecosystem wanting FP DISCIPLINE without a new runtime | full-stack projects sharing domain types between client and server | the most common first encounter with FP for many engineers
• Not ideal for: teams wanting the compiler to ENFORCE purity (TS purity is a convention, any/as can always escape it) | CPU-bound work needing a stricter type system's exhaustiveness guarantees

[STARTER] Quick Setup

mkdir order-service && cd order-service && npm init -y
npm install typescript fp-ts --save
npm install --save-dev fast-check vitest
npx tsc --init --strict

Test runner: Vitest or Jest. "strict": true in tsconfig.json is load-bearing -- without it, null/ undefined silently widen every type and most patterns below lose their guarantee.

[STARTER] Type System for Domain Modeling

Choice Types (Discriminated Unions)

type PaymentMethod =
  | { readonly _tag: "CreditCard"; readonly cardNumber: string; readonly expiryDate: string }
  | { readonly _tag: "BankTransfer"; readonly accountNumber: string }
  | { readonly _tag: "Cash" };

The shared _tag field is the discriminant: a switch on _tag narrows the union per-branch, and with "strict": true a missing case is a compile error via the never exhaustiveness check below.

Record Types and Branded Domain Wrappers

interface Customer {
  readonly customerId: CustomerId;
  readonly customerName: CustomerName;
  readonly customerEmail: EmailAddress;
}

type OrderId = number & { readonly __brand: "OrderId" };
type EmailAddress = string & { readonly __brand: "EmailAddress" };

TypeScript has no true newtype -- a branded intersection type is the idiomatic substitute: OrderId is structurally a number at runtime (zero cost) but the __brand phantom field stops the compiler from accepting a bare number where an OrderId is expected.

[INTERMEDIATE] Smart Constructors

import { Either, left, right } from "fp-ts/Either";

function mkEmailAddress(raw: string): Either<ValidationError, EmailAddress> {
  return raw.includes("@")
    ? right(raw as EmailAddress)
    : left({ _tag: "InvalidEmail", raw });
}

The brand cast (as EmailAddress) happens ONLY inside the smart constructor -- callers never cast directly, so every EmailAddress in scope has passed mkEmailAddress.

[INTERMEDIATE] Composition Style

pipe for Left-to-Right Composition

import { pipe } from "fp-ts/function";
import * as E from "fp-ts/Either";

const placeOrder = (raw: RawOrder): Either<OrderError, Confirmation> =>
  pipe(raw, validateOrder, E.chain(priceOrder), E.chain(confirmOrder));

E.chain is Either's bind -- short-circuits on Left, threads the Right value forward.

Effect (the newer alternative to fp-ts)

import { Effect } from "effect";

const placeOrder = (raw: RawOrder): Effect.Effect<Confirmation, OrderError> =>
  Effect.gen(function* () {
    const validated = yield* validateOrder(raw);
    const priced = yield* priceOrder(validated);
    return yield* confirmOrder(priced);
  });

Generator-based Effect.gen reads imperatively while remaining a pure DESCRIPTION of the computation until Effect.runPromise/runSync actually executes it -- the same "build a value, run it at the edge" discipline Haskell's IO enforces at the type level.

Accumulating Validation Errors

import { validate } from "fp-ts/Apply";

const validateCustomer = (raw: RawCustomer) =>
  pipe(
    { name: validateName(raw.name), email: validateEmail(raw.email) },
    ({ name, email }) => E.Do.pipe(E.apS("name", name), E.apS("email", email)),
  );

apS/Applicative validation collects ALL failures via Either's Semigroup on the error side, unlike chain which stops at the first.

[INTERMEDIATE] Effect Management

Plain TypeScript has no purity enforcement -- Effect's own Effect<A, E, R> type is the closest analogue to Haskell's IO: a VALUE describing a computation, not the computation itself, so a function returning Effect<Order, never, never> is provably side-effect-free by its type alone (no ambient R requirement, no E to fail with) in a way a bare Order return never was.

