A utility for defining eminently readable parameterized (or example-based) tests in Elixir's ExUnit, inspired by example tests in Cucumber.
Parameterized tests let you define variables along a number of dimensions
and re-run the same test body (including all setup) for each
combination of variables.
A simple example:
setup context do
# context.permissions gets set by the param_test below
permissions = Map.get(context, :permissions, nil)
user = AccountsFixtures.user_fixture(permissions: permissions)
%{user: user}
end
param_test "users with editor permissions or better can edit posts",
"""
| permissions | can_edit? | description |
|-------------|-----------|---------------------------------|
| :admin | true | Admins have max permissions |
| :editor | true | Editors can edit (of course!) |
| :viewer | false | Viewers are read-only |
| nil | false | Anonymous viewers are read-only |
""",
%{user: user, permissions: permissions, can_edit?: can_edit?} do
assert Posts.can_edit?(user) == can_edit?, "#{permissions} permissions should grant edit rights"
endThat test will run 4 times, with the variables from from the table being applied
to the test's context each time (and therefore being made
available to the setup handler). These variables are:
:permissions:can_edit?- the special
:descriptionvariable (see About test names, and improving debuggability for how this is used)
Thus, under the hood this generates four unique tests, equivalent to doing something like this:
setup context do
permissions = Map.get(context, :permissions, nil)
user = AccountsFixtures.user_fixture{permissions: permissions}
%{user: user}
end
for {permissions, can_edit?, description} <- [
{:admin, true, "Admins have max permissions"},
{:editor, true, "Editors can edit (of course!)"},
{:viewer, false, "Viewers are read-only"},
{nil, false, "Anonymous viewers are read-only"}
] do
@permissions permissions
@can_edit? can_edit?
@description description
@tag permissions: @permissions
@tag can_edit?: @can_edit?
@tag description: @description
test "users with at least editor permissions can edit posts — #{@description}", %{user: user} do
assert Posts.can_edit?(user) == @can_edit?
end
end
endAs you can see, even with only 3 variables (just 2 that impact the test semantics!),
the for comprehension comes with a lot of boilerplate. But the param_test
macro supports an arbitrary number of variables, so you can describe complex
business rules like "users get free shipping if they spend more than $100,
or if they buy socks, or if they have the right coupon code":
param_test "grants free shipping based on the marketing site's stated policy",
"""
| spending_by_category | coupon | ships_free? | description |
|-------------------------------|-------------|-------------|------------------|
| %{shoes: 19_99, pants: 29_99} | | false | Spent too little |
| %{shoes: 59_99, pants: 49_99} | | true | Spent over $100 |
| %{socks: 10_99} | | true | Socks ship free |
| %{pants: 1_99} | "FREE_SHIP" | true | Correct coupon |
| %{pants: 1_99} | "FOO" | false | Incorrect coupon |
""",
%{
spending_by_category: spending_by_category,
coupon: coupon,
ships_free?: ships_free?
} do
shipping_cost = ShippingCalculator.calculate(spending_by_category, coupon)
if ships_free? do
assert shipping_cost == 0
else
assert shipping_cost > 0
end
endThe package also provides a second macro, param_feature, which wraps
Wallaby's feature tests the same way param_test wraps ExUnit's test.
(While you can use the plain param_test macro in a test module that
contains use Wallaby.Feature, doing so will break some Wallaby features
including screenshot generation on failure.)
Parameterized testing reduces toil associated with writing tests that cover a wide variety of different example cases. It also localizes the test logic into a single place, so that at a glance you can see how a number of different factors affect the behavior of the system under test.
As a bonus, a table of examples (with their expected results) often matches how the business communicates the requirements of a system, both internally and to customers—for instance, in a table describing shipping costs based on how much a customer spends, where they're located, whether they've bought a promotional product, etc. This means parameterized tests can often be initially created by pulling directly from a requirements document that your product folks provide, and the product folks can later read the tests (or at least the parameters table) if they want to verify the behavior of the system.
Parameterized tests are superficially similar to property-based tests. Both allow you to write fewer tests while covering more of your system's behavior. This library is not a replacement for property tests, but rather complimentary to them.
There are a few reasons you might choose to write a parameterized test rather than a property test:
-
When describing policies, not invariants: Much of a system's business logic comes down to arbitrary choices made by a product team. For instance, there's nothing in the abstract description of a shipping calculator that says buying socks or spending $100 total should grant you free shipping. Those aren't principles that every correctly implemented shipping system would implement. Instead, they're choices made by someone (maybe a product manager) which will in all likelihood be fiddled with over time.
Contrast that with the classic use cases for property tests: every (correct) implementation of, say,
List.sort/1will always have the dual properties of:- every element of the input being represented in the output, and
- every element being "less than" the element after it.
These sorting properties are invariants of the sorting function, and therefore are quite amenable to property testing.
-
Ease of writing: Property tests take a lot of practice to get good at writing. They're often quite time consuming to produce, and even when you think you've adequately described the parameters to the system.
-
For communication with other stakeholders: The table of examples in a parameterized test can be made readable by non-programmers (or non-Elixir programmers), so they can be a good way of showing others in your organization which behaviors of the system you've verified. Because they can compactly express a lot of test cases, they're much more suitable for this than saying "go read the title of every line in this file that starts with
test." -
For verifying the exact scenarios described by other stakeholders: Sometimes the edges of a particular behavior may be fuzzy—not just to you, but in the business domain as well. Hammering out hard-and-fast rules may not be necessary or worth it, so property tests that exercise the boundaries would be overkill. In contrast, when your product folks produce a document that describes the behavior of particular scenarios, you can encode that in a table and ensure that for the cases that are well-specified, the system behaves correctly.
