July 2026 Summaries

Data governance serves as an organizational framework ensuring secure, compliant, and efficient handling of data across an organization's analytics ecosystem, comparable to an operating system managing access and operations on a computer. It establishes decision rights and accountability for data creation, consumption, and control, addressing critical questions about data access, security, and auditability, which are essential for the success of data initiatives. As modern data platforms become more complex, integrating diverse systems such as data lakes and discovery tools, governance becomes crucial not just for security but also for maintaining data lineage, quality, and consent management, especially in AI and ML contexts. Proper governance prevents technical debt and facilitates faster, compliant data operations by providing clear interfaces to data assets. Many organizations face challenges in implementing governance due to technical complexities, performance overheads, and identity management across disparate systems, often resulting in governance drift. Successful governance requires practical, integrated systems that align with existing tools and workflows, starting small and expanding systematically to ensure consistent enforcement and auditability without compromising performance.

Predicate pushdown is a query optimization technique that enhances data processing efficiency by shifting filtering conditions from the query engine to the data source, thus reducing the amount of data transferred over the network. This method is crucial in modern data architectures, particularly within federated query engines and data lakehouses, as it minimizes scan costs, network I/O, and latency by applying filters at the data source level. Beyond simple predicates, pushdown can extend to projections, aggregations, joins, and dereferences, although its effectiveness hinges on the specific connectors and data sources used. The technique is vital for cost control and performance in environments like Amazon Athena and BigQuery, where charges are based on data volume processed or scanned. Columnar file formats such as Parquet and ORC further enhance pushdown effectiveness by allowing data skipping based on predicates, thus accelerating processes like feature engineering and dashboard operations. However, successful implementation requires understanding connector limitations, optimizing data layouts, and continuously validating pushdown effectiveness through EXPLAIN plans. Advanced strategies for complex scenarios include leveraging parallel extraction paths, using cached or materialized views, and implementing governance controls to ensure compliance and security.

AI governance and data governance, while related, serve distinct purposes and require separate frameworks to ensure comprehensive oversight. Data governance focuses on managing the quality, security, lineage, and lifecycle of data assets, ensuring accurate and unaltered inputs. In contrast, AI governance monitors model behavior post-data consumption, addressing issues like explainability, model drift, and bias. The assumption that AI governance is merely an extension of data governance is flawed because data governance lacks mechanisms for monitoring model outputs. The rise of federated data environments complicates governance further, requiring governance to be applied at the query layer rather than relying on centralized systems. This shift necessitates that data products, which include built-in metadata, quality standards, and access policies, act as operational bridges between the two governance disciplines. Compliance with regulations such as the EU AI Act highlights the urgent need for organizations to develop AI governance structures alongside existing data governance programs to ensure that both data inputs and model outputs are reliably governed.

AI agents have transformed the landscape of enterprise data usage, compelling data products to bridge the gap between machine processing and human-derived insights. Unlike humans, AI agents rely solely on the data product as their source of truth and lack the ability to interpret context or resolve ambiguities through inquiry, making the precision and completeness of data products crucial. AI-ready data products must embed metric logic, instructional metadata, and maintain a rigorous version history, ensuring that agents can operate without needing additional context or clarification. This involves treating metadata with the same care as application code and balancing AI's speed with human oversight to certify data interpretations. The success of AI in enterprises hinges on this disciplined approach, as data products become the fundamental building blocks of a comprehensive enterprise context layer, able to deliver trusted insights independently of human intervention.

In the evolving realm of enterprise AI, success hinges on overcoming the persistent challenge of data access rather than relying on centralization strategies that have historically fallen short. As AI agents, especially those powered by large language models, become integral to decision-making, they require access to a diverse range of contextual data spread across various sources, including legacy systems and modern tools. This challenge is exacerbated by the fragmented, unstructured nature of data and the need for stringent governance measures. The traditional centralized approach, which has failed in past data initiatives such as data warehouses and lakes, is impractical due to the exponential growth of data and its dispersed nature. Instead, a model of universal data access is proposed, allowing data to be accessed where it resides, thereby enabling faster and more efficient data utilization. Starburst’s Enterprise Intelligence Platform exemplifies this approach by offering federated access, integrated governance, and contextual understanding, facilitating AI's ability to reach and process data effectively across disparate environments without the need for exhaustive data migration.

