Azure Cosmos DB – HTAP using Azure Synapse Link

I haven’t done much blogging in the last 6 months. My last article was posted in January, on Data Governance in the Covid era. After changing jobs in March, I prioritized getting settled into my new role. That said, I have had the writer’s itch for a while and thought what better topic to break the rut with than Cosmos DB. I am a big fan of Cosmos DB – Microsoft’s multi-model NoSQL database service in Azure, and I have written a few articles on the topic.

Recent stats of my blog show more interest in data movement out of Cosmos DB, possibly for reporting purposes. Check out my 2019 article on real-time data movement from Cosmos DB using Change Feed and Azure Functions, where I explained how to persist container changes to SQL Server. Well, that is one way to enable real-time analytics. Cosmos DB has evolved a lot since then, and there are new and smarter ways to achieve similar results. Last year (Dec 2020), Microsoft announced the general availability of Azure Synapse Analytics – the unified analytics service in Azure.

sqlroadie – 2021 blog stats

HTAP capabilities of Cosmos DB using Synapse Analytics

In this article, we will explore the HTAP capabilities of Cosmos DB. The goal is to derive real-time insights from transactional changes made to Cosmos DB, in a cost-effective manner with minimal overhead. I want the solution to be scalable, reliable and overall simple to maintain.

What is HTAP?

HTAP stands for hybrid transaction/analytical processing. Traditionally, data systems can be classified as either OLTP – Online Transaction Processing or OLAP – Online Analytical Processing. Amazon shopping website is a good example of an OLTP system, as it manages millions of customer orders daily in real-time. However, Amazon would have billions of historical customer order details in their OLAP data warehouse. This helps them to create product recommendations by analyzing historical order data, and serve it to the OLTP system through analytics APIs.

Ordinarily, transaction data has to be copied or replicated from OLTP -> OLAP databases through ETL processes, and reporting/insights/recommendations generated in the OLAP system at a latency that made business sense. This is a tedious process and involves a lot of overhead. Majority of enterprises still use this approach. However, advancements in computing have enabled a new age HTAP architecture.

Gartner defines HTAP as “a hybrid transaction/analytical processing (HTAP) architecture is best enabled by in-memory computing techniques and technologies to enable analytical processing on the same data store that is used to perform transaction processing. By removing the latency associated with moving data from operational databases to data warehouses and data marts for analytical processing, this architecture enables real-time analytics and situation awareness on live transaction data as opposed to after-the-fact analysis on stale data.

This is exciting stuff, and quite a game-changer in the world of databases. Enough said, how do we go ahead and design a solution using Cosmos DB’s HTAP capabilities!

Synapse Link for Cosmos DB

Synapse Link for Cosmos DB is a cloud-native HTAP capability that creates a tight seamless integration between Cosmos DB and Synapse Analytics. It gives you the ability to gain near real-time analytics and insights over operational data in Cosmos DB with no ETL and no impact to the performance of OLTP transactions in Cosmos DB containers.

Let us break that down. Cloud-native (if you didn’t click the link) is the modern approach, where you design systems using cloud services to scale and perform well consistently. Synapse Analytics (if you are not familiar) can be thought of as Microsoft’s cloud data warehouse, for starts. And that is because Synapse is a lot more than that, but we have to start somewhere, don’t we?

So, Synapse Link creates a seamless, automagic integration between Cosmos DB and Synapse Analytics, thus removing the need for ETL. How does that work? Because a picture is worth a thousand words, I have referred to a diagram from Microsoft’s official doco on Synapse Link. The highlighted portion is new and works behind the scenes to deliver near-real time analytics with no additional ETL. When we create a container in an Azure Cosmos DB account that has Synapse Link enabled, an analytical store is automatically provisioned. This analytical store helps to ensure that there is no impact to OLTP transactions against the Cosmos DB container.

