Azure Purview deployment best practices

This article identifies common tasks that can help you deploy Purview into production. These tasks can be completed in phases, over the course of a month or more. Even organizations who have already deployed Purview can use this guide to ensure they're getting the most out of their investment.

A well-planned deployment of a data governance platform (such as Azure Purview), can give the following benefits:

  • Better data discovery
  • Improved analytic collaboration
  • Maximized return on investment.

Prerequisites

  • Access to Microsoft Azure with a development or production subscription
  • Ability to create Azure resources including Purview
  • Access to data sources such as Azure Data Lake Storage or Azure SQL in test, development, or production environments
    • For Data Lake Storage, the required role to scan is Reader Role
    • For SQL, the identity must be able to query tables for sampling of classifications
  • Access to Azure Security Center or ability to collaborate with Security Center Admin for data labeling

Identify objectives and goals

Many organizations have started their data governance journey by developing individual solutions that cater to specific requirements of isolated groups and data domains across the organization. Although experiences may vary depending on the industry, product, and culture, most organizations find it difficult to maintain consistent controls and policies for these types of solutions.

Some of the common data governance objectives that you might want to identify in the early phases, include:

  • Maximizing the business value of your data
  • Enabling a data culture where data consumers can easily find, interpret, and trust data
  • Increasing collaboration amongst various business units to provide a consistent data experience
  • Fostering innovation by accelerating data analytics to reap the benefits of the cloud
  • Decreasing time to discover data through self-service options for various skill groups
  • Reducing time-to-market for the delivery of analytics solutions that improve service to their customers
  • Reducing the operational risks that are due to the use of domain-specific tools and unsupported technology

The general approach is to break down those overarching objectives into various categories and goals. Some examples are:

Category Goal
Discovery Admin users should be able to scan Azure and non-Azure data sources (including on-premises sources) to gather information about the data assets automatically.
Classification The platform should automatically classify data based on a sampling of the data and allow manual override using custom classifications.
Consumption The business users should be able to find information about each asset for both business and technical metadata.
Lineage Each asset must show a graphical view of underlying datasets so that the users understand the original sources and what changes have been made.
Collaboration The platform must allow users to collaborate by providing additional information about each data asset.
Reporting The users must be able to view reporting on the data estate including sensitive data and data that needs additional enrichment.
Data governance The platform must allow the admin to define policies for access control and automatically enforce the data access based on each user.
Workflow The platform must have the ability to create and modify workflow so that it is easy to scale out and automate various tasks within the platform.
Integration Other third-party technologies such as ticketing or orchestration must be able to integrate into the platform via script or REST APIs.

Top questions to ask

Once your organization agrees on the high-level objectives and goals, there will be many questions from multiple groups. It’s crucial to gather these questions in order to craft a plan to address all of the concerns. Some example questions that you may run into during the initial phase:

  1. What are the main organization data sources and data systems?
  2. For data sources that are not supported yet by Purview, what are my options?
  3. How many Purview instances do we need?
  4. Who are the users?
  5. Who can scan new data sources?
  6. Who can modify content inside of Purview?
  7. What process can I use to improve the data quality in Purview?
  8. How to bootstrap the platform with existing critical assets, glossary terms, and contacts?
  9. How to integrate with existing systems?
  10. How to gather feedback and build a sustainable process?

While you might not have the answer to most of these questions right away, it can help your organization to frame this project and ensure all “must-have” requirements can be met.

Include the right stakeholders

To ensure the success of implementing Purview for the entire enterprise, it’s important to involve the right stakeholders. Only a few people are involved in the initial phase. However, as the scope expands, you will require additional personas to contribute to the project and provide feedback.

