Services + svcs Dependency Injection

This project separates Django's framework concerns from business logic using a plain service layer, wired with the svcs service locator. The result:

• Views are one-liners. They pull a wired service and call a method.
• Services contain the business logic. They take repositories (and other services) via __init__, call methods on them, and return DTOs.
• Services never import Django ORM or models. Every test can run without a database.
• One registry, one get[T]() helper. The same call works in views, tasks, commands, anywhere.

Why svcs Instead of Module-Level Singletons or a Custom Container

• svcs is a tiny, typed, well-maintained service locator — no metaclasses, no decorators, no framework coupling.
• Factories are lazy: a service is constructed only when something asks for it.
• Generic get[T](type[T]) -> T preserves types through IDE/type-checker inference.
• Swapping an implementation in tests is a one-line factory override.
• No import-order gymnastics: the registry is populated once at startup and then used by name.

The Registry — src/project/services.py

import svcs

from products.repositories.product import ProductRepository
from products.services.product import ProductService
from orders.repositories.order import OrderRepository
from orders.services.order import OrderService

registry = svcs.Registry()

# --- Repositories ---------------------------------------------------------
registry.register_factory(ProductRepository, ProductRepository)
registry.register_factory(OrderRepository, OrderRepository)


# --- Services (factories pull repos from the container) ------------------
def _product_service_factory(container: svcs.Container) -> ProductService:
    repo = container.get(ProductRepository)
    return ProductService(repo)


def _order_service_factory(container: svcs.Container) -> OrderService:
    repo = container.get(OrderRepository)
    product_repo = container.get(ProductRepository)
    return OrderService(repo, product_repo)


registry.register_factory(ProductService, _product_service_factory)
registry.register_factory(OrderService, _order_service_factory)


def get[T](service_type: type[T]) -> T:
    """Resolve a service from the registry. Works anywhere — views, tasks, commands, tests."""
    return svcs.Container(registry).get(service_type)

Patterns to follow:

• Repositories register themselves. Use register_factory(Repo, Repo) — the class is its own factory because repositories take no constructor arguments.
• Services register via a named factory. _<entity>_service_factory(container) resolves every dependency from the container and hands it to the service's __init__. No hidden imports, no module-level singletons.
• Register in dependency order. Repos before services, lower-level services before higher-level ones. svcs doesn't enforce this, but it keeps the file readable.
• One registry per project. Don't create ad-hoc registries — everything goes through project.services.registry.

Writing a Service

File: src/<app>/services/<entity>.py

from decimal import Decimal
from typing import List

from ..dtos.product import ProductDTO
from ..repositories.product import ProductRepository


class ProductService:
    def __init__(self, repo: ProductRepository):
        self.repo = repo

    def create_product(self, name: str, price: Decimal, stock: int) -> ProductDTO:
        return self.repo.create(name=name, price=price, stock=stock)

    def get_product(self, product_id: str) -> ProductDTO:
        return self.repo.get_by_id(product_id)

    def list_products(self) -> List[ProductDTO]:
        return self.repo.list_all()

Rules:

• Dependencies come in through __init__. The service never instantiates its own repositories or services. If a service needs another service, pass it in.
• Zero ORM. No .objects, no F() / Q(), no model imports, no select_related. All database access goes through a repository.
• Every public method returns a DTO or list[DTO]. Never a model instance, never a queryset.
• ID arguments are str. See the prefixed-ulids skill.
• Business rules live here. Validation, orchestration across repositories, invariant checks, error raising — all of it.
• Services are stateless. They hold references to their dependencies and nothing else. No caches, no counters, no module-level state.
• Raise plain exceptions. Use ValueError, PermissionError, domain-specific exceptions — not Http404 or anything Django-flavored. The view layer turns them into HTTP responses.

Cross-Entity Logic

When a service method touches more than one aggregate — e.g. creating an order that decrements product stock — inject both repositories and orchestrate them. Example from OrderService:

class OrderService:
    def __init__(self, repo: OrderRepository, product_repo: ProductRepository):
        self.repo = repo
        self.product_repo = product_repo

    def create_order(self, items: List[Dict[str, Any]]) -> OrderDTO:
        for item in items:
            product = self.product_repo.get_by_id(item["product_id"])
            if product.stock < item["quantity"]:
                raise ValueError(
                    f"Insufficient stock for product {product.name}: "
                    f"requested {item['quantity']}, available {product.stock}"
                )

        order = self.repo.create(items=items)

        for item in items:
            self.product_repo.decrement_stock(item["product_id"], item["quantity"])

        return order

Notes:

• The service orchestrates two repositories but touches zero ORM.
• If a multi-repo write needs atomicity, wrap it in with transaction.atomic(): — that's one of the very few django.db imports allowed in a service.
• Never call another service from inside a service unless that service is explicitly injected. No hidden get(...) calls inside service methods.

