I am reading some examples how to implement control of concurrency in DDD.
At first, I thought that it is responsability of the infraestrcuture, or the repository, because the domain only should to be responsible of the core bussiness. But I have seen many examples in which the domain entity has a property for the version.
Then, in the consumer, in the application layer, it is seen in this way:
public void MyApplicationMethod()
{
Order myOrder = _orderRepository.Get(1);
myOrder.UpdateComments("another comments");
//If I arrive here, I could update, so I can increase the version.
myOrder.IncreaseVersion();
_orderRepository.Commit();
}
But I don't know why is the consumer of the domain which has to increase the version. Perhaps the reason is because the domain isn't the responsability to control de concurrence or control the version, because it is not really a rule of the bussiness. But if this would be true, then why to have a property and a method to increase the version?
But if it is allowed or not bad solution that the domain knows about the version and has the method. Why to delegate the increase version to the client? Why not to increase the version in the update method of the entity? Something like that:
public class Order
{
private int _version;
bool _isVersionIncreased = false;
private void IncreaseVersion()
{
if(_isVersionIncreased == false)
{
_version += _version;
_isVersionIncreased = true;
}
}
public UpdateComments(string paramComments)
{
_comments = paramComments;
IncreaseVersion();
}
}
In this way I can simplify the code in the consumer. Also, if the entity is modified many times, with the variable _isVersionIncreased I can ensure I will increase only once the version, if I update many properties with different methods. The drawback is that I want to update again the same entity after a commit, I would have to rest the variable or load the entity again from the repository. But normally I use to create a new context for each action, so it is not a big problem.
In sumary, I would like to know which is a good way to control concurrency in DDD, and what problems could have my solution, to increase the version in the entity method, instead of do int in the consumer (application layer).
Thanks.
Related
I'm begginer in multithreading. I recently started to writing something like multithreaded observer. I need some clarification.
Let's say I'm working with Subject, and I'm changing its state. Then Observers (in example - GUI widgets) have to be notified, so they could perform the update() method.
And there is my question: how am i handling those getValue() performed by many Observers? If it's just a getter for some variable, do i have to run it in new thread? Does it require any locking?
Or mayby there is a metod to just send those new value to GUI thread, and letting widgets there access those value. And again, can it be a single loop, or do i have to create another threads for every widget to get those value?
That's a difficult subject. Here are couple of things that will guide and help you with it.
Embrace eventual consistency. When one object updates on one thread, others will receive change notifications and update to the correct state eventually. Don't try to keep everything in sync all the time. Don't expect everything to be up to date all the time. Design your system to handle these situations. Check this video.
Use immutability especially for collections. Reading and writing to a collection from multiple threads can result in disasters. Don't do it. Use immutable collections or use snapshotting. Basically one object that will called from multiple thread will return a snapshot of the state of the collection. when a notification for a change is received, the reader (GUI in your case) will request a snapshot of the new state and update it accordingly.
Design rich Models. Don't use AnemicModels that have only setters and getters and let others manipulate them. Let the Model protect it's data and provide queries for it's state. Don't return mutable objects from properties of an object.
Pass data that describes changes with change notifications. This way readers (GUI) may sync their state only from the change data without having to read the target object.
Divide responsibility. Let the GUI know that it's single threaded and received notifications from the background. Don't add knowledge in your Model that it will be updated on a background thread and to know that it's called from the GUI and give it the responsibility of sending change requests to a specific thread. The Model should not care about stuff like that. It raises notifications and let subscribers handle them the way they need to. Let the GUI know that the change notification will be received on the background so it can transfer it to the UI thread.
Check this video. It describes different way you can do multithreading.
You haven't shown any code or specified language, so I'll give you an example in pseudo code using a Java/C# like language.
public class FolderIcon {
private Icon mIcon;
public Icon getIcon() { return mIcon; }
public FolderIcon(Icon icon) {
mIcon = icon;
}
}
public class FolderGUIElement : Observer {
private Folder mFolder;
private string mFolderPath;
public FolderGUIElement(Folder folder) {
mFolder = folder;
mFolderPath = mFolder.getPath();
folder.addChangeListener(this);
}
public void onSubjectChanged(change c) {
if(c instanceof PathChange) {
dispatchOnGuiThread(() => {
handlePathChange((PathChange)change);
});
}
}
handlePathChange(PathChange change) {
mFolderPath = change.NewPath;
}
}
public class Folder : Subject {
private string mPath;
private FolderIcon mIcon;
public string getPath() { return mPath; }
public FolderIcon getIcon() { return mIcon; }
public void changePath(string newPath) {
mPath = patnewPath;
notifyChanged(new PathChange(newPath));
}
public void changeIcon(FolderIcon newIcon) {
mIcon = newIcon;
notifyChanged(new IconChange(newIcon));
}
}
Notice couple of things in the example.