Some key stakeholders that you may want to include:

Persona Roles
Chief Data Officer The CDO oversees a range of functions that may include data management, data quality, master data management, data science, business intelligence, and creating data strategy. They can be the sponsor of the Purview implementation project.
Domain/Business Owner A business person who influences usage of tools and has budget control
Data Analyst Able to frame a business problem and analyze data to help leaders make business decisions
Data Architect Design databases for mission-critical line-of-business apps along with designing and implementing data security
Data Engineer Operate and maintain the data stack, pull data from different sources, integrate and prepare data, set up data pipelines
Data Scientist Build analytical models and set up data products to be accessed by APIs
DB Admin Own, track, and resolve database-related incidents and requests within service-level agreements (SLAs); May set up data pipelines
DevOps Line-of-Business application development and implementation; may include writing scripts and orchestration capabilities
Data Security Specialist Assess overall network and data security, which involves data coming in and out of Purview

Identify key scenarios

Purview can be used to centrally manage data governance across an organization’s data estate spanning cloud and on-premises environments. To have a successful implementation, you must identify key scenarios that are critical to the business. These scenarios can cross business unit boundaries or impact multiple user personas either upstream or downstream.

These scenarios can be written up in various ways, but you should include at least these five dimensions:

  1. Persona – Who are the users?
  2. Source system – What are the data sources such as Azure Data Lake Storage Gen2 or Azure SQL Database?
  3. Impact Area – What is the category of this scenario?
  4. Detail scenarios – How the users use Purview to solve problems?
  5. Expected outcome – What is the success criteria?

The scenarios must be specific, actionable, and executable with measurable results. Some example scenarios that you can use:

Scenario Detail Persona
Catalog business-critical assets I need to have information about each data sets to have a good understanding of what it is. This scenario includes both business and technical metadata data about the data set in the catalog. The data sources include Azure Data Lake Storage Gen2, Azure Synapse DW, and/or Power BI. This scenario also includes on-premises resources such as SQL Server. Business Analyst, Data Scientist, Data Engineer
Discover business-critical assets I need to have a search engine that can search through all metadata in the catalog. I should be able to search using technical term, business term with either simple or complex search using wildcard. Business Analyst, Data Scientist, Data Engineer, Data Admin
Track data to understand its origin and troubleshoot data issues I need to have data lineage to track data in reports, predictions, or models back to its original source and understand the changes and where the data has resided through the data life cycle. This scenario needs to support prioritized data pipelines Azure Data Factory and Databricks. Data Engineer, Data Scientist
Enrich metadata on critical data assets I need to enrich the data set in the catalog with technical metadata that is generated automatically. Classification and labeling are some examples. Data Engineer, Domain/Business Owner
Govern data assets with friendly user experience I need to have a Business glossary for business-specific metadata. The business users can use Purview for self-service scenarios to annotate their data and enable the data to be discovered easily via search. Domain/Business Owner, Business Analyst, Data Scientist, Data Engineer

Deployment models

If you have only one small group using Purview with basic consumption use cases, the approach could be as simple as having one Purview instance to service the entire group. However, you may also wonder whether your organization needs more than one Purview instance. And if using multiple Purview instances, how can employees promote the assets from one stage to another.

Determine the number of Purview instances

In most cases, there should only be one Purview account for the entire organization. This approach takes maximum advantage of the “network effects” where the value of the platform increases exponentially as a function of the data that resides inside the platform.

However, there are exceptions to this pattern:

  1. Testing new configurations – Organizations may want to create multiple instances for testing out scan configurations or classifications in isolated environments. Although there is a “versioning” feature in some areas of the platform such as glossary, it would be easier to have a “disposable” instance to freely test.
  2. Separating Test, Pre-production and Production – Organizations want to create different platforms for different kinds of data stored in different environments. It is not recommended as those kinds of data are different content types. You could use glossary term at the top hierarchy level or category to segregate content types.
  3. Conglomerates and federated model – Conglomerates often have many business units (BUs) that operate separately, and, in some cases, they won't even share billing with each other. In those cases, the organization will end up creating a Purview instance for each BU. This model is not ideal, but may be necessary, especially because BUs are often not willing to share billing.
  4. Compliance – There are some strict compliance regimes, which treat even metadata as sensitive and require it to be in a specific geography. If a company has multiple geographies, the only solution is to have multiple Purview instances, one for each geography.

