Pathway encapsulates your business logic into simple operation objects (AKA application services on the DDD lingo).
$ gem install pathway
Pathway helps you separate your business logic from the rest of your application; regardless of is an HTTP backend, a background processing daemon, etc. The main concept Pathway relies upon to build domain logic modules is the operation, this important concept will be explained in detail in the following sections.
Pathway also aims to be easy to use, stay lightweight and extensible (by the use of plugins), avoid unnecessary dependencies, keep the core classes clean from monkey patching and help yield an organized and uniform codebase.
As mentioned earlier the operation is an essential concept Pathway is built around. Operations not only structure your code (using steps as will be explained later) but also express meaningful business actions. Operations can be thought of as use cases too: they represent an activity -to be performed by an actor interacting with the system- which should be understandable by anyone familiar with the business regardless of their technical expertise.
Operations shouldn't ideally contain any business rules but instead, orchestrate and delegate to other more specific subsystems and services. The only logic present then should be glue code or any data transformations required to make interactions with the inner system layers possible.
Operations work as function objects, they are callable and hold no state, as such, any object that responds to call
and returns a result object can be a valid operation and that's the minimal protocol they need to follow.
The result object must follow its protocol as well (and a helper class is provided for that end) but we'll talk about that in a minute.
Let's see an example:
class MyFirstOperation
def call(input)
result = Repository.create(input)
if result.valid?
Pathway::Result.success(result)
else
Pathway::Result.failure(:create_error)
end
end
end
result = MyFirstOperation.new.call(foo: 'foobar')
if result.success?
puts result.value.inspect
else
puts "Error: #{result.error}"
end
Note first, we are not inheriting from any class nor including any module. This won't be the case in general as pathway
provides classes to help build your operations, but it serves to illustrate how little is needed to implement one.
Also, let's ignore the specifics about Repository.create(...)
, we just need to know that is some backend service from which a value is returned.
We then define a call
method for the class. It only checks if the result is available and then wraps it into a successful Result
object when is ok, or a failing one when is not.
And basically, that's all is needed, you can then call the operation object, check whether it was completed correctly with success?
and get the resulting value.
By following this protocol, you will be able to uniformly apply the same pattern on every HTTP endpoint (or whatever means your app has to communicate with the outside world). The upper layer of the application will offload all the domain logic to the operation and only will need to focus on the HTTP transmission details.
Maintaining always the same operation protocol will also be very useful when composing them.
As should be evident by now an operation should always return either a successful or failed result. These concepts are represented by following a simple protocol, which Pathway::Result
subclasses comply with.
As we've seen before, by querying success?
on the result we can see if the operation we just ran went well, or call to failure?
to see if it failed.
The actual result value produced by the operation is accessible at the value
method and the error description (if there's any) at error
when the operation fails.
To return wrapped values or errors from your operation you must call Pathway::Result.success(value)
or Pathway::Result.failure(error)
.
It is worth mentioning that when you inherit from Pathway::Operation
you'll have helper methods at your disposal to create result objects easily. For instance, the previous section's example could be written as follows:
class MyFirstOperation < Pathway::Operation
def call(input)
result = Repository.create(input)
result.valid? ? success(result) : failure(:create_error)
end
end
Pathway::Error
is a helper class to represent the error description from a failed operation execution (and also supports pattern matching as we'll see later).
Its use is completely optional but provides you with a basic schema to communicate what went wrong. You can instantiate it by calling new
on the class itself or using the helper method error
provided by the operation class:
class CreateNugget < Pathway::Operation
def call(input)
validation = Validator.call(input)
if validation.ok?
success(Nugget.create(validation.values))
else
error(:validation, message: 'Invalid input', details: validation.errors)
end
end
end
As you can see error(...)
expects the type
as the first parameter (and only the mandatory) then message:
and details
keyword arguments; these 2 last ones can be omitted and have default values. The type parameter must be a Symbol
, message:
a String
and details:
can be a Hash
or any other structure you see fit.
Finally, the Error
object have three accessors available to get the values back:
result = CreateNugget.new.call(foo: 'foobar')
if result.failure?
puts "#{result.error.type} error: #{result.error.message}"
puts "Error details: #{result.error.details}"
end
Mind you, error(...)
creates an Error
object wrapped into a Pathway::Failure
so you don't have to do it yourself.
