Data-Oriented Architecture

There’s a little-known pattern in software architecture that deserves more attention. Data-Oriented Architecture was first described by Rajive Joshi in a 2007 whitepaper at RTI , and again in 2017 by Christian Vorhemus and Erich Schikuta at the University of Vienna in this iiWAS paper . DOA is an inversion of the traditional dichotomy between a monolithic binary and data store (monolithic architecture) on the one hand, and small, distributed, independent binaries each with their own data stores (microservices, and service-oriented architecture) on the other. In data-oriented architecture, a monolithic data store is the sole source of state in the system, which is being acted on by loosely-coupled, stateless microservices.

I was lucky that my former employer also fell upon this unusual architectural choice. It was a reminder that things can be done differently. Data-oriented architecture isn’t a silver bullet by any means; it has its own unique set of costs and benefits. What I did find, though, is that a lot of large companies and ecosystems are stuck at exactly the type of bottleneck that data-oriented architecture is meant to resolve.

A quick note on Monolithic Architecture

Since a lot of architectures are often defined in contrast to monolithic architectures, it’s worth spending some time describing it. It is, after all, the fabled state of nature of server-side software development.

In a monolithic service, the bulk of server-side code is in one program that is talking to one or more databases, handling multiple aspects of a functional computation. Imagine a trading system that receives requests from customers to buy or sell some security, prices them, and fills their orders.

Within a monolithic server, code could still be componentized and separated into individual modules, but there’s no forced API boundary between the different components of the program. The only rigidly defined APIs in the program are typically either (a) between the UI and the server (in whatever REST/HTTP protocol they decide on), (b) between the server and the data stores (in whatever query language they decide on), or (c) between the server and its external dependencies.

Service-Oriented Architecture and microservices

Service-oriented architectures(SOA), on the other hand, break up monolithic programs into services for each independent, componentized function. In our trading app, we might have a separate service as the external API receiving requests and handling responses to the customer, a second system receiving prices and other information about the market, a third system tracking orders, risk, etc. The interface between each of these services is a formally-defined API layer. Services typically communicate one-on-one through RPCs, although other techniques like message-passing and pubsub are common.

Service oriented architectures allow different services to be developed and reasoned about independently (and in parallel), if needed. The services are loosely-coupled, which means that a totally new service can now reuse the other services.

As each service in an SOA defines its own API, each service can be independently accessed and interacted with. Developers debugging or mocking individual pieces can call individual components separately, and new flows can re-compose these individual services to enable new behaviors.

Microservicesare a type of service-oriented architecture. Depending on who you ask, they might differ from SOA because the services are meant to be especially small and lightweight, or that they’re just a synonym of SOA altogether.

Problems of Scale

In SOA, individual components communicate directly with each other through a specific API defined by each of the components. To communicate, each component is individually addressable (i.e. using an IP address, service address, or some other internal identifier to send requests/messages back and forth). This means that each component in the architecture needs to know about its dependencies, and needs to integrate specifically with them.

Depending on the topology of the architecture, this can mean that an additional component might need to know about all the previous components. Also, this can mean that replacing an individual service that N other components already talk to can be challenging: you’d need to take care in preserving whatever ad hoc API you defined, and make sure you have a migration plan for moving each of the components from addressing the old service to addressing the new one. Since service-to-service APIs are ad hoc , it often means that RPCs between components can be arbitrarily complicated, which potentially increases the surface area of possible API changes in the future. Each API change in a service depended on by many others is a significant undertaking.

What I’m getting at here is that as a microservices ecosystem grows, it starts being susceptible to the following problems at scale:

1. N ² growth in complexity of integration as the number of components grow,
2. The shape of a network becomes hard to reason about a priori ; i.e. creating or maintaining a testing environment or sandbox will require a lot of reasoning to make sure no component within a graph has an external dependency

A few friends volunteered their own problems with Service-oriented architecture at scale:

Another problem with growth of SOA that I’ve seen is dependency cycles between services. Because you roll them out independently and rarely bring up the whole system from scratch, its easy to introduce cycles and break the DAG.

Another problem with SOAs at scale that may be worth calling out is that they require you to know all of your future customer workflows ahead of time. If you don’t—and you segregate data for a single workflow across multiple verticals—then you’re stuck with either performance problems trying to guarantee transactionality across multiple persistent stores or redefining which vertical masters duplicate (cached but really persisted in a db) data.

Data-Oriented Architecture

In Data-Oriented Architecture (DOA), systems are still organized around small, loosely-coupled components, as in SOA microservices. But DOA departs from microservices in two key ways:

1. Components are always stateless

   Rather than componentizing and federating data stores for each relevant component, DOA mandates describing the data or state-layer in terms of a centrally managed global schema.

2. Component-to-component interaction is minimized , instead favoring interaction through the data layer

   In our trading system, a component receiving prices for different securities is just publishing prices in a canonical form in our data store. A system can consume these prices by querying the data layer for prices, rather than request prices from a specific service (or set of services) through a specific API.

   Here, the integration cost is linearized. A DOA schema change means that up to N components might need to be updated, rather up to N ² connections between them.

Where this really shines is when individual high-level data types are populated by different providers. If we replace one service with one table, then we haven’t simplified things much. The benefit is if there are multiple sources of the same general data type. If a trading system is connecting to multiple marketplaces, who each publish requests from customers to an RFQ table, then downstream systems can query this one table and not worry about where a customer request is coming from.

Types of Component Communication

Since component-to-component interaction is minimized in DOA, how would one replace the inter-component communication in SOA today with interaction through the data layer?

1. Data Produces and Consumes

Organizing components into producers and consumers of data is the main way to design a DOA system.