Image courtesy: Microsoft

This is a robust mechanism to perform large-scale analytics over data in Cosmos DB and implement an archival policy efficiently to keep the containers sleek without losing access to valuable historical data. It is worth noting that the transactional and analytical stores get their own, separate TTL (Time to Live) properties. TTL is one of my favorite features of Cosmos DB. Self-destruction can be so powerful sometimes! I will cover this in more detail in my sample solution.

Sample Solution

Confucius says, “I hear and I forget. I see and I remember. I do and I understand”. So, I built a sample solution that would ingest data into Cosmos DB at scale, and I was impressed at just how quickly data was available downstream in Synapse Analytics. The sample solution features a Python Notebook, running on Databricks and ingesting the popular New York City taxi trip data into a Cosmos DB container using Cosmos DB Spark 3 OLTP connector for SQL API. The trip data that became available in Synapse Analytics was used to build a Power BI dashboard, thus effectively producing near real-time analytics with no ETL and minimal code. I will give a step-by-step breakdown of the sample solution. Refer to this guide by Microsoft for more details.

Sample Cosmos DB HTAP solution
Sample solution using NYC Taxi trip data

1. Cosmos DB container

Firstly, we want to enable Synapse Link in the Cosmos DB account using the Settings blade. It is worth noting that at the time of writing this article (Oct 2021), Synapse Link and analytical store is supported only for SQL and MongoDB APIs.

Enable Azure Synapse Link

At the moment, analytical store can be turned ON only for new containers. If you wish to make use of Cosmos DB’s HTAP capabilities for an existing container, you will need to migrate data to a new container. Note that containers cannot be renamed. Auto-sync latency is within 2 minutes, as per Microsoft documentation. In my tests, I did not see a delay of more than a minute or so for data to be available downstream in Synapse Analytics.

New Cosmos DB container with Analytical store turned ON

Analytical store does not need request units (RU/throughput) to be allocated, nor does it consume the RU assigned to the transactional store – 1000 in the snip above. Analytical store follows a consumption-based pricing model, which is based on data storage and analytical read/write operations and queries executed. Storage can be optimized by setting the Analytical Store Time To Live property in the Data Explorer blade under Scale & Settings option. This dictates for how long data would be retained in the analytical store, and is independent from the container TTL property. You should set the Analytical Store TTL to a value that gives you sufficient time to process the data and persist or derive insights from it as required. At the time of writing this article (Oct 2021), analytical store does not support backup and restore, i.e. if a Cosmos DB backup is restored, only the transactional store will be recovered. It is also worth noting that schema representation in the analytical store could be different from the container schema.

Analytical Store Time to Live

2. Databricks – Python notebook to write NYC Taxi trip data to Cosmos DB

In the sample solution, I used a Python notebook to write data to Cosmos DB using the Spark 3 OLTP Connector for SQL API. This was just to simulate incoming OLTP application requests.

Python notebook
Sample data

3. Synapse Analytics

In the Synapse Workspace, Cosmos DB analytical store may be accessed using the Spark pool or built-in serverless SQL pool. I used the serverless SQL pool, which is a pay per query distributed data processing system built for big data and computational functions.

In the Synapse Studio, I created a linked service by connecting to the external Cosmos DB resource.

Connect to Azure Cosmos DB

In this case, I chose Managed Identity as the authentication method.

New Linked Service – Cosmos DB SQL API

In a little while, the linked service was added. At this stage, we are almost ready to query the underlying analytical store of GreenTaxiTrip container.

The next step was to create a user database in the serverless SQL pool, so that a CREDENTIAL could be created to access the Cosmos DB database securely. You cannot create the credential in master database.

4. Data Ingestion

I scheduled a Databricks job to execute the Python notebook to ingest NYC Taxi trip data from DBFS to Cosmos DB using the Spark 3 OLTP connector for SQL API. You may also execute it manually. Once data became available in Synapse Analytics through the analytical store, I wrote a couple of queries over the data using familiar T-SQL and converted them to views.

