tonbo

We are looking for great developers to join us! tonbo.io/careers

@kylemathews Thanks. Working from the DS spec was tractable for a small team familiar with data infra. The protocol design absorbed a lot of complexity that would've been ours to invent otherwise

Very impressive!

@s2_streamstore appreciated, totally fair on the roadmap

@tonboio congrats on launching. multi-node s2-lite is not on the current roadmap :) for elastic scalability we recommend the managed service s2.dev

Just open-sourced Ursula: an HTTP event streams server. Built on @ElectricSQL's Durable Streams Protocol, Ursula keeps a combo we have not seen elsewhere: { Apache 2.0, quorum-replicated, sub-50ms p99, S3 economics }. github.com/tonbo-io/ursula

@s2_streamstore i'd be happy to share repro details on the catch-up case

@tonboio good to know of the catch-up breakage, we'll investigate and fix! s2-lite is much younger than the cloud service, it's the same API but different backends

Hey @s2_streamstore, when's the multi-node version of open-source S2 shipping? Asking for a friend who wants quorum durability + self-hosting. (Ursula benchmarks 18× vs S2 Lite, but single-process isn't a fair fight. Would love a real one. 😉)

@ElectricSQL On 3 × c7g.4xlarge with Raft quorum (baselines: 1 × c7g.4xlarge): - 5.9x writes vs Durable Streams (Node.js reference) - 5.2x writes vs S2 Lite - 6.1ms p99 SSE fan-out (160x vs DS, 18x vs S2 Lite) - 300/300 conformance suite ursula.tonbo.io/benchmark

Use tonbo to persist states?🤔🤔🤔

☺️🥰🥰🥰🥰

In edge and serverless systems, the ideal state is: FEWER READS, which only possible when schema and metadata are trustworthy. Once they are, pruning can happen before any data is read. This is what tonbo + aisle are built on.

In serverless, fewer reads mean lower cost. That’s why Tonbo is designed to do pruning before reading, not after. Aisle explores this idea as a standalone component: metadata-driven Parquet pruning for Rust. github.com/tonbo-io/aisle…

How do you coordinate writes in a distributed database? Traditional answer: Run etcd or ZooKeeper 24/7. Tonbo's answer: Use S3 ETags. Any stateless function can write, perfect for serverless, zero infrastructure. Read more -> github.com/tonbo-io/fusio…

Hey! Merry Christmas!

@snappercayt Python wrapper is on our roadmap. Stay tuned!🫰

@tonboio This is cool, now I want a python wrapper for this

Tonbo supports streaming Arrow RecordBatches directly from Parquet on S3. Process millions of rows without loading them all into memory. Async-native. Zero intermediate copies.

Tonbo uses typed-arrow to make declarative schemas super easy. Schemas are derived at compile time, so there’s zero runtime overhead. ORM-like with support for deeply nested Arrow schemas all in idiomatic Rust.

@kiwicopple 🤔who's using rust?

Agent workloads change the shape of data. Execution stretches across steps. Run state has to live between them. Observability will be becoming part of the loop. That's the shape Tonbo is designed for. 👉 tonbo.io/blog/tonbo-man…

This is exactly the shape Tonbo is built for. Tonbo has native, metadata-driven MVCC and snapshots. Creating and maintaining thousands of snapshots is cheap, and time travel is fast enough to sit on the execution path. Not a backup feature. Not an afterthought.  (3/4)

Agent workloads are non-deterministic, multi-step, and constantly revising their own state. That naturally pushes data toward: - append-only writes - cheap snapshots - and fast time travelState stops being overwritten and starts becoming a trace.  (2/4)

Data versioning is much more important now. In traditional apps, versioning was about data management. Useful, but rarely urgent. In agent systems, it becomes part of execution semantics. Without versions, the system can’t explain what it just did.  (1/4)