If you decide to use Pathway::Error.new(...)
directly, you will have to pass all the arguments as keywords (including type:
), and you will have to wrap the object before returning it.
It was previously mentioned that operations should work like functions, that is, they don't hold state and you should be able to execute the same instance all the times you need, on the other hand, there will be some values that won't change during the operation lifetime and won't make sense to pass as call
parameters, you can provide these values on initialization as context data.
Context data can be thought of as 'request data' on an HTTP endpoint, values that aren't global but won't change during the execution of the request. Examples of this kind of data are the current user, the current device, a CSRF token, other configuration parameters, etc. You will want to pass these values on initialization, and probably pass them along to other operations down the line.
You must define your initializer to accept a Hash
with these values, which is what every operation is expected to do, but as before, when inheriting from Operation
you have the helper class method context
handy to make it easier for you:
class CreateNugget < Pathway::Operation
context :current_user, notify: false
def call(input)
validation = Validator.call(input)
if validation.valid?
nugget = Nugget.create(owner: current_user, **validation.values)
Notifier.notify(:new_nugget, nugget) if @notify
success(nugget)
else
error(:validation, message: 'Invalid input', details: validation.errors)
end
end
end
op = CreateNugget.new(current_user: user)
op.call(foo: 'foobar')
In the example above context
is defining :current_user
as a mandatory argument (it will raise an error if not provided) and :notify
as an optional config argument, since it has a default value. Note that any extra non-defined value provided will be simply ignored.
Both of these parameters are available through accessors (and instance variables) inside the operation. Also, there is a context
private method you use to get all the initialization values as a frozen hash, in order to pass them along easily.
If you don't care about keeping the operation instance around you can execute the operation directly on the class. To do so, use call
with the initialization context first and then the remaining parameters:
user = User.first(session[:current_user_id])
context = { current_user: user }
CreateNugget.call(context, params[:nugget]) # Using 'call' on the class
Also, you have Ruby's alternative syntax to invoke the call
method: CreateNugget.(context, params[:nugget])
. In both cases, you'll get the operation result like when invoking call
on the operation's instance.
Mind you that a context must always be provided for this syntax, if you don't need any initialization use an empty hash.
There's also a third way to execute an operation, made available through a plugin, that will be explained later.
Finally, the steps are the heart of the Operation
class and the main reason you will want to inherit your own classes from Pathway::Operation
.
So far we know that every operation needs to implement a call
method and return a valid result object, pathway
provides another option: the process
block DSL, this method will define call
behind the scenes for us, while also providing a way to define a business-oriented set of steps to describe our operation's behavior.
Every step should be cohesive and focused on a single responsibility, ideally by offloading work to other subsystems. Designing steps this way is the developer's responsibility but is made much simpler by the use of custom steps provided by plugins as we'll see later.
Let's start by showing some actual code:
# ...
# Inside an operation class body...
process do
step :authorize
step :validate
set :create_nugget, to: :nugget
step :notify
end
# ...
To define your call
method using the DSL just call to process
and pass a block, inside it, the DSL will be available.
Each step
(or set
) call is referring to a method inside the operation class, superclasses, or available through a plugin, these methods will be eventually invoked by call
.
All of the steps constitute the operation use case and follow a series of conventions in order to carry the process state along the execution process.
When you run the call
method, the auto-generated code will save the provided argument at the input
key within the execution state. Subsequent steps will receive this state and will be able to modify it, setting the result or auxiliary values, in order to communicate with the next steps on the execution path.
Each step (as the operation as a whole) can succeed or fail, when the latter happens execution is halted, and the operation call
method returns immediately.
To signal a failure you must return a failure(...)
or error(...)
in the same fashion as when defining call
directly.
If you return a success(...)
or anything that's not a failure the execution carries on but the value is ignored. If you want to save the result value, you must use set
instead of step
at the process block, which will save your wrapped value, into the key provided at to:
.
Also, non-failure return values inside steps are automatically wrapped so you can use success
for clarity's sake but it's optional.
If you omit the to:
keyword argument when defining a set
step, the result key will be used by default, but we'll explain more on that later.
