Ready to dive into the lake?
lakeFS is currently only
available on desktop.

For an optimal experience, provide your email below and one of our lifeguards will send you a link to start swimming in the lake!

lakeFS Community
The lakeFS Team
The lakeFS Team Author

lakeFS is on a mission to simplify the lives of...

Published on August 3, 2023

4 Metadata Management Challenges And How To Solve Them

Modern data architectures support an increasing number and variety of business use cases. Product creation, tailored customer experiences, fraud detection, regulatory compliance, and data monetization are some examples.

To enable such use cases, a data-driven business needs modern solutions for accessing, managing, and processing data. One of them is a data lake, a centralized repository for storing, processing, and securing large volumes of structured, semi-structured, and unstructured data. 

One of the key issues to consider here is metadata and how you manage it. Since a data lake is not a database, the amount of metadata it provides is much lower than what you would get from a database. 

This is just one of the several challenges teams encounter when managing metadata in data lakes.

Metadata management in data lakes: why is it so tricky?

Metadata provides context to the content of data sets and is a crucial component in making data comprehensible and accessible in applications. 

However,, since raw data is frequently fed into a data lake, many organizations fail to incorporate the procedures required to verify the data or apply organizational data standards to it. Because of the lack of effective metadata management, the data in a data lake is less helpful for analytics.

To manage data in a data lake efficiently, you need a framework for recording technical, operational, and business data so that you can identify and exploit your data for multiple use cases. 

A data lake management platform is one approach to automating metadata management. This type of platform can automate the collection of metadata on arrival while performing transformations and relate it to specific meanings, such as those in an enterprise business lexicon, which assures that all users are consistently reading the same data according to a set of rules and concepts—and it can be automatically updated as your data changes.

Solid metadata management capabilities simplify and automate routine data administration. A poor metadata architecture can prevent data lakes from progressing from an analytical sandbox or proof of concept (POC) with limited data sets and one use case to a production-ready, enterprise-wide data platform with many users and multiple use cases – i.e., a modern data architecture. It also dramatically impacts the performance of any data analysis with the use of data sets stored in the data lake.

Let’s dive into the challenges of managing metadata in data lakes and take a look at the potential approaches that help teams solve the problem.

4 challenges of managing metadata in a data lake

1. Data siloization

In many organizations, teams keep data in separate buckets, and those buckets are partitioned according to the path to the data. The same applies to metadata.

Teams often lack file system capabilities or hierarchy that enable efficient usage of metadata. Keeping data in separate silos that don’t align has serious implications for the organization’s ability to leverage metadata to support its most important use cases.

2. Immutability vs. Mutability

While data lakes are immutable, the data itself is not. The nature of data is transient, as it constantly changes. Examples of changing data are:

  • backfills in operational data, 
  • re-processing of data due to improvements in ML models or other algorithms, 
  • or just data that describes the world, such as sensor data, that changes because the world changes. 

Consider a map or a set of patient data over time – they are bound to change.

Metadata is needed to manage the tracking of changes in a data set over time and the different versions of the data to allow concurrency. To do that, teams need a layer of metadata that isolates the data set as it’s currently used from the changes that are accumulating and expose the changes to consumers in an atomic action.

3. Versioning

Since data sets change over time and organizations manage a large number of data sets, metadata is required to manage the changes over time of more than just one data set (as presented in the immutability section) but rather to maintain the consistency between all the data sets ingested and derived within the data lake. 

Versions are an essential requirement for efficient metadata management in a data lake. Data version control ensures consistency and opens the door to full reproducibility and auditing.

Without data versioning, it is difficult to find the data sets that were involved in a certain analysis as they were at the time of the analysis.

4. Need for an abstraction layer

Data lakes are implemented by using object stores. Object stores manage files without even considering them as files but rather as objects. The consumers of data from the data lake are often data engineers using SQL who are used to querying tables in a database. 

How can we serve them well in a data lake? When managing data that can be presented as a table, teams need an abstraction that will allow them to access a set of files as if they were accessing a table in a database. Turning a set of files into a table is something one does with metadata. 