We are using immutable objects from Folder. That means that the GUI elements cannot get the value of Path or FolderIcon and change it thus affecting Folder. When changing the icon we are creating a brand new FolderIcon object instead of modifying the old one. Folder itself is mutable, but it uses immutable objects for it's properties. If you want you can use fully immutable objects. A hybrid approach works well.
When we receive change notification we read the NewPath from the PathChange. This way we don't have to call the Folder again.
We have changePath and changeIcon methods instead of setPath and setIcon. This captures the intent of our operations better thus giving our model behavior instead of being just a bag of getters and setters.
If you haven't read Domain Driven Design I highly recommend it. It's not about multithreading, but on how to design rich models. It's in my list of books that every developer should read. On concept in DDD is ValueObject. It's immutable and provide a great way to implement models and is especially useful in multithreaded systems.
Assume scenario where the service requires some global configuration to handle some request.
For example when user wants to do something it requires some global configuration to check whether the user is permited todo so.
I realize that in axon i can have command handlers that could handle commands without specified target aggregate so the handling part isn't a problem.
Problem is where i would like to have persistent storage on top of that and some invariants when trying to change the configuration. The whole idea of the configuration is that it should be consistent like aggregate in axon.
ConfigService {
#Inject
configRepository;
#Inject
eventGateway;
#CommandHandler
handle(changeConfig){
let current = configRepository.loadCurrent;
//some checks
//persist here?
eventGateway.send(confgChanged)
}
#EventHandler
on(configChanged){
//or persist here?
configRepository.saveCurrent(configChanged.data)
}
}
If I do persistance on the command handler I think I shouldn't use event handler because it would save it twice. But then when i somehow lose the config repository data i can rebuild it based on the events.
Im not sure what im missing here in the understanding of the DDD concepts, to put it simply i would like to know where to put command handler for something that is neither an aggregate nor entity.
Maybe i should create command handler that calls the Config service instead making config service the command handler.
Are you using Axon without event sourcing here?
In Axon framework it is generally good practice only to change the state of an aggregate with events. If you are going to mix state or configuration loaded from a repository with state from the event store, how will you be able to guarantee that when you replay the same events, the resulting state will be the same? The next time the aggregate is loaded, there may be different state in your configRepository, resulting in a different state and different behavior of your aggregate.
Why is this bad? Well, those same events may have been handled by eventprocessors, they may have filled query tables, they may have sent messages to other systems or done other work based on the state the system had at the time. You will have a disagreement between your query database and your aggregate.
A concrete example: Imagine your aggregate processed a command to switch an email service on. The aggregate did this by applying an EmailServiceEnabledEvent and changing its own state to 'boolean emailEnabled = true'. After a while, the aggregate gets unloaded from memory. Now you change that configurationRepository to disable switching the email service on. When the aggregate is loaded again, events from the event store are applied, but this time it loads the configuration from your repository that says it shouldn't switch the email service on. The 'boolean emailEnabled' state is left false. You send a disable email service command to the aggregate, but the command handler in the aggregate thinks the email is already disabled, and doesn't apply an EmailServiceDisabledEvent. The email service is left on.
In short: I would recommend using commands to change the configuration of your aggregate.
It seems to me that you your global configuration is either a specification or a set of rules like in a rules engine.
Unlike the patterns described in GOF book, in DDD, some building blocks/patterns are more generic and can apply to different types of object that you have.
For example an Entity is something that has a life-cycle and has an identity. The stages in the life-cycle usually are: created, persisted, reconstructed from storage, modified and then it's life cycle ends by being deleted, archived, completed etc.
A Value Object is something that doesn't have identity, (most of the time) is immutable, two instances can be compared by the equality of their properties. Value Object represent important concepts in our domains like: Money in system that deal with accounting, banking etc., Vector3 and Matrix3 in systems that do mathematical calculations and simulations like modeling systems (3dsMax, Maya), video games etc. They contain important behavior.
So everything that you need to track and has identity can be an Entity.
You can have a Specification that is an entity, a Rule that is an entity, an Event can also be an entity if it has a unique ID assigned to it. In this case you can treat them just like any another entity. You can form aggregates, have repositories and services and use EventSourcing if necessary.
On the other hand a Specification, a Rule, an Event or a Command can also be Value Objects.
Specifications and Rules can also be Domain Services.
One important thing here is also the Bounded Context. The system that updates these rules is probably in a different Bounded context than the system that applies there rules. It's also possible that this isn't the case.
