Postgres FM turbopuffer
9 snips
Sep 12, 2025 
Join Simon Eskildsen, CEO of TurboPuffer and former Shopify infrastructure engineer, as he dives deep into the intricacies of database scaling. He shares fascinating insights on ANN indexing types and the crucial trade-offs for search workloads. Discover when moving search out of Postgres might be beneficial. Simon also contrasts TurboPuffer's low-latency vector search capabilities with S3's archival strengths, discusses the architectural nuances of local caching, and unpacks the benefits of SPFresh clustering for large datasets.
AI Snips
Chapters
Transcript
Episode notes
Shopify Shaping Database Philosophy
- Simon shared a decade at Shopify where he scaled databases and handled sharding and on-call duties.
- That experience shaped TurboPuffer's design priorities around simplicity and operational durability.
Tiered Storage Lowers Vector Cost
- TurboPuffer targets search workloads by using an LSM and tiered storage to cut storage costs versus triple-replicated relational DBs.
- This trade-off sacrifices some predictable latency and fast commit times but drastically lowers per-GB costs.
Offload Large Vector Workloads
- Move very large vector collections out of Postgres when memory and triple-replication costs become prohibitive.
- Keep transactional data in Postgres and offload search/vector workloads to specialized storage to postpone sharding.