To operate with the execution state, every step method receives a structure representing the current state. This structure is similar to a Hash
and responds to its main methods (:[]
, :[]=
, :fetch
, :store
, :include?
and to_hash
).
When an operation is executed, before running the first step, an initial state is created by copying all the values from the initialization context (and also including input
).
Note that these values can be replaced in later steps but it won't mutate the context object itself since is always frozen.
A state object can be splatted on method definition in the same fashion as a Hash
, thus, allowing us to cherry-pick the attributes we are interested in any given step:
# ...
# This step only takes the values it needs and doesn't change the state.
def send_emails(state)
user, report = state[:user], state[:report]
ReportMailer.send_report(user.email, report)
end
# ...
On each step, you can access or change the result the operation will produce on a successful execution.
The value will be stored at one of the attributes within the state.
By default, the state's key :value
will hold the result, but if you prefer to use another name you can specify it through the result_at
operation class method.
Let's now go through a fully defined operation using steps:
class CreateNugget < Pathway::Operation
context :current_user
process do
step :authorize
step :validate
set :create_nugget
step :notify
end
result_at :nugget
def authorize(*)
unless current_user.can? :create, Nugget
error(:forbidden)
end
end
def validate(state)
validation = NuggetForm.call(state[:input])
if validation.ok?
state[:params] = validation.values
else
error(:validation, details: validation.errors)
end
end
def create_nugget(state)
Nugget.create(owner: current_user, **state[:params])
end
def notify(state)
Notifier.notify(:new_nugget, state[:nugget])
end
end
In the example above the operation will produce a nugget (whatever that is...).
As you can see in the code, we are using the previously mentioned methods to indicate we need the current user to be present on initialization: context: current_user
, a call
method (defined by process do ... end
), and the result value should be stored at the :nugget
key (result_at :nugget
).
Let's delve into the process
block: it defines three steps using the step
method and create_nugget
using set
, as we said before, this last step will set the result key (:nugget
) since the to:
keyword argument is absent.
Now, for each of the step methods:
:authorize
doesn't need the state so just ignores it, then checks if the current user is allowed to run the operation and halts the execution by returning a:forbidden
error type if is not, otherwise does nothing and the execution goes on.:validate
gets the state, checks the validity of the:input
value which as we said is just thecall
method input, returns anerror(...)
when there's a problem, and if the validation is correct it updates the state but saving the sanitized values in:params
. Note that on success the return value isstate[:params]
, but is ignored like on:authorize
, since this method was also specified usingstep
.:create_nugget
first takes the:params
attribute from the state, and callscreate
on theNugget
model with the sanitized params and the current user. The return value is saved to the result key (:nugget
in this case) as the step is defined usingstep
withoutto:
.:notify
grabs the:nugget
from the state, and simply emits a notification with it, it has no meaningful return value, so is ignored.
The previous example goes through all the essential concepts needed for defining an operation class. If you can grasp it, you already have a good understanding on how to implement one. There are still some very important bits to cover (like testing), and we'll tackle them in the latter sections.
On a final note, you may be thinking the code could be a bit less verbose; also, shouldn't very common stuff like validation or authorization be simpler to use? and why always specify the result key name? maybe is possible to infer it from the surrounding code. We will address all those issues in the next section using plugins, pathway
's extension mechanism.
Pathway operations can be extended with plugins. They are very similar to the ones found in Roda or Sequel. So if you are already familiar with any of those gems you shouldn't have any problem with pathway
's plugin system.
To activate a plugin just call the plugin
method on the operation class:
class BaseOperation < Pathway::Operation
plugin :foobar, qux: 'quz'
end
class SomeOperation < BaseOperation
# The :foobar plugin will also be activated here
end
The plugin name must be specified as a Symbol
(or also as the Module
where is implemented, but more on that later), and it can take parameters next to the plugin's name.
When activated it will enrich your operations with new instance and class methods plus extra customs step for the process
DSL.
Mind you, if you wish to activate a plugin for a number of operations you can activate it for all of them directly on the Pathway::Operation
class, or you can create your own base operation and all its descendants will inherit the base class' plugins.