Here's an example.
Let's have a system, where a Customer can buy stuff. This sytem will also have Discounts on Orders that have specific Rules.
Let's say we have rule that says that: if a Customer has made an Order with more than 5 LineItems he get's a discount. If that Order has a total price of some amount (say 1000$) he gets discount.
The percentage of the discounts can be changed by the Sales team. The Sales system has OrderDicountPolicy aggregates that it can modify. On the other hand the Ordering system only reads OrderDicountPolicy aggregates and won't be able to modify them as this is the responsibility of the Sales team.
The Sales system and the Ordering system can be part of two separate Bounded Contexts: Sales and Orders. The Orders Bounded Context depends on Sales Bounded Context.
Note: I'll skip the most implementation details and add only the relevant things to shorten and simplify this example. If it's intent is not clear, I'll edit and add more details. UUID, DiscountPercentage and Money are value objects that I'll skip.
public interface OrderDiscountPolicy {
public UUID getID();
public DiscountPercentage getDiscountPercentage();
public void changeDiscountPercentage(DiscountPercentage percentage);
public bool canApplyDiscount(Order order);
}
public class LineItemsCountOrderDiscountPolicy implements OrderDiscountPolicy {
public int getLineItemsCount() { }
public void changeLineItemsCount(int count) { }
public bool canApplyDiscount(Order order) {
return order.getLineItemsCount() > this.getLineItemsCount();
}
// other stuff from interface implementation
}
public class PriceThresholdOrderDiscountPolicy implements OrderDiscountPolicy {
public Money getPriceThreshold() { }
public void changePriceThreshold(Money threshold) { }
public bool canApplyDiscount(Order order) {
return order.getTotalPriceWithoutDiscount() > this.getPriceThreshold();
}
// other stuff from interface implementation
}
public class LineItem {
public UUID getOrderID() { }
public UUID getProductID() { }
public Quantity getQuantity { }
public Money getProductPrice() { }
public Money getTotalPrice() {
return getProductPrice().multiply(getQuantity());
}
}
public enum OrderStatus { Pending, Placed, Approced, Rejected, Shipped, Finalized }
public class Order {
private UUID mID;
private OrderStatus mStatus;
private List<LineItem> mLineItems;
private DscountPercentage mDiscountPercentage;
public UUID getID() { }
public OrderStatus getStatus() { }
public DscountPercentage getDiscountPercentage() { };
public Money getTotalPriceWithoutDiscount() {
// return sum of all line items
}
public Money getTotalPrice() {
// return sum of all line items + discount percentage
}
public void changeStatus(OrderStatus newStatus) { }
public List<LineItem> getLineItems() {
return Collections.unmodifiableList(mLineItems);
}
public LineItem addLineItem(UUID productID, Quantity quantity, Money price) {
LineItem item = new LineItem(this.getID(), productID, quantity, price);
mLineItems.add(item);
return item;
}
public void applyDiscount(DiscountPercentage discountPercentage) {
mDiscountPercentage = discountPercentage;
}
}
public class PlaceOrderCommandHandler {
public void handle(PlaceOrderCommand cmd) {
Order order = mOrderRepository.getByID(cmd.getOrderID());
List<OrderDiscountPolicy> discountPolicies =
mOrderDiscountPolicyRepository.getAll();
for (OrderDiscountPolicy policy : discountPolicies) {
if (policy.canApplyDiscount(order)) {
order.applyDiscount(policy.getDiscountPercentage());
}
}
order.changeStatus(OrderStatus.Placed);
mOrderRepository.save(order);
}
}
public class ChangeOrderDiscountPolicyPercentageHandler {
public void handle(ChangeOrderDiscountPolicyPercentage cmd) {
OrderDiscountPolicy policy =
mOrderDiscountRepository.getByID(cmd.getPolicyID());
policy.changePercentage(cmd.getDiscountPercentage());
mOrderDiscountRepository.save(policy);
}
}
You can use EventSourcing if you think that it's appropriate for some aggregates. The DDD book has a chapter on global rules and specifications.
Let's take a look what whould we do in the case of a distributed application for example using microservices.
Let's say we have 2 services: OrdersService and OrdersDiscountService.
There are couple of ways to implement this operation. We can use:
Choreography with Events
Orchestration with explicit Saga or a Process Manager
Here's how we can do it if we use Choreography with Events.
CreateOrderCommand -> OrdersService -> OrderCreatedEvent
OrderCreatedEvent -> OrdersDiscountService -> OrderDiscountAvailableEvent or OrderDiscountNotAvailableEvent
OrderDiscountAvailableEvent or OrderDiscountNotAvailableEvent -> OrdersService -> OrderPlacedEvent
In this example to place the order OrdersService will wait for OrderDiscountNotAvailableEvent or OrderDiscountNotAvailableEvent so it can apply a discount before changing the status of the order to OrderPlaced.
