Can you trust the data in your data lake? Many companies are guilty of dumping data into the data lake without a strategy for keeping track of what’s being ingested. This leads to a murky, swampy repository. If you don’t have transparency into your lake so that you can feel confident using the data, what’s the point of deploying a data lake in the first place?
You know Hadoop is a different animal than the data warehouse, requiring distinct technologies and skill sets. Unlike relational databases, Hadoop is little help when it comes to quality control. Without incorporating additional tools into your data lake architecture you have no way to apply metadata to your data as it is ingested. You can’t automate metadata management so that you can scale to the volume and velocity of big data. You also have no way to customize rules for different data types from different sources.
How do you design your data lake architecture and what functionality do you need to validate the data in your data lake? From our years of experience deploying data lakes for leading companies across heavily regulated industries like financial services and healthcare, we’ve developed some best practices to help companies clean up and derive more value from their data lakes.
Create Data Zones
Managing data ingestion requires thinking about where the data should land in your lake and where it goes after it’s ingested, in line with your data lifecycle management strategy. We recommend creating zones in the file system of your data lake, dedicated for specific uses; namely, “transient,” “raw,” “trusted” and “refined” zones. By building a rule-based architecture tied to the metadata that’s applied upon ingestion, you can automate validating the data as you move it from zone to zone. You also may want to incorporate a discovery sandbox “zone,” moving trusted data there for wrangling, discovery and exploratory analysis.
- Transient zone: This is the loading zone, where you can perform basic quality checks using MapReduce or Spark.
- Raw zone: Next, data is loaded into the raw data zone where it can be masked or tokenized to protect sensitive data, such as personally identifiable information (PII), personal health information (PHI) or payment card industry (PCI) information. This is where raw datasets exist for business analysts and data scientists to access.
- Trusted zone: This zone functions as the single source of truth for data after it has been cleaned and validated. It can contain both master data and reference data. We define master data as basic data sets, such as basic customer information (e.g., names, addresses), which needs to be kept up to date using change data capture (CDC) mechanisms. Reference data consists of more complex datasets, such as datasets created with merged information from multiple sources – for example, more detailed customer profiles.
- Refined zone: In the refined zone, data is enriched, prepared, and curated for analysis. This is where data can be used by applications, business analysts or business intelligence tools.
Give Business Users a Say
We’ve found it’s a good idea to employ a rule-based data policy engine to enable business users to define the business rules used for data quality validation, as they typically are most familiar with the data. Also, by allowing business users to associate these rules with the metadata for the ingested datasets, it becomes part of the workflow from an orchestration perspective. For example, as the data leaves the raw zone, you’re able to execute these rules before you make the data available in the trusted zone.
Rate Your Data
The beauty of Hadoop is that it captures all raw data, unlike relational databases, which will only capture data if it meets your schema criteria. Although this presents challenges for data quality, you have the option to determine the level of acceptable data quality, allowing you to potentially get more value from your data. In other words, Hadoop can store “incomplete” datasets, which may be useful for users who don’t need certain information. For example, some customer records may be missing gender data, but contain what the user needs, such as state of residence. To provide a clear picture of what data may be incomplete or of a certain quality, it’s important to separate the “good” records from the “bad” records and create thresholds that specify how many bad records are tolerable before a dataset is rejected.
Go Back to the Source
Automating data quality checks as data moves from one zone to another is key. Another area to automate is reporting back to data producers. If “bad” records – according to the criteria you’ve established – are ingested, you can automate reporting back to the data producer so that they can rectify those records. Then you can re-ingest them back into the data lake.
The backbone of successful data validation in the data lake? Operationalizing data validation in the data lake ultimately requires an underlying data management platform that manages, tracks and governs metadata, data quality and security. Of course, when we work with our clients, we build an architecture based on the Zaloni Data Platform, which gives businesses the control and governance they need to feel confident in the quality of their data.