This plugin provides integration with the dry-validation gem. I won't explain in detail how to use this library since is already extensively documented on its official website, but instead, I'll assume certain knowledge of it, nonetheless, as you'll see in a moment, its API is pretty self-explanatory.
dry-validation
provides a very simple way to define contract objects (conceptually very similar to form objects) to process and validate input. The provided custom :validate
step allows you to run your input through a contract to check if your data is valid before carrying on. When the input is invalid it will return an error object of type :validation
and the reasons the validation failed will be available at the details
attribute. Is usually the first step an operation runs.
When using this plugin we can provide an already defined contract to the step to use or we can also define it within the operation. Let's see a few examples:
class NuggetContract < Dry::Validation::Contract
params do
required(:owner).filled(:string)
required(:price).filled(:integer)
end
end
class CreateNugget < Pathway::Operation
plugin :dry_validation
contract NuggetContract
process do
step :validate
step :create_nugget
end
# ...
end
As is shown above, the contract is defined first, then configured to be used by the operation by calling contract NuggetContract
, and validate the input at the process block by placing the step step :validate
inside the process
block.
class CreateNugget < Pathway::Operation
plugin :dry_validation
contract do
params do
required(:owner).filled(:string)
required(:price).filled(:integer)
end
end
process do
step :validate
step :create_nugget
end
# ...
end
Now, this second example is equivalent to the first one, but here we call contract
with a block instead of an object parameter; this block will be used as the definition body for a contract class that will be stored internally. Thus keeping the contract and operation code in the same place, this is convenient when you have a rather simpler contract and don't need to reuse it.
One interesting nuance to keep in mind regarding the inline block contract is that, when doing operation inheritance, if the parent operation already has a contract, the child operation will define a new one inheriting from the parent's. This is very useful to share validation logic among related operations in the same class hierarchy.
As a side note, if your contract is simple enough and has parameters and not extra validations rules, you can call the params
method directly instead, the following code is essentially equivalent to the previous example:
class CreateNugget < Pathway::Operation
plugin :dry_validation
params do
required(:owner).filled(:string)
required(:price).filled(:integer)
end
process do
step :validate
step :create_nugget
end
# ...
end
If you are familiar with dry-validation
you probably know it provides a way to inject options before calling the contract.
In those scenarios, you must either set the auto_wire: true
plugin argument or specify how to map options from the execution state to the contract when calling step :validate
.
Lets see and example for the first case:
class CreateNugget < Pathway::Operation
plugin :dry_validation, auto_wire: true
context :user_name
contract do
option :user_name
params do
required(:owner).filled(:string)
required(:price).filled(:integer)
end
rule(:owner) do
key.failure("invalid owner") unless user_name == values[:owner]
end
end
process do
step :validate
step :create_nugget
end
# ...
end
Here the defined contract needs a :user_name
option, so we tell the operation to grab the attribute with the same name from the state by activating :auto_wire
, afterwards, when the validation runs, the contract will already have the user name available.
Mind you, this option is false
by default, so be sure to set it to true
at Pathway::Operation
if you'd rather have it enabled for all your operations.
On the other hand, if for some reason the name of the contract's option and state attribute don't match, we can just pass with: {...}
when calling to step :validate
, indicating how to wire the attributes, the following example illustrates just that:
class CreateNugget < Pathway::Operation
plugin :dry_validation
context :current_user_name
contract do
option :user_name
params do
required(:owner).filled(:string)
required(:price).filled(:integer)
end
rule(:owner) do
key.failure("invalid owner") unless user_name == values[:owner]
end
end
process do
step :validate, with: { user_name: :current_user_name } # Inject :user_name to the contract object with the state's :current_user_name
step :create_nugget
end
# ...
end
The with:
parameter can always be specified for step :validate
, and allows you to override the default mapping regardless if auto-wiring is active or not.
Pathway supports the dry-validation
gem down to version 0.11
(inclusive) in case you still have unmigrated code. When using versions below 1.0
the concept of contract is not present and instead of calling the contract
method to set up your validation logic, you must use the form
method. Everything else remains the same except, obviously, that you would have to use dry-definition
's old API which is a bit different from the current one.
This very simple plugin adds a custom step called :authorize
, that can be used to check for permissions and halt the operation with a :forbidden
error when they aren't fulfilled.
