CQRS Event Sourcing: Validate UserName uniqueness

问题:

Let's take a simple "Account Registration" example, here is the flow:

• User visit website
• Click "Register" button and fill form, click "Save" button
• MVC Controller: Validate UserName uniqueness by reading from ReadModel
• RegisterCommand: Validate UserName uniqueness again (here is the question)

Of course, we can validate UserName uniqueness by reading from ReadModel in MVC controller to improve performance and user experience. However, we still need to validate the uniqueness again in RegisterCommand, and obviously, we should NOT access ReadModel in Commands.

If we do not use Event Sourcing, we can query the domain model, so that's no a problem. But if we're using Event Sourcing, we are not able to query domain model, so how can we validate UserName uniqueness in RegisterCommand?

Notice: User class has an Id property, and UserName is not the key property of User class. We can only get the domain object by Id when using event sourcing.

BTW: In the requirement, if the entered UserName is already taken, the website should show error message "Sorry, the user name XXX is not available" to the visitor. It's not acceptable to show a message, say, "We are creating your account, please wait, we will send the registration result to you via Email later", to the visitor.

Any ideas? Many thanks!

[UPDATE]

A more complex example:

Requirement:

When placing an order, the system should check the client's ordering history, if he is a valuable client (if the client placed at least 10 orders per month in the last year, he is valuable), we make 10% off to the order.

Implementation:

We create PlaceOrderCommand, and in the command, we need to query the ordering history to see if the client is valuable. But how can we do that? We shouldn't access ReadModel in command! As Mikael said, we can use compensating commands in the account registration example, but if we also use that in this ordering example, it would be too complex, and the code might be too difficult to maintain.

回答1:

If you validate the username using the read model before you send the command, we are talking about a race condition window of a couple of hundred milliseconds where a real race condition can happen, which in my system is not handled. It is just too unlikely to happen compared to the cost of dealing with it.

However, if you feel you must handle it for some reason or if you just feel you want to know how to master such a case, here is one way:

You shouldn't access the read model from the command handler nor the domain when using event sourcing. However, what you could do is to use a domain service that would listen to the UserRegistered event in which you access the read model again and check whether the username still isn't a duplicate. Of course you need to use the UserGuid here as well as your read model might have been updated with the user you just created. If there is a duplicate found, you have the chance of sending compensating commands such as changing the username and notifying the user that the username was taken.

That is one approach to the problem.

As you probably can see, it is not possible to do this in a synchronous request-response manner. To solve that, we are using SignalR to update the UI whenever there is something we want to push to the client (if they are still connected, that is). What we do is that we let the web client subscribe to events that contain information that is useful for the client to see immediately.

Update

For the more complex case:

I would say the order placement is less complex, since you can use the read model to find out if the client is valuable before you send the command. Actually, you could query that when you load the order form since you probably want to show the client that they'll get the 10% off before they place the order. Just add a discount to the PlaceOrderCommand and perhaps a reason for the discount, so that you can track why you are cutting profits.

But then again, if you really need to calculate the discount after the order was places for some reason, again use a domain service that would listen to OrderPlacedEvent and the "compensating" command in this case would probably be a DiscountOrderCommand or something. That command would affect the Order Aggregate root and the information could be propagated to your read models.

For the duplicate username case:

You could send a ChangeUsernameCommand as the compensating command from the domain service. Or even something more specific, that would describe the reason why the username changed which also could result in the creation of an event that the web client could subscribe to so that you can let the user see that the username was a duplicate.

In the domain service context I would say that you also have the possibility to use other means to notify the user, such like sending an email which could be useful since you cannot know if the user is still connected. Maybe that notification functionality could be initiated by the very same event that the web client is subscribing to.

When it comes to SignalR, I use a SignalR Hub that the users connects to when they load a certain form. I use the SignalR Group functionality which allows me to create a group which I name the value of the Guid I send in the command. This could be the userGuid in your case. Then I have Eventhandler that subscribe to events that could be useful for the client and when an event arrives I can invoke a javascript function on all clients in the SignalR Group (which in this case would be only the one client creating the duplicate username in your case). I know it sounds complex, but it really isn't. I had it all set up in an afternoon. There are great docs and examples on the SignalR Github page.

回答2:

I think you are yet to have the mindset shift to eventual consistency and the nature of event sourcing. I had the same problem. Specifically I refused to accept that you should trust commands from the client that, using your example, say "Place this order with 10% discount" without the domain validating that the discount should go ahead. One thing that really hit home for me was something that Udi himself said to me (check the comments of the accepted answer).

Basically I came to realise that there is no reason not to trust the client; everything on the read side has been produced from the domain model, so there is no reason not to accept the commands. Whatever in the read side that says the customer qualifies for discount has been put there by the domain.