We can also use an explicit Saga to do Orchestration between services.
This Saga will containt the sequence of steps for the process so it can execute it.
PlaceOrderCommand -> Saga
Saga asks OrdersDiscountService to see if a discount is available for that Order.
If discount is available, Saga calls OrdersService to apply a discount
Saga calls OrdersService to set the status of the Order to OrderPlaced
Note: Steps 3 and 4 can be combined
This raises the question: *"How OrdersDiscountService get's all the necessary information for the Order to calculate discounts?"*
This can either be achieved by adding all of the information of the order in the Event that this service will receive or by having OrdersDiscountService call OrdersService to get the information.
Here's a Great video from Martin Folwer on Event Driven Architectures that discusses these approaches.
The advantage of Orchestration with a Saga is that the exact process is explicitly defined in the Saga and can be found, understood and debugged.
Having implicit processes like in the case of the Choreography with Events can be harder to understand, debug and maintain.
The downside of having Sagas is that we do define more things.
Personally, I tend to go for the explicit Saga especially for complex processes, but most of the systems I work and see use both approaches.
Here are some additional resources:
https://blog.couchbase.com/saga-pattern-implement-business-transactions-using-microservices-part/
https://blog.couchbase.com/saga-pattern-implement-business-transactions-using-microservices-part-2/
https://microservices.io/patterns/data/saga.html
The LMAX Architecture is very interesting read. It's not distributed system, but is event driven and records both incomming events/commands and outgoint events. It's an interesting way to capture everything that happend in a system or a service.
Hi I have a maybe a common problem that I think not entirely can be solved by Autofac or any IoC container. It can be a design problem that I need some fresh input on.
I have the classic MVC web solution with EF 6. Its been implemented in a true DDD style with Anti-corruption layer, three bounded contexts, cross-cutting concerns movers out to infrastructure projects. It has been a real pleasure to see all pieces fall in to place in good way. We also added Commands to CUD operations into Domain.
Now here is the problem. Customer want a change log that tracks every entities property and when updates are done we need to save into change log values before and after update. We have implemented that successful in a ILoggerService that wraps a Microsoft test utility that we uses to detect changes. But I, my role is Software Architect, took the decision to Decorate our generic repositories with a ChangeTrackerRepository that have a dependency on ILoggerService. This works fine. The Decorator track methods Add(…) and Modify(…) in our IRepository<TEntity>.
The problem is that we have Repositories that have custom repositories that have custom queries like this:
public class CounterPartRepository : Repository<CounterPart>, ICounterPartRepository
{
public CounterPartRepository(ManagementDbContext unitOfWork)
: base(unitOfWork)
{}
public CounterPart GetAggregate(Guid id)
{
return GetSet().CompleteAggregate().SingleOrDefault(s => s.Id == id);
}
public void DeleteCounterPartAddress(CounterPartAddress address)
{
RemoveChild(address);
}
public void DeleteCounterPartContact(CounterPartContact contact)
{
RemoveChild(contact);
}
}
We have simple repositories that just closes the generic repository and get proper EF Bounded context injected into it (Unit Of Work pattern):
public class AccrualPeriodTypeRepository : Repository<AccrualPeriodType>, IAccrualPeriodTypeRepository
{
public AccrualPeriodTypeRepository(ManagementDbContext unitOfWork)
: base(unitOfWork)
{
}
}
The problem is that when decorating AccrualPeriodTypeRepository with AutoFac through generic Decorator we can easily inject that repo into CommandHandler actor like this
public AddAccrualPeriodCommandHandler(IRepository<AccrualPeriod> accrualRepository)
This works fine.
But How do we also decorate CounterPartRepository???
I have gone through several solutions in my head and they all end up with a dead-end.
1) Manually decorate every custom repository generate to many custom decorators that it will be near unmaintainable.
2) Decorate the closed Repository Repository with extended custom queries. This smells bad. Should be part of that repository?
3) If we consider 2… maybe Skip our Services and only rely on IRepository for operating on our Aggregate Roots and IQueryHandler (see article https://cuttingedge.it/blogs/steven/pivot/entry.php?id=92)
I need some fresh input to a common problem I think, when it comes to decorating your repositories when you have custom closed repositories and simple repositories also closed but both inherit from same Repository
Have you consider decorating command handlers instead of decorating repositories?
Repos are too low level, and it is not their responsibility to know what should be logged and how.