In order to use it you must define a boolean predicate to check for permissions, by passing a block to the authorization
method:
class MyOperation < Pathway::Operation
plugin :simple_auth
context :current_user
authorization { current_user.is_admin? }
process do
step :authorize
step :perform_some_action
end
# ...
end
The sequel_models
plugin helps integrate operations with the Sequel ORM, by adding a few custom steps.
This plugin expects you to be using Sequel
model classes to access your DB. In order to exploit it, you need to indicate which model your operation is going to work with, hence you must specify said model when activating the plugin with the model:
keyword argument, or later using the model
class method.
This configuration will then be used on the operation class and all its descendants.
class MyOperation < Pathway::Operation
plugin :sequel_models, model: Nugget, search_by: :name, set_result_key: false
end
# Or...
class MyOperation < Pathway::Operation
plugin :sequel_models
# This is useful when using inheritance and you need different models per operation
model Nugget, search_by: :name, set_result_key: false
process do
step :authorize
step :perform_some_action
end
end
As you can see above you can also customize the search field (:search_by
) and indicate if you want to override or not the result key (:set_result_key
) when calling the model
method.
These two options aren't mandatory, and by default, Pathway will set the search field to the class model primary key, and override the result key to a snake-cased version of the model name (ignoring namespaces if contained inside a class or module).
Let's now take a look at the provided extensions:
This step will fetch a model from the DB, by extracting the search field from the call
method input parameter stored at :input
in the execution state. If the model cannot be fetched from the DB it will halt the execution with a :not_found
error, otherwise it will simply save the model into the result key (which will be :nugget
for the example below).
You can later access the fetched model from that attribute and if the operation finishes successfully, it will be used as its result.
class UpdateNugget < Pathway::Operation
plugin :sequel_models, model: Nugget
process do
step :validate
step :fetch_model
step :fetch_model, from: User, using: :user_id, search_by: :pk, to: :user # Even the default class can also be overrided with 'from:'
step :update_nugget
end
# ...
end
As a side note, and as shown in the 3rd step, :fetch_model
allows you to override the search column (search_by:
), the input parameter to extract from input
(using:
), the attribute to store the result (to:
) and even the default search class (from:
). If the current defaults don't fit your needs and you'll have these options available. This is commonly useful when you need some extra object, besides the main one, to execute your operation.
These two are a bit special since they aren't actually custom steps but just new methods that extend the process DSL itself.
These methods will take a block as an argument within which you can define inner steps.
Keeping all that in mind the only thing transaction
and after_commit
really do is surround the inner steps with SEQUEL_DB.transaction { ... }
and SEQUEL_DB.after_commit { ... }
blocks, respectively.
class CreateNugget < Pathway::Operation
plugin :sequel_models, model: Nugget
process do
step :validate
transaction do
step :create_nugget
step :attach_history_note
after_commit do
step :send_emails
end
end
end
# ...
end
When won't get into the details for each step in the example above, but the important thing to take away is that :create_nugget
and :attach_history_note
will exists within a single transaction and send_mails
(and any steps you add in the after_commit
block) will only run after the transaction has finished successfully.
Another nuance to take into account is that calling transaction
will start a new savepoint, since, in case you're already inside a transaction, it will be able to properly notify that the transaction failed by returning an error object when that happens.
This plugin extends the call
class method on the operation to accept a block. You can then use this block for flow control on success, failure, and also different types of failures.
There are two ways to use this plugin: by discriminating between success and failure, and also discriminating according to the specific failure type.
In each case you must provide the action to execute for every outcome using blocks:
MyOperation.plugin :responder # 'plugin' is actually a public method
MyOperation.(context, params) do
success { |value| r.halt(200, value.to_json) } # BTW: 'r.halt' is a Roda request method used to exemplify
failure { |error| r.halt(403) }
end
In the example above we provide a block for both the success and the failure case. On each block, the result value or the error object error will be provided at the blocks' argument, and the result of the corresponding block will be the result of the whole expression.
Lets now show an example with the error type specified:
MyOperation.plugin :responder
MyOperation.(context, params) do
success { |value| r.halt(200, value.to_json) }
failure(:forbidden) { |error| r.halt(403) }
failure(:validation) { |error| r.halt(422, error.details.to_json) }
failure(:not_found) { |error| r.halt(404) }
end
As you can see is almost identical to the previous example only that this time you provide the error type on each failure
call.
