Skip to content

'From scratch to online in production' in a single day, with Django - Part 2

Quick summary of the previous part

In part 1 we saw a way to set up a Django project in a quite generic way, that can be used for any kind of project:

  • Dependencies are managed by Poetry
  • Settings that differ from an environment to another come from environment variables, and are populated from optional .env files in "local development" mode, powered by django-environ.
    SECRET_KEY = env.str("SECRET_KEY")
        "default": env.db_url("DATABASE_URL"),
    On top of this we also have a Python module for each type of environment - so we'll use the DJANGO_SETTINGS_MODULE environment variable to point to project.settings.development during local development , project.settings.heroku on Heroku, project.settings.test when running our test suite, etc.


Having such multiple settings files is a best practice I learned by reading the great book Two Scoops of Django - which itself cites this talk from Jacob Kaplan-Moss.

  • A src/ folder contains all the application code, within 2 top-level packages for our Python modules:
    • project, which contains the Django settings and the WSGI/ASGI HTTP entrypoints
    • apps for the Django apps - so each of them is free to have any name we want, without any risks of collisions with an existing 3rd-party package.

Next step: we have to organise the business logic inside our Django apps

Django has some built-in recommendations about how to structure a project.
Its applications concept for example is a good start to split the code into smaller units, and some basic rules like this one are good guidelines:


If there are 20+ models in a single app, think about ways to break it down into smaller apps, as it probably means your app is doing too much.
In practice, we like to lower this number to no more than five to ten models per app.

From Two Scoops of Django

Now, let's take a look at an aspect that may be a mystery for other developers who enter this new ecosystem they're not familiar with yet - like me a few years ago, when I switched from "PHP + Symfony ๐Ÿ˜" to "Python + Django ๐Ÿ" :

How to organise the business logic inside each Django app?

Me, as I was learning Django ๐Ÿ˜…

Well, I guess "the Django way" would probably be to follow the Active Record pattern and write most of the code for this in the Models themselves - as well as in their Managers (which are similar to "repositories" or "entity managers" when using the Data Mapper pattern)

However, in my own (subjective) case I generally find that doing so doesn't scale very well as the project grows, as we end up having all the business logic grouped in a bunch of huge classes.
Which is why, as I was learning Django, I was quite eager to find another way to structure my code... ๐Ÿค”

An efficient pattern for the business logic: simplicity that scales well ๐Ÿ‘Œ

When I started to learn Django I was lucky enough to stumble upon this Django Styleguide, published on GitHub by the software development company HackSoft.
And more specifically this part, where they explain "services" and "selectors":

My own adaptation of the "services" and "selectors" concept from HackSoft

As I'm mostly working with GraphQL in my day-to-day job, I opted for a terminology that rings a bell a bit more to my ears than "services" and "selectors": mutations and queries.
The former is a package that contains code that alter a database (adding, modifying or deleting data from it), while the latter is specialised in fetching data.

Here is how it looks like applied to my "Gin Rummy leaderboard" mini project:
(the parts of the tree that don't matter in this case are replaced with three dots)

