I have a general question about the CQRS paradigm in general.
I understand that a CommandBus and EventBus will decouple the domain model from our Query-side datastore, the merits of eventual consistency, and being able to denormalize the storage on the Query side to optimize reads, etc. That all sounds great.
But I wonder as I begin to expand the number of the components on the Query side responsible for updating the Query datastore, if they wouldn't start to contend with one another to perform their updates?
In other words, if we tried to use a pub/sub model for the EventBus, and there were a lot of different subscribers for a particular event type, couldn't they start to contend with one another over updating various bits of denormalized data? Wouldn't this put us in the same boat as we were before CQRS?
As I've heard it explained, it sounds like CQRS is supposed to do away with this contention all together, but is this just an ideal, and in reality we're only really minimizing it? I feel like I could be missing something here, but can't put my finger on it.
it all depends on how you have designed the infrastructure. Strictly speaking, CQRS in itself doesn't say anything about how the Query models are updated. Using Events is just a one of the options you have. CQRS doesn't say anything about dealing with contention either. It's just an architectural pattern that leaves you with more options and choices to deal with things like concurrency. In "regular" architectures, such as the layered architecture, you often don't have these options at all.
If you have scaled your command processing component out on multiple machines, you can assume that they can produce more events than a single event handling component can handle. That doesn't have to be a bad thing. It may just mean that the Query models will be updated with a slightly bigger delay during peak times. If it is a problem for you, then you should consider scaling out the query models too.
The Event Handler component themselves will not be contending with each other. They can safely process events in parallel. However, if you design the system to make them all update the same data store, your data store could be the bottleneck. Setting up a cluster or dividing the query model over different data sources altogether could be a solution to your problem.
Be careful not to prematurely optimize, though. Don't scale out until you have the figures to prove that it will help in your specific case. CQRS based architectures allow you to make a lot of choices. All you need to do is make the right choice at the right time.
So far, in the application's I am involved with, I haven't come across situations where the Query model was a bottleneck. Some of these applications produce more than 100mln events per day.
Related
I've read a few posts relating to this, but i still can't quite grasp how it all works.
Let's say for example i was building a site like Stack Overflow, with two pages => one listing all the questions, another where you ask/edit a question. A simple, CRUD-based web application.
If i used CQRS, i would have a seperate system for the read/writes, seperate DB's, etc..great.
Now, my issue comes to how to update the read state (which is, after all in a DB of it's own).
Flow i assume is something like this:
WebApp => User submits question
WebApp => System raises 'Write' event
WriteSystem => 'Write' event is picked up and saves to 'WriteDb'
WriteSystem => 'UpdateState' event raised
ReadSystem => 'UpdateState' event is picked up
ReadSystem => System updates it's own state ('ReadDb')
WebApp => Index page reads data from 'Read' system
Assuming this is correct, how is this significantly different to a CRUD system read/writing from same DB? Putting aside CQRS advantages like seperate read/write system scaling, rebuilding state, seperation of domain boundaries etc, what problems are solved from a persistence standpoint? Lock contention avoided?
I could achieve a similar advantage by either using queues to achieve single-threaded saves in a multi-threaded web app, or simply replicate data between a read/write DB, could i not?
Basically, I'm just trying to understand if i was building a CRUD-based web application, why i would care about CQRS, from a pragmatic standpoint.
Thanks!
Assuming this is correct, how is this significantly different to a CRUD system read/writing from same DB? Putting aside CQRS advantages like seperate read/write system scaling, rebuilding state, seperation of domain boundaries etc, what problems are solved from a persistence standpoint? Lock contention avoided?
The problem here is:
"Putting aside CQRS advantages …"
If you take away its advantages, it's a little bit difficult to argue what problems it solves ;-)
The key in understanding CQRS is that you separate reading data from writing data. This way you can optimize the databases as needed: Your write database is highly normalized, and hence you can easily ensure consistency. Your read database in contrast is denormalized, which makes your reads extremely simple and fast: They all become SELECT * FROM … effectively.
Under the assumption that a website as StackOverflow is way more read from than written to, this makes a lot of sense, as it allows you to optimize the system for fast responses and a great user experience, without sacrificing consistency at the same time.
Additionally, if combined with event-sourcing, this approach has other benefits, but for CQRS alone, that's it.
