Facilitating the Spread of Knowledge and Innovation in Professional Software Development

Write for InfoQ


Choose your language

InfoQ Homepage Presentations Eventual Consistency – Don’t Be Afraid!

Eventual Consistency – Don’t Be Afraid!



Susanne Braun shares her experiences from different case studies with industry clients, and open access design guidelines developed using action research.


Susanne Braun is a software architect located in Berlin. She works as Principal Tech Lead at SAP Signavio and is mostly interested in data-intensive systems, distributed systems, and data engineering.

About the conference

Software is changing the world. QCon empowers software development by facilitating the spread of knowledge and innovation in the developer community. A practitioner-driven conference, QCon is designed for technical team leads, architects, engineering directors, and project managers who influence innovation in their teams.


Braun: I would like to directly start with a quote from Pat Helland, who said in 2009, there is an interesting connection between so-called application-based eventual consistency, and certain quality attributes. If you look at the quality attributes that he mentions, and also at the ones that are closely related, you will realize these are all quality attributes that we commonly try to achieve in modern software architecture design. Let it be cloud native, let it be reactive systems, let it be offline-first, you name it. Therefore, I believe it's really important to not only understand the main concept of eventual consistency, but also its pitfalls. What I quite often observe is that the discussion around eventual consistency is a bit one-sided. Because if you look at the product context, when it was heavily promoted by people such as Eric Brewer, you will see that there's actually more to it. Eric Brewer proposed BASE as a counter concept to ACID transaction guarantees, and he said, we will not only have to forfeit the consistency of ACID, but also the isolation of ACID. This is quite significant.


Let me very briefly contrast ACID versus BASE again. The important part around ACID consistency is that it was defined in a time when we used to design our systems with a single database that was very central, that was a single source of truth. Here, ACID consistency was all about that single source of truth always being in a consistent state, so that in particular, also, all invariants are met at all times. Eventual consistency, on the other hand, as promoted by BASE, the first thing is, eventual consistency is only applicable in the context of a distributed system. Second, it also only makes sense if we replicate data in the system, for example, in order to increase availability or something like that. What eventual consistency then says is that, actually, most of the time, the replicated data is not in a consistent or identical state. Eventual consistency guarantees that, over time, the different copies converge state-wise towards each other. This is very much about convergence actually. Isolation, this is, of course, all about concurrency control, a tough topic, as you probably agree.

Consistency vs. Isolation

Let me just summarize when we talk about consistency in distributed systems, it's very often and very much about masking actually the distributed nature of the system and rather make it again appear as one system. When it comes to isolation, here, this is all about trying to mask the effects of concurrent execution and providing the illusion to the application developer that the programs he is writing are currently the only user of the system so that he does not have to care about the different pitfalls around concurrency. Let me also bring the official definition of eventual consistency. The term was not invented by Eric Brewer. Actually, it is much older and was already in the '90s coined by Douglas Terry, who is a very famous distributed systems researcher, and also has basically worked for most of the big hyperscalers. He originally said, a system providing eventual consistency guarantees that replicas would eventually converge to a mutually consistent state. Meaning that as soon as any update activity has ceased, then we will eventually reach identical content in a different system. This definition is a bit abstract. It's not very hard to grasp and to really understand the implications for your system. I think he agreed, because a few years later he provided a more pragmatic definition of eventual consistency. He then said, ok, a system provides eventual consistency if, first of all, each update operation is eventually received by each replica. Second, what we need for the convergence guarantee is that any operations that are not commutative, meaning that they are order dependent, they of course need to be executed in the same order at each replica.

The third thing that we need to achieve convergence is of course determinism, meaning that each update operation needs to produce the same result deterministically at each replica. Remember, whenever you have to deal with eventual consistency, for example, because you are using a database management system that only provides eventual consistency, the only thing that you get is this convergence guarantee. This has quite some implications for our applications. Because you will have to handle data that is potentially outdated. If you update outdated data, you have update conflicts. If you're not able to properly resolve these update conflicts, in retrospect, you can have quite severe concurrency anomalies such as lost updates, for example. Of course, events and operations might come out of order and you also have to deal with that. You probably agree that this is potentially a huge source of human error.