โ”œโ”€โ”€ src/                                                                                                                                                                                                   
โ”‚     โ”œโ”€โ”€ apps/                                                                                                                                                                                              
โ”‚     โ”‚     โ”œโ”€โ”€ authentication/                                                                                                                                                                                
โ”‚     โ”‚     โ”‚     โ””โ”€โ”€ ...                                                                                                                                                                              
โ”‚     โ”‚     โ”œโ”€โ”€ gin_scoring/                                                                                                                                                                                   
โ”‚     โ”‚     โ”‚     โ”œโ”€โ”€ domain/                                                                                                                                                                                    
โ”‚     โ”‚     โ”‚     โ”‚     โ”œโ”€โ”€ mutations/ # (1)
โ”‚     โ”‚     โ”‚     โ”‚     โ”‚     โ”œโ”€โ”€
โ”‚     โ”‚     โ”‚     โ”‚     โ”‚     โ””โ”€โ”€
โ”‚     โ”‚     โ”‚     โ”‚     โ”œโ”€โ”€ queries/ # (2)
โ”‚     โ”‚     โ”‚     โ”‚     โ”‚     โ”œโ”€โ”€
โ”‚     โ”‚     โ”‚     โ”‚     โ”‚     โ”œโ”€โ”€
โ”‚     โ”‚     โ”‚     โ”‚     โ”‚     โ”œโ”€โ”€
โ”‚     โ”‚     โ”‚     โ”‚     โ”‚     โ””โ”€โ”€
โ”‚     โ”‚     โ”‚     โ”‚     โ””โ”€โ”€
โ”‚     โ”‚     โ”‚     โ”œโ”€โ”€ jinja2/
โ”‚     โ”‚     โ”‚     โ”‚     โ””โ”€โ”€ ...
โ”‚     โ”‚     โ”‚     โ”œโ”€โ”€ migrations/
โ”‚     โ”‚     โ”‚     โ”‚     โ””โ”€โ”€ ...
โ”‚     โ”‚     โ”‚     โ”œโ”€โ”€
โ”‚     โ”‚     โ”‚     โ”œโ”€โ”€
โ”‚     โ”‚     โ”‚     โ”œโ”€โ”€
โ”‚     โ”‚     โ”‚     โ”œโ”€โ”€
โ”‚     โ”‚     โ”‚     โ”œโ”€โ”€
โ”‚     โ”‚     โ”‚     โ”œโ”€โ”€
โ”‚     โ”‚     โ”‚     โ””โ”€โ”€
โ”‚     โ”‚     โ””โ”€โ”€
โ”‚     โ”œโ”€โ”€ project/
โ”‚     โ”‚     โ”œโ”€โ”€ settings/
โ”‚     โ”‚     โ”‚     โ””โ”€โ”€ ...
โ”‚     โ”‚     โ”œโ”€โ”€ # (3)
โ”‚     โ”‚     โ”œโ”€โ”€
โ”‚     โ”‚     โ”œโ”€โ”€
โ”‚     โ”‚     โ””โ”€โ”€
โ”‚     โ””โ”€โ”€*
โ”œโ”€โ”€ tests/
โ”‚     โ””โ”€โ”€ ...
โ”œโ”€โ”€ docker-compose.yml
โ”œโ”€โ”€ Dockerfile
โ”œโ”€โ”€ Makefile
โ”œโ”€โ”€ poetry.lock
โ””โ”€โ”€ pyproject.toml

  1. In this package we have the code that alters data
  2. In this package we have the code that fetches data
  3., and were created by the startproject command. is where I quickly configure Jinja, following the Django documentation:

We can see that for each Django app we have 2 sub-packages:

  • domain.mutations is where the code that alter data lives
  • domain.queries is where the code that fetch data lives

They're both structured the same way, following these principles from HackSoft's styleguide:

  • Each of their module exposes only one public function - and optionally some types if it has to, in order to describe the shape of its input and/or output.
  • Each of these one-per-module-functions only accepts keyword arguments.
  • These functions' signatures should be fully "type hinted".
  • Each module is free to use as many private functions it needs to achieve the job described by its single public function.
  • For the mutations, these functions' name should start with a verb, since they're the reflection of a business logic action
  • Each module's name is prefixed with an underscore, to emphasise that it should not be imported directly
  • The file is in charge of exposing the public function of each module to the "outside world" - i.e. the Python code that doesn't live in the same package.


