Structured streaming with Azure Databricks into Power BI & Cosmos DB
In this blog we’ll discuss the concept of Structured Streaming and how a data ingestion path can be built using Azure Databricks to enable the streaming of data in near-real-time. We’ll touch on some of the analysis capabilities which can be called from directly within Databricks utilising the Text Analytics API and also discuss how Databricks can be connected directly into Power BI for further analysis and reporting. As a final step we cover how streamed data can be sent from Databricks to Cosmos DB as the persistent storage.
Structured streaming is a stream processing engine which allows express computation to be applied on streaming data (e.g. a Twitter feed). In this sense it is very similar to the way in which batch computation is executed on a static dataset. Computation is performed incrementally via the Spark SQL engine which updates the result as a continuous process as the streaming data flows in.
The above architecture illustrates a possible flow on how Databricks can be used directly as an ingestion path to stream data from Twitter (via Event Hubs to act as a buffer), call the Text Analytics API in Cognitive Services to apply intelligence to the data and then finally send the data directly to Power BI and Cosmos DB.
All data which arrives from the data stream is treated as an unbounded input table. For each new data within the data stream, a new row is appended to the unbounded input table. The entirety of the input isn’t stored, but the end result is equivalent to retaining the entire input and executing a batch job.
The input table allows us to define a query on itself, just as if it were a static table, which will compute a final result table written to an output sink. This batch-like query is automatically converted by Spark into a streaming execution plan via a process called incremental execution.
Incremental execution is where Spark natively calculates the state required to update the result every time a record arrives. We are able to utilize built in triggers to specify when to update the results. For each trigger that fires, Spark looks for new data within the input table and updates the result on an incremental basis.
Queries on the input table will generate the result table. For every trigger interval (e.g. every three seconds) new rows are appended to the input table, which through the process of Incremental Execution, update the result table. Each time the result table is updated, the changed results are written as an output.
The output defines what gets written to external storage, whether this be directly into the Databricks file system, or in our example CosmosDB.
To implement this within Azure Databricks the incoming stream function is called to initiate the StreamingDataFrame based on a given input (in this example Twitter data).
