Summary
Databricks now offers model units, a VM-like abstraction for allocating and scaling GPU resources per customer, enabling cost-aware load balancing and autoscaling that saved over 80% in GPU costs. Runtime reliability mechanisms like black-box health checks and multimodal bottleneck profiling further improve throughput and recover from silent failures automatically.
Summary generated by brickster.ai. For the full article, follow the source link above.
More from Databricks Blog
BI Serving Pointers; Maximizing for Performance and TCO
Databricks now offers Unity Catalog Metric Views for a headless semantic layer, enabling governed business metrics across all BI tools and AI agents. Maximize performance and TCO by structuring your physical layer with star schemas, liquid clustering, and Predictive Optimization, and leverage aggregate-aware materialization for OLAP-style performance.
How the lakebase architecture stays resilient to cloud failures
Lakebase's architecture is built for resilience to cloud failures, not patched for it, by using stateless Postgres compute on zone-redundant storage and separating hot-path control-plane operations. This approach, validated through chaos testing and per-database availability tracking, addresses the unique reliability demands of agent workloads that start tens of millions of databases daily.
Introducing Always-On pricing: automatic savings for Databricks Lakebase
Databricks Lakebase now offers Always-On pricing, providing serverless flexibility with a 25% lower price on baseline capacity for established production workloads. Activate with a single toggle to disable scale-to-zero and set an autoscaling range, then after 24 hours of continuous use, baseline capacity bills at the Always-On rate while spikes bill at standard Autoscaling rates.
Announcing Lakebase Change Data Feed (CDF)
Lakebase Change Data Feed (CDF) is now in Public Preview, eliminating pipeline sprawl from operational databases by exposing every table's changes through Unity Catalog Managed Tables. This enables native CDC governed end-to-end without sidecar infrastructure, allowing operational data to function as the native Bronze layer in the medallion architecture.
Building a FHIR-native health data platform on Databricks Lakebase
Health Samurai's Aidbox now runs natively on Databricks Lakebase, providing a FHIR-native health data platform that standardizes clinical data at ingestion and makes it instantly available for Spark, ML, and AI. This architecture inherently delivers compliance with CMS-0057 and ONC mandates, eliminating the need for separate compliance workstreams.
AI readiness in telecommunications
Telco AI initiatives stall at production scale due to data debt, not model quality; Databricks Unity Catalog provides the semantic layer and governance needed to bridge this gap. It unifies disparate systems via Lakehouse Federation, offering AI agents rich context and enabling end-to-end governance for regulatory compliance and accurate operational tasks.