Databricks Job
Cosmos DB query

Note the number of records, and the corresponding timestamp in the transactional store. When the record count query completed in Synapse Analytics 9 seconds later, the number of records were out by only 545! Auto-sync latency is 2 minutes, as per Microsoft doco. It is worth noting that even if a document had a TTL set to lower than 2 minutes, it would still appear in the analytical store however, I have not tested it.

Cosmos DB – transactional store record count

My test dataset had 7 million records, and I left the job running to see how well the analytical store would perform. I was quite satisfied with the performance of Synapse Link. In this case, I used only one container but it would be interesting to see how well auto-sync holds up at scale.

5. Near real-time analytics using Power BI

Having validated that trip data written to the Cosmos DB container were visible in Synapse Analytics in near real-time, I proceeded to create 2 simple views using the OPENROWSET function, and a Power BI dashboard that used the views as its data source. Please excuse my lazy visualization efforts, but the point is you can easily get near real-time insights over operational data at scale using Synapse Link. All without any ETL!

Views created in Synapse serverless SQL pool over data in analytical store

Other alternatives

When it comes to enabling analytics over operational data in Cosmos DB, I cannot think of a better way to do it. The Change Feed way of doing it is not as scalable or easily maintainable as using Synapse Link. The obvious difference is that you will need a Synapse Analytics account to make use of Synapse Link. If you are a Microsoft house, this would make sense as Synapse is the unified analytics service and integrates directly with several services such as Azure Machine Learning, Cognitive Services and Power BI.


At the time of writing this article (Oct 2021), Synapse Link for Cosmos DB is only supported for the SQL and Mongo DB APIs. I think it will be really cool to have it enabled for the Gremlin API. Graph databases are getting increasingly popular, and graph-enabled analytics such as fraud detection and money laundering is bound to become more common place.

More limitations are available here. The most glaring one is the lack of support for backup and restore of data in the analytical store.

The road ahead

HTAP is made possible by the power of cloud computing. I believe it will not be too long before HTAP capability is available for Azure SQL databases. That will be quite a game changer!

Aside from the basic Business Intelligence feature I demonstrated in the sample solution, Synapse Analytics also enables batch scoring using native scoring for models trained using RevoscalePy and RevoscaleR packages, AutoML to develop a Regression/Classification/Time Series Forecasting model using the Azure Machine Learning linked service, and multi language support for advanced analytics using the Spark pool.

Cosmos DB’s HTAP capability is fascinating, and I can think of several use cases to enable real-time decision intelligence at scale. I hope you had a good read. Please leave a comment if you have questions or any feedback.

References – Jovan Popovic – Microsoft Mechanics

Azure Cosmos DB: real-time data movement using Change Feed and Azure Functions

Many people who read my Cosmos DB articles are looking for an effective way to export data to SQL, either on-demand or in real-time. After performing a search term analysis for my blog earlier this year, I had made up my mind about posting a solid article on exporting data from Cosmos DB to SQL Server.

Note that this serverless and event-based architecture may be used to not only persist Cosmos DB changes to SQL Server, but trigger alternate actions such as stream processing or loading to blob/Data Lake.


Real-time ETL using Cosmos DB Change Feed and Azure Functions

In this article, we will focus on creating a data pipeline to ETL (Extract, Transform and Load) Cosmos DB container changes to a SQL Server database. My main requirements or design considerations are:

  • Fault-tolerant and near real-time processing
  • Incur minimum additional cost
  • Simple to implement and maintain

Cosmos DB Change Feed

Cosmos DB Change Feed listens to Cosmos DB containers for changes and outputs the list of items that were changed in the chronological order of their modification. Cosmos DB Change Feed enables building efficient and scalable solutions for the following use cases:

  • Triggering a notification or calling an API
  • Real-time stream processing
  • Downstream data movement or archiving