The next thing is, you do not get any isolation guarantees. I'm not even talking about serializable. It's also, you don't get basic stuff, such as Repeatable Read, for example. I believe Repeatable Read is something we're constantly relying on when we write transaction programs. For example, very often, we first read some data out of the database, we evaluate some conditions on that data we have just loaded. Then, depending on whether or not our condition evaluates successfully, we do an update of the data or not. When doing so we, of course, assume that the data we have read in the beginning does not change in the course of our transaction because of other concurrent users. If you would have to care about that, then everything becomes much more complicated. This means that we have to deal with much more concurrency control things. You all know how it is. First of all, it's hard to test. If there are any issues, mostly they first emerge in production, so the first time when there's real heavy load on the system. Then, of course, it's very hard to reproduce, and even harder to debug. This is also quite a huge source of human error.

When it comes to eventual consistency, it's on one hand, very interesting and an important topic. On the other hand, it's also quite challenging. You would assume there's a lot of research going on to mitigate some of the toughest challenges, and this is also true. There's a dozen of different variants of weak consistency available. Eventual consistency is just one variant, which is probably the most popular one. However, I worked at Fraunhofer as a researcher for several years, and even me, at the time when I was a researcher, it was hard to keep up with all the different variants and things going on there. What is even worse is that these are very often just very minor improvements and nothing like the big step. What I think is rather valuable in this context is also something that Pat Helland said in this regard. He said also in 2009 that he believes that it is time for us to move past the examination of eventual consistency only in terms of updates and storage systems, but the real action comes, the interesting things come when we start to examine eventual consistency in the context of concrete domain and operational semantics.

DDD Layered Architecture

This is not straightforward what he means. Let me very quickly show you the typical three-layered architecture as promoted, for example, by domain driven design. This could be a three-layered architecture of a microservice, for example. Typically, in the top layer, we have the application layer that takes care of controlling overall application flow. We have application services there. Here, the application service methods are usually protected by transaction boundaries. Usually, when you start such an application service method, at the same time, also a database transaction is started in the infrastructure layer. When it comes to concrete domain logic, this is usually delegated to the domain layer. The application service will, for example, use a repository to query and aggregate. Aggregate again is a domain driven design concept where you cluster domain objects that have similar update behaviors or domain objects that usually need to be updated together. You load these domain objects via the repositories, and then the actual domain logic is implemented in methods of these aggregates. An aggregate method will probably be invoked, and in the course of that method, the state of the aggregate also might change. These changes, of course, need then, again, to be propagated into the infrastructure layer and translated into, for example, SQL statements. What Pat Helland was referring to is that when we deal with eventual consistency, we usually try to handle it completely in the infrastructure layer. What he believes is that it could be very beneficial if he would try to also consider it in the domain layer itself, because this is the part where the actual domain logic lives. This is the idea he was talking about.

Recap of Concurrency Control in Relational DBs

This is also not straightforward, why this should be beneficial by any means. In order to understand that, I need to very quickly recap how concurrency control is usually implemented within relational database management systems. The database management system usually tracks, on a very low-level, basic read and write operations on data items. Then it defines so-called conflict relationships between these read and write operations. For example, when two transactions write the same data item, then this is a conflict relationship, both operations are conflicting. The same is true for reading and writing the same database item by two different transactions. The only thing that is basically compatible are pure read operations. The database monitors these operations and uses the conflict relationships to build a so-called conflict graph. You all know probably that if this conflict graph does not have any cycle, this means that the concurrent execution of the transactions has actually been equivalent to a serializable or has been equivalent to a serial execution, which is commonly referred to as conflict serializability. If we know the execution of concurrent transactions is equivalent to a serial execution, we all know that by no means there can have been any concurrency anomalies because it was equivalent to serial execution. That's the idea.

