Introducing ZeptoDB

Time-series systems make different trade-offs around query language, storage, latency, ecosystem, and operations. ZeptoDB focuses on a specific combination: live time-series evidence, agent-scoped memory, prompt cache, replay, and an in-process Python path in one operational layer.

ZeptoDB is built for that combination.

Today, that means more than fast tick or sensor queries. ZeptoDB turns live operational timelines into memory for AI agents: raw events, retrieved context, cache hits, model calls, tool calls, and decisions stay replayable with SQL.

The Architecture

ZeptoDB is a C++20 in-memory columnar database built from scratch for microsecond-latency time-series workloads. The core pipeline:

Feed Handler → MPMC Ring Buffer → Column Store → SIMD/JIT Query Engine
                    ↓
                   WAL (crash recovery)
                    ↓
               Parquet HDB (S3/NFS)

Every layer is designed to minimize copying and allocation on the hot path.

Ingestion: Lock-free MPMC ring buffer with Highway SIMD batch copy — 5.52M events/sec
Storage: Arena allocator with a columnar layout designed to limit hot-path allocation
Execution: LLVM JIT compiles SQL to native code, Highway SIMD for vectorized aggregation
Python: pybind11 zero-copy — NumPy arrays reference engine memory directly (522ns)

How It Fits

ZeptoDB’s project benchmarks describe ZeptoDB runs; they are not automatically comparable with numbers published for another system under different conditions. The current Benchmarks page contains no selection-grade cross-vendor performance ranking.

Use the architecture guides for current, source-linked orientation: kdb+, ClickHouse, InfluxDB 3, and TimescaleDB. For a selection decision, reproduce the target schema, retention, durability, concurrency, recovery, and failure behavior on the same hardware.

Who Is It For?

Quant Researchers

ASOF JOIN, VWAP, EMA, xbar — the temporal operations you need, in standard SQL. Python zero-copy for notebooks.

Trading Desks

Evaluate the measured ingestion path and the FIX, ITCH, Kafka, or WebSocket connector actually included and tested in your selected build.

IoT / Robotics

10KHz sensor ingestion, Window JOIN for sensor fusion, Parquet HDB for historical replay.

What’s Next

ZeptoDB is source-available under BUSL-1.1, with production use allowed by the Additional Use Grant except commercial DBaaS or managed-service resale before the 2030-04-01 Apache-2.0 change date. We’re actively developing:

Shard migration dual-write/catch-up for Agent Memory
AgentOps dashboards for retrievals, cache events, LLM calls/errors, context traces, replay windows, and tool calls
Additional industrial, Physical AI, and edge connector examples
JDBC/ODBC drivers for BI tool integration
Additional deployment recipes and reproducible operator examples

Try it: Quick Start →