4 approaches to solving the challenges of managing data in data lakes

1. Metastores

Metadata includes technical metadata about a data set’s structure, data type, and statistical information about the values in each column. You need this data to assemble and run analytic queries like SQL statements.

Technical metadata is especially important when discussing data lakes. That’s because, unlike integrated database repositories like RDBMS that contain built-in technical metadata, technical metadata is a distinct component in a data lake that teams must deliberately set up and maintain. The use of a metastore can assist in the first and fourth challenges listed above for your structured data.

The most widely used metastore interface is Hive Metastore, and a wide range of large data query and processing engines and libraries support it. The data in a Hive Metastore is just as significant as the data in the data lake and must be treated as such. This implies that its metadata must be kept permanent, highly accessible, and included in any disaster recovery configuration.

2. Open table formats

When migrating structured data from a relational database to object storage, you lose many of the typical database guarantees. Open table formats (OTFs) address this need by offering a table abstraction that lets you create, insert, update, and remove information. It also helps in the management of table schema evolution and allows for some concurrency and transactionality. The OTFs available today, all OSS based, are Apache Hudi, Apache Iceberg, and Delta Lake

While each format operates somewhat differently, the basic premise is the same.

While the table’s data is stored in immutable Parquet files, changes are saved in extra data files, and information about how to utilize those delta files is stored in log files. 

All readers and authors must agree on the order of the log in order to prevent corruption or discrepancies.

One interesting side effect of open table formats is time travel. They let you examine different versions of a given table, indicating changes over time by iterating through the sorted log. You can switch between versions of a table by reading up to a specific log point, assuming the table wasn’t compacted.

OTFs are a good solution to the second challenge of immutability vs. mutability, mentioned above.

3. Metastore for open table formats

There are solutions on the market that bring both metastore and open table formats together to deliver a wide range of metadata that is missing in the lake, with control capabilities such as permissions, auditing, and governance… Tabular and DataBricks Unity Catalog are good examples of that. 

Tabular is a platform for cloud-native warehousing and automation that serves as a centralized repository for all of your analytic data. Tabular storage works similarly to a typical data warehouse. You can build tables with well-defined schemas, read and write data using their preferred tool, and govern data access with a robust Role-Based Access governance system.

Unity Catalog provides a centralized location for managing data access controls that apply to all workspaces and personas. Unity Catalog’s security model is based on standard ANSI SQL and allows administrators to issue rights at the catalog, database (also known as schema), table, and view levels in their current data lake using familiar syntax.

Unity Catalog records user-level audit logs that document data access, collect lineage data, and allows you to categorize and record data assets while also providing a search interface to assist data consumers in finding data. And it comes with top-level metastores offering a three-level namespace that organizes your data.

4. Data Version Control systems for data lakes

Live data systems constantly absorb new data as different users experiment on the same datasets. This may easily lead to many versions of the same dataset, which becomes a management challenge.

Data version control helps teams address it by bringing a well-established approach to versioning source code to the world of data. Many data versioning tools are available as open-source, so teams can start experimenting with them right away.

lakeFS is an open-source version management system that operates on top of a data lake and is built on Git-like semantics. Data engineers and data scientists can use it to version control their data while building and maintaining data pipelines and ML models to ensure reproducibility, collaboration, and top-notch results. 

Another example is Nessie Project, an open-source solution that provides greater control and consistency. It draws inspiration from GitHub, a platform where developers create, test, release, and update software versions. By extending analogous development processes and concepts to data, Nessie enables data engineers to update, restructure, and correct datasets while maintaining a consistent version of reality, opening the door to DataOps.

Wrap up

A data lake’s metadata is critical for its operations, and its management provides teams with capabilities that will put the data lake on par with a database experience for lake users.

To get the most out of a data lake, you must be able to assure data quality and dependability while also democratizing data access. Democratizing access entails giving more users across the company access and making it easier for users to find the data they need. 

All of this vital functionality relies on implementing a strong, scalable system for capturing and managing information. Metadata is unquestionably the most important component of a successful next-generation data architecture.

Git for Data – lakeFS

  • Get Started
    Get Started