Topical Edition on Databases - Track 2
summary
Firebolt admits it started as a ClickHouse fork but now ships a rebuilt planner, full distributed query execution, Iceberg and DuckLake read support, Postgres for metadata, CDC from Postgres, and a hybrid storage strategy to avoid vendor lock‑in; launch is scheduled for later this summer and a public Docker image appears already. Also presented: Amethyst, a 3,000‑line Go prototype that does segment-level adaptation, switching between leveled and tiered compaction via a metadata-driven FSM to cut a four‑phase workload to 770s; space reclamation remains unresolved.
AI-written summary · about
chapters (20)
- 0:00Firebolt's database evolution
- 1:03Open formats Iceberg DuckLake
- 3:12Single binary vs microservices
- 8:56Performance optimizations analytics
- 11:02AI-driven agentic database workloads
- 29:51Amethyst adaptive compaction intro
- 33:27Amethyst leveled tiered compaction
- 41:03Amethyst performance four-phase workload
- 47:10Amethyst future limits
- 56:11Q&A Amethyst overhead
- 1:00:12Group photo lunch transition
- 3:02:06Panel AI database workloads
- 3:06:00AI workloads databases need
- 3:13:00AI context guardrails challenges
- 3:21:40AI reshaping databases in 5 years
- 3:30:18DB changes for AI workloads
- 3:40:40Vector DB scalability latency
- 3:50:20Bespoke DBs specialized workloads
- 4:00:07Specialized vs general-purpose DBs
- 4:11:46DBs consuming AI closing
related talks
56mBreakout Session
27mBuilding a PostgreSQL data warehouse
05mMaking Iceberg walk and talk like Postgres
30mThe database trends that are transforming your database infrastructure forever
01h 04mHow We Build Firebolt
21mAnalytics in Postgres –A decade in the making
34mAnalytics and AI with PostgreSQL - Mission Possible!
34mCatalogs, Iceberg, and Delta: Building the Foundation for Lakehouse Analytics