Introduction
Timelake is the open time-series lakehouse. It keeps time series in your own bucket (S3 and S3-compatible stores, Google Cloud Storage, Azure Blob Storage) or a local directory, keeps them laid out for time-series queries without maintenance jobs, and reads them with a query engine that uses that layout. Today tables are stored as Apache Iceberg tables with Parquet data files, so other engines can read them too.
It is a Rust library with a command-line tool (timelake), a Python package and a TypeScript client. There is no server: the catalog is files in the warehouse, commits are atomic through conditional writes, and maintenance runs inside the writer or as one stateless command.
What it does
- Time-series tables. You declare a time column and a series key (such as
device_id), like a TimescaleDB hypertable. Data is partitioned by hour, day or month and every file is sorted by(series key, time). - Streaming-friendly writes. Each append is one atomic Iceberg snapshot. Retries can be made idempotent with a batch id, and concurrent writers are safe.
- Built-in maintenance. Small files from frequent appends are compacted inline, duplicates are removed (the latest write wins), old data is dropped by retention, and old snapshots are expired.
- Fast latest values. A last-value index holds the newest row of every series; queries for the latest row are rewritten to read it.
- Open format. Plain DuckDB reads Timelake tables with
iceberg_scan; no Timelake software is needed to read.
When to use it
Timelake targets telemetry and other append-heavy time series (IoT and industrial data first) where object storage is the system of record and you would otherwise run S3 Tables or self-managed Iceberg plus scheduled compaction jobs. It is not an OLTP database; the freshest data is visible at commit granularity (each append), not per row.
Status
This is v0, an early open-source release (Apache 2.0). It covers the write path, maintenance and the reader. A REST catalog, a hot tier for sub-commit freshness, deletes and GDPR erasure, and continuous aggregates are planned; see Roadmap & limitations.
Next: Quick start.