Architecting for the Cloud

The biggest difference between cloud-based applications and the applications running in your data center is scalability. The cloud offers scalability on demand, allowing you to expand and contract your application as load fluctuates. This scalability is what makes the cloud appealing, but it can’t be achieved by simply lifting your existing application to the cloud. In order to take advantage of what the cloud has to offer, you need to re-architect your application around scalability. The other business benefit comes in terms of price, as in the cloud costs scale linearly with demand.

Sample Architecture of a Cloud-Based Application

Designing an application for the cloud often requires re-architecting your application around scalability. The figure below shows what the architecture of a highly scalable cloud-based application might look like.

The Client Tier: The client tier contains user interfaces for your target platforms, which may include a web-based user interface, a mobile user interface, or even a thick client user interface. There will typically be a web application that performs actions such as user management, session management, and page construction. But for the rest of the interactions the client makes RESTful service calls into the server.

Services: The server is composed of both caching services, from which the clients read data, that host the most recently known good state of all of the systems of record, and aggregate services that interact directly with the systems of record for destructive operations (operations that change the state of the systems of record).

Systems of Record: The systems of record are your domain-specific servers that drive your business functions. These may include user management CRM systems, purchasing systems, reservation systems, and so forth. While these can be new systems in the application you’re building, they are most likely legacy systems with which your application needs to interact. The aggregate services are responsible for abstracting your application from the peculiarities of the systems of record and providing a consistent front-end for your application.

ESB: When systems of record change data, such as by creating a new purchase order, a user “liking” an item, or a user purchasing an airline ticket, the system of record raises an event to a topic. This is where the idea of an event-driven architecture (EDA) comes to the forefront of your application: when the system of record makes a change that other systems may be interested in, it raises an event, and any system interested in that system of record listens for changes and responds accordingly. This is also the reason for using topics rather than using queues: queues support point-to-point messaging whereas topics support publish-subscribe messaging/eventing. If you don’t know who all of your subscribers are when building your application (which you shouldn’t, according to EDA) then publishing to a topic means that anyone can later integrate with your application by subscribing to your topic.

Whenever interfacing with legacy systems, it is desirable to shield the legacy system from load. Therefore, we implement a caching system that maintains the currently known good state of all of the systems of record. And this caching system utilizes the EDA paradigm to listen to changes in the systems of record and update the versions of the data it hosts to match the data in the systems of record. This is a powerful strategy, but it also changes the consistency model from being consistent to being eventually consistent. To illustrate what this means, consider posting an update on your favorite social media site: you may see it immediately, but it may take a few seconds or even a couple minutes before your friends see it. The data will eventually be consistent, but there will be times when the data you see and the data your friends see doesn’t match. If you can tolerate this type consistency then you can reap huge scalability benefits.

NoSQL: Finally, there are many storage options available, but if your application needs to store a huge amount of data it is far easier to scale by using a NoSQL document store. There are various NoSQL document stores, and the one you choose will match the nature of your data. For example, MongoDB is good for storing searchable documents, Neo4J is good at storing highly inter-related data, and Cassandra is good at storing key/value pairs. I typically also recommend some form of search index, such as Solr, to accelerate queries to frequently accessed data.

Let’s begin our deep-dive investigation into this architecture by reviewing service-oriented architectures and REST.

REpresentational State Transfer (REST)

The best pattern for dividing an application into tiers is to use a service-oriented architecture (SOA). There are two main options for this, SOAP and REST. There are many reasons to use each protocol that I won’t go into here, but for our purposes REST is the better choice because it is more scalable.

REST was defined in 2000 by Roy Fielding in his doctoral dissertation and is an architectural style that models elements as a distributed hypermedia system that rides on top of HTTP. Rather than thinking about services and service interfaces, REST defines its interface in terms of resources, and services define how we interact with these resources. HTTP serves as the foundation for RESTful interactions and RESTful services use the HTTP verbs to interact with resources, which are summarized as follows:

• GET: retrieve a resource

• POST: create a resource

• PUT: update a resource

• PATCH: partially update a resource

• DELETE: delete a resource

• HEAD: does this resource exist OR has it changed?

• OPTIONS: what HTTP verbs can I use with this resource

For example, I might create an Order using a POST, retrieve an Order using a GET, change the product type of the Order using a PATCH, replace the entire Order using a PUT, delete an Order using a DELETE, send a version (passing the version as an Entity Tag or eTag) to see if an Order has changed using a HEAD, and discover permissible Order operations using OPTIONS. The point is that the Order resource is well defined and then the HTTP verbs are used to manipulate that resource.

