Dieser Artikel ist auch auf Deutsch verfügbar
Data Product vs. Data as a Product
The term product comes from the product thinking approach that has found its way into software development in recent years. Zhamak Dehghani applied the term in second core principle of data mesh: data as a product. It means that software, or now data, is always designed from the consumer’s point of view in a way they get the best user experience. Just like a physical product, these should be consequently developed for the needs of a consumer. They are explained to the customer in a comprehensible way (intuitively or through an instruction manual), they are optimized to be easily accessible in a way that fits best for the user, and perhaps also be advertised within the organization to show the potential. And consequently, they may also have a price that consumers are willing to pay. Data is now seen as valuable for the company company’s and is no longer just a by-product of software development.
The term data product is derived from the data as a product principle and follows its ideas, but is not to be taken synonymously. Let’s try with a definition:
A data product is a logical unit that contains all components to process and store domain data for analytical or data-intensive use cases and makes them available to other teams via output ports.Jochen Christdatamesh-architecture.com
So, a data product is something technical that is implemented by data product developers. It uses data technologies to store and process large data sets, often millions of entries and more. The size of a data product is designed to cover coherent domain concepts or use cases that are valuable on their own. The max size is defined by the scope one team can handle. Data products can be roughly compared to microservices or self-contained systems, but using data technologies and serving analytical needs. Despite the term product, data product consumers are usually other internal teams, not external customers.
Data Product Examples
- The team Product Search offers a data product Search Queries that contains all queries that users entered into the search bar, the number of results, and information about the entry, the user has clicked.
- The team Article Management offers a data product Articles with article master data, both the current state and the history.
- The team Checkout offers a data product Orders with all orders since 2020. It has two output ports: one with PII included, one with PII redacted.
- The team Fulfillment has a data product Shelf Warmers with all articles that haven’t been sold during the last 3 months.
- The team Management Support uses other data products to create a Realtime Business Dashboard data product for the CEO. It does not share data and has no output ports.
- The team Recommendations uses other data products to train an ML-model for recommendations. The ML model is shared as a Tensorflow SavedModel directory on an object store. The Marketing team uses this model to make customer-specific recommendations in the newsletter.
Data Product Architecture
A data product has several components that build a coherent unit, they are usually defined in one Git repository. The following diagram shows the typical components of a data product.
A data product applies the design principle of information hiding. There are interfaces to the outside and internal components. The actual implementation of the components may vary on the use case and data platform.
The output ports represent the main API of a data product: They represent read-only access to structured data sets in the form of tables, files, or topics. A data product can have multiple output ports: They may provide the same data set in different technologies or different datasets in the same technology, e.g., one output port that contains PII data and a second output port with PII redacted. A new output port can also be added, when a structural change is necessary to evolve a data product over time.
The primary interface technology for an output port is SQL. It allows simple access to large data sets and is supported by practically all analytical tools. An output port is often implemented as a SQL view as an abstraction layer that makes it possible to change the underlying data structure without affecting data consumers. Other interface technologies for output ports are files, or topics for stream processing or as an asynchronous API to operational systems.
Output ports define the model of the provided data set. This model is defined in a schema with all tables, attributes, and types. Typical technologies are SQL DDL, dbt models, Protobuf, Avro, or JSON Schema. The model may also be described in a data catalog entry.
Output ports are optional or may be private, if a data product only serves for team-internal analytical use-cases.
Access to output ports is governed through data contracts.
A data product can have two types of data sources: operational systems or other data products.
In Data Mesh, the teams that develop the operational systems also make their relevant domain data available in data products. Often this is realized via asynchronous topics, preferably by using defined domain events. Ultimately, however, it is up to the domain team to decide how their domain data is ingested into their data products.
Data products may also use other data products over their output port, when they have an agreed data contract. They can be owned by the same team or by other teams. This is typical for consumer-aligned data products or aggregated data products, but also source-aligned data products may link other domain data when useful, e.g., to lookup master data.
Data consumers need to find the data products that are relevant for them. As data usually has a domain-specific meaning, it is important to provide an extensive description of the data model’s semantics.
Further metadata, such as contact details, the maturity level, data product usage through others, the data quality tests, or the service-level objectives, are important for data consumers so that they can decide whether a data product is trustworthy and suitable for their use-case.
It is a good practice to use CI/CD pipeline steps to automatically publish metadata to a data catalog and the data product inventory, such as the Data Mesh Manager.
A data product is developed and maintained by a single team that understands the business domain, the business processes and the data. The team is responsible for providing the promised data quality and service-level objectives. A data product has one dedicated contact person, the team’s product owner, who is ultimately responsible for the data product and its quality.