Let me give you an example. Here we have two transactions that run concurrently. We have a green transaction and a blue transaction. Here we can see that the green transaction reads a data item a, and this data item a is also later written by the blue transaction. This means that if we assume a serial execution, the green transaction must have run before the blue transaction. This is why we have this dependency in the conflict graph, green transaction must happen before the blue transaction. We can do this the same way with all other operations to determine the conflict relationships. The thing here is that also the blue transaction reads a data item b that is afterwards written by the green transaction. Again, as the blue transaction does not see the changes made by the green transaction later, the blue transaction must have run before the green transaction. We get another dependency in our conflict graph. This is, of course, a contradiction. The green transaction cannot run before the blue, and at the same time, the blue transaction run before the green one. Here we would have a circle. We would know that this is not equivalent to serial execution. We don't have conflict serializability, and so, in this case, the database needs to assume that there might have been the worst case of a concurrency anomaly. A relational database management system would prevent this concurrent schedule either by means of locks or by rolling back one of the two transactions.

Business Semantics - Banking

The interesting part is now if we look at what has actually happened in the domain layer, what happened on a semantic level. Let's do that. We will see that these transactions have been two transactions from the banking domain. Both transactions withdraw some amount of money from account a and deposits some other amount of money to account b. The thing is, if we look at this in more detail, you will realize that at a semantic level, nothing bad has happened. No concurrency anomaly or something whatsoever. This is mainly due to the fact that operations in the domain layer, withdraw and deposit, these operations are commutative. They commute, it doesn't matter whether we first execute both withdraw operations and then both deposit operations, or if we execute them in an interleaved manner, as these operations are commutative, and all operations have been executed atomically, we don't have any issues there. This concurrent execution of transactions could have actually been run and there would not have happened anything bad. However, the database management system does not know anything about the semantics and the domain semantics, so it has to assume the worst case. This problem happens in practice quite often. A lot of concurrent executions could actually run in parallel without any issues, but the database management system always has to assume the worst case because it's not aware of the domain semantics and the domain logic.

Multilevel Transactions

This problem was known for quite a long time. Researchers already tried to solve that in the '90s with so-called multilevel transactions. The idea here was to exploit the semantics of operations in higher levels of our architecture. The ultimate goal was to increase concurrency, of course. What they do is that they basically decompose transactions first into operations, operations again into suboperations, until we end up with low-level operations in the infrastructure layer. Typically, we only have three layers, which pretty well matches to our three-layered architecture. Usually, we have transactions at the top. We have domain operations or business operations in the middle layer. We have low-level read and write operations in the infrastructure layer. Then at each level, also a conflict relationship is defined. In the infrastructure layer, we keep the existing conflict relationships, as I just explained before, whereas in the domain layer, we consider domain operations that are not commutative to be conflicting. Then, we just have to build a conflict graph at each level. If every graph is acyclic, then what we get is something called multilevel serializability. The interesting thing is that, in practice, multilevel serializability is nearly as good in terms of guarantees as standard serializability.

Multilevel Transactions Example

Let's look at an example. We have, again, our banking transactions. The approach is actually quite similar. We first start to analyze whether the domain operations are conflicting. We basically apply the same approach as before. We just treat domain operations as if the domain operations would be transactions. We analyze, the conflict relationship is quite similar, but this time only between domain operations and not transactions. What we get then is the conflict graph as depicted on the upper right part of the slide. We have two dependencies, but we don't have a cycle. Then, if we go one level upper or one level above, we would like to understand the conflicts between our transactions. We will see that we actually don't have domain operations there that are conflicting because all domain operations commute. We don't have any operation at all that is conflicting with another one, meaning that our conflict graph is basically empty. This also means that it doesn't have any cycles. The conflict graphs are acyclic at each level. If we had multilevel transactions, this concurrent execution of transactions could have run, and we would have a much higher concurrency as with standard serializability.