Resolving a Service

From a view (django-ninja)

from project.services import get
from project.types import AuthedRequest


@products_router.post("/", response={201: ProductDTO})
def create_product(request: AuthedRequest, payload: CreateProductIn):
    service = get(ProductService)
    return Status(201, service.create_product(**payload.dict()))

Annotating request as AuthedRequest (defined in src/project/types.py) makes the auth contract explicit and narrows request.user to a guaranteed-authenticated Django User. This is a typing contract, not runtime enforcement — auth is still expected to be wired via middleware or ninja's auth= parameter.

From a Celery task

from celery import shared_task

from project.services import get
from products.services.product import ProductService


@shared_task
def reprice_product(product_id: str, new_price: str) -> None:
    service = get(ProductService)
    service.update_price(product_id, Decimal(new_price))

From a management command

from django.core.management.base import BaseCommand

from project.services import get
from products.services.product import ProductService


class Command(BaseCommand):
    def handle(self, *args, **options):
        service = get(ProductService)
        for dto in service.list_products():
            self.stdout.write(dto.name)

The same get() call works in all three contexts because the registry is global and the container is cheap to construct.

Testing

Services are tested without a database. Pass in a MagicMock for each repository, configure its return values, and assert on the service's behavior.

from decimal import Decimal
from unittest.mock import MagicMock

import pytest

from products.dtos.product import ProductDTO
from products.services.product import ProductService


def test_create_product_delegates_to_repo():
    repo = MagicMock()
    expected = ProductDTO(id="prd_fake", name="Widget", price=Decimal("9.99"), stock=5)
    repo.create.return_value = expected

    service = ProductService(repo)
    result = service.create_product(name="Widget", price=Decimal("9.99"), stock=5)

    assert result is expected
    repo.create.assert_called_once_with(name="Widget", price=Decimal("9.99"), stock=5)


def test_create_order_rejects_insufficient_stock():
    order_repo = MagicMock()
    product_repo = MagicMock()
    product_repo.get_by_id.return_value = ProductDTO(
        id="prd_fake", name="Widget", price=Decimal("9.99"), stock=1
    )

    service = OrderService(order_repo, product_repo)

    with pytest.raises(ValueError, match="Insufficient stock"):
        service.create_order(items=[{"product_id": "prd_fake", "quantity": 5}])

    order_repo.create.assert_not_called()

This is the most important test layer — it proves the business logic is correct independently of Django, migrations, fixtures, or the database. If a service's tests need @pytest.mark.django_db, something has leaked: find the ORM call and push it back into a repository.

Overriding a Service in Tests

For integration tests that go through the API, override a factory to substitute a fake or a stub:

from project.services import registry
from products.services.product import ProductService


@pytest.fixture
def fake_product_service():
    fake = MagicMock(spec=ProductService)
    registry.register_factory(ProductService, lambda _: fake)
    yield fake
    # svcs re-registers on next call; reset to the real factory if other tests need it.

Common Mistakes

• Importing models in a service. If you see from app.models import X, the service is doing ORM work. Move it to the repository.
• Calling SomeRepository() inside a service method. Inject it via __init__ and hold the reference.
• Returning querysets or model instances from a service. Always return DTOs.
• Putting business logic in the view. The view should only decode input, call get(SomeService).method(...), and pass the result back.
• Registering a service with register_factory(Service, Service). That only works for repositories because they take no arguments. Services need a factory that resolves their dependencies.
• Reaching into request.user from the service. Pass the caller's identity as an explicit argument (user_id: str) so the service stays framework-agnostic.

Verify

• Every service under src/<app>/services/ has an __init__ that takes its dependencies explicitly.
• No file under src/<app>/services/ imports from django.db.models, <app>.models, or uses .objects.
• Every service in the project is registered in src/project/services.py with a factory that resolves its dependencies from the container.
• Service tests do not use @pytest.mark.django_db.

uv run ruff check src
uv run pyrefly check src
uv run pytest