Create a process to move to production

Some organizations may decide to keep things simple by working with a single production version of Purview. They probably don’t need to go beyond discovery, search, and browse scenarios. If some assets have incorrect glossary terms, it’s quite forgiving to let people self-correct. However, most organizations that want to deploy Purview across various business units will want to have some form of process and control.

Another important aspect to include in your production process is how classifications and labels can be migrated. Purview has over 90 system classifiers. You can apply system or custom classifications on file, table, or column assets. Classifications are like subject tags and are used to mark and identify content of a specific type found within your data estate during scanning. Sensitivity labels are used to identify the categories of classification types within your organizational data, and then group the policies you wish to apply to each category. It makes use of the same sensitive information types as Microsoft 365, allowing you to stretch your existing security policies and protection across your entire content and data estate. It can scan and automatically classify documents. For example, if you have a file named multiple.docx and it has a National ID number in its content, Purview will add classification such as EU National Identification Number in the Asset Detail page.

In Purview, there are several areas where the Catalog Administrators need to ensure consistency and maintenance best practices over its life cycle:

  • Data assets – Data sources will need to be rescanned across environments. It’s not recommended to scan only in development and then regenerate them using APIs in Production. The main reason is that the Purview scanners do a lot more “wiring” behind the scenes on the data assets, which could be complex to move them to a different Purview instance. It’s much easier to just add the same data source in production and scan the sources again. The general best practice is to have documentation of all scans, connections, and authentication mechanisms being used.
  • Scan rule sets – This is your collection of rules assigned to specific scan such as file type and classifications to detect. If you don’t have that many scan rule sets, it’s possible to just re-create them manually again via Production. This will require an internal process and good documentation. However, if you rule sets change on the daily or weekly basis, this could be addressed by exploring the REST API route.
  • Custom classifications – Your classifications may not also change on a regular basis. During the initial phase of deployment, it may take some time to understand various requirements to come up with custom classifications. However, once settled, this will require little change. So the recommendation here is to manually migrate any custom classifications over or use the REST API.
  • Glossary – It’s possible to export and import glossary terms via the UX. For automation scenarios, you can also use the REST API.
  • Resource set pattern policies – This functionality is very advanced for any typical organizations to apply. In some cases, your Azure Data Lake Storage has folder naming conventions and specific structure that may cause problems for Purview to generate the resource set. Your business unit may also want to change the resource set construction with additional customizations to fit the business needs. For this scenario, it’s best to keep track of all changes via REST API, and document the changes through external versioning platform.
  • Role assignment – This is where you control who has access to Purview and which permissions they have. Purview also has REST API to support export and import of users and roles but this is not Atlas API-compatible. The recommendation is to assign an Azure Security Group and manage the group membership instead.

Plan and implement different integration points with Purview

It’s likely that a mature organization already has an existing data catalog. The key question is whether to continue to use the existing technology and sync with Purview or not. To handle syncing with existing products in an organization, Purview provides Atlas REST APIs. Atlas APIs provide a powerful and flexible mechanism handling both push and pull scenarios. Information can be published to Purview using Atlas APIs for bootstrapping or to push latest updates from another system into Purview. The information available in Purview can also be read using Atlas APIs and then synced back to existing products.

For other integration scenarios such as ticketing, custom user interface, and orchestration you can use Atlas APIs and Kafka endpoints. In general, there are four integration points with Purview:

