a database is nothing without its guarantees as part of our recent patch_by_filter launch, we clarified the ACID properties users can expect from turbopuffer read more here: turbopuffer.com/docs/guaran...

new: `patch_by_filter` updates all documents that match the provided filter, making bulk updates a breeze!

new: rank by filter! boost scores when docs match a condition (e.g. scale > 1pib). plugs straight into rank_by, and works alongside full-text search

september changelog turbopuffer.com/docs/roadmap

tpuf ANN v3 can search 100 billion vectors with a p99 of 200ms simplicity scales [v3 in beta, unfiltered search, 1024D, k=10, 92% recall]

it was driving us mad that a 96 core box could only drive 13 Gbps from S3 turns out eBPF probes from our network monitoring imposed more overhead than anyone expected 5-10x improvement in network throughput → faster puffin'

august changelog turbopuffer.com/docs/roadmap

"Having our data in turbopuffer makes us feel ready for whatever comes next in AI-enabled software. It gives us the confidence to adapt quickly." — Tristan Homsi, co-founder of Readwise turbopuffer.com/customers/r...

new: native facet support with group_by

july changelog turbopuffer.com/changelog

every few months we re-run the landing page benchmarks against prod vector search p90 down ~50%

ripgrep on object storage?

our ruby client just dropped 💎

by popular demand, now introducing one-dimensional vectors 🎈

"turbopuffer makes it too easy to build state-of-the-art AI apps" – Notion turbopuffer.com/customers/n...

new GCP regions. now puffin' in Toronto and Singapore turbopuffer.com/docs/regions

june changelog turbopuffer.com/docs/roadmap

we reduced query pricing for large namespaces by up to 80% puff more, pay less

S3’s conditional writes were the key ingredient that made our architecture possible now it's our time to pay it forward

"Their responsiveness and shipping velocity make us feel like we are their only customer." – Tom Moor, Head of Engineering at Linear turbopuffer.com/customers/l...

changelog for May 🐡🎉 turbopuffer.com/docs/roadmap

turbopuffer is generally available. petabytes in prod. turbopuffer.com/

🇨🇦 🏔️ 🗽 turbopuffer.com/about

changelog for April ✨🐡 turbopuffer.com/docs/roadmap

modern NVMe drives are blazing fast, but software needs to be architected from the ground up to take full advantage of them we recently rolled out a new I/O scheduler that uses Direct I/O to maximize performance of our SSD caches, avoiding page cache cpu and locking overhead

"After rolling out Turbopuffer, the searchDocs step in our retrieval pipeline went from being the slowest to the fastest." 🚀 🐡 Read how turbopuffer improved search response time for ref.tools: ref.tools/use-case/tur...

changelog for March 🐡 💨 turbopuffer.com/docs/roadmap

yesterday we processed 10B+ vector/document writes (~40TB) that's 100,000+ writes per second sustained

when we said "coming soon" we really meant it now puffin' in aws-us-east-1 and aws-eu-central-1

now available: turbopuffer AWS regions ☁️ come find us in us-west-2 and ap-southeast-2

changelog for february! 🐠 turbopuffer.com/docs/roadmap

GCP released a new ARM machine type with NVMe SSDs recently 70% increase in end-to-end production indexing throughput at 20-33% lower machine cost!

You can now delete 100,000s of records in a single roundtrip with `delete_by_filter` turbopuffer.com/docs/upsert...

every time we re-run the landing page benchmarks, they improve slightly. (bars are off because these are simply website screenshots) turbopuffer.com/

Summary of what we shipped in January 🐡 turbopuffer.com/docs/roadmap

welcoming two excellent additions to team tpuf this week: @arash11gt as a customer engineer, and @nikhilbenesch to the DB team from pasts at Materialize & Cockroach. we're locking in a p999 eng team to build the best search engine for scale.

changelog for november & december 📉🔢💫 turbopuffer.com/docs/roadmap

Serious latency improvement on a big namespace with lots of reads and writes for a customer. I/O and storage optimizations: p50: 100ms → 70ms p90: 350ms → 140ms (pictured) p99: 650ms → 300ms (mix of hot/cold data) more performance improvements coming.