Shameless plug: My team and I have created a comprehensive introduction to CQRS, DDD and event-sourcing, maybe this helps to improve understanding as well. See this website for details.
A good starting point would be to review Greg Young's 2010 essay, where he tries to clarify the limited scope of the CQRS pattern.
CQRS is simply the creation of two objects where there was previously only one.... This separation however enables us to do many interesting things architecturally, the largest is that it forces a break of the mental retardation that because the two use the same data they should also use the same data model.
The idea of multiple data models is key, because you can now begin to consider using data models that are fit for purpose, rather than trying to tune a single data model to every case that you need to support.
Once we have the idea that these two objects are logically separate, we can start to consider whether they are physically separate. And that opens up a world of interesting trade offs.
what problems are solved from a persistence standpoint?
The opportunity to choose fit for purpose storage. Instead of supporting all of your use cases in your single read/write persistence store, you pull documents out of the key value store, and run graph queries out of the graph database, and full text search out of the document store, events out of the event stream....
Or not! if the cost benefit analysis tells you the work won't pay off, you have the option of serving all of your cases from a single store.
It depends on your applications needs.
A good overview and links to more resources here: https://learn.microsoft.com/en-us/azure/architecture/patterns/cqrs
When to use this pattern:
Use this pattern in the following situations:
Collaborative domains where multiple operations are performed in parallel on the same data. CQRS allows you to define commands with
enough granularity to minimize merge conflicts at the domain level
(any conflicts that do arise can be merged by the command), even when
updating what appears to be the same type of data.
Task-based user interfaces where users are guided through a complex process as a series of steps or with complex domain models.
Also, useful for teams already familiar with domain-driven design
(DDD) techniques. The write model has a full command-processing stack
with business logic, input validation, and business validation to
ensure that everything is always consistent for each of the aggregates
(each cluster of associated objects treated as a unit for data
changes) in the write model. The read model has no business logic or
validation stack and just returns a DTO for use in a view model. The
read model is eventually consistent with the write model.
Scenarios where performance of data reads must be fine tuned separately from performance of data writes, especially when the
read/write ratio is very high, and when horizontal scaling is
required. For example, in many systems the number of read operations
is many times greater that the number of write operations. To
accommodate this, consider scaling out the read model, but running the
write model on only one or a few instances. A small number of write
model instances also helps to minimize the occurrence of merge
conflicts.
Scenarios where one team of developers can focus on the complex domain model that is part of the write model, and another team can
focus on the read model and the user interfaces.
Scenarios where the system is expected to evolve over time and might contain multiple versions of the model, or where business rules
change regularly.
Integration with other systems, especially in combination with event sourcing, where the temporal failure of one subsystem shouldn't
affect the availability of the others.
This pattern isn't recommended in the following situations:
Where the domain or the business rules are simple.
Where a simple CRUD-style user interface and the related data access operations are sufficient.
For implementation across the whole system. There are specific components of an overall data management scenario where CQRS can be
useful, but it can add considerable and unnecessary complexity when it
isn't required.
We have microservices, each generating events that are being stored by a event-sourcing repository. We use Cassandra to store the event data.
As you may know, the order of the events is important.
When we generate these events from different services running in different machines, how to manage the time (timestamp) going out of sync across these thereby resulting in an event order mismatch.
As you may know, the order of the events is important.
In some cases - but you'll want to be careful not to confuse time, order, and correlation.
When we generate these events from different services running in different machines, how to manage the time (timestamp) going out of sync across these thereby resulting in an event order mismatch.
Give up the idea that there is an "order" to events that are happening in different places. There is no now.
Udi Dahan on race conditions in the business world:
A microsecond difference in timing shouldn’t make a difference to core business behaviors.
If your micro service boundaries are correct, then events happening in two difference services at about the same time are coincident -- there isn't one correct ordering of them, because (to stretch an analogy) they are in different light cones. The only ordering that is inherently real is that within a single aggregate event history.
What can make real sense is tracking causation; these changes in this book of record are a reaction to those changes in that book of record.
One simple form of this is to track happens-before, which is where ideas like vector clocks begin to appear.
In most discussions that I have seen, this information would be passed along as meta data of the recorded events.
This is typically done via vector clocks:
A vector clock is an algorithm for generating a partial ordering of events in a distributed system and detecting causality violations.
If I understand your problem correctly, you're trying to guard writes, i.e. to make sure that a microservice instance is up to date with all the relevant events before making another write.