  • Data Asset – This enables Purview to scan a store’s assets in order to enumerate what those assets are and collect any readily available metadata about them. So for SQL this could be a list of DBs, tables, stored procedures, views and config data about them kept in places like sys.tables. For something like Azure Data Factory (ADF) this could be enumerating all the pipelines and getting data on when they were created, last run, current state.
  • Lineage – This enables Purview to collect information from an analysis/data mutation system on how data is moving around. For something like Spark this could be gathering information from the execution of a notebook to see what data the notebook ingested, how it transformed it and where it outputted it. For something like SQL, it could be analyzing query logs to reverse engineer what mutation operations were executed and what they did. We support both push and pull based lineage depending on the needs.
  • Classification – This enables Purview to take physical samples from data sources and run them through our classification system. The classification system figures out the semantics of a piece of data. For example, we may know that a file is a Parquet file and has three columns and the third one is a string. But the classifiers we run on the samples will tell us that the string is a name, address, or phone number. Lighting up this integration point means that we have defined how Purview can open up objects like notebooks, pipelines, parquet files, tables, and containers.
  • Embedded Experience – Products that have a “studio” like experience (such as ADF, Synapse, SQL Studio, PBI, and Dynamics) usually want to enable users to discover data they want to interact with and also find places to output data. Purview’s catalog can help to accelerate these experiences by providing an embedding experience. This experience can occur at the API or the UX level at the partner’s option. By embedding a call to Purview, the organization can take advantage of Purview’s map of the data estate to find data assets, see lineage, check schemas, look at ratings, contacts etc.

Phase 1: Pilot

In this phase, Purview must be created and configured for a very small set of users. Usually, it is just a group of 2-3 people working together to run through end-to-end scenarios. They are considered the advocates of Purview in their organization. The main goal of this phase is to ensure key functionalities can be met and the right stakeholders are aware of the project.

Tasks to complete

Task Detail Duration
Gather & agree on requirements Discussion with all stakeholders to gather a full set of requirements. Different personas must participate to agree on a subset of requirements to complete for each phase of the project. 1 Week
Set up the starter kit Go through Purview Quick Start and set up the Purview Starter Kit to demo Purview to all stakeholders. 1 Day
Navigating Purview Understand how to use Purview from the home page. 1 Day
Configure ADF for lineage Identify key pipelines and data assets. Gather all information required to connect to an internal ADF account. 1 Day
Scan a data source such as Azure Data Lake Storage Add the data source and set up a scan. Ensure the scan successfully detects all assets. 2 Day
Search and browse Allow end users to access Purview and perform end-to-end search and browse scenarios. 1 Day

Acceptance criteria

  • Purview account is created successfully in organization subscription under the organization tenant.
  • A small group of users with multiple roles can access Purview.
  • Purview is configured to scan at least one data source.
  • Users should be able to extract key values of Purview such as:
    • Search and browse
    • Lineage
  • Users should be able to assign asset ownership in the asset page.
  • Presentation and demo to raise awareness to key stakeholders.
  • Buy-in from management to approve additional resources for MVP phase.

Phase 2: Minimum viable product

Once you have the agreed requirements and participated business units to onboard Purview, the next step is to work on a Minimum Viable Product (MVP) release. In this phase, you will expand the usage of Purview to more users who will have additional needs horizontally and vertically. There will be key scenarios that must be met horizontally for all users such as glossary terms, search, and browse. There will also be in-depth requirements vertically for each business unit or group to cover specific end-to-end scenarios such as lineage from Azure Data Lake Storage to Azure Synapse DW to Power BI.

Tasks to complete

Task Detail Duration
Scan Azure Synapse Analytics Start to onboard your database sources and scan them to populate key assets 2 Days
Create custom classifications and rules Once your assets are scanned, your users may realize that there are additional use cases for more classification beside the default classifications from Purview. 2-4 Weeks
Scan Power BI If your organization uses Power BI, you can scan Power BI in order to gather all data assets being used by Data Scientists or Data Analysts which have requirements to include lineage from the storage layer. 1-2 Weeks
Import glossary terms In most cases, your organization may already develop a collection of glossary terms and term assignment to assets. This will require an import process into Purview via .csv file. 1 Week
Add contacts to assets For top assets, you may want to establish a process to either allow other personas to assign contacts or import via REST APIs. 1 Week
Add sensitive labels and scan This might be optional for some organizations, depending on the usage of Labeling from M365. 1-2 Weeks
Get classification and sensitive insights For reporting and insight in Purview, you can access this functionality to get various reports and provide presentation to management. 1 Day
Onboard additional users using Purview managed users This step will require the Purview Admin to work with the Azure Active Directory Admin to establish new Security Groups to grant access to Purview. 1 Week