What about the following:
1) You have your command handlers in a way:
public class DeleteCounterPartAddressHandler : IHandle<DeleteCounterPartAddressCommand>
{
//this might be set by a DI container, or passed to a constructor
public ICounterPartRepository Repository { get; set; }
public void Handle(DeleteCounterPartAddressCommand command)
{
var counterpart = repository.GetPropertyById(command.CounterPartId);
// in DDD you always want to read and aggregate
// and save an aggregate as a whole
property.DeleteAdress(command.AddressId);
repository.Save(counterpart)
}
}
2) Now you can simply use Chain Of Responsibility pattern to "decorate" your handlers with logging, transactions, whatever:
public class LoggingHandler<T> : IHandler<T> {
private readonly IHandler<T> _innerHandler;
public LoggingHandler(IHandler<T> innerHandler) {
_innerHandler = innerHandler;
}
public void Handle(T command)
{
//Obviously you do it properly, but you get the idea
_log.Info("Before");
_innerHandler.Handle(command);
_log.Info("After");
}
}
Now you have just one piece of code responsible for logging and you can compose it with any command handler, so if you ever want to log a particular command then you just "wrap" it with the logging handler, and it is still your IHandle<T> so the rest of the system is not impacted.
And you can do it with other concerns too (threading, queueing, transactions, multiplexing, routing, etc.) without messing around and plumbing this stuff here and there.
Concerns are very well separated this way.
It is also much better (to me) because you log on a real operation (business) level, rather than on low-level repository.
Hope it helps.
P.S. In DDD you really want your repositories to only expose aggregate-level methods because Aggregates suppose to take care of their invariants (and nothing else, no services, no repositories), and because Aggregate represents transaction boundary.
Really, it is up to the Repository how to get the Aggregate from persisted storage and how to persist it back, outside it should look like you ask someone for an object and it gives you an object you can call behaviors on.
So normally you would only get an aggregate from the repository, call its behavior(s) and then save it back. Which really means that your repositories would mostly have GetById and Save methods, not some internals like "UpdateThatPartOfAnAggregate".
I've taken a read to the Advantages of message based web services article and am wondering if there is there a recommended style/practice to versioning Restful resources in ServiceStack? The different versions could render different responses or have different input parameters in the Request DTO.
I'm leaning toward a URL type versioning (i.e /v1/movies/{Id}), but I have seen other practices that set the version in the HTTP headers (i.e Content-Type: application/vnd.company.myapp-v2).
I'm hoping a way that works with the metadata page but not so much a requirement as I've noticed simply using folder structure/ namespacing works fine when rendering routes.
For example (this doesn't render right in the metadata page but performs properly if you know the direct route/url)
/v1/movies/{id}
/v1.1/movies/{id}
Code
namespace Samples.Movies.Operations.v1_1
{
[Route("/v1.1/Movies", "GET")]
public class Movies
{
...
}
}
namespace Samples.Movies.Operations.v1
{
[Route("/v1/Movies", "GET")]
public class Movies
{
...
}
}
and corresponding services...
public class MovieService: ServiceBase<Samples.Movies.Operations.v1.Movies>
{
protected override object Run(Samples.Movies.Operations.v1.Movies request)
{
...
}
}
public class MovieService: ServiceBase<Samples.Movies.Operations.v1_1.Movies>
{
protected override object Run(Samples.Movies.Operations.v1_1.Movies request)
{
...
}
}
Try to evolve (not re-implement) existing services
For versioning, you are going to be in for a world of hurt if you try to maintain different static types for different version endpoints. We initially started down this route but as soon as you start to support your first version the development effort to maintain multiple versions of the same service explodes as you will need to either maintain manual mapping of different types which easily leaks out into having to maintain multiple parallel implementations, each coupled to a different versions type - a massive violation of DRY. This is less of an issue for dynamic languages where the same models can easily be re-used by different versions.
Take advantage of built-in versioning in serializers
My recommendation is not to explicitly version but take advantage of the versioning capabilities inside the serialization formats.
E.g: you generally don't need to worry about versioning with JSON clients as the versioning capabilities of the JSON and JSV Serializers are much more resilient.
Enhance your existing services defensively
With XML and DataContract's you can freely add and remove fields without making a breaking change. If you add IExtensibleDataObject to your response DTO's you also have a potential to access data that's not defined on the DTO. My approach to versioning is to program defensively so not to introduce a breaking change, you can verify this is the case with Integration tests using old DTOs. Here are some tips I follow:
Never change the type of an existing property - If you need it to be a different type add another property and use the old/existing one to determine the version
Program defensively realize what properties don't exist with older clients so don't make them mandatory.