Going a bit deeper now, we'll explain how to implement your own plugins. As was mentioned before pathway
follows a very similar approach to the Roda or Sequel plugin systems, which is reflected at its implementation.
Each plugin must be defined in a file placed within the pathway/plugins/
directory of your gem or application, so pathway
can require the file; and must be implemented as a module inside the Pathway::Plugins
namespace module. Inside your plugin module, three extra modules can be defined to extend the operation API ClassMethods
, InstanceMethods
and DSLMethods
; plus a class method apply
for plugin initialization when needed.
If you are familiar with the aforementioned plugin mechanism (or others as well), the function of each module is probably starting to feel evident: ClassMethods
will be used to extend the operation class, so any class methods should be defined here; InstanceMethods
will be included on the operation so all the instance methods you need to add to the operation should be here, this includes every custom step you need to add; and finally DSLMethods
will be included on the Operation::DSL
class, which holds all the DSL methods like step
or set
.
The apply
method will simply be run whenever the plugin is included, taking the operation class on the first argument and all then arguments the call to plugin
received (excluding the plugin name).
Let's explain with more detail using a complete example:
# lib/pathway/plugins/active_record.rb
module Pathway
module Plugins
module ActiveRecord
module ClassMethods
attr_accessor :model, :pk
def inherited(subclass)
super
subclass.model = self.model
subclass.pk = self.pk
end
end
module InstanceMethods
delegate :model, :pk, to: :class
# This method will conflict with :sequel_models so you mustn't load both plugins in the same operation
def fetch_model(state, column: pk)
current_pk = state[:input][column]
result = model.first(column => current_pk)
if result
state.update(result_key => result)
else
error(:not_found)
end
end
end
module DSLMethods
# This method also conflicts with :sequel_models, so don't use them at once
def transaction(&steps)
transactional_seq = -> seq, _state do
ActiveRecord::Base.transaction do
raise ActiveRecord::Rollback if seq.call.failure?
end
end
around(transactional_seq, &steps)
end
end
def self.apply(operation, model: nil, pk: nil)
operation.model = model
opertaion.pk = pk || model&.primary_key
end
end
end
end
The code above implements a plugin to provide basic interaction with the ActiveRecord gem. Even though is a very simple plugin, it shows all the essentials to develop more complex ones.
As is pointed out in the code, some of the methods implemented here (fetch_model
and transmission
) collide with methods defined for :sequel_models
, so as a consequence, these two plugins are not compatible with each other and cannot be activated for the same operation (although you can still do it for different operations within the same application).
You must be mindful about colliding method names when mixing plugins since Pathway
can't bookkeep compatibility among every plugin that exists or will ever exist.
Is a good practice to document known incompatibilities on the plugin definition itself when they are known.
The whole plugin is completely defined within the ActiveRecord
module inside the Pathway::Plugins
namespace, also the file is placed at the load path in pathway/plugin/active_record.rb
(assuming lib/
is listed in $LOAD_PATH
). This will ensure when calling plugin :active_record
inside an operation, the correct file will be loaded and the correct plugin module will be applied to the current operation.
Moving on to the ClassMethods
module, we can see the accessors model
and pk
are defined for the operation's class to allow configuration.
Also, the inherited
hook is defined, this will simply be another class method at the operation and as such will be executed normally when the operation class is inherited. In our implementation, we just call to super
(which is extremely important since other modules or parent classes could be using this hook) and then copy the model
and pk
options from the parent to the subclass in order to propagate the configuration downwards.
At the end of the ActiveRecord
module definition, you can see the apply
method. It will receive the operation class and the parameters passed when the plugin
method is invoked. This method is usually used for loading dependencies or just setting up config parameters as we do in this particular example.
InstanceMethods
first defines a few delegator methods to the class itself for later use.
Then the fetch_model
step is defined (remember steps are but operation instance methods). Its first parameter is the state itself, as in the other steps we've seen before, and the remaining parameters are the options we can pass when calling step :fetch_model
(mind you, this is also valid for steps defined in operations classes). Here we only take a single keyword argument: column: pk
, with a default value; this will allow us to change the look-up column when using the step and is the only parameter we can use, passing other keyword arguments or extra positional parameters when invoking the step will raise errors.