When would you write a property test instead of an example tests?
- When you can specify true invariants about the desired behavior
- When you want the absolute highest confidence in your code
- When the correctness of a piece of code is important enough to merit a large time investment in getting the tests right
- When the system's behavior at the edges is well specified
And of course there's nothing wrong with using a mix of normal tests, parameterized tests, and property tests for a given piece of functionality.
-
Add
parameterized_testto yourmix.exsdependencies:def deps do [ {:parameterized_test, "~> 0.3", only: [:test]}, ] end
-
Run
$ mix deps.getto download the package -
Write your first example test by adding
import ParameterizedTestto the top of your test module, and using theparam_testmacro.You can optionally include a separator between the header and body of the table (like
|--------|-------|), and adescriptioncolumn to improve the errors you get when your test fails (see About test names, and improving debuggability for more on descriptions).The header of your table will be parsed as atoms to pass into your test context. The body cells of the table can be any valid Elixir expression, and empty cells will produce a
nilvalue.A dummy example:
defmodule MyApp.MyModuleTest do use ExUnit.Case, async: true import ParameterizedTest param_test "behaves as expected", """ | variable_1 | variable_2 | etc | | ---------------- | -------------------- | ------- | | %{foo: :bar} | div(19, 3) | false | | "bip bop" | String.upcase("foo") | true | | ["whiz", "bang"] | :ok | | | | nil | "maybe" | """, %{ variable_1: variable_1, variable_2: variable_2, etc: etc } do assert MyModule.valid_combination?(variable_1, variable_2, etc) end end
ExUnit requires each test in a module to have a unique name. By default,
without a description for the rows in your parameters table,
ParameterizedTest appends a stringified version of the parameters
passed to your test to the name you give the test. Consider this test:
param_test "checks equality",
"""
| val_1 | val_2 |
| :a | :a |
| :b | :c |
""",
%{val_1: val_1, val_2: val_2} do
assert val_1 == val_2
endUnder the hood, this produces two tests with the names:
"checks equality ([val_1: :a, val_b: :a])""checks equality ([val_1: :b, val_b: :c])"
And if you ran this test, you'd get an error that looks like this:
1) test checks equality ([val_1: :b, val_2: :c]) (MyModuleTest)
test/my_module_test.exs:4
Assertion with == failed
code: assert val_1 == val_2
left: :b
right: :c
stacktrace:
test/my_module_test.exs:11: (test)
You can improve the names in the failure cases by providing a description
column. When provided, that column will be used in the name. You may want
to use this to explain why this combination of values should produce
the expected outcome; for instance:
param_test "grants free shipping for spending $99 or more, or with coupon FREE_SHIP",
"""
| total_cents | coupon | free? | description |
| ----------- | ----------- | ----- | --------------------------- |
| 98_99 | | false | Spent too little |
| 99_00 | | true | Min for free shipping |
| 99_01 | | true | Spent more than the minimum |
| 1_00 | "FREE_SHIP" | true | Had the right coupon |
| 1_00 | "FOO" | false | Unrecognized coupon |
""", %{total_cents: total_cents, coupon: coupon, free?: gets_free_shipping?} do
shipping_cost = ShippingCalculator.calculate(total_cents, coupon)
free_shipping? = shipping_cost == 0
assert free_shipping? == gets_free_shipping?
endSuppose in your ShippingCalculator implementation, you mistakenly set
the free shipping threshold to be greater than $99.00, when your web site's
state policy was $99 or more. You'd get an error when running this test that
looks like this (note the first line ends with "Spent the min. for free shipping" from the description column):
1) test grants free shipping for spending $99 or more, or with coupon FREE_SHIP - Spent the min. for free shipping (ShippingCalculatorTest)
test/shipping/shipping_calculator_test.exs:34
Assertion with == failed
code: assert free_shipping? == gets_free_shipping?
left: false
right: true
stacktrace:
test/shipping/shipping_calculator_test.exs:47: (test)
Both this package and the
new, built-in parameterization
do similar things: they re-run the same test body with a different set of
parameters. However, the built-in parameterization works by re-running
your entire test module with each parameter set, so it's primarily aimed
at cases where the tests should work the same regardless of the parameters.
(The docs give the example of testing the Registry module and expecting
it to behave the same regardless of how many partitions are used.)
In contrast, the param_test macro is designed to use different parameters
on a per-test basis, and it's expected that the parameters will cause different
behavior between the test runs (and you'd generally expect to see one column
that describes what the results should be).
Finally, of course, there's the format of a param_test. The tabular, often
quite human-friendly format encourages collaboration with less technical
people on your team; your product manager may not be able to read a for
comprehension, but if you link them to a Markdown table on GitHub that shows
the test cases you've covered, they can probably make sense of them.
No sweat, you don't have to use it. You can instead pass a hand-rolled list of
parameters to the param_test macro, like this:
param_test "shipping policy matches the web site",
[
# Items in the parameters list can be either maps...
%{spending_by_category: %{pants: 29_99}, coupon: "FREE_SHIP"},
# ...or keyword lists
[spending_by_category: %{shoes: 19_99, pants: 29_99}, coupon: nil]
],
%{spending_by_category: spending_by_category, coupon: coupon} do
...
endJust make sure that each item in the parameters list has the same keys.
The final option is to pass a path to a file that contains your test parameters (we currently support .md/.markdown, .csv, and .tsv files), like this:
param_test "pull test parameters from a file",
"test/fixtures/params.md",
%{
spending_by_category: spending_by_category,
coupon: coupon,
gets_free_shipping?: gets_free_shipping?
} do
...
end