Types of operations

  • Change feed tracks inserts and updates. Deletes are not tracked yet
  • Cannot control change feed to track only one kind of operation, for example only inserts
  • For tracking deletes in the Change Feed, workaround is to soft-delete and assign a small TTL (Time To Live) value of “n” to automatically delete the item after “n” seconds
  • Change Feed can be read for historic items, as long as the items have not been deleted
  • Change Feed items are available in order of their modification time (_ts system attribute), per logical partition key, and tagged with the same _lsn (system attribute) value for all items modified in the same transaction

Read more about Azure Cosmos DB Change Feed from Microsoft docs to gain a thorough understanding. Change Feed can be processed using Azure Functions or Change Feed Processor Library. In this article, we will use Azure Functions.

Azure Functions

Azure Functions is an event-driven, serverless compute platform for easily running small pieces of code in Azure. Key points to note are:

  • Write specific code for a problem without worrying about an application or the infrastructure to run it
  • Use either C#, F#, Node.js, Java, or PHP for coding
  • Pay only for the time your code runs and trust Azure to scale
  • As of July 2019, the Azure Functions trigger for Cosmos DB is supported for use with the Core (SQL) API only

Read more from Microsoft docs to understand full capabilities of Azure Functions.

If you use Consumption plan pricing, it includes a monthly free grant of 1 million requests and 400,000 GBs of resource consumption per month per subscription in pay-as-you-go pricing across all function apps in that subscription, as per MS docs.

Compare hosting plans and check out pricing details for Azure Functions at the Functions pricing page to gain a thorough understanding of pricing options.

Real-time data movement using Change Feed and Azure Functions

The following architecture will allow us to listen to a Cosmos DB container for inserts and updates, and copy changes to a SQL Server Table. Note that Change Feed is enabled by default for all Cosmos DB containers.

I will create a Cosmos DB container and add an Azure Function to listen to the Cosmos DB container. I will then modify the Azure Function code to parse modified container items and save them to a SQL Server table.

1. First, I navigated to Azure portal, Cosmos DB blade and created a container called reservation in my Cosmos DB database. As it is purely for the purposes of this demo, I assigned lowest throughput of 400 RU/s



2. Now that the container is ready, proceed to create an Azure Function App. The Azure Function will be hosted in the Azure Function app




3. Add an Azure Function within the newly created Azure Function App. Azure Function trigger for Cosmos DB utilizes the scaling and event-detection functionalities of Change Feed processor, to allow creation of small reactive Azure Functions that will be triggered on each new input to the Cosmos DB container.





4. Configure the trigger. Leases container may be manually created. Alternately, check the box that says “Create lease collection if it does not exist”. Please note that you would incur cost for storage and compute for leases container.


I got this error that read – “The binding type(s) ‘cosmosDBTrigger’ are not registered. You just need to install the relevant extension. I saw many posts about this, so it will most likely be fixed soon.


Sort out the error by installing the extension for Azure Cosmos DB trigger.



5. Once the function is up and running, add an item to the reservations container that we are monitoring. And we have a working solution!



6. Trigger definition may be modified to achieve different things, in our case we will parse the feed output and persist changes to SQL server. You can download the csx file I used.




We have successfully implemented a serverless, event-based low cost architecture that is built to scale. Bear in mind that you would still end up paying for Azure Function and the underlying leases collection, but there will be minimum additional RU cost incurred from reading your monitored container(s) as you are tapping into the Change Feed.

You can monitor the function and troubleshoot errors.


I hope you found the article useful. Add a comment if you have feedback for me. If you have any question, drop me a line on LinkedIn. I’ll be happy to help 🙂 Happy coding!


Azure Databricks – Introduction (Free Trial)

Microsoft’s Azure Databricks is an advanced Apache Spark platform that brings data and business teams together. In this introductory article, we will look at what the use cases for Azure Databricks are, and how it really manages to bring technology and business teams together.


Before we delve deeper into Databricks, it is good to have a general understanding of Apache Spark.