Domain Operation Design

This is, of course, a good thing. We can apply this principle not only locally to one microservice. Of course, we can also increase concurrency or concurrent execution of domain operations or update operations that run at different replicas or nodes of a system. That's why it's of course interesting to think about how we can optimize the design of our domain operations, so that these operations can run concurrently and conflict-free at different nodes of a system with eventual consistency. Let's look into that in more detail. I would like to again start with a statement from Pat Helland. He also said in 2009 that we should start to design our domain operations to be ACID 2.0 compliant, meaning that these operations should be associative, commutative, and idempotent. What we get then are truly distributed operations that can be executed at any node of the system in any order, and it still yields, in the end, the same consistent state. We also don't have to do complex conflict resolution and things like that.

An interesting application of this ACID 2.0 principle are so-called conflict-free replicated data types. These are special data types with built-in conflict resolution, because these types, they not only have normal update method or operations, but they also ship with a commutative merge operation, that is at the same time designed to be a least upper bound of two conflicting versions. I know this sounds very complex. CRDTs are grounded in algebraic theories of monotonic semilattices and things like that. What I can do is I can provide you an easy-to-understand example so that you can better understand how the mechanics of CRDTs are, let's put it like that. A very good example is actually how Amazon resolves conflicts on their shopping cart service. Assume that we have two conflicting versions of the same shopping cart, maybe this is because of a network partition. The user has been directed because of that to another replica, so now at the end, we have two versions, and we need to reconcile this in some way. What Amazon actually does, it just takes the two shopping carts, and it builds the union of these two shopping carts. As you might realize now, this union operation is an operation that is associative, commutative, and idempotent. Further, the union of two sets is also always the least upper bound of these two sets, just to give you a bit of background on the mathematics stuff.

If you look at it like this, it's quite simple. It's also very powerful. It guarantees convergence without complicated conflict resolution. It also has some drawbacks. Following this approach, it means that if the user has deleted items from the shopping cart such as, for example, the soap, could be the case that he has deleted it, then that deleted item might reappear. This is a drawback that Amazon accepts, because, for Amazon, it is so important that this shopping cart service is highly available. If it's not available for only just a few seconds or minutes, Amazon loses really a very big amount of money. For Amazon, this is a better compromise. They accept that from time to time, they might have to manually resolve these things, or that the user returns items that he has ordered.

Let me also give you a bit of a better understanding of what is a distributed operation all about. If you look at the image, I think it's very apparent that if we have multiple replicas executing domain operations concurrently, then there needs to be a point when these replicas synchronize and exchange update operations. Then of course, these operations from the other replica, or remote operations are executed, this of course means that the operations are executed in different orders at different replicas. When you have a truly distributed operation, this doesn't matter. You get the same result at each replica, independent of the execution order. Very often, this is not complicated to realize. In a lot of cases, it's actually no big deal. There are also cases where it can be quite challenging.

I just want to give you one example when it's not straightforward, so that you have a better feeling about it and know what I mean. This is an example with collaborative text editing. Assume we have two replicas, a green one and a red one. Both replicas execute operations concurrently without direct synchronization. Of course, at some point, as already mentioned, they will exchange update operations. The green operation at some point will also be executed on the red replica, but it will run on a different state than on a green replica because we have concurrent operations run before. Let me go into the details here. We assume we have a very simple text here, it's just a very brief hello world text with some typos of course. Let's assume that at the green replica, the typo in the second word was fixed. Here, an r has been inserted at position 6 of the text. On the second replica, the red one, the first word has been corrected. An e and an l has been inserted. If a node just simply re-executes the insert of the r on position 6, this is of course not what we are intending to get as a result. This is just to give you an example. This is not easy to fix, however, there exists already quite a bunch of libraries that implement CRDTs exactly aimed at collaborative text editing. That's it regarding domain operation design. Of course, we will not be able to design all our domain operations to be ACID 2.0 compliant.

