Bond
Design By Contract for Elixir.
Bond lets you attach preconditions and postconditions to your functions and verify them at runtime. A contract is a plain Elixir boolean expression with optional labels:
defmodule Account do
use Bond
@pre positive_amount: amount > 0
@post non_negative_balance: result >= 0
def withdraw(balance, amount), do: balance - amount
end
When a contract fails, Bond raises a Bond.PreconditionError or
Bond.PostconditionError with the failing assertion's label, expression,
location, and the local binding — telling you exactly what went wrong and
where.
Bond is an implementation of the Design By Contract methodology (also called programming by contract), introduced by Bertrand Meyer with the Eiffel language. See the About guide for background.
Usage
use Bond in any module to enable the @pre, @post, check/1, and
check/2 annotations. Contracts may use any Elixir expression that returns
a boolean (or a truthy value).
defmodule Math do
use Bond
@pre numeric_x: is_number(x), non_negative_x: x >= 0
@post float_result: is_float(result),
non_negative_result: result >= 0.0,
"sqrt of 0 is 0": (x == 0) ~> (result === 0.0),
"sqrt of 1 is 1": (x == 1) ~> (result === 1.0),
"x > 1 implies result smaller than x": (x > 1) ~> (result < x)
def sqrt(x), do: :math.sqrt(x)
end
@pre and @post accept one or more labelled assertions. Preconditions
have access to the function's parameters; postconditions also have access
to the result variable (bound to the function's return value) and
old(...) expressions that snapshot a value before the function runs (see
old expressions below).
Bond also provides a check/1,2 macro for placing assertions at arbitrary
points inside a function body — useful for sanity checks during development.
check honours the :bond, :checks config (see
Conditional compilation) and is safe to
disable in production builds.
When to use
check{: .warning}Don't use
checkfor input validation, validating data from external systems, or anything else that protects the integrity of your code. If the check were removed (or compiled out via config), the system must still behave correctly. Use ordinary control flow for that.
use Bond{: .info}
use BondoverridesKernel.@/1so that@pre,@post, and@docannotations can be intercepted and recorded, and installs@on_definition,@before_compile, and@after_compilecompiler hooks that wrap functions with contracts viadefoverridableat the end of module compilation. Yourdefs anddefps are otherwise left alone.
use Bondalso imports theBondmodule so thecheck/1andcheck/2macros are available, and importsBond.Predicatesso the predicate functions and operators defined there (such as~>and|||) can be used in assertions.Bond.Predicatescan be explicitly imported elsewhere if you want the operators outside of contract expressions.
Assertion syntax
An assertion is a boolean (or truthy) Elixir expression, optionally paired with a label. Labels are atoms or strings; they appear in error messages and generated documentation.
The recommended form is the keyword list, even for a single assertion:
@pre positive_x: x > 0
@post non_decreasing: result >= old(result)
@pre numeric_x: is_number(x), non_negative_x: x >= 0
For a bare assertion where a label adds no information, the bare form is also fine:
@pre is_number(x)
@post is_float(result)
For symmetry with ExUnit's assert(value, message) and assert message, value
patterns, the check/2 macro also accepts a label before or after the
expression:
check is_number(x)
check x_is_number: is_number(x)
check "x is a number", is_number(x)
check is_number(x), "x is a number"
Bond also provides the Bond.Predicates module with operators that are often
useful in assertions — notably ~> (logical implication) and <~ (pattern
match). Bond.Predicates is automatically imported into assertion
expressions, so you can use these operators directly:
@post (x == 0) ~> (result == 0.0)
@post {:ok, _} <~ result
See Bond.Predicates for the full list.
old expressions
old expressions allow postconditions to access the value of any arbitrary
expression prior to execution of the function body. Postconditions are
"pre-compiled" in such a way that any old expressions that appear in
assertions are resolved to the value that they had at the start of function
execution.
While this facility is not particularly relevant for purely functional code, it can be useful for stateful components of an application.
For example, imagine a simple, stateful Counter module that uses an Agent
to store the current count (some Agent code omitted for brevity):
defmodule Counter do
use Bond
def get_count(agent) do
Agent.get(agent, & &1)
end
@post count_incremented_by_1: get_count(agent) == old(get_count(agent)) + 1
def increment_count(agent) do
Agent.update(agent, &(&1 + 1))
end
end
Notice how the old expression captures the value of get_count/1 prior to
execution of the function, and this value is used to verify that the value of
get_count/1 has been updated as expected.