The product owner is responsible for the life-cycle and evolution of a data product, incorporating the requirements of the (potential) consumers and domain-internal analytical needs. They also set the price that is charged for the use of a data product.
Data needs to be cleaned, aggregated, composed, and transformed to implement the output ports schema or to answer analytical questions.
What technology is used, and how code is organized internally, is an implementation detail of a data product. It depends on the data platform and implementation details are up to the decision of the development teams.
In many cases, SQL queries are used for simple transformations, and Apache Spark for complex pipelines.
A scheduling and orchestration tool, such as Airflow, is used to run the transformation code.
A data product usually needs to store a significant amount of data in some kind of data storage, such as tables or files in an object store. Data storage is provided as a self-service through the data platform. A data product has its own private realm that is isolated from other data products.
What technology is used and how data is organized internally, is an implementation detail of a data product. It depends on the data platform and is up to the decision of the development teams. In many cases, column-originated storage technologies are used.
Data products provide managed and high-quality data sets, so as in any software engineering discipline, tests are essential. There are different types of tests:
Unit tests test the transformation code itself. They use a fixed input data and define the expected output data.
Expectation tests run during deployment on the real data models and verify that the source data from the input ports, intermediary models, and the output port meet the defined expectations.
Quality tests run regularly on real data to monitor the service-level objectives.
When domain data is shared with other teams, it is important to describe the semantics of the data and the business context business context of how the data was created.
Besides a description of the data model attributes, a good documentation also gives an introduction, what to expect from the data sets and first hints, which data may be interesting and how to access them.
A good way to implement a documentation is to provide an interactive notebook (Jupyter, Google Collab, Databricks Notebook) with example queries.
Data technologies quickly become expensive when used at scale. Therefore, it is important to monitor the costs of data products. They may be the foundations for the price that is billed to the data consumers, as agreed in data contracts.
Policies as Code
Global policies are the rules of play in the data mesh and defined by the federated governance group, such as naming conventions, data classification schemes, or access control.
While most of the policies should be implemented on a data platform level, some policies need to be configured on data product level, especially when domain knowledge is required or product owners need to decide on permissions. Examples are column-level classification of domain data, PII tagging, or access control.
A data product has its own CI/CD pipeline and infrastructure resource definitions. The CI/CD pipeline is triggered when the transformation code or the data model changes, tests are executed, and the data product is deployed to the data platform, inline with global policies. The data platform team may provide modules or templates for the data product teams to use.
A data product can have additional ports and capabilities that are not directly used by data consumers, but are important for the operation of the data product. These include ports for monitoring, logging, and admin functions.
Data Product Specification
To describe the metadata of a data product, it is common to have a data product specification file in the repository of the data product, or it is generated through the deployment process by the data platform.
A formal data product specification can act as a foundation for automation and to provide metadata to other systems, such as an enterprise catalog or a data product catalog.
Data Product Implementation
Now let’s look at an example, how an actual data product can be implemented using the AWS S3 and Athena tech stack.
In this example, the data platform team provides a Terraform module, that provisions all necessary services on the data platform to run a data product, that are in line with the policies and conventions defined in the governance group.
The data product developers have one Git repository per data product. They use the provided Terraform module and configure it for their data product. In the same repository, they define the transformation code as a SQL query and a JSON schema file with the model of the output port and a detailed description of the data model.
terraform apply, usually triggered through the CI/CD-pipeline, all required resources are provisioned, such as S3 buckets, AWS Athena resources, and lambda functions. Also, permissions are created in AWS IAM.
The pipeline also pushed metadata to the data catalog and data product inventory.
Find an example for a Terraform module implementation on GitHub.
Data Mesh Manager
n our data mesh projects, we noticed, that good tooling was missing to discover, manage, and govern data products. That’s why we created the Data Mesh Manager. It uses the data products metadata to build a comprehensive data product inventory. This inventory is a powerful entry point to navigate through a complex data mesh and find trustworthy data products that are relevant for data consumers. A data mesh map visualizes the data mesh and helps to understand the data mesh topology.
Data Mesh Manager also supports the full life-cycle of data contracts as a self-service: Request and accept the access to a data product to agree on a bilateral data contract. Regular reassessment dates allow data contracts to be renewed or terminated, for example, when a data product version is deprecated.
Through the REST-API, the Data Mesh Manager is fully integrable with all data platforms and triggers the automated creation of IAM permissions in the data platform, once a data contract is created or updated.