Pigeonhole

SQLite owns local OLTP and DuckDB owns local OLAP. Pigeonhole targets the missing quadrant: local data that is sparse, versioned, and scanned by row. Think feature stores, time series keyed by entity, crawl and event caches, graph adjacency, and per-user state.

cargo add pigeonhole, open a file, and you get rows of arbitrary sparse columns grouped into families, with versions, TTLs, and prefix and range scans. There is no server.

use pigeonhole::{Pigeonhole, Options, Family};

let db = Pigeonhole::open("crawl.phdb", Options::default())?;
let pages = db.table("pages")?
    .family("meta", Family::default().max_versions(1))
    .family("links", Family::default().bloom_bits(10))
    .create_if_missing()?;

pages.mutate(b"com.example/a")
    .put("meta", b"status", b"200")
    .put("links", b"com.example/b", b"")
    .commit()?;

Memtables flush into the file and compact, so data is bounded by disk rather than memory, and commits are crash-safe through a write-ahead log. A clean close leaves a single file.

Pigeonhole is experimental 0.x. The on-disk format and API may change before 1.0, and it isn’t recommended for production yet. The current focus is latency: io_uring, direct I/O, and an async API.