Apache Spark is an open-source, unified analytics engine for big data processing, maintained by the Apache Software Foundation. Spark and its RDDs were developed in 2012 in response to limitations of MapReduce

Key factors that make Spark ideal for big data processing are:

  • Speed – up to 100X faster
  • Ease of use – code in Java, Scala, Python, R and SQL
  • Generality – use SQL, streaming and complex analytics
Apache Spark Ecosystem.jpg
Pic courtesy: Microsoft

Databricks – the company – was founded by creators of Apache Spark. Databricks provides a web-based platform for working with Spark, with automated cluster management and IPython-style notebooks. It is aimed at unifying data science and engineering across the Machine Learning (ML) life cycle from data preparation, to experimentation and deployment of ML applications. Databricks, by virtue of its big data processing capabilities, also facilitates big data analytics. Databricks, as the name implies, thus lets you build solutions using bricks of data.

Azure Databricks

Azure Databricks combines Databricks and Azure to allow easy set up of streamlined workflows and an interactive work space that lets data teams and business collaborate. If you’ve been following data products on Azure, you’d be nodding your head along, imagining where Microsoft is going with this 🙂

Azure Databricks enables integration across a variety of Azure data stores and services such as Azure SQL Data Warehouse, Azure Cosmos DB, Azure Data Lake Store, Azure Blob storage, and Azure Event Hub. Add rich integration with Power BI, and you have a complete solution.

Azure Databricks Overview
Pic courtesy: Microsoft

Why use Azure Databricks?

By now, we understand that Azure Databricks is an Apache Spark-based analytics platform that has big data processing capabilities and brings data and business teams together. How exactly does it do that, and why would someone use Azure Databricks?

  1. Fully managed Apache Spark clusters: With the serverless option, create clusters easily without having to set up your own data infrastructure. Dynamically auto-scale clusters up and down, and auto-terminate inactive clusters after a predefined period of inactivity. Share clusters with your teams, reduce time spent on infrastructure management and improve iteration time.

  2. Interactive workspace: Streamline data processing using secure workspaces, assign relevant permissions to different teams. Mix languages within a notebook – use your favorite out of R, Python, Scala and SQL. Explore, model and execute data-driven applications by letting Data Engineers prepare and load data, Data Scientists build models, and business teams analyze results. Visualize data in a few clicks using familiar tools like Matplotlib, ggplot or take advantage of the rich integration with Power BI.

  3. Enterprise security: Use SSO through Azure Active Directory integration to run complete Azure-based solutions. Roles-based access control enables fine-grained user permissions for notebooks, clusters, jobs, and data.

  4. Schedule notebook execution: Build, train and deploy AI models at scale using GPU-enabled clusters. Schedule notebooks as jobs, using runtime for ML that comes preinstalled and preconfigured with deep learning frameworks and libraries such as TensorFlow and Keras. Monitor job performance and stay on top of your game.

  5. Scale seamlessly: Target any amount of data or any project size using a comprehensive set of analytics technologies including SQL, Streaming, MLlib and GraphX. Configure number of threads, select number of cores and enable autoscaling to dynamically scale processing capabilities leveraging a Spark engine that is faster and performant through various optimizations at the I/O layer and processing layer (Databricks I/O).

Of course, all of this comes at a price. If this article has piqued your interest, hop over to Azure Databricks homepage and avail the 14 day free trial!

Azure Databricks - Free Trial 14 days.jpg

Suggested learning path:

  1. Read more about Azure Databricks –
  2. Create a Spark cluster and run a Spark job on Azure Databricks –
  3. ETL using Azure Databricks –
  4. Stream data into Azure Databricks using Event Hubs –
  5. Sentiment analysis on streaming data using Azure Databricks –

I hope you found the article useful. Share your learning experience with me. My next article will be on Real-time analytics using Azure Databricks.

Azure Databricks - Real time analytics.jpg
Azure Databricks