Azure Cosmos DB is rapidly growing in popularity, and for good reason. Microsoft’s globally distributed, multi-model database service has massively scalable storage and throughput, provides a sliding scale for consistency, is fully and automatically indexed, and it exposes multiple APIs. Throw in a server-side programming model for ACID transactions with stored procedures, triggers, and user-defined functions, along with 99.999% SLAs on availability, throughput, latency, and consistency, and it’s easy to see why Cosmos DB is fast winning the hearts of developers and solution architects alike.
Yet still today, one of the most overlooked capabilities in Cosmos DB is its change feed. This little gem sits quietly behind every container in your database, watches for changes, and maintains a persistent record of them in the order they occur. This provides you with a reliable mechanism to consume a continuous and incremental feed of changes, as documents are actively written or modified in any container.
There are numerous use cases for this, and I’ll call out a few of the most common ones in a moment. But all of them share a need to respond to changes made to a Cosmos DB container. And the first thought that comes to the mind of a relational database developer is to use a trigger for this. Cosmos DB supports triggers as part of its server-side programming model, so it could be natural to think of using this feature to consume changes in real time when you need to.
But with change feed, you’ve got a reliable mechanism for retrieving changes made to any container, all the way back to the beginning of time. You can write code (in your preferred language) that consumes the change feed to process it as needed and deploy that code to run on Azure. This paves an easy path for you to build many different solutions for many different scenarios.
Scenarios for Change Feed
Some of the more common use cases for change feed include:
- Replicating containers for multiple partition keys
- Denormalizing a document data model across containers
- Triggering API calls for an event-driven architecture
- Real time stream processing and materialized view patterns
- Moving or archiving data to secondary data stores
Each of these deserves their own focused blog post (and will hopefully get one). For the broader context of this overview post, however, I’ll discuss them each at high level.
Replicating containers for multiple partition keys
One of the most (perhaps the most) important things you need to do when creating a container is to decide on an appropriate partition key – a single property in your data that the container will be partitioned by.
Now sometimes this is easy, and sometimes it is not. In order to settle on the correct choice, you need a clear understanding of how your data is used (written to and queried), and how horizontal partitioning works. You can read all about this in my previous blog post, Horizontal Partitioning in Azure Cosmos DB.
But what do you do when you can’t decide? What if there are two properties that make good choices, one for write performance, and another that’s better for query performance? This can lead to “analysis paralysis,” a scary condition that is fortunately and easily remedied using change feed.
All you do is create two containers, each partitioned by a different partition key. The first container uses a partition key that’s optimized for writes (it may also be appropriate for certain types of queries as well), while the second one uses a partition key optimized for most typical queries. Simply use change feed to monitor changes made to the container as the writes occur and replicate the changes out to the second container.
Your application then writes to the first container and queries from the second container, simple as that! I’ll show you a detailed walkthrough of exactly how to implement this using C# and the Change Feed Processor Library with Azure Functions in my next post.
Denormalizing a document data model across containers
Developers with a background in relational database design often struggle initially with the denormalized approach to data modeling in the NoSQL world of JSON documents. I personally empathize; from my own experience, I know that it can be difficult at first to embrace concepts that run contrary to deeply engrained practices that span decades of experience in the field.
Data duplication is a case in point, where this is considered a big no-no in the normalized world of relational databases with operational workloads. But with NoSQL, we often deliberately duplicate data in order to avoid expensive additional lookups. There is no concept of a JOIN in any NoSQL database engine, and we can avoid having to perform our own “manual” joins if we simply duplicate the same information across documents in different containers.
This is a somewhat finer-grained version of the previous scenario, which replicates entire documents between two containers. In this case, we have different documents in each container, with data fragments from changed documents in one container being replicated into other (related) documents in another container.
But how do you ensure that the duplicated data remains in sync as changes occur in the source container? Why, change feed of course! Just monitor the source container and update the target container. I’ll show you exactly how to do this in a future post.
Triggering API calls for an event-driven architecture
In this scenario, you source events to a set of microservices, each with a single responsibility. For instance, an ecommerce website with a large-scale order processing pipeline. The pipeline is broken up into a set of smaller microservices, each of which can be scaled out independently. Each microservice is responsible for a single task in the pipeline, such as calculating tax on each order, generating tax audit records, processing each order payment, sending orders off to a fulfillment center, and generating shipping notifications.