In addition to keeping application resources and interactions clean, using the HTTP verbs can greatly enhance performance. Specifically, if you define a time-to-live (TTL) on your resources, then HTTP GETs can be cached by the client or by an HTTP cache, which offloads the server from constantly rebuilding the same resource.

REST defines three maturity levels, affectionately known as the Richardson Maturity Model (because it was developed by Leonard Richardson):

1. Define resources

2. Properly use the HTTP verbs

3. Hypermedia Controls

Thus far we have reviewed levels 1 and 2, but what really makes REST powerful is level 3. Hypermedia controls allow resources to define business-specific operations or “next states” for resources. So, as a consumer of a service, you can automatically discover what you can do with the resources. Making resources self-documenting enables you to more easily partition your application into reusable components (and hence makes it easier to divide your application into tiers).

Sideline: you may have heard the acronym HATEOAS, which stands for Hypermedia as the Engine of Application State. HATEOAS is the principle that clients can interact with an application entirely through the hypermedia links that the application provides. This is essentially the formalization of level 3 of the Richardson Maturity Model.

RESTful resources maintain their own state so RESTful web services (the operations that manipulate RESTful resources) can remain stateless. Stateless-ness is a core requirement of scalability because it means that any service instance can respond to any request. Thus, if you need more capacity on any service tier, you can add additional virtual machines to that tier to distribute the load. To illustrate why this is important, let’s consider a counter-example: the behavior of stateful servers. When a server is stateful then it maintains some client state, which means that subsequent requests by a client to that server need to be sent to that specific server instance. If that tier becomes overloaded then adding new server instances to the tier may help new client requests, but will not help existing client requests because the load cannot be easily redistributed.

Furthermore, the resiliency requirements of stateful servers hinder scalability because of fail-over options. What happens if the server to which your client is connected goes down? As an application architect, you want to ensure that client state is not lost, so how to we gracefully fail-over to another server instance? The answer is that we need to replicate client state across multiple server instances (or at least one other instance) and then define a fail-over strategy so that the application automatically redirects client traffic to the failed-over server. The replication overhead and network chatter between replicated servers means that no matter how optimal the implementation, scalability can never be linear with this approach.

Stateless servers do not suffer from this limitation, which is another benefit to embracing a RESTful architecture. REST is the first step in defining a cloud-based scalable architecture. The next step is creating an event-driven architecture.

Deploying to the Cloud

This paper has presented an overview of a cloud-based architecture and provided a cursory look at REST and EDA. Now let’s review how such an application can be deployed to and leverage the power of the cloud.

Deploying RESTful Services

RESTful web services, or the operations that manage RESTful resources, are deployed to a web container and should be placed in front of the data store that contains their data. These web services are themselves stateless and only reflect the state of the underlying data they expose, so you are able to use as many instances of these servers as you need. In a cloud-based deployment, start enough server instances to handle your normal load and then configure the elasticity of those services so that new server instances are added as these services become saturated and the number of server instances is reduced when load returns to normal. The best indicator of saturation is the response time of the services, although system resources such as CPU, physical memory, and VM memory are good indicators to monitor as well. As you are scaling these services, always be cognizant of the performance of the underlying data stores that the services are calling and do not bring those data stores to their knees.

The above graphics shows that the services that interact with Document Store 1 can be deployed separately, and thus scaled independently, from the services that interact with Document Store 2. If Service Tier 1 needs more capacity then add more server instances to Service Tier 1 and then distribute load to the new servers.

Deploying an ESB

The choice of whether or not to use an ESB will dictate the EDA requirements for your cloud-based deployment. If you do opt for an ESB, consider partitioning the ESB based on function so that excessive load on one segment does not take down other segments.

 The importance of segmentation is to isolate the load generated by System 1 from the load generated by System 2. Or stated another way, if System 1 generates enough load to slow down the ESB, it will slow down its own segment, but not System 2’s segment, which is running on its own hardware. In our initial deployment we had all of our systems publishing to a single segment, which exhibited just this behavior! Additionally, with segmentations, you are able to scale each segment independently by adding multiple servers to that segment (if your ESB vendor supports this).

Cloud-based applications are different from traditional applications because they have different scalability requirements. Namely, cloud-based applications must be resilient enough to handle servers coming and going at will, must be loosely-coupled, must be as stateless as possible, must expect and plan for failure, and must be able to scale from a handful of server to tens of thousands of servers.

There is no single correct architecture for cloud-based applications, but this paper presented an architecture that has proven successful in practice making use of RESTful services and an event-driven architecture. While there is much, much more you can do with the architecture of your cloud application, REST and EDA are the basic tools you’ll need to build a scalable application in the cloud.