Domain Data Design

The second thing that we can also do is we can also try to optimize the design of our domain objects. Also here, I would like to start with a quote from Pat Helland who said in 2016, immutability changes everything. Correspondingly, immutability is also an important classification criterion in a taxonomy that I have developed as part of my PhD at Fraunhofer. What I basically say is that when it comes to data, or aggregates, or domain objects, whatever, I basically say, there's trivial ones, and there's non-trivial ones. Of course, the trivial ones are rather easy to deal with under eventual consistency, whereas the non-trivial ones are quite challenging in relation to eventual consistency. Let me introduce the taxonomy. The first thing you'll need to ask is, is my data actually immutable? If this is the case, you have immutable aggregates in your domain model, which is a good thing, because that also means you cannot have updates, which means you cannot have update conflicts. Typical examples for immutable aggregates are domain events. Also, think of all the machine learning use cases, a lot of time-series data such as machine sensor data, market data, and so on, is actually immutable.

The next frame you need to ask yourself is, my data might not be immutable, but maybe it is the case that I can just derive or calculate this data on demand at any time out of other data. That of course means you also don't need to update these aggregates, you just calculate them whenever you need them. Here, let me give you also some examples. Of course, any kind of aggregated data comes to mind such as data warehouse reports, KPIs, and things like that. Here, again, the machine learning use case, think about your recommendations in your online store, or the timeline in your social media application, these are all generated data. The last thing that you can ask is, my data might not be immutable, I also cannot derive it, but it might be the case that there is just a single updater that from a domain perspective is allowed to make updates to that object or aggregate. Then you have a dedicated aggregate. Of course, then it's also very improbable you have conflicts, because that would mean that the same user working with different devices and doing things, this is also something that is rather not possible. Here, examples are any kind of crowd data. For example, your reviews in online stores, your social media posts, but also dedicated master data such as user profiles, account settings, and so on. All other data is rather non-trivial in relation to eventual consistency.

Still, we also divide these non-trivial aggregates into three different subclasses, depending on how challenging they are. In order to classify that, we use update frequency during peak times. Correspondingly, the probability for simultaneous updates during peak times and again corresponding the probability for concurrency anomalies. On the lowest level, we have so-called reference aggregates. This is data that is rather long-lived, that is rarely updated, but that's often also very business critical, and referenced a lot from other data. Examples here are master data such as CRM data, resources, products, things like that. Then we have activity aggregates. These aggregates capture the state of activities with multiple actors. Think about classical business processes, workflows, and things like that. Here, of course, depending on the state of the activity, we might have peak times with more requests. Correspondingly, here we have a certain probability that anomalies occurred. Here, you should probably think about a feasible strategy for conflict resolution. The most challenging ones are the collaboration result aggregates. These capture the result of collaborative knowledge work, such as, for example, multiple people offering a text. It can also be a whiteboard diagram, or a CAP model. Of course, here, the probability for conflict is very high. Correspondingly, you should have a strategy that can automatically resolve conflicts, because, otherwise, this might have a very negative impact on your user experience.

This is one thing that you might ask, which data do I have in my domain model? Another thing that you can ask is, how severe is it actually, that my data is corrupted because of conflicts and concurrency anomalies. This is something I think we need to discuss with the domain experts. It's not something we can decide on our own. Also, here, given the Amazon example, I think it's perfectly valid to accept a certain degree of corruption, if on the other hand availability is crucial for your business. The third thing I would like to ask is, how often do these different classes actually occur in practice? How many instances of them do we have? I believe, if we look at traditional information systems, such as ERP systems, CRM systems, and so on, I believe we have a comparatively high degree of non-trivial aggregates. Maybe this is also an explanation why when we designed these systems, we commonly preferred to use relational database management systems, which here provide the highest guarantees in terms of consistency and isolation.