Thus, you potentially have N microservices communicating with up to N-1 other microservices, which adds significant complexity to the larger architecture. The design can be greatly simplified if, instead, all these microservices communicate through a bus; that is, a persistent event store. And Cosmos DB serves as an excellent persistent event store, because the change feed makes it easy to broker and source these events to each microservice. Furthermore, because the events themselves are persisted, the order processing pipeline itself is very robust and incredibly resilient to failures. You can also query and navigate the individual events, so that this data can be surfaced out through a customer care API.
Check out this link for a real-world case study of how Jet.com implemented an event-sourced architecture using the Cosmos DB change feed with hundreds of microservices: https://customers.microsoft.com/en-us/story/jet-com-powers-innovative-e-commerce-engine-on-azure-in-less-than-12-months. There is also a video on this from Microsoft Build 2018 here: https://channel9.msdn.com/Events/Build/2018/BRK3602
Real time stream processing and materialized view patterns
The change feed can also be used for performing real time stream processing and analytics. In the order processing pipeline scenario, for example, this would enable you to take all the events and materialize a single view for tracking the order status. You could then easily and efficiently present the order status through a user-facing API.
Other examples of so called “lambda architectures” include performing real time analytics on IoT telemetry or building a scoring leader board for a massively multiplayer online video game.
Moving or archiving data to secondary data stores
Another common scenario for using the change feed involves replicating data from Cosmos DB as your primary (hot) store to some other secondary (cold) data store. Cosmos DB is a superb hot store because it can sustain heavy write ingestion, and then immediately serve the ingested records back out to a user-facing API.
Over time as the volume of data mounts, however, you may want to offload older data to cold storage for archival. Once again, change feed is a wonderful mechanism to implement a replication strategy that does just that.
Consuming the Change Feed
So how do you actually work with the change feed? There are several different ways, and I’ll conclude this blog post by briefly explaining three of them. (Don’t worry, I’ll drill deeper into all three next post!)
First, you can query the change feed directly using the SDK. This raw approach works but is the hardest to implement at large scale. Essentially, you first need to discover all the container’s partitions, and then you query each of them for their changes. You also need to persist state metadata; for example, a timestamp for when the change feed was last queried, and a continuation token for each partition. Plus, you’ll want to optimize performance by spawning multiple tasks across different partitions so that they get processed in parallel.
If all this sounds like a lot of work, it is. Which is why you’ll almost certainly want to leverage the Change Feed Processor (CFP) library instead.
Change Feed Processor (CFP) Library
The CFP library provides a high-level abstraction over direct access that greatly simplifies the process of reading the change feed from all the different partitions of a container. This is a separate NuGet package that you pull into your project, and it handles all the aforementioned complexity for you. It will automatically persist state, track all the partitions of the container, and acquire leases so that you can scale out across many consumers.
To make this work, the CFP library persists a set of leases as documents in another dedicated Cosmos DB container. Then, when you spin up consumers, they attempt to acquire leases as they expire.
All you do is write an observer class that implements IChangeFeedObserver. The primary method of this interface that you need to implement is ProcessChangesAsync, which receives the change feed as a list of documents that you can process as needed. No partitions to worry about, no timestamps or continuation tokens to persist, and no scale-out needs to concern yourself with.
However, you still need to write your own host, and deploy the DLL with your observer class to an Azure app service. Although the process is straightforward, going with Azure Functions instead provides an even easier deployment model.
The simplest way to consume the change feed is by using Azure Functions with a Cosmos DB trigger. If you’re not already familiar with Azure Functions, they let you write individual methods (functions), which you deploy for execution in a serverless environment hosted on Azure. The term “serverless” here means without also having to write a host and deploy an Azure app service to run your code.
Azure Functions are invoked by one of any number of triggers. Yes, Azure Functions also uses the term triggers, but unlike triggers in Cosmos DB, an Azure Functions trigger always fires when its conditions are met. There are several different triggers available, including the one that we care about here, the Cosmos DB trigger. This Azure Functions trigger binds to configuration that points to the container you want to monitor changes on, and the lease collection that gets managed by the CFP library under the covers.
From there, the process is identical to using the CFP library, only the deployment model is dramatically simpler. The Azure Functions method that is bound using the Cosmos DB trigger receives a parameter with the same list of documents that a CFP library’s observer class does in its ProcessChangesAsync method, and you process change feed data the way you need to just the same.
Hopefully, this blog post opened your eyes to the change feed in Azure Cosmos DB. This powerful feature creates all sorts of exciting possibilities across a wide range of use cases. In the next installment, I’ll focus on container replication for multiple partition keys, and walk you step by step through the process of building a change feed solution using direct access, CFP library, and Azure Functions – all with working code samples. So, stay tuned!