Data Lakehouses for Generative AI: SQream, Domo, Snowflake, Matillion
In a 2024 AI Connect webinar, panelists from SQream, Domo, Snowflake and Matillion explained why data lakehouses suit generative AI. They covered migration and interoperability, avoiding a data junk drawer, governance, GPU costs and where architectures are headed.
By The AI Enterprise Magazine Team4 min read
Key takeaways
- One copy, many workloads. Balkenende said a lakehouse lets BI, AI and warehousing use the same data without replication.
- Pick an open standard. Interoperability across clouds and engines is the top technical challenge, he said; he pointed to Iceberg and Delta Lake.
- Avoid the junk drawer. Rowley urged specific, project-by-project adoption instead of dumping all data in.
- Govern from day one. Lineage, literacy and observability help cut unused data, improving performance and cost.
- Expect composable architectures. Panelists predicted plug-and-play components, GPU processing and stronger cataloging.
In a 2024 AI Connect webinar for data executives, four vendor and practitioner panelists explained how data lakehouses support generative AI. The panel was made up of Yaniv Leven, VP of Market Strategy at SQream, Kristie Rowley, PhD, Principal Data Scientist and Senior Director of AI Labs at Domo, Arun Agarwal, Senior Product Marketing Manager for AI/ML at Snowflake, and Mark Balkenende, VP of Product Marketing at Matillion. A AI Connect co-founder moderated, and Snowflake, Matillion, Domo and SQream were named as sponsors in the introduction.
What is a data lakehouse and why does it matter for AI?
Agarwal described three trends: data is growing in volume, sources and formats, consumption now ranges from analytics to generative AI, and architectures are changing as a result. A lakehouse combines the adaptability and low cost of a data lake with the performance, governance and quality of a data warehouse, he said, by decoupling storage and compute.
Rowley said it brings structured data (what fits in a spreadsheet) together with unstructured data such as images, text and email, so teams can "triangulate" findings. Leven said the lakehouse adds a management layer and ACID capabilities to a lake, making it more manageable.
Balkenende said the point is to keep data in one place, in mostly raw form, and let different workloads such as BI, AI and warehousing use it without copying it between environments.
How do lakehouses help generative AI projects?
Leven said the answer depends on the project. Fine-tuning an existing model can need gigabytes and weeks, he said, while building a model from scratch can need hundreds of terabytes to petabytes and months. For fine-tuning, he said, a lakehouse is generally the best choice, since transactions after training require management features and traditional warehouses struggle with unstructured data.
Balkenende said lakehouses open unstructured data to people beyond highly skilled data scientists, and generative AI lowers the bar further. Agarwal described the path from raw data to a language model as a series of steps, from extraction and cleaning to prediction and generation. A unified platform helps keep lineage across them, he said, and Snowflake talks about bringing AI to the data rather than data to the model.
What are the biggest challenges in adopting a lakehouse?
Balkenende named interoperability. Data may sit across several clouds in formats such as Parquet, JSON and CSV, and not every engine can read all of it. He pointed to open table formats like Iceberg and Delta Lake and urged teams to "pick a standard" so they do not have to reformat data later.
Leven said the hardest part is migrating data while keeping quality, and that data movement never ends. He said governance, cataloging and workflow tools are better bought off the shelf than built.
Agarwal used a people, process and technology lens. He suggested a small, well-scoped first project, such as a marketing use case, then scaling, and building a center of excellence. Rowley warned about the "junk drawer" problem. "I would counsel people to not make more messes while you're anxious to adopt new technology," she said, adding that teams do not need to boil the ocean.
How do you protect performance and cost?
Balkenende said not to repeat the Hadoop-era mistake of throwing in all the data. Start with governance, literacy and lineage, he said, and use data observability to find terabytes nobody uses, which improves performance and cuts cost.
Leven, speaking from SQream's perspective, expected GPU processing to become cheaper and more commoditized, and said lakehouses should be built to use GPUs as well as CPUs. Agarwal said teams should reduce data movement for latency and security reasons and avoid proprietary formats and copies.
What does the panel expect next?
Rowley said everyone will start thinking about data cleanliness and readiness the way data scientists do, because chatbots need well-described data and data dictionaries. Agarwal pointed to GPU performance and to trust and ethics, including cataloging and semantic layers.
Balkenende expected composable, interoperable architectures, and said he does not think generative AI will replace traditional BI reporting. Leven said generative AI moves computing "from a world of ones and zeros to a world of probability," and predicted that once ACID is easier in lakehouses, warehousing will be "a thing of the past." That is his view, not a consensus of the panel.
On hype, Rowley said generative AI will not answer the most important business questions, but "it's not magic, it's math." Agarwal listed batch text processing, business intelligence and document assistants as common use cases. Rowley cited text-to-SQL as a success at Domo, and Balkenende described a customer support model that drafts responses for staff to review.
Where can you hear more from data and AI leaders?
AI Connect, formerly AI Connect, hosts 24 free live webinars on its 2027 schedule, including the AI × Enterprise Panel Series. Registration is coming soon. See /webinars/.
Frequently asked questions
What is a data lakehouse?
Panelists described it as combining a data lake's low cost and flexibility with a data warehouse's performance, governance and quality, by decoupling storage and compute.
Why do lakehouses matter for generative AI?
They hold structured and unstructured data in one place and support fine-tuning and retrieval workflows, according to the panel.
What is the biggest challenge in adopting a lakehouse?
Balkenende named interoperability across clouds and formats; Leven named migrating data while keeping quality.
Will generative AI replace traditional BI?
Balkenende said he does not expect generative AI to replace traditional BI reporting, and expected composable architectures instead.
This article is based on remarks by Yaniv Leven, VP of Market Strategy, SQream; Kristie Rowley, PhD, Principal Data Scientist and Sr. Director, AI Labs, Domo; Arun Agarwal, Senior Product Marketing Manager, AI/ML, Snowflake; Mark Balkenende, VP of Product Marketing, Matillion at AI Connect webinar (Generative AI for Data Executives) on June 18, 2024. More coverage: Webinars.
Filed under: Data & AI Infrastructure, AI Governance & Trust, AI Use Cases & Applications