The goal of the last 2 points is to avoid this kind of imports, where we have to repeat the same name twice - once of the module name and once for the function itself:

from .domain.mutations.save_game_result import save_game_result
So instead of this repetition we can simply do:
from .domain.mutations import save_game_result

A concrete example, with the Gin Rummy leaderboard app

Ok, enough theory - let's see how that works in the context of this mini project! ๐Ÿ™‚

We have a single Django model, that looks like this:

class GameResult(models.Model):
   player_north_name = models.CharField(max_length=50)
   player_south_name = models.CharField(max_length=50)
   outcome = models.CharField(max_length=10, choices=[(outcome, outcome) for outcome in GAME_OUTCOME.__args__])  # type: ignore
   # These 2 ones can be `null` when the outcome is `draw`:
   winner_name = models.CharField(max_length=50, null=True)
   deadwood_value = models.PositiveSmallIntegerField(null=True)
   # Computed from the previous `outcome` and `deadwood_value` fields:
   winner_score = models.PositiveSmallIntegerField(null=True)

   created_at = models.DateTimeField(

   def is_draw(self) -> bool:
       return self.outcome == "draw"

   def loser_name(self) -> str | None:
       if self.is_draw:
          return None
       return [name for name in (self.player_north_name, self.player_south_name) if name != self.winner_name][0]

   def __str__(self) -> str:
       return f"{self.player_north_name.title()} vs {self.player_south_name.title()}, on {self.created_at.strftime('%a %d %b at %H:%M')}"

The domain.mutations package of our Django app

For this minimalist project we need only one mutation, which is triggered when the user submits the "New game result" HTML form:

# file: src/apps/gin_scoring/domain/
from ...domain.gin_rummy import GAME_OUTCOME, calculate_round_score
from ...models import GameResult

def save_game_result(
    *, # (1)
    player_north_name: str,
    player_south_name: str,
    outcome: GAME_OUTCOME, # (2)
    winner_name: str | None,
    deadwood_value: int,
) -> GameResult:
    is_draw = outcome == "draw"

    winner_score = None
    if is_draw:
        winner_name = None
        winner_score = calculate_round_score(game_outcome=outcome, deadwood_value=deadwood_value)

    game_result_model = GameResult(
        # (3)
    # (4)

    return game_result_model

  1. We force this function to be used only with the "keyword arguments" syntax
  2. GAME_OUTCOME is just a literal type, described later on in this same article ๐Ÿ™‚
  3. To my knowledge there's no equivalent of the Shorthand property names of ES2015 in Python - which I guess must be on purpose, since readability is almost always the top priority of the language?
    For that reason we have to repeat the [field name]=[arg name] pattern, but in my opinion it's not really an issue ๐Ÿ™‚
  4. The model also has a created_at field, but it will automatically be set by the Django ORM since we've used when we defined the models.DateTimeField field ๐Ÿ™‚

And then, we expose that function to the rest of the Python code:

# file: src/apps/gin_scoring/domain/mutations/
from ._save_game_result import save_game_result

All we have to do now is to use that mutation from a Django view.
There are several ways to do this, but here is an example:

# file: src/apps/gin_scoring/
from .domain.mutations import save_game_result

# ...

def post_game_result(request: HttpRequest) -> HttpResponse:
        game_result_payload = GameResultPayload(**request.POST.dict())
    except pydantic.ValidationError:
        return HttpResponseBadRequest()

        # (1)

    return redirect("index")

  1. Note that we could also have opted for a more concise way to transfer data from the "validation and normalisation" data structure to the mutation: ๐Ÿ™‚
    save_game_result( **game_result_payload.dict() )

Validating the input of our Django views

There are multiple ways to validate and normalise the input of our Django views, before passing its data to the "domain" layer.
In this case I chose to use Pydantic.

# file: src/apps/gin_scoring/
from typing import Any

import pydantic

from .domain.gin_rummy import GAME_OUTCOME
from .helpers import normalize_player_name

class GameResultPayload(pydantic.BaseModel):
    player_north_name: str
    player_south_name: str
    outcome: GAME_OUTCOME
    winner_name: str | None
    deadwood_value: int

    def normalize_player_names(cls, values: dict[str, Any]):
        # In order to have consistent recording when players "Rachel" and "Olivier" add a game result, whether
        # "Rachel" is "north" and "Olivier" is "south" or vice-versa, we sort their names alphabetically
        # and then always set the "north" player to the first one and the "south" one to the second one:
        player_north_name, player_south_name = sorted(
        values["player_north_name"] = player_north_name
        values["player_south_name"] = player_south_name
        return values

