Langfuse
Langfuse shipped 24 releases over the past 90 days, progressively hardening its data layer while expanding evaluation and experimentation capabilities. The team migrated trace and observation tables to event-based storage, a foundational shift visible in v3.149, v3.155, and subsequent releases that optimized queries and reduced memory usage through schema restructuring. In parallel, observation-level LLM-as-a-judge evaluations graduated from experimental to open beta in v3.150, then expanded to support categorical outputs, filter logic, and historical backfill in v3.161 and v3.163. Experiments infrastructure matured through pagination, run dialogs, and breadcrumb navigation across v3.162 and v3.163. Supporting this, blob storage exports grew richer with session IDs, dataset run IDs, gzip compression, and error tracking across v3.161–v3.163, while dashboard query performance improved via server-sent event streaming for progress visibility in v3.162–v3.163.
March focused on dashboard responsiveness, blob storage exports, and evaluation tooling. Server-side event streaming for ClickHouse query progress shipped alongside gzip compression for exports. LLM-as-a-judge expanded to filter by tool names and counts, while experiment workflows gained a run dialog, and prompt management added folder duplication—also shipping event replay, categorical eval outputs, and direct insertion into the events_full table for improved ingest performance.