// Pure domain logic -- ordinary function, no Effect wrapper needed
const calculateDiscount = (order: Order): Discount =>
  order.lines.length > 10 ? 0.1 : 0.0;

// Effectful boundary -- Effect<A, E, R> makes the DB dependency explicit in R
const saveOrder = (order: Order): Effect.Effect<void, DbError, OrderRepository> =>
  Effect.gen(function* () {
    const repo = yield* OrderRepository;
    yield* repo.save(order);
  });

[ADVANCED] Ports as Interfaces, Adapters as Layers (Hexagonal Architecture)

class OrderRepository extends Context.Tag("OrderRepository")<
  OrderRepository,
  { readonly findOrder: (id: OrderId) => Effect.Effect<Option<Order>>; readonly save: (o: Order) => Effect.Effect<void> }
>() {}

const PostgresOrderRepository = Layer.succeed(OrderRepository, {
  findOrder: (id) => Effect.tryPromise(() => db.query("SELECT * FROM orders WHERE id = $1", [id])),
  save: (order) => Effect.tryPromise(() => db.insert("orders", order)),
});

Context.Tag is the port; Layer provides the adapter. Tests substitute a Layer.succeed backed by an in-memory map, with no framework or DI container needed.

[INTERMEDIATE] Testing

Frameworks: fast-check (property-based, integrates with Vitest/Jest) | Vitest (fast, ESM-native) | Effect's own @effect/vitest for testing Effect values directly. See pbt-typescript for detailed PBT patterns.

Property Test Example

import fc from "fast-check";
import { test } from "vitest";

test("round-trips through serialization", () =>
  fc.assert(fc.property(orderArbitrary, (order) => deserialize(serialize(order)) === order)));

test("validated orders always have a positive total", () =>
  fc.assert(
    fc.property(rawOrderArbitrary, (raw) => {
      const result = validateOrder(raw);
      return E.isLeft(result) || result.right.total > 0;
    }),
  ));

Custom Arbitrary

const emailArbitrary: fc.Arbitrary<EmailAddress> = fc
  .tuple(fc.stringMatching(/^[a-z]{1,10}$/), fc.stringMatching(/^[a-z]{1,8}$/))
  .map(([user, domain]) => `${user}@${domain}.com` as EmailAddress);

[ADVANCED] Idiomatic Patterns

Exhaustiveness Checking via never

function describe(method: PaymentMethod): string {
  switch (method._tag) {
    case "CreditCard": return `Card ending ${method.cardNumber.slice(-4)}`;
    case "BankTransfer": return `Transfer from ${method.accountNumber}`;
    case "Cash": return "Cash";
    default: {
      const _exhaustive: never = method; // compile error if a case is missing
      return _exhaustive;
    }
  }
}

Option Instead of null/undefined

import { Option, some, none, fromNullable } from "fp-ts/Option";

const findCustomer = (id: CustomerId): Option<Customer> =>
  fromNullable(database.get(id));

Composes with map/chain the same way Either does, avoiding a null-check tree at every call site.

Maturity and Adoption

• fp-ts is in low-maintenance mode: Effect is the actively developed successor (built by overlapping authors) and is where new projects should start; fp-ts remains fine for existing codebases.
• Purity is a convention, not enforced: nothing stops a "pure" function from calling Date.now() or Math.random() inline; code review and the effect-typing discipline above are the only guard.
• Effect's learning curve is steep: Effect<A, E, R>'s three type parameters and generator syntax take longer to internalize than fp-ts's Either/TaskEither alone; introduce incrementally.

Common Pitfalls

1. as casts defeating brands: a branded type (OrderId) is only as strong as the discipline never to cast an arbitrary number into it outside the smart constructor.
2. Mixing Promise and Effect/TaskEither: pick one effect representation per boundary; interop (Effect.tryPromise, TE.tryCatch) at the edge, never scattered through domain logic.
3. interface merging surprises: TypeScript's structural typing means two unrelated interfaces with the same shape are interchangeable; branded types (above) are the fix, not a naming convention alone.
4. Non-exhaustive switch without strict: the never exhaustiveness check in Idiomatic Patterns only fires under "strict": true; without it a missing case silently falls through.