Let's now examine the fetch_model
step body, it's not really that much different from other steps, here we extract the model primary key from state[:input][column]
and use it to perform a search. If nothing is found an error is returned, otherwise the state is updated in the result key, to hold the model that was just fetched from the DB.
We finally see a DSLMethods
module defined to extend the process DSL.
For this plugin, we'll define a way to group steps within an ActiveRecord
transaction, much in the same way the :sequel_models
plugin already does for Sequel
.
To this end, we define a transaction
method to expect a steps block and pass it down to the around
helper below which expects a callable (like a Proc
) and a step list block. As you can see the lambda we pass on the first parameter makes sure the steps are being run inside a transaction or aborts the transaction if the intermediate result is a failure.
The around
method is a low-level tool available to help extend the process DSL and it may seem a bit daunting at first glance but its usage is quite simple, the block is just a step list like the ones we find inside the process
call; and the parameter is a callable (usually a lambda), that will take 2 arguments, an object from which we can run the step list by invoking call
(and is the only thing it can do), and the current state. From here we can examine the state and decide upon whether to run the steps, how many times (if any), or run some code before and/or after doing so, like what we need to do in our example to surround the steps within a DB transaction.
As of right now, only rspec
is supported, that is, you can obviously test your operations with any framework you want, but all the provided matchers are designed for rspec
.
In order to load Pathway's operation matchers you must add the following line to your spec_helper.rb
file, after loading rspec
:
require 'pathway/rspec'
Pathway provides a few matchers in order to test your operation easier. Let's go through a full example:
# create_nugget.rb
class CreateNugget < Pathway::Operation
plugin :dry_validation
params do
required(:owner).filled(:string)
required(:price).filled(:integer)
optional(:disabled).maybe(:bool)
end
process do
step :validate
set :create_nugget
end
def create_nugget(state)
Nugget.create(state[:params])
end
end
# create_nugget_spec.rb
describe CreateNugget do
describe '#call' do
subject(:operation) { CreateNugget.new }
context 'when the input is valid' do
let(:input) { owner: 'John Smith', value: '11230' }
it { is_expected.to succeed_on(input).returning(an_instace_of(Nugget)) }
end
context 'when the input is invalid' do
let(:input) { owner: '', value: '11230' }
it { is_expected.to fail_on(input).
with_type(:validation).
message('Is not valid').
and_details(owner: ['must be present']) }
end
end
describe '.contract' do
subject(:contract) { CreateNugget.build_contract }
it { is_expected.to require_fields(:owner, :price) }
it { is_expected.to accept_optional_field(:disabled) }
end
end
This first matcher works on the operation itself and that's why we could set subject
with the operation instance and use is_expected.to succeed_on(...)
on the example.
The assertion it performs is simply that the operation was successful, also you can optionally chain returning(...)
if you want to test the returning value, this method allows nesting matchers as is the case in the example.
This second matcher is analog to succeed_on
but it asserts that operation execution was a failure instead. Also if you return an error object, and you need to, you can assert the error type using the type
chain method (aliased as and_type
and with_type
); the error message (and_message
, with_message
or message
); and the error details (and_details
, with_details
or details
). Mind you, the chain methods for the message and details accept nested matchers while the type
chain can only test by equality.
Finally, we can see that we are also testing the operation's contract (or form), implemented here with the dry-validation
gem.
Two more matchers are provided: require_fields
(aliased require_field
) to test when a contract is expected to define a required set of fields, and accept_optional_fields
(aliased accept_optional_field
) to test when a contract must define a certain set of optional fields, both the contract class (at operation class method contract_class
) or an instance (operation class method build_contract
) can be provided.
These matchers are only useful when using dry-validation
(on every version newer or equal to 0.11.0
) and will probably be extracted to their own gem in the future.
After checking out the repo, run bin/setup
to install dependencies. Then, run rake spec
to run the tests. You can also run bin/console
for an interactive prompt that will allow you to experiment.
To install this gem onto your local machine, run bundle exec rake install
. To release a new version, update the version number in version.rb
, and then run bundle exec rake release
, which will create a git tag for the version, push git commits and tags, and push the .gem
file to rubygems.org.
Bug reports and pull requests are welcome on GitHub at https://github.com/pabloh/pathway.
The gem is available as open source under the terms of the MIT License.