Langfuse Shifts to Observations-First Data Model, Achieving 10x Dashboard Performance Improvements
Key Takeaways
- ▸Langfuse moved from a trace-and-observations data model to a single, wide, immutable observations table in ClickHouse, eliminating costly read-time joins and deduplication
- ▸Performance improvements are substantial: initial table loads improved from seconds to milliseconds, dashboard load times improved by 10x+ for large projects
- ▸Adoption of OpenTelemetry's immutable span model (now 60% of all observations) enabled architectural simplifications and eliminated expensive update operations
Summary
Langfuse, an open-source LLM engineering platform, has announced a major architectural shift to an observations-centric data model built on ClickHouse, moving away from its previous trace-and-observations separation. The redesign, now in beta on Langfuse Cloud, eliminates costly read-time joins and deduplication operations, dramatically improving performance metrics: initial table loads have improved from seconds to milliseconds, and dashboard load times for large projects have improved by at least 10x for longer durations.
The migration addresses scaling challenges that emerged in mid-2025 when the previous Postgres-inspired approach became increasingly expensive at scale. Large users were limited to viewing only a few days of historical data, and trace listing operations took seconds. By adopting OpenTelemetry's immutable span model—now accounting for approximately 60% of all observations in Langfuse Cloud—the team eliminated the need for complex update logic and expensive deduplication operations that had plagued ClickHouse's ReplacingMergeTree implementation.
This redesign represents a fundamental rethinking of how observability data should be stored and accessed in production LLM applications. The shift from mutable to immutable data patterns, combined with OpenTelemetry adoption, creates a foundation for handling petabyte-scale observability workloads while maintaining sub-second query performance.
- The previous ReplacingMergeTree update pattern became prohibitively expensive at scale, requiring full-row replacements and complex deduplication logic during reads
Editorial Opinion
Langfuse's migration to an observations-first model is a pragmatic case study in database design evolution. The shift from mutable Postgres patterns to ClickHouse's immutable data semantics highlights how infrastructure choices must evolve with product scale. By embracing OpenTelemetry's immutability as a feature rather than fighting it with expensive deduplication, Langfuse has created a template for building scalable observability platforms—demonstrating that sometimes the best performance optimization is reconsidering your fundamental data model.