In that case, have a look at lightweight transactions, which can be used to implement optimistic locking in Cassandra.
This talk by Christopher Batey is a very good start.
I have seen ORM use a unit of work to commit multiple repositories in a single step.
I have also seen DDD and the use of aggregate roots saved via repositories, when using event stores persistence conceptually becomes quite clear to understand.
I always need to write data access code and whilst I am familiar with ORM, I am new to domain driven design and event sourcing - event sourcing is great, but does come with a lot of infrastructure.
Ultimately I would like to some rules to help decide at what point (code size, number of database entities) when DDD+ES becomes worth the extra effort over CRUD systems.
To help decide my questions are as follows:
I haven't seen aggregate roots combined in to a single unit of work, is this avoided? If so what problems can this cause?
In DDD a customer entity may have addresses and phones embedded within it (value objects), whereas in ORM there is a unit of work with customer, phone and address repositories. What is the best way to explain and understand these different approaches?
Can ORM use multiple different unit of works (each referencing relevant and related repositories/tables) to represent an aggregate root?
What are the pain/warning signs to look out for with impedance mismatch from my domain to ORM, at which point we may consider switching to an event store?
An aggregate defines a consistency boundary. In NoSQL databases, it is usually not possible to commit multiple entities per transaction. Therefore, in DDD with NoSQL, it is desirable to only have a single aggregate in a unit of work while updates to entities external to the aggregate at hand are delivered in an eventually consistent manner.
If addresses and phones are value objects then they shouldn't have repositories. In the ORM, they would be mapped as components of a parent entity not a separate mapping.
I'm not sure what you'd achieve this way?
One pain point that naturally leads to event sourcing is the need to preserve all state changes in an aggregate. Furthermore, event sourcing and the concept of domain events in general provide a different domain modelling methodology focused on behavior rather than state. I'd consider ES when there is potential business value in preserving all state changes. If you are willing to make the initial infrastructure investment, ES can in many ways be simpler by avoiding ORM madness. Think of CRUD as event sourcing with only 4 event types, or even 2 (read, update). Beyond the most basic domains, it is desirable to have more context beyond changes to data which leads you to ES.
I'll admit that I am still quite a newbie with DDD and even more so with CQRS. I also realize that DDD and/or CQRS might not be the right approach to every problem. Nevertheless, I like the principals but have some questions in the context of a current project.
The solution is a simulator that generates performance data based on the current configuration. Administrators can create and modify the specifications for simulations. Testers set some environmental conditions and run the simulator. The results are captured, aggregated and reported.
The solution consists of 3 component areas each with their own use-cases, domain logic and supporting data structure. As a result, a modular designed seems appealing as a way to segregate logic and separate concerns.
The first area would be the administrative aspect which allows users to create and modify the specifications. This would be a CRUD heavy 'module'.
The second area would be for executing the simulations. The domain model would be similar to the first area but optimized for executing the simulation as opposed to providing a convenient model for editing.
The third area is reporting.
From this I believe that I have three Bounding Contexts, yes? I have three clear entry points into the application, three sets of domain logic and three different data models to support the domain logic.
My first instinct is to follow these lines and create three modules (assemblies) that encapsulate the domain layer for each area. Should I also have three separate databases? Maybe more than three to support write versus read?
I gather this may be preferred for CQRS but am not sure how to go about it. It appears to me that CQRS suggests a set of back-end processes that move data around. But if that's the case, and data persistence is cross-cutting (as DDD suggests), then doesn't my data access code need awareness of all of the domain objects? If so, then is there a benefit to having separate modules?
Finally, something I failed to mention earlier is that specifications are considered 'drafts' until published, which makes then available for simulation. My PublishingService needs to have knowledge of the domain model for both the first and second areas so that when it responds to the SpecificationPublishedEvent, it can read the specification, translate the model and persist it for execution. This makes me think I don't have three bounding contexts after all. Or am I missing something in my analysis?
You may have a modular UI for this, but I don't see three separate domains in what you are describing necessarily.
First off, in CQRS reporting is not directly a domain model concern, it is a facet of the separated Read Model which takes on the responsibility of presenting the domain state optimized for reporting.
Second just because you have different things happening in the domain is not necessarily a reason to bound them away from each other. I'd take a read through the blue DDD book to get a bit better feel for what BCs look like.
