We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Lobsters·May 21·5 min read
Ursula: thread-per-core, multi-Raft Rust runtime for HTTP event streams
TL;DR
Distributed HTTP event streaming systems need low-latency writes and fault tolerance across multiple nodes. Ursula, a self-hosted server, implements thread-per-core architecture (one OS thread per CPU core) with multi-Raft consensus (a replication protocol ensuring data agreement) for in-memory writes and S3 cold storage.
✦ Why It Matters
Engineers can deploy low-latency, fault-tolerant event streaming without cloud lock-in or complex distributed systems expertise.
Key Takeaways
- Evaluate Ursula for self-hosted event streaming needing sub-50ms write latency and multi-node fault tolerance.
- Thread-per-core architecture reduces CPU context switching overhead in high-throughput scenarios.
- Multi-Raft enables parallel consensus across independent replication groups, improving throughput.
- Combine in-memory hot tier with S3 cold tier to balance performance and storage cost.
Full Summary
Event streaming platforms must balance write latency, data durability, and multi-node consistency. Ursula addresses this by combining three architectural choices: thread-per-core execution (assigning one thread to each CPU core to minimize context switching), multi-Raft consensus (running multiple independent Raft instances for parallel replication), and a hot-cold storage tier (in-memory ring buffer for active data, S3 for archival).
Built on ElectricSQL's Durable Streams Protocol, Ursula achieves sub-50ms commit latency when writes reach quorum (majority agreement across nodes). The in-memory hot ring optimizes the write path, while S3 integration provides cost-effective long-term storage without sacrificing performance.
This design targets self-hosted deployments requiring both speed and reliability.