Personally, I like setting up a docker environment.

An example from a pet project:

docker-compose.yml

version: '3.3'

services:

  app:
    build:
      context: .
      dockerfile: Dockerfile
    working_dir: /app
    command: 'mix phx.server'
    ports:
      - '4000:4000'
    volumes:
      - '~/.bash-history:/.bash-history'
      - '.:/app:cached'
      - '/tmp:/tmp:delegated'
      - '.bashrc:/root/.bashrc:cached'
      - '/var/run/docker.sock:/var/run/docker.sock:cached'
      - '/var/tmp:/var/tmp:cached'
      - '/tmp:/tmp:cached'
    depends_on:
      - db

  db:
    image: postgres:11-alpine
    volumes:
      - 'pgdata:/var/lib/postgres/data'

volumes:
  pgdata:

Makefile

start:
	docker-compose up -d

stop:
	docker-compose down

build:
	docker-compose build app

shell:
	docker-compose run --rm app bash

install:
	docker-compose run --rm app mix deps.get

install-assets:
	docker-compose run --rm app bash -c "cd assets && npm install && node node_modules/webpack/bin/webpack.js --mode development"

compile:
	docker-compose run --rm app bash -c "mix do compile, phx.digest"

db-setup:
	docker-compose run --rm app mix ecto.setup

db-reset:
	docker-compose run --rm app mix ecto.reset

start-interactive:
	docker-compose run --rm --service-ports app iex -S mix phx.server

test:
	docker-compose run --rm app mix test

setup: build install install-assets compile db-setup

.PHONY: test

The difficulty lies in the details.

If you can assert correctness purely based on the query alone you could use a mock, but at least for SQL this is really hard as even simple changes in the order of building up a query are likely to make the logic of comparing a query fail, but could very well still result in the same proper results when handled by an actual db. Such differences defeat the purpose of a testsuite allowing you to refactor safely; Or you’d need to basically reimplement the db’s query engine. This is different to e.g. a mocked API (which is the subject of the blogpost), as the inputs there are likely to be way less complex to assert against than sql.

This is very normal, and is also the way that people test when using Postgres as a database.

I do not mean to hijack this topic, but has anybody have tried to incorporate pgtap into their workflow (with elixir projects) and reached any meaningful conclusion, successes or failures in doing so?