Keep a single global namespace (only relevant for XML/SOAP endpoints)
I do this by using the [assembly] attribute in the AssemblyInfo.cs of each of your DTO projects:
[assembly: ContractNamespace("http://schemas.servicestack.net/types",
ClrNamespace = "MyServiceModel.DtoTypes")]
The assembly attribute saves you from manually specifying explicit namespaces on each DTO, i.e:
namespace MyServiceModel.DtoTypes {
[DataContract(Namespace="http://schemas.servicestack.net/types")]
public class Foo { .. }
}
If you want to use a different XML namespace than the default above you need to register it with:
SetConfig(new EndpointHostConfig {
WsdlServiceNamespace = "http://schemas.my.org/types"
});
Embedding Versioning in DTOs
Most of the time, if you program defensively and evolve your services gracefully you wont need to know exactly what version a specific client is using as you can infer it from the data that is populated. But in the rare cases your services needs to tweak the behavior based on the specific version of the client, you can embed version information in your DTOs.
With the first release of your DTOs you publish, you can happily create them without any thought of versioning.
class Foo {
string Name;
}
But maybe for some reason the Form/UI was changed and you no longer wanted the Client to use the ambiguous Name variable and you also wanted to track the specific version the client was using:
class Foo {
Foo() {
Version = 1;
}
int Version;
string Name;
string DisplayName;
int Age;
}
Later it was discussed in a Team meeting, DisplayName wasn't good enough and you should split them out into different fields:
class Foo {
Foo() {
Version = 2;
}
int Version;
string Name;
string DisplayName;
string FirstName;
string LastName;
DateTime? DateOfBirth;
}
So the current state is that you have 3 different client versions out, with existing calls that look like:
v1 Release:
client.Post(new Foo { Name = "Foo Bar" });
v2 Release:
client.Post(new Foo { Name="Bar", DisplayName="Foo Bar", Age=18 });
v3 Release:
client.Post(new Foo { FirstName = "Foo", LastName = "Bar",
DateOfBirth = new DateTime(1994, 01, 01) });
You can continue to handle these different versions in the same implementation (which will be using the latest v3 version of the DTOs) e.g:
class FooService : Service {
public object Post(Foo request) {
//v1:
request.Version == 0
request.Name == "Foo"
request.DisplayName == null
request.Age = 0
request.DateOfBirth = null
//v2:
request.Version == 2
request.Name == null
request.DisplayName == "Foo Bar"
request.Age = 18
request.DateOfBirth = null
//v3:
request.Version == 3
request.Name == null
request.DisplayName == null
request.FirstName == "Foo"
request.LastName == "Bar"
request.Age = 0
request.DateOfBirth = new DateTime(1994, 01, 01)
}
}
Framing the Problem
The API is the part of your system that exposes its expression. It defines the concepts and the semantics of communicating in your domain. The problem comes when you want to change what can be expressed or how it can be expressed.
There can be differences in both the method of expression and what is being expressed. The first problem tends to be differences in tokens (first and last name instead of name). The second problem is expressing different things (the ability to rename oneself).
A long-term versioning solution will need to solve both of these challenges.
Evolving an API
Evolving a service by changing the resource types is a type of implicit versioning. It uses the construction of the object to determine behavior. Its works best when there are only minor changes to the method of expression (like the names). It does not work well for more complex changes to the method of expression or changes to the change of expressiveness. Code tends to be scatter throughout.
Specific Versioning
When changes become more complex it is important to keep the logic for each version separate. Even in mythz example, he segregated the code for each version. However, the code is still mixed together in the same methods. It is very easy for code for the different versions to start collapsing on each other and it is likely to spread out. Getting rid of support for a previous version can be difficult.
Additionally, you will need to keep your old code in sync to any changes in its dependencies. If a database changes, the code supporting the old model will also need to change.
A Better Way
The best way I've found is to tackle the expression problem directly. Each time a new version of the API is released, it will be implemented on top of the new layer. This is generally easy because changes are small.
It really shines in two ways: first all the code to handle the mapping is in one spot so it is easy to understand or remove later and second it doesn't require maintenance as new APIs are developed (the Russian doll model).
The problem is when the new API is less expressive than the old API. This is a problem that will need to be solved no matter what the solution is for keeping the old version around. It just becomes clear that there is a problem and what the solution for that problem is.