I don't really understand your domain well enough but I'll try to give some general suggestions.
Start with where you talked about your PublishingService. I see a Specification aggregate root which takes a few commands that probably look like CreateNewSpecification, UpdateSpecification and PublishSpecification.
The events look similar and probably feel redundant: SpecificationCreated, SpecificationUpdated, SpecificationPublished. Which kind of sucks but a CRUD heavy model doesn't have very interesting behaviors. I'd also suggest finding an automated way to deal with model/schema changes on this aggregate which will be tedious if you don't use code generation, or handle the changes in a dynamic *emphasized text*way that doesn't require you to build new events each time.
Also you might just consider not using event sourcing for such an aggregate root since it is so CRUD heavy.
The second thing you describe seems to be about starting a simulation which will run based on a Specification and produce data during that simulation (I assume). An event driven architecture makes sense here to decouple updating the reporting data from the process that is producing the data. This has huge benefits if you are producing large amounts of data to process.
However it doesn't sound like a Simulation is necessarily the kind of AR that would benefit from Event Sourcing either. For a couple reasons:
Simulation really takes only one Command which is something like StartSimulation
Simulation then produces events over it's life-time which represent what is happening internally with the simulation
Simulation doesn't seem to ever receive any other Commands that could depend on the current state of the Simulation
Simulation is not interacted with by multiple clients/users simultaneously and as we pointed out it isn't really interacted with at all
In general, domain modeling is very specific to each individual project so it's hard to give you all the information you need to build your domain model. It will come as a result of spending a great deal of time trying to understand your user's needs and the problem they are trying to solve with the software. It likely will go through multiple refinements as you develop insights into their process.
Our Domain has a need to deal with large amounts (possibly more than 1000 records worth) of objects as domain concepts. This is largely historical data that Domain business logic needs do use. Normally this kind of processing depends on a Stored Procedure or some other service to do this kind of work, but since it is all intimately Domain Related, and we want to maintain the validity of the Model, we'd like to find a solution that allows the Aggregate to manage all of the business logic and rules required to work with the data.
Essentially, we're talking about past transaction data. Our idea was to build a lightweight class and create an instance for each transaction we need to work with from the database. We're uncomfortable with this because of the volume of objects we'd be instantiating and the potential performance hit, but we're equally uncomfortable with offloading this Domain logic to a stored procedure since that would break the consistency of our Model.
Any ideas on how we can approach this?
"1000" isn't really that big a number when it comes to simple objects. I know that a given thread in the system I work on may be holding on to tens of thousands of domain objects at a given time, all while other threads are doing the same at the same time. By the time you consider all of the different things going on in a reasonably complicated application, 1000 objects is kind of a drop in the bucket.
YMMV depending on what sort of resources those objects are holding on to, system load, hard performance requirements, or any number of other factors, but if, as you say, they're just "lightweight" objects, I'd make sure you actually have a performance problem on your hands before you try getting too fancy.
Lazy loading is one technique for mitigating this problem and most of the popular object-relational management solutions implement it. It has detractors (for example, see this answer to Lazy loading - what’s the best approach?), but others consider lazy loading indispensable.
Pros
Can reduce the memory footprint of your aggregates to a manageable level.
Lets your ORM infrastructure manage your units of work for you.
In cases where you don't need a lot of child data, it can be faster than fully materializing ("hydrating") your aggregate root.
Cons
Chattier that materializing your aggregates all at once. You make a lot of small trips to the database.
Usually requires architectural changes to your domain entity classes, which can compromise your own design. (For example, NHibernate just requires you to expose a default constructor make your entities virtual to take advantage of lazy loading - but I've seen other solutions that are much more intrusive).
By contrast, another approach would be to create multiple classes to represent each entity. These classes would essentially be partial aggregates tailored to specific use cases. The main drawback to this is that you risk inflating the number of classes and the amount of logic that your domain clients need to deal with.
When you say 1000 records worth, do you mean 1000 tables or 1000 rows? How much data would be loaded into memory?
It all depends on the memory footprint of your objects. Lazy loading can indeed help, if the objects in question references other objects which are not of interest in your process.
If you end up with a performance hog, you must ask yourself (or perhaps your client) if the process must run synchronously, or if it can be offloaded to a batch process somewhere else.
Using DDD, How Does One Implement Batch Processing?