If on the other hand we look at social media, we have the perfect opposite. The largest share of data is actually trivial. This might also be the reason why social media very often pioneered eventual consistency concepts, and might also be the reason why eventual consistency is chosen very often here. The last thing is to look at the systems that we are building right now or in the future. I believe these systems are heavily data intensive systems. I also believe that in these data intensive systems, we have all kinds of aggregate classes. For me, this is a very strong indicator that we can no longer continue with general decisions with regards to consistency, but we have to much more come to the point where we make much more situational decisions and decisions based on the context, the concrete domain semantics.

Best Practices

What I would like to share now are some best practices with you. Of course, the first thing I would recommend is to use trivial aggregates whenever feasible. Of course, it's perfectly clear we won't have a lot of examples where the whole domain model can be built out of trivial aggregates. On the other hand, I think we can more often apply them, than we usually might believe at first sight. A typical example here is stock management. You can model your domain model for stock management as given on the left-hand side with the antipattern. You have here a stock item with a counter that counts the current number of items in stock. This of course also means, if you have a lot of incoming and outgoing goods, a lot of movement here, you will have a lot of concurrency and probably you will then have conflicts on this counter attribute that are very hard to resolve. On the left-hand side, this is an activity aggregate that is, of course, not trivial, but you could design your domain model differently and only build it with trivial aggregates. One thing you could do is you could keep a history of all goods receipts and all order confirmations, these are of course immutable. Then, based on these goods receipts and order confirmations, you could in theory at any time calculate the current number of items that are in stock. That would be then a derived aggregate.

Another thing I would like to highlight is that, if you look at domain driven design and the aggregate pattern, domain driven design recommends you to cluster domain objects with a similar update behavior and put them under the same aggregate boundary. That same update behavior is very fuzzy. Here, you can see, this is a domain model from smart farming. This is about operations to be executed on fields such as, for example, a harvesting operation. Here we have fields where the operation is to be executed. We have assignment of staff and machinery. We also have documentation records. This is actually a German thing. We have a lot of rules here. One thing is that the operator of a machine needs to document, for example, how much chemistry has been put on the field. These documentation records, if you look at our taxonomy, this is actually a dedicated object, because this documentation record can only be created by the operator on the machine and only that operator would be allowed to update it, because we do not want to allow manipulation afterwards. As this is a dedicated aggregate, it's much less that we have conflicts on these domain objects. It makes a lot of sense to separate these different domain objects and have one aggregate for the dedicated data and keep the operation data in a separate aggregate as well.

We also believe that this can be generalized. This should be a general rule. It makes perfect sense to when you realize you have domain objects that are actually part of different classes of our taxonomy, then these domain objects should also be separated and put into separate aggregates with separate aggregate boundaries. Here the example, a field is, of course, master data. It's rarely updated. It is therefore a reference aggregate and should be kept separately. With that, we can of course also increase the share of trivial aggregates within our domain model. We know that a certain part of the domain model is secure, and we don't need to care about it. Here, we have one non-trivial aggregate left, which is the operation aggregate and, of course, we need to look at that in more detail and find out whether or not we need some appropriate conflict management.

The last thing I would like to share with you is, in case you really need strong transactional guarantees, please consider a pattern called primary copy replication. It's very commonly also applied in the context of microservice architectures. The idea is to have here dedicated microservice for certain aggregates or entities, for example, the order service that is the only service with the permission to update order data. All other microservices might, of course, replicate orders, but they are not allowed to update them. In case they need to update they need to delegate to the order service.


We have this documented in a guide that is freely available on GitHub, If you like what I just presented, please have a look. You can find there a much more comprehensive documentation with code examples, and text further explaining the ideas behind it. I also would like to mention that we not only developed these best practices and guidelines based on our experience in collaborations with industry clients, but we also did a number of empirical evaluations. We evaluated these methods in collaboration with industry partners, collected data systematically in the course of an actual research study. The result of the study you can find in this paper, It's freely available. It's also a good read. We got the best paper award at one of the top software engineering conferences for that. At the moment, I'm no longer with Fraunhofer as I submitted my PhD last year. I joined SAP Signavio in Berlin.


See more presentations with transcripts


Recorded at:

Dec 20, 2023