If you can, at a high level, write your business logic as a series of map , filter , reduce , flatMap , and other monadic operations, you can write your DOA system as a series of components , each which queries or subscribes to its inputs and produces its outputs. The challenge in DOA is that these intermediate steps are visible, queryable data—which means that it needs to be well-encapsulated, well-represented, and corresponds to a particular business-logic concept. The advantage, though, is that the behavior of the system is externally observable, traceable, and auditable.

In a SOA trading system, a component taking orders from a marketplace might make RPCs to figure out how to price, quote, or trade on an order. In DOA, a microservice takes requests from marketplaces (usually in a SOA manner) and produces RFQs, while other producers are producing pricing data, etc. Another microservice queries for RFQs, joined with all their pricing and outputs quotes, orders, or whatever custom response datums are needed.

2. Triggering Actions and Behaviors

Sometimes, the simplest way to think about communication between components is as an RPC. While a well-designed DOA systemshould see a majority of its inter-component communication replaced by producer/consumer paradigms, you might still need direct ways for component X to tell Y to do Z.

First, it’s important to consider if RPCs can be reorganized as events and their effects . I.e., asking if, rather than component X sending RPCs to component Y where event E happens, can X instead produce events E , and have component Y drive the responses by consuming these events?

This approach, which I’ll call data-based events , can be a powerful inversion of how we typically have component communication. The reason it is so powerful is because it allows us to take the term loosely-coupled to the next level. Systems don’t need to know who is consuming their events (rather than an RPC caller who absolutely needs to know who they’re calling), and producers don’t need to worry about where the events are coming from, just the business-logic semantic meaning of those events.

There is, of course, a naïve way to implement data-based events, where each event is persisted to a database in its own table corresponding 1:1 with a serialized version of a the RPC request. In that case, data-based events don’t decouple a system at all. For data-based events to work, translating a request/response into persisted events require them to be meaningful business-logic constructs.

Data-based events might not be a good match. For example, if you actually want to trigger a behavior in a specific component. In those cases, leaving a limited number of actual component-to-component RPCs is probably still desirable.

Case Studies where Data-oriented architecture shines

High Integration Problem Spaces

Part of why I keep mentioning trading/finance software as an example is because finance often requires a large integration surface area. A typical sell-side firm allowing smaller customers to trade often integrate with many marketplaces to interact with customers, and many liquidity providers to get prices and place orders. The business logic that needs to happen between when a request comes in on a marketplace and a response comes out to the customer is a complex, multi-stage process.

In a high-integration problem space, individual services might need to know about a lot of other services. To avoid an O(N ² ) integration cost, and to avoid complex individual services with a high fan-out ratio, rearchitecting a system around data producers and consumers allows integrations to be simpler. If a new integration comes along, rather than having to edit N new systems, or one system which has a complex fan-out to N other systems, the integration process can involve writing one adaptor that produces data in the common DOA schema, and consumes the final output and renders it in the right wire format.

Implicitly, there’s a new kind of complexity in integration: thinking about the schema. Any new integration should feel native to your system, and your schema should be extended without adding shims, hacks, and special cases. This, in and of itself, is a hard exercise. When the number of integrations is sufficiently high, however, the difficulty amortizes and is often worthwhile.

Sandboxing Data, and Reasoning about Data Isolation

If you’re prototyping or testing things manually, you’re hopefully doing it outside of production. Yet the way some SOA ecosystems are architected often means that it’s not easy to know what environment a service is in, or if a particular environment is self-contained at all.

An environment is an internally-consistent, consistently-connected collection of services, usually/ideally structured in the same topology as production. Since SOA services are typically independently addressable, environment consistency predicates that every service must agree with every other service in an environment on which address to call for what. The RPC, pubsub, and data flow must not leak from one environment to another.

There are obviously ways around this in SOA, like shifting to service registries that generate the right configuration for services, or, when services are accessed through a URI, hiding direct service addresses in favor of different paths under an environment prefix.

In DOA, however, the concept of an environment is much simpler. Knowing which data store layer a component connects to is sufficient to describe what environment it is in. Since all components store no state internally, data is isolated by definition. There’s no danger of leaking data from one environment to another, as components only communicate via the data store.

Data-oriented Architecture is closer than you think

There are a lot of common examples that approximate data-oriented architecture today. A data monolith where all (or most) data is persisted in one large data store, often signals that a system’s architecture is approaching DOA.

Knowledge Graphs , for example, are a generalized data monolith. That said, they’re often not generic enough; with a lot of state related to business logic potentially missing.

GraphQL is often used as a normalized datastore layer, like a data monolith. The degree to which GraphQL can be a successful backend to a DOA system is more about the system’s choices of schema design: choosing generalizable schema and tables related to business logic concepts, rather than schema and tables specific to the specific source of this data.

It’s all about the trade-off

This architecture is not a magic bullet. Where data-oriented architecture erases classes of problems, new ones arise: It requires the designer to think hard about data ownership. Situations where multiple-writers are modifying the same record can be tenuous, often encouraging systems to carefully partition record write ownership. And because inter-component APIs are encoded in the data, there is one shared, global schema that must be well-thought out.

I’m reminded of Google’s Protocol Buffers documentation, which, in a discussion of marking fields in a schema as required warns: Required Is Forever . At Broadway Technology, CTO Joshua Walsky used to say something similar about DOA schema: Data Is Forever . For reasons similar to the Protobuf warning, it turns out, removing a column from a table in a loosely-coupled distributed system is really , really hard.

My two cents: If you find yourself worrying about your architecture scaling horizontally , you might think about designing it with a data monolith in its center.