What is data observability? 4 pillars of data observability.
Observability has its roots in control theory, providing a measure of how well the state and behavior of software can be predicted from outputs. With the rise of DevOps, observabiliy gained traction, becoming a crucial concept for software engineering. At this time, companies like Datadog and New Relic paved the way for software observability. With the ever-increasing complexity of cloud infrastructure, software observability tools have become essential in software engineering. Software observability gives engineers a full view of their entire system architecture and empowers them to quickly address bugs or performance problems by precisely pinning down root causes.
Organizations increasingly recognize the importance of maintaining high-quality data in today's data-driven world. Organizations are now coming up with processes and protocols for ensuring data quality.
Data observability is a rapidly developing technology that seeks to eliminate data uncertainty and empower organizations with real-time intelligence. By allowing data consumers the ability to monitor their data, quickly assess potential problems, and get context for resolution, data observability provides an invaluable tool in the fight against data entropy. While this may require an initial investment into tools and resources, it pays off long-term due to the reduced probability of bad data impacting the business.
To sum up, data observability aims to monitor data to quickly assess where issues might have happened to give teams the context to solve them promptly.
Data observability has its roots in software engineering, but it’s crucial to understand the distinctions between them. While Metrics, Traces, and Logs provide the foundation for software observability, they do not sufficiently summarize data observability.
When it comes to understanding data observability, one must understand the four key pillars that comprise the concept, which are: metrics, metadata, lineage, and logs. Here we describe each pillar and the importance of each when it comes to mitigating data uncertainty.
The 4 pillars of data observability
1. Metrics - The internal characteristics of the data
Data can be characterized by certain properties and metrics, which vary based on the type of data. For numeric datasets, summary statistics such as mean, standard deviation, and skewness are used to describe their distribution. Categorical data, on the other hand, relies upon summary statistics like the number of groups and uniqueness.
Some general metrics, regardless of business type, include:
Data Freshness/Timeliness: Is the information current?
Metadata Completeness/Volume: Are there any gaps in the information, such as missing values or incomplete records?
Data Duplication: Are there any duplicate records?
Metadata Schema: Has the data structure changed, resulting in pipeline disruptions or outdated dashboards?
Data Accuracy: Does the data fall within a specific range or conform to a specific format?
2. Metadata - The data about the data
Metadata can be defined as data that provides information about other data, generally a dataset. In other words, metadata is data that has the purpose of defining and describing the data object it is linked to. Examples of metadata include titles and descriptions, tags and categories, information on who created or modified the data, and who can access or update the data.
Many people underestimate the importance of metadata compared to digital data, but it is crucial for optimizing the use of data. Proper handling of metadata can enhance the search, retrieval, and discovery of pertinent information, allowing users to maximize their data. Metadata has various applications in business, including:
Classification: Metadata is crucial in classifying and organizing data.
Findability: Metadata can enhance search and retrieval processes, enabling users to quickly locate the information they need.
Information security: Metadata can be utilized to regulate data distribution by verifying access and editing rights.
Having a strong metadata management plan in place guarantees that an organization's data is consistent, accurate, and of high quality across different systems. Companies that employ an all-encompassing metadata management approach are more likely to base their business decisions on dependable data compared to those without any metadata management solutions.
3. Data Lineage - The dependencies between data
Metrics and metadata can be used to describe a single dataset sufficiently. However, datasets are not isolated. In fact, datasets are related to each other in intricate ways. This is where lineage becomes essential. Data lineage reveals how different file types and systems are interrelated, giving you a clear understanding of where something came from and its possible destinations.
Lineage models allow tracing anomalous data back to its source, ensuring that any issues can be quickly identified and rectified. These models connect with applications, jobs, and orchestrators that have generated such information - creating a cohesive system for troubleshooting efficiency.
Knowing downstream dependencies is essential to break down silos between different teams within an organization. For example, consider a company with separate software engineering and data teams that rarely communicate with each other. The former may not realize how their updates could affect the latter. Through data lineage, teams can access downstream dependencies and overcome communication barriers.
4. Logs - The interactions between data and the external world
Metrics describe the inner qualities, metadata describes its external characteristics, and lineage traces dependencies. But how does the data interact with the “outside world”? This is where logs come into play. Logs can be machine-generated or human-generated interactions.
On the one hand, machine-generated interactions with data can include data movement - like data being replicated from external sources to a data warehouse.
Machine-generated interactions also include data transformation - like dbt transforming a source table into a derived table.
On the other hand, machine-human interactions include interactions between the data and its users —like data engineers working on new models and data scientists creating machine-learning models.
Together, these four pillars help you better understand your data infrastructure and ensure your data is in a good state at every stage of the data life cycle.
To conclude, there are four means to comprehensively depict the state of our data at any given time: metrics for internal properties, metadata for external properties, lineage for dependencies, and logs for interactions. If any of these pillars is absent, your capability to reconstruct the data's state will be insufficient.