The example from mythz's example in this style is:
namespace APIv3 {
class FooService : RestServiceBase<Foo> {
public object OnPost(Foo request) {
var data = repository.getData()
request.FirstName == data.firstName
request.LastName == data.lastName
request.DateOfBirth = data.dateOfBirth
}
}
}
namespace APIv2 {
class FooService : RestServiceBase<Foo> {
public object OnPost(Foo request) {
var v3Request = APIv3.FooService.OnPost(request)
request.DisplayName == v3Request.FirstName + " " + v3Request.LastName
request.Age = (new DateTime() - v3Request.DateOfBirth).years
}
}
}
namespace APIv1 {
class FooService : RestServiceBase<Foo> {
public object OnPost(Foo request) {
var v2Request = APIv2.FooService.OnPost(request)
request.Name == v2Request.DisplayName
}
}
}
Each exposed object is clear. The same mapping code still needs to be written in both styles, but in the separated style, only the mapping relevant to a type needs to be written. There is no need to explicitly map code that doesn't apply (which is just another potential source of error). The dependency of previous APIs is static when you add future APIs or change the dependency of the API layer. For example, if the data source changes then only the most recent API (version 3) needs to change in this style. In the combined style, you would need to code the changes for each of the APIs supported.
One concern in the comments was the addition of types to the code base. This is not a problem because these types are exposed externally. Providing the types explicitly in the code base makes them easy to discover and isolate in testing. It is much better for maintainability to be clear. Another benefit is that this method does not produce additional logic, but only adds additional types.
I am also trying to come with a solution for this and was thinking of doing something like the below. (Based on a lot of Googlling and StackOverflow querying so this is built on the shoulders of many others.)
First up, I don’t want to debate if the version should be in the URI or Request Header. There are pros/cons for both approaches so I think each of us need to use what meets our requirements best.
This is about how to design/architecture the Java Message Objects and the Resource Implementation classes.
So let’s get to it.
I would approach this in two steps. Minor Changes (e.g. 1.0 to 1.1) and Major Changes (e.g 1.1 to 2.0)
Approach for minor changes
So let’s say we go by the same example classes used by #mythz
Initially we have
class Foo { string Name; }
We provide access to this resource as /V1.0/fooresource/{id}
In my use case, I use JAX-RS,
#Path("/{versionid}/fooresource")
public class FooResource {
#GET
#Path( "/{id}" )
public Foo getFoo (#PathParam("versionid") String versionid, (#PathParam("id") String fooId)
{
Foo foo = new Foo();
//setters, load data from persistence, handle business logic etc
Return foo;
}
}
Now let’s say we add 2 additional properties to Foo.
class Foo {
string Name;
string DisplayName;
int Age;
}
What I do at this point is annotate the properties with a #Version annotation
class Foo {
#Version(“V1.0")string Name;
#Version(“V1.1")string DisplayName;
#Version(“V1.1")int Age;
}
Then I have a response filter that will based on the requested version, return back to the user only the properties that match that version. Note that for convenience, if there are properties that should be returned for all versions, then you just don’t annotate it and the filter will return it irrespective of the requested version
This is sort of like a mediation layer. What I have explained is a simplistic version and it can get very complicated but hope you get the idea.
Approach for Major Version
Now this can get quite complicated when there is a lot of changes been done from one version to another. That is when we need to move to 2nd option.
Option 2 is essentially to branch off the codebase and then do the changes on that code base and host both versions on different contexts. At this point we might have to refactor the code base a bit to remove version mediation complexity introduced in Approach one (i.e. make the code cleaner) This might mainly be in the filters.
Note that this is just want I am thinking and haven’t implemented it as yet and wonder if this is a good idea.
Also I was wondering if there are good mediation engines/ESB’s that could do this type of transformation without having to use filters but haven’t seen any that is as simple as using a filter. Maybe I haven’t searched enough.
Interested in knowing thoughts of others and if this solution will address the original question.
I have read Evans, Nilsson and McCarthy, amongst others, and understand the concepts and reasoning behind a domain driven design; however, I'm finding it difficult to put all of these together in a real-world application. The lack of complete examples has left me scratching my head. I've found a lot of frameworks and simple examples but nothing so far that really demonstrates how to build a real business application following a DDD.
Using the typical order management system as an example, take the case of order cancellation. In my design I can see an OrderCancellationService with a CancelOrder method which accepts the order # and a reason as parameters. It then has to perform the following 'steps':
Verify that the current user has the necessary permission to cancel an Order
Retrieve the Order entity with the specified order # from the OrderRepository
Verify that the Order may be canceled (should the service interrogate the state of the Order to evaluate the rules or should the Order have a CanCancel property that encapsulates the rules?)
Update the state of the Order entity by calling Order.Cancel(reason)
Persist the updated Order to the data store
Contact the CreditCardService to revert any credit card charges that have already been processed
Add an audit entry for the operation
Of course, all of this should happen in a transaction and none of the operations should be allowed to occur independently. What I mean is, I must revert the credit card transaction if I cancel the order, I cannot cancel and not perform this step. This, imo, suggests better encapsulation but I don't want to have a dependency on the CreditCardService in my domain object (Order), so it seems like this is the responsibility of the domain service.
