Open-sourced ZenithDB, Fastest DB for Agent Traces in Rust

Hey Guys!

I just open-sourced Zenith, a columnar database engine purpose-built for AI agent traces.

why I built it

Got mad with Clickhouse and couldn't find a OSS alternative for traces. tbh Agent traces are weird. they're long, sparse, high-cardinality JSON with bursty ingest, late annotations, and fat text fields. existing backends are built for short structured spans and pay a 10-100x cost on this workload.

Built for things ppl acc do with Agent Traces:

1. load a full trace tree (B1): pull every span for a trace_id, e.g., "show me everything that happened in this user's session". 571μs p95 at 1B rows
2. filter by attribute (B2): "find spans where model=claude-opus and tokens > 4000". 4.3k μs p95 at 1B rows
3. full-text search over prompts and outputs (B3): grep across millions of tool calls, prompts, and completions
4. JSONPath into nested attributes (B6): query arbitrary JSON shapes without pre-defining a schema
5. group by model / tool / user (B8): cost and token aggregations across a billion spans in 6.8ms

try it (30 seconds)

curl -fsSL https://raw.githubusercontent.com/Polarityinc/zenith/main/install.sh | sh
zen serve --config examples/zenithdb.dev.toml

links

• repo: github.com/Polarityinc/zenith
• Benchmarks Zenith

it's alpha. core engine is feature-complete but on-disk format and wire protocols may shift before 1.0. stars, issues, and PRs all welcome :)

Comments

[Shane Barakat]

Sick!