Building trustworthy AI data products for enterprise use hinges on creating reliable, well-documented, and governed data layers rather than focusing solely on the AI model itself. The process begins with the careful curation of data products, which involves transforming raw data into structured, meaningful views that align with real business questions. This requires collaboration between data engineers and business experts to ensure accuracy and relevance, along with the meticulous creation and maintenance of metadata to provide context and prevent misinterpretations. Governance and access control are critical, as they ensure that sensitive information is protected and only accessible to appropriate users. While this foundational work may seem slow initially, it establishes a robust framework that accelerates future developments and instills confidence in the AI's outputs, making it a worthwhile investment for decision-makers.

Starburst and Dell's partnership aims to address the critical issue of enterprise data fragmentation by creating a robust data foundation necessary for AI production success. Over the past two years, their collaboration has evolved into a deep co-engineering effort, introducing the Dell Data Analytics Engine powered by Starburst, which decouples compute from storage and utilizes open standards to prevent vendor lock-in. This architecture enhances data accessibility and integrity, enabling faster insights and reduced costs compared to traditional cloud migration. The partnership has also developed a unified context layer that integrates with NVIDIA technologies, ensuring AI systems have high-performance data retrieval and reasoning capabilities while maintaining business logic and security constraints. By eliminating mandatory ETL processes, this joint solution reduces data pipeline latency, a primary bottleneck for AI, allowing for efficient AI inference across hybrid environments. Additionally, their efforts have led to significant performance milestones with GPU acceleration, showcasing remarkable speed improvements in benchmarks such as TPC-H and ClickBench, ultimately making AI infrastructure economically viable. As they continue to innovate, Starburst and Dell are committed to expanding their high-performance engine to support future agentic workloads, ensuring enterprises have a governed path to the AI-driven future.

In the context of enterprise AI, universal data access has shifted from a mere advantage to a crucial requirement, especially when comparing data platforms like Starburst and Dremio. Starburst, built for data federation, offers extensive reach across over 50 data sources with pushdown optimization, providing a significant price-performance edge and lower total cost of ownership compared to Dremio, which is focused on lakehouse architectures with around 20 connectors. The architectural differences are highlighted by the ability of Starburst to handle high-concurrency, fault-tolerant workloads through its open-source Trino engine, while Dremio's performance is optimized for a single lakehouse environment. This distinction is crucial in the AI era, where agents require real-time access to diverse data sources to function effectively. Starburst's approach, including its support for multiple open table formats, offers greater flexibility and efficiency in data querying, making it a more suitable choice for enterprises with complex, multi-system data landscapes. While Dremio may suffice for single lakehouse operations with predictable workloads, Starburst's comprehensive access capabilities provide a more robust solution for the demands of modern AI-driven analytics.

Query optimization is the process by which database engines and query processors select the most efficient execution plan for SQL statements, relying heavily on cost-based strategies that estimate computational costs. This process is crucial in modern data landscapes where queries may span various platforms like object storage, cloud warehouses, and streaming services. Effective optimization enhances performance and reduces costs by minimizing network transfers and efficiently managing data movement. The rise of cloud-native analytics has shifted optimization focus beyond traditional CPU and memory usage to include factors like cross-cloud latency and storage scan costs. Challenges in query optimization arise from the heterogeneous nature of modern data ecosystems, where different source systems offer varying levels of optimization support and maintain statistics inconsistently. In federated environments, cross-system joins can trigger costly data migrations if not optimized correctly. Strategies for successful query optimization include choosing table formats like Apache Iceberg that support metadata-driven pruning, maintaining quality statistics, implementing fault tolerance for critical workflows, and using materialized views to transform complex queries into faster lookups. By systematically building optimization capabilities, organizations can ensure efficient, reliable, and scalable data pipelines that adapt to evolving data patterns and business needs.

Rippling enhances its HR, IT, and Finance analytics capabilities by utilizing Starburst's open, federated data platform to manage data across diverse sources without centralization, thus maintaining flexibility and reducing costs. The integration of Starburst enables Rippling to query data where it resides, employing a foundation built on Apache Iceberg and Trino, which avoids the pitfalls of traditional data centralization, such as latency and lock-in. Rippling's platform, powered by the Rippling Query Language (RQL), extends Starburst's capabilities with native support for business-specific semantics and custom functions, allowing for efficient multi-currency analytics and real-time querying of live service data through virtual tables. This collaboration has resulted in a robust architecture that supports millions of queries daily with minimal latency and operational overhead, showcasing a scalable and adaptable solution that meets the demands of diverse business contexts and customer-specific schemas.