BTW: In the requirement, if the entered UserName is already taken, the website should show error message "Sorry, the user name XXX is not available" to the visitor. It's not acceptable to show a message, say, "We are creating your account, please wait, we will send the registration result to you via Email later", to the visitor.

If you are going to adopt event sourcing & eventual consistency, you will need to accept that sometimes it will not be possible to show error messages instantly after submitting a command. With the unique username example the chances of this happening are so slim (given that you check the read side before sending the command) its not worth worrying about too much, but a subsequent notification would need to be sent for this scenario, or perhaps ask them for a different username the next time they log on. The great thing about these scenarios is that it gets you thinking about business value & what's really important.

UPDATE : Oct 2015

Just wanted to add, that in actual fact, where public facing websites are concerned - indicating that an email is already taken is actually against security best practices. Instead, the registration should appear to have gone through successfully informing the user that a verification email has been sent, but in the case where the username exists, the email should inform them of this and prompt them to login or reset their password. Although this only works when using email addresses as the username, which I think is advisable for this reason.

回答3:

There is nothing wrong with creating some immediately consistent read models (e.g. not over a distributed network) that get updated in the same transaction as the command.

Having read models be eventually consistent over a distributed network helps support scaling of the read model for heavy reading systems. But there's nothing to say you can't have a domain specific read model thats immediately consistent.

The immediately consistent read model is only ever used to check and receive data before issuing a command (really its a service to the command), you should never use it for directly displaying read data to a user (i.e. from a GET web request or similar). Use eventually consitent, scaleable read models for that.

回答4:

I think for such cases, we can use a mechanism like "advisory lock with expiration".

Sample execution:

• Check username exists or not in eventually consistent read model
• If not exists; by using a redis-couchbase like keyvalue storage or cache; try to push the username as key field with some expiration.
• If successful; then raise userRegisteredEvent.
• If either username exists in read model or cache storage, inform visitor that username has taken.

Even you can use an sql database; insert username as a primary key of some lock table; and then a scheduled job can handle expirations.

回答5:

Like many others when implementing a event sourced based system we encountered the uniqueness problem.

At first I was a supporter of letting the client access the query side before sending a command in order to find out if a username is unique or not. But then I came to see that having a back-end that has zero validation on uniqueness is a bad idea. Why enforce anything at all when it's possible to post a command that would corrupt the system ? A back-end should validate all it's input else you're open for inconsistent data.

What we did was create an index at the command side. For example, in the simple case of a username that needs to be unique, just create a UserIndex with a username field. Now the command side can check if a username is already in the system or not. After the command has been executed it's safe to store the new username in the index.

Something like that could also work for the Order discount problem.

The benefits are that your command back-end properly validates all input so no inconsistent data could be stored.

A downside might be that you need an extra query for each uniqueness constraint and you are enforcing extra complexity.

回答6:

About uniqueness, I implemented the following:

• A first command like "StartUserRegistration". UserAggregate would be created no matter if user is unique or not, but with a status of RegistrationRequested.

• On "UserRegistrationStarted" an asynchronous message would be sent to a stateless service "UsernamesRegistry". would be something like "RegisterName".

• Service would try to update (no queries, "tell don't ask") table which would include a unique constraint.

• If successful, service would reply with another message (asynchronously), with a sort of authorization "UsernameRegistration", stating that username was successfully registered. You can include some requestId to keep track in case of concurrent competence (unlikely).

• The issuer of the above message has now an authorization that the name was registered by itself so now can safely mark the UserRegistration aggregate as successful. Otherwise, mark as discarded.

Wrapping up:

• This approach involves no queries.

• User registration would be always created with no validation.

• Process for confirmation would involve two asynchronous messages and one db insertion. The table is not part of a read model, but of a service.

• Finally, one asynchronous command to confirm that User is valid.

• At this point, a denormaliser could react to a UserRegistrationConfirmed event and create a read model for the user.

回答7:

Have you considered using a "working" cache as sort of an RSVP? It's hard to explain because it works in a bit of a cycle, but basically, when a new username is "claimed" (that is, the command was issued to create it), you place the username in the cache with a short expiration (long enough to account for another request getting through the queue and denormalized into the read model). If it's one service instance, then in memory would probably work, otherwise centralize it with Redis or something.

Then while the next user is filling out the form (assuming there's a front end), you asynchronously check the read model for availability of the username and alert the user if it's already taken. When the command is submitted, you check the cache (not the read model) in order to validate the request before accepting the command (before returning 202); if the name is in the cache, don't accept the command, if it's not then you add it to the cache; if adding it fails (duplicate key because some other process beat you to it), then assume the name is taken -- then respond to the client appropriately. Between the two things, I don't think there'll be much opportunity for a collision.

If there's no front end, then you can skip the async look up or at least have your API provide the endpoint to look it up. You really shouldn't be allowing the client to speak directly to the command model anyway, and placing an API in front of it would allow you to have the API to act as a mediator between the command and read hosts.