Acceptance criteria

  • Successfully onboard a larger group of users to Purview (50+)
  • Scan business critical data sources
  • Import and assign all critical glossary terms
  • Successfully test important labeling on key assets
  • Successfully met minimum scenarios for participated business units’ users

Phase 3: Pre-production

Once the MVP phase has passed, it’s time to plan for pre-production milestone. Your organization may decide to have a separate instance of Purview for pre-production and production, or keep the same instance but restrict access. Also in this phase, you may want to include scanning on on-premises data sources such as SQL Server. If there is any gap in data sources not supported by Purview, it is time to explore the Atlas API to understand additional options.

Tasks to complete

Task Detail Duration
Refine your scan using scan rule set Your organization will have a lot of data sources for pre-production. It’s important to pre-define key criteria for scanning so that classifications and file extension can be applied consistently across the board. 1-2 Days
Assess region availability for scan Depending on the region of the data sources and organizational requirements on compliance and security, you may want to consider what regions must be available for scanning. 1 Day
Understand firewall concept when scanning This step requires some exploration of how the organization configures its firewall and how Purview can authenticate itself to access the data sources for scanning. 1 Day
Understand Private Link concept when scanning If your organization uses Private Link, you must lay out the foundation of network security to include Private Link as a part of the requirements. 1 Day
Scan on-premises SQL Server This is optional if you have on-premises SQL Server. The scan will require setting up Self-hosted Integration Runtime and adding SQL Server as a data source. 1-2 Weeks
Use Purview REST API for integration scenarios If you have requirements to integrate Purview with other 3rd party technologies such as orchestration or ticketing system, you may want to explore REST API area. 1-4 Weeks
Understand Purview pricing This step will provide the organization important financial information to make decision. 1-5 Days

Acceptance criteria

  • Successfully onboard at least one business unit with all of users
  • Scan on-premises data source such as SQL Server
  • POC at least one integration scenario using REST API
  • Complete a plan to go to production which should include key areas on infrastructure and security

Phase 4: Production

The above phases should be followed to create an effective information governance, which is the foundation for better governance programs. Data governance will help your organization prepare for the growing trends such as AI, Hadoop, IoT, and blockchain. It is just the start for many things data and analytics, and there is plenty more that can be discussed. The outcome of this solution would deliver:

  • Business Focused - A solution that is aligned to business requirements and scenarios over technical requirements.
  • Future Ready - A solution will maximize default features of the platform and use standardized industry practices for configuration or scripting activities to support the advancements/evolution of the platform.

Tasks to complete

Task Detail Duration
Scan production data sources with Firewall enabled If this is optional when firewall is in place but it’s important to explore options to hardening your infrastructure. 1-5 Days
Enable Private Link If this is optional when Private Link is used. Otherwise, you can skip this as it’s a must-have criterion when Private is enabled. 1-5 Days
Create automated workflow Workflow is important to automate process such as approval, escalation, review and issue management. 2-3 Weeks
Operation documentation Data governance is not a one-time project. It is an ongoing program to fuel data-driven decision making and creating opportunities for business. It is critical to document key procedure and business standards. 1 Week

Acceptance criteria

  • Successfully onboard all business unit and their users
  • Successfully meet infrastructure and security requirements for production
  • Successfully meet all use cases required by the users

Platform Hardening

Additional hardening steps can be taken:

  • Increase security posture by enabling scan on firewall resources or use Private Link
  • Fine-tune scope scan to improve scan performance
  • Use REST APIs to export critical metadata and properties for backup and recovery
  • Use workflow to automate ticketing and eventing to avoid human errors

Next steps