    def validate_winner_name(cls, v: str, values: dict[str, Any]) -> str | None:
        is_draw = values["outcome"] == "draw"

        if is_draw:
            return None  # No winner name for "draw" games

        if not v:
            raise ValueError(f"non-draw games must have a winner name")
        winner_name = normalize_player_name(v)
        player_names = (values["player_north_name"], values["player_south_name"])
        if winner_name not in player_names:
            raise ValueError(f"winner name {v} is not part of the players' names '{','.join(player_names)}'")

        return winner_name

The domain.queries package of our Django app

For this simple app we need only 3 queries:

  • One to get the global "Hall of fame", where we determine the ranking of players based on all the games played so far
  • One to get the monthly "Hall of fame", which does the same but with a distinct ranking for each month
  • One that returns exhaustive data for the last 10 games that were played

Let's take a look at the second one, for example:

# file: src/apps/gin_scoring/domain/queries/
from collections import defaultdict
from datetime import datetime
from typing import NamedTuple

from django.db.models import Count, Sum
from django.db.models.functions import TruncMonth

from ...models import GameResult  # (1)

class HallOfFameMonthResult(NamedTuple): # (2)
    month: datetime
    winner_name: str
    game_counts: int
    win_counts: int
    score_delta: int

def hall_of_fame_monthly() -> list[HallOfFameMonthResult]:
    # โš ๏ธ Probably not the very best way to achieve this...
    # But this is a project I gave myself one single day to build,
    # so that will do the job ๐Ÿ˜…

    # @link
    win_counts = Count("winner_score")
    total_score = Sum("winner_score")

    raw_results = (
        .annotate(month=TruncMonth("created_at")) # (3)
        .values("month", "winner_name")
        .annotate(win_counts=win_counts, total_score=total_score)
        # Each won round is worth 25 points:
        .annotate(grand_total=(win_counts * 25) + total_score)
        .order_by("-month", "-grand_total")
    raw_results_per_month: dict[datetime, list[dict]] = defaultdict(list)
    for raw_result in raw_results:

    returned_results: list[HallOfFameMonthResult] = []
    for month, month_results in raw_results_per_month.items():
        winner_result = month_results[0]
        winner_grand_total = winner_result["grand_total"] or 0
        second_best_grand_total = 0 if len(month_results) < 2 else (month_results[1]["grand_total"] or 0)
        games_count = sum([res["win_counts"] for res in month_results])

                score_delta=winner_grand_total - second_best_grand_total,

    return returned_results

  1. I like using relative imports for things we import that live in the same Django app.
    In all other cases I'd use absolute imports.
  2. We're not returning ActiveRecord items directly from a Django QuerySet there, but aggregated results.
    There are several data structures we can use in Python for that kind of "value objects", but I generally opt for typing.NamedTuple
  3. This is where we group the database rows by month
  4. Yes, I could probably have used itertools.groupby instead ^_^


Note that we could have split this into several functions - in which case they would all be private functions (their name would start with an underscore), and only the "domain" one would be public ๐Ÿ™‚

And similarly, the file is in charge of exposing only what the rest of the Python

# file: src/apps/gin_scoring/domain/queries/
from ._hall_of_fame import hall_of_fame
from ._hall_of_fame_monthly import (
    HallOfFameMonthResult, # (1)
from ._last_game_results import last_game_results # (2)

  1. Sometimes it's useful to not only export the one public function of the module, but also a dedicated type it's using for its input or output - so other Python modules can also use type hints when interacting with the "domain" layer
  2. But most of the time, all we need is to expose the public function of the module ๐Ÿ˜„

But what about the domain logic that is neither a mutation nor a query?

We still have to put somewhere some parts of the domain don't fall in either categories. ๐Ÿค”
For example:

  • Constants, enums, literal types...
  • Data structures describing some aspects of the business logic, that can be used by both mutations and queries.
  • Various forms of "memory-only" computations.