A data catalog is a crucial component in modern data architectures, acting as a control plane that maintains metadata about data locations, structures, ownership, and relationships, rather than storing the data itself. Key cloud providers offer catalog solutions, such as AWS Glue Data Catalog and Google Cloud's Dataplex Universal Catalog, which enhance metadata management within their ecosystems. These catalogs are vital for powering analytics engines, enforcing policies, tracking data lineage, and supporting AI and machine learning workflows by providing governance and data quality assurance. Data catalogs help eliminate data duplication, ensuring analysts access current and properly governed information, and significantly reduce the time spent on data discovery and preparation. Despite their benefits, implementing data catalogs can be complex due to cloud platform fragmentation, metadata staleness, and permission mismatches. Successful implementation requires a methodical approach, including unifying identity and access models, automating metadata updates, and designing for performance and cost optimization. Integrating technical catalogs with enterprise discovery tools and ensuring end-to-end discovery and lineage are essential for maintaining metadata accuracy and facilitating self-service analytics, AI initiatives, and robust data governance.

Dremio's acquisition by SAP highlights a significant shift in the data platform landscape, where traditional lakehouse engines are increasingly challenged by the demands of AI. These engines, designed to centralize data into a single lake for optimized querying, are becoming insufficient as AI requires broad data reach and real-time responsiveness across diverse systems without precomputed constraints. AI agents operate by generating dynamic, on-the-fly queries that require immediate access to up-to-date data from multiple sources, which traditional lakehouse architectures cannot adequately support. This shift underscores the need for platforms that can federate data access and provide consistent governance across various data sources while handling high concurrency and unpredictable workloads. Consequently, the industry is moving toward architectures that prioritize total data reach, real-time access, and contextual relevance over precomputed efficiency, reshaping how vendors like Starburst approach data platform design.

AI is transforming the landscape of Business Intelligence (BI) by offering fast, conversational, and dynamic data interactions that traditional BI tools struggle to match due to their slow and static nature. The emergence of AI highlights the need for a robust data architecture capable of providing the necessary context for accurate and timely AI responses, which the data lakehouse model efficiently delivers. This architecture combines the flexibility of data lakes with the performance of data warehouses, creating a semantic backbone crucial for AI workloads. Critical features of an AI-ready data lakehouse include data federation, support for unstructured and structured data, strong query performance, and the integration of Apache Iceberg for advanced data management. Additionally, the establishment of a semantic layer through data products is essential for documenting data lineage, ensuring data quality, and providing the context required for AI to function effectively. As AI continues to evolve, data lakehouses are positioned as a central component in facilitating enterprise-wide data access and enhancing AI capabilities.

Enterprise data architecture is undergoing a transformative shift from centralized systems to flexible, distributed foundations, driven by the increasing complexity of AI-driven workloads. Traditional centralized approaches, which placed all data in one rigid platform, are giving way to hybrid cloud architectures that can accommodate data spread across multiple platforms and environments. This shift is prompting the development of new deployment models, such as Starburst's Bring Your Own Cloud (BYOC), which balances the simplicity of SaaS with the control of self-managed environments. As data environments become more distributed, the need for robust resiliency and resource management grows, leading to innovations like Starburst's coordinator high availability and intelligent multi-cluster routing. Additionally, enhanced observability tools are essential for optimizing performance across clusters and workloads, as highlighted by the Starburst console's capabilities. The Starburst Icehouse architecture exemplifies this new era, offering open, high-performance data lakehouse solutions built on Apache Iceberg, which facilitate efficient analytics and AI workloads. Managed operations through features like Starburst LakeOps reduce the operational burden, allowing teams to focus on delivering value rather than managing infrastructure. This evolution represents a shift toward continuously operating, AI-ready data systems that integrate existing technologies into cohesive, manageable solutions.

An article from Berkeley’s EPIC Data Lab titled "Intelligence Is Free. Now What?" proposes a shift in focus from generating intelligence to managing it as inference becomes cheaper, requiring new data systems for AI agents to support long-running operations, coordination, and knowledge accumulation. Concurrently, a paper detailing Trellis, a database architecture centered around the agent experience graph, suggests treating the entire search history of an agent as a primary database abstraction rather than disposable logs. This approach, already partially implemented by Meta, aligns with Berkeley's vision by enabling crash recovery, collaboration, and continuous learning through a shared, durable data system that supports stateless agents and cross-agent coordination. Trellis redefines memory management in AI systems by promoting collective intelligence and treating search processes as database workloads, ultimately positioning itself as a response to Berkeley's research agenda for future data systems in AI.

