Matt Weingarten is a Senior Data Engineer who writes about his work and perspectives on the data space on his Medium blog—we highly encourage you to go check it out!
I’m happy to be collaborating with the folks at Validio again, after a very successful series we did on data quality last year. Validio just recently put out their latest whitepaper on Deep Data Observability, and one of the key points they touched upon in that paper was the rise of different data formats and how that relates to the ongoing pursuit of full end-to-end observability.
Structured data is always nice, but it’s no longer realistic to expect that in standard data processing. As a result, we need to adapt to support different formats, from both an ingestion and an observability standpoint.
The first question you might be asking yourself is, “What is semi-structured data?” Essentially, this is data that does not strictly conform to the tabular structure we’re all familiar with in traditional databases, but it still has separators in place to identify which field is which. With the rise of APIs and other mechanisms by which more and more teams are consuming data, semi-structured is starting to become more of the norm.
One of the first projects I worked on when I joined Disney was for user preferences in ESPN (essentially, your favorite teams, sports, athletes, and more so that your feed is catered towards your favorites). The data we receive from upstream for this product is in JSON form, so I learned how to deal with semi-structured data rather quickly. Our processing in turn parses the JSON and explodes it into a more structured form, which makes it easier for our consumers to query our data.
That last point might be the biggest hindrance preventing further adoption of semi-structured data. Most data analysts or consumers don’t want to waste time with complicated SQL to retrieve data, which is why structured data is preferred. That’s also why self-serve BI that can do all of this SQL in the backend is starting to become more mainstream as well, so additional efforts aren’t needed.
The nested data structure is the de facto standard for backend applications, whereas structured data, which usually involves flattening, has been the standard in the generation of warehouse analytics, as the example below shows. That’s all changing.
Traditionally, structured data is meant for the data warehouse. But as we talked about with data quality, data is not just in warehouses anymore. Data lakes, data streams, and data lakehouses are all becoming increasingly popular with data processing, and Deep Data Observability will be able to look at all data sources for proper monitoring.
For all of our processing that involves multiple layers, the consumption layer is what we expose to stakeholders. However, all the layers that lead up to that consumption layer should have proper checks in place as well. What if something goes wrong along the way and the final data will get impacted as a result? You’d definitely want to know about that so that you can alert consumers beforehand.
That’s why our team tries to build data quality checks into every hop of our process, so that we’re on top of everything. We’d want to do the exact same thing with data observability, and that’s what a proper platform should be capable of doing if you want to truly be tapping into the full power of your datasets. In other words, you want to observe and validate all dimensions of data quality, such as the ones below, in the entire data pipeline (not just the warehouse).
While we just focused on semi-structured data and how it relates to data warehouses and other data technologies here, there’s plenty to discuss when it comes to Deep Data Observability. I highly recommend checking out the rest of Validio’s whitepaper (and perhaps will be writing some more posts on this topic in the future) if you’re interested in this area.