Note, however, that there is a potential race condition in the above code.
Since Agents are inherently concurrent, it is possible that another call to
increment_count/1 is interleaved between execution of the function body and
the call to get_count/1 that appears in the postcondition. In this scenario
the postcondition would fail because the new value of get_count/1 would be
at least 2 greater than the old value captured in the postcondition, rather
than exactly 1 greater as specified in the count_incremented_by_1 assertion.
As a first attempt to alleviate this race condition we can update the
increment_count/1 function so that it returns the updated count as its result
and use that result in the postcondition directly:
@post returns_updated_count: result == old(get_count(agent)) + 1
def increment_count(agent) do
Agent.get_and_update(agent, fn count ->
new_count = count + 1
{new_count, new_count}
end)
end
In this version we utilize Agent.get_and_update/3 to update the counter and
return the updated counter value in one operation. The new counter value is the
result of the function which can be used in postconditions. The
returns_updated_count assertion compares this result to the old value of
get_count/1 to ensure that it was incremented by exactly 1.
However, as you may have noticed, it is still possible for another call to
increment_count/1 to be interleaved between the call to get_count/1 in the
old expression of the postcondition and the call to Agent.get_and_update/3
in the function body. Alas, there is no way to "lock" an Agent over multiple
operations to ensure that there are no concurrent updates to the Agent state.
Therefore, our only choice is to soften the guarantee made by our
postcondition:
@post count_increased: get_count(agent) > old(get_count(agent))
def increment_count(agent) do
Agent.update(agent, &(&1 + 1))
end
The count_increased assertion in the postcondition now guarantees only that
the new value of get_count/1 is strictly greater than the old value. This
assertion always holds true regardless of the number of concurrent state
updates to the counter.
Although this assertion is not as strong as the count_incremented_by_1
assertion in the original version, it is the strongest we can provide given
the possibility of concurrent state updates.
Future versions of Bond may provide stronger support for stateful contracts in the form of invariants for structs and/or stateful processes, although this is still a subject of research.
Documenting contracts
Contracts are part of a module's public interface, in the same way that
function signatures and typespecs are. Bond treats them that way: every
function with a contract gets a #### Preconditions and/or
#### Postconditions section appended to its @doc, formatted as the
original assertion source. The sections appear in ex_doc output and in
editors that show function docs on hover (VS Code, Vim's K, etc.).
Auto-generated contract sections appear whether or not you wrote a @doc
yourself — Bond synthesises one when needed.
Conditional compilation and docs {: .info}
When a function has all of its contracts compile-disabled (see Conditional compilation), the function runs with zero contract overhead and its auto-generated contract sections are also suppressed. If you want the contract documentation visible in production builds, leave at least one of
:preconditionsor:postconditionsenabled, or write an explicit@docfor the function.
Conditional compilation
Contracts are evaluated on every call by default. For hot paths or production builds you can compile contracts out entirely via three application-config keys, read at compile time:
# config/prod.exs
config :bond,
preconditions: false,
postconditions: false,
checks: false
Each key defaults to true. When set to false:
:preconditions—@preevaluation is omitted from generated override clauses, and the#### Preconditionssection is omitted from the auto-generated docs.:postconditions— same for@post.:checks— everycheck/1,2macro call expands to:okand the wrapped expression is not evaluated. Don't put side effects insidecheck.
When both:preconditions and :postconditions are disabled for a
function, Bond emits no override clause at all. The function runs exactly
as you wrote it, with no per-call overhead.
Because Application.compile_env/3 is used to read the config, changing
these values requires recompilation (mix deps.compile bond --force, or in
practice MIX_ENV=prod mix compile --force). The Elixir compiler tracks
the dependency for you in normal incremental builds.
A typical pattern: enable contracts in dev and test, disable in prod.
# config/config.exs
import Config
# Default: everything enabled.
# Specific environments may override below.
# config/prod.exs
import Config
config :bond,
preconditions: false,
postconditions: false,
checks: false
Installation
bond can be installed by adding it to your list of dependencies in mix.exs:
def deps do
[
{:bond, "~> 0.10.0"}
]
end
Documentation
Documentation is generated with ExDoc and published on HexDocs and be found at https://hexdocs.pm/bond/Bond.html.