The enterprise context layer is an essential component for organizations aiming to leverage AI effectively by bridging the gap between raw data and meaningful insights. Traditionally, human analysts have played a crucial role in translating business questions into actionable data queries, but the adoption of AI and text-to-SQL tools necessitates a structured context layer to replace this human intuition. This layer organizes and governs data definitions, meaning, and governance, ensuring that AI systems can accurately interpret and apply business logic. By curating data products, embedding metric logic, and enriching metadata, businesses can create AI-ready data environments that facilitate accurate, auditable insights. The context layer enables trustworthy AI operations by grounding them in a shared semantic structure that evolves with the business, thus transforming AI strategy into actionable business execution and reducing reliance on static dashboards.

Dremio, a data platform known for its SQL analytics on data lakehouses, has been acquired by SAP, raising concerns about its future direction and potential limitations under new ownership. Starburst, a competitor to Dremio, argues that this acquisition could lead to vendor lock-in, changes in priorities, and architectural constraints that impact existing and potential customers. The acquisition may hinder Dremio's independent development and introduce challenges for AI and agentic workloads, particularly concerning concurrency and governance. Starburst positions itself as a robust alternative, emphasizing its open-source Trino foundation, scalability, and governance capabilities through its Context Layer, offering consistent business logic enforcement across queries. Starburst highlights its flexibility and independence from vendor constraints as significant advantages for organizations evaluating their data platform strategies in light of Dremio's acquisition.

The text discusses the integration of SQL and graph analytics on Apache Iceberg within data lakehouse architectures, highlighting the use of Trino and PuppyGraph to perform analytics without duplicating data. Apache Iceberg offers efficient metadata handling and query performance, forming the foundation for large-scale data analysis. Trino serves as a distributed SQL engine, enabling interactive analytics on Iceberg tables, while PuppyGraph facilitates graph analytics by allowing graph models to be defined on relational data without requiring separate graph databases. This unified approach enables organizations to leverage both SQL and graph queries for comprehensive data analysis, such as tracing transaction flows and identifying relationships, all while maintaining a single data layer. Starburst enhances this ecosystem with platforms like Starburst Enterprise and Starburst Galaxy, offering production-ready environments and extending the Icehouse architecture with features like managed Iceberg ingestion and performance optimizations. A demonstration using synthetic financial data illustrates how both SQL and graph analytics can be applied to the same dataset, showcasing the capabilities of this integrated architecture.

Guardrails is a feature introduced by Starburst to enhance governance and security for their AI agent, AIDA, by providing a set of administrative controls that ensure the agent behaves as intended within an organization's guidelines. Built on a governed query engine, AIDA already adheres to strict access controls, row and column security, and ensures queries remain within the platform. Guardrails adds a behavioral layer with four controls: agent protection, data product protection, prompt limiting, and topic filtering, each designed to prevent misuse or unintended actions by AIDA. These controls are easily configurable via the Starburst UI and provide clear feedback when a request is blocked, ensuring transparency for users and administrators. The current controls focus on behavioral instructions, but future developments will introduce a structural detection layer using external ML-based classification for enhanced security, allowing organizations to define varied policies for different contexts, thereby maintaining robust security and operational efficiency. Guardrails is available on Starburst Enterprise and Starburst Galaxy, with conservative defaults allowing organizations to start safely and adjust settings to their specific needs.

The concept of a context layer is essential for the successful deployment of AI agents in production, bridging the gap between prototypes and real-world applications by providing the necessary business intelligence. AI agents often falter not because they lack access to data but because they lack contextual understanding of the business environment, which can be remedied by implementing an architectural context layer. This layer integrates federated data access and data products, enabling AI to access and interpret business-specific information from diverse and scattered data sources. Data products package data with metadata, business logic, and governance, transforming raw data into actionable insights. This context layer serves as an agentic control plane, delivering consistent and scalable business intelligence, which is crucial for making accurate, responsible decisions. By centralizing reusable intelligence across domains rather than data, the context layer facilitates faster, more reliable AI-driven solutions, ultimately reducing time to market and enhancing the quality and effectiveness of AI agents.

GPUs, originally designed for graphics rendering, have emerged as effective tools for accelerating certain SQL workloads due to their ability to perform operations in parallel on large datasets. Historically, the use of GPUs in database processing was limited due to challenges like data transfer bottlenecks, but advancements in GPU architecture and programming, such as NVIDIA's CUDA, have enabled broader applications. While GPUs can significantly enhance SQL query performance by leveraging their high memory bandwidth, the actual benefits depend on factors like data storage speed and the nature of the SQL operations. For optimal performance, data should ideally reside in GPU memory, and both SQL vendors and users must collaborate to ensure efficient data handling and operator support. This collaboration involves managing memory on the GPU and optimizing data partitioning and sorting to mitigate bottlenecks, particularly those related to PCIe connections.