I am looking for someone to show me code examples how this could/should be "assembled". The thought-process behind the code would be helpful in getting me to connect all of the dots for myself. Thx!
Your domain service may look like this. Note that we want to keep as much logic as possible in the entities, keeping the domain service thin. Also note that there is no direct dependency on credit card or auditor implementation (DIP). We only depend on interfaces that are defined in our domain code. The implementation can later be injected in the application layer. Application layer would also be responsible for finding Order by number and, more importantly, for wrapping 'Cancel' call in a transaction (rolling back on exceptions).
class OrderCancellationService {
private readonly ICreditCardGateway _creditCardGateway;
private readonly IAuditor _auditor;
public OrderCancellationService(
ICreditCardGateway creditCardGateway,
IAuditor auditor) {
if (creditCardGateway == null) {
throw new ArgumentNullException("creditCardGateway");
}
if (auditor == null) {
throw new ArgumentNullException("auditor");
}
_creditCardGateway = creditCardGateway;
_auditor = auditor;
}
public void Cancel(Order order) {
if (order == null) {
throw new ArgumentNullException("order");
}
// get current user through Ambient Context:
// http://blogs.msdn.com/b/ploeh/archive/2007/07/23/ambientcontext.aspx
if (!CurrentUser.CanCancelOrders()) {
throw new InvalidOperationException(
"Not enough permissions to cancel order. Use 'CanCancelOrders' to check.");
}
// try to keep as much domain logic in entities as possible
if(!order.CanBeCancelled()) {
throw new ArgumentException(
"Order can not be cancelled. Use 'CanBeCancelled' to check.");
}
order.Cancel();
// this can throw GatewayException that would be caught by the
// 'Cancel' caller and rollback the transaction
_creditCardGateway.RevertChargesFor(order);
_auditor.AuditCancellationFor(order);
}
}
A slightly different take on it:
//UI
public class OrderController
{
private readonly IApplicationService _applicationService;
[HttpPost]
public ActionResult CancelOrder(CancelOrderViewModel viewModel)
{
_applicationService.CancelOrder(new CancelOrderCommand
{
OrderId = viewModel.OrderId,
UserChangedTheirMind = viewModel.UserChangedTheirMind,
UserFoundItemCheaperElsewhere = viewModel.UserFoundItemCheaperElsewhere
});
return RedirectToAction("CancelledSucessfully");
}
}
//App Service
public class ApplicationService : IApplicationService
{
private readonly IOrderRepository _orderRepository;
private readonly IPaymentGateway _paymentGateway;
//provided by DI
public ApplicationService(IOrderRepository orderRepository, IPaymentGateway paymentGateway)
{
_orderRepository = orderRepository;
_paymentGateway = paymentGateway;
}
[RequiredPermission(PermissionNames.CancelOrder)]
public void CancelOrder(CancelOrderCommand command)
{
using (IUnitOfWork unitOfWork = UnitOfWorkFactory.Create())
{
Order order = _orderRepository.GetById(command.OrderId);
if (!order.CanBeCancelled())
throw new InvalidOperationException("The order cannot be cancelled");
if (command.UserChangedTheirMind)
order.Cancel(CancellationReason.UserChangeTheirMind);
if (command.UserFoundItemCheaperElsewhere)
order.Cancel(CancellationReason.UserFoundItemCheaperElsewhere);
_orderRepository.Save(order);
_paymentGateway.RevertCharges(order.PaymentAuthorisationCode, order.Amount);
}
}
}
Notes:
In general I only see the need for a domain service when a command/use case involves the state change of more than one aggregate. For example, if I needed to invoke methods on the Customer aggregate as well as Order, then I'd create the domain service OrderCancellationService that invoked the methods on both aggregates.
The application layer orchestrates between infrastructure (payment gateways) and the domain. Like domain objects, domain services should only be concerned with domain logic, and ignorant of infrastructure such as payment gateways; even if you've abstracted it using your own adapter.
With regards to permissions, I would use aspect oriented programming to extract this away from the logic itself. As you see in my example, I've added an attribute to the CancelOrder method. You can use an intercepter on that method to see if the current user (which I would set on Thread.CurrentPrincipal) has that permission.
With regards to auditing, you simply said 'audit for the operation'. If you just mean auditing in general, (i.e. for all app service calls), again I would use interceptors on the method, logging the user, which method was called, and with what parameters. If however you meant auditing specifically for the cancellation of orders/payments then do something similar to Dmitry's example.