Well... In my case, I find that in each Django app the domain package itself is a very good place to welcome these! ๐Ÿ™‚

For this mini project, for example, I chose to have a single Python module (apps.gin_scoring.domain.gin_rummy) to store some "business-logic-related" stuff that is specific to the Gin Rummy game , and that are neither a mutation nor a query:

# file: src/apps/gin_scoring/domain/
from typing import Literal

# Possible outcomes of a Gin Rummy game: # (1)
GAME_OUTCOME = Literal["knock", "gin", "big_gin", "undercut", "draw"]

def calculate_round_score(*, game_outcome: GAME_OUTCOME, deadwood_value: int) -> int:
    # @link
    match game_outcome:
        case "draw":
            return 0
        case "knock":
            return deadwood_value
        case "gin":
            return 25 + deadwood_value
        case "big_gin":
            return 31 + deadwood_value
        case "undercut":
            return 15 + deadwood_value
        case _:
            raise ValueError(f"Invalid game outcome value '{game_outcome}'")

  1. More on that Literal type below ๐Ÿ‘‡

We can see indeed that calculating the score of a round, depending on its outcome and the value of its deadwood, is neither a mutation nor a query: it's just a standalone computation, that does not depend on anything we would have in a database.

And the same goes for enumerating the possible outcomes of a Gin Rummy game.

Literals or Enums?

To express the outcome of a Gin Rummy game I could of course have used a Python enum instead:

class GameOutcome(enum.Enum):
    KNOCK = "knock" # or ``
    GIN = "gin"
    BIG_GIN = "big_gin"
    UNDERCUT = "undercut"
    DRAW = "draw"
As for me, I must admit that I have no strict rules when I have to choose between one or another way to describe that kind of data ๐Ÿคท

I was mostly using Enums until a few years ago, but as I was using TypeScript more and more I realised that I really liked using literal types there - the TypeScript equivalent of that Python Literal would be:

export type GAME_OUTCOME = "knock" | "gin" | "big_gin" | "undercut" | "draw"

I appreciate the concision of literals, and tend to use them when I have the feeling that having such "literal values" spread in the code wouldn't cause any issue later on if I have to change their values ๐Ÿ™‚
Enums are certainly easier to handle in case of refactorings, but so far I've never come across a case where during a refactoring I regretted having opted for a literal rather than an Enum - fingers crossed, it won't be the case anytime soon! ๐Ÿ˜„ ๐Ÿคž

And that's it! ๐Ÿ™‚

As we can see the pattern is very simple to implement, and its few principles are a very good guideline for developers when they have to add some code.

  • Is it code that creates, updates or deletes data in a database?
    ๐Ÿ‘‰ Let's create a new module in the domain.mutations package of the related Django app, that will expose one single "kwargs-only" function - its name will start with a verb.
  • Is it code that reads data from a database?
    ๐Ÿ‘‰ Let's create a new module in the domain.queries package of the related Django app, that will expose one single "kwargs-only" function.
  • Is it code that expresses the business logic but neither alters nor reads data from a database?
    ๐Ÿ‘‰ Let's put that in a module of the domain package of the related Django app.

The beauty of that pattern is that it really scales very well, despite its simplicity: my former teams and I used it for years on ever-growing code bases without having ever faced a case where the pattern would show a limitation. ๐Ÿ‘Œ


The 3rd (and last) article of this series will be a quick one, about how I hosted this app - for free โœŒ - at the end of that single day of work.


Thank you so much HackSoft for your Django Styleguide! ๐Ÿ’š

I would also like to thank Audrey and Daniel Roy Greenfeld for their book Two Scoops Of Django, which was a very helpful resource for me when I started to learn Django and tried to see what the best practices could be in this ecosystem - definitely worth the purchase! ๐Ÿ™‚

And thanks again to my friend Yann - - for his careful review ๐Ÿ‘“ and useful feedback on this article ๐Ÿค—

Part 3

UPDATE: Part 3 is online ๐Ÿ™‚