rawdb

Non-transactional embedded storage engine with a filesystem-like API.

It features:

Multiple named regions in one file
Automatic space reclamation via hole punching
Regions grow and move automatically as needed
Zero-copy mmap access
Thread-safe with concurrent reads and writes
Page-aligned allocations (4KB)
Crash-consistent with explicit flush
Foundation for higher-level abstractions (e.g., vecdb)

It is not:

A transactional database (no ACID, transactions, or rollback)
A query engine (no SQL, indexes, or schemas)

Install

cargo add rawdb

Usage

use rawdb::{Database, Result};

fn main() -> Result<()> {
    // open database
    let temp_dir = tempfile::TempDir::new()?;
    let db = Database::open(temp_dir.path())?;

    // create regions
    let region1 = db.create_region_if_needed("region1")?;
    let region2 = db.create_region_if_needed("region2")?;

    // write data (buffered in mmap, not yet durable)
    region1.write(&[0, 1, 2, 3, 4])?;
    region2.write_at(&[5, 6, 7, 8, 9], 0)?;

    // flush to disk for durability
    db.flush()?;
    
    // read via mmap (data is immediately visible)
    {
        let reader = region1.create_reader();
        let _data = reader.read_all();
    } // reader dropped here, releasing its reference

    // remove region (space becomes reusable hole after flush)
    region1.remove()?;

    // punch holes in the file
    // db.compact()?; // doesn't work with doc-tests

    Ok(())
}

Sparse files

rawdb relies on sparse file support. Files grow via set_len() which creates logical size without allocating physical blocks, and compact() punches holes to reclaim unused blocks. This means:

The filesystem must support sparse files (ext4, XFS, APFS, ZFS, Btrfs — most modern filesystems do)
du and ls -s show actual disk usage; ls -l shows the larger logical size
Copying with cp may "densify" the file unless you use cp --sparse=always (Linux) or a sparse-aware tool
Backups should use sparse-aware tools to avoid inflating archives

Durability

Operations become durable after calling flush(). Before flush, writes are visible in memory but not guaranteed to survive crashes.

Design:

4KB metadata entries: Atomic page-sized writes per region with embedded IDs
Single metadata file: Rebuilt into HashMap on startup for O(1) lookups
No WAL: Simple design with lazy flushing for consistency
Lazy writes: Both data and metadata are written to mmaps immediately but not synced until flush

Write model:

Data writes modify the data mmap immediately (visible but not durable)
Metadata changes modify the metadata mmap immediately (visible but not durable)
Holes from moves/removes are marked as pending (not reusable until flush)
flush() syncs both mmaps (data → metadata → file size), then promotes pending holes
Ensures metadata never points to unflushed data and old locations aren't reused prematurely

Region operations:

Expand in-place when possible (last region or adjacent hole)
Copy-on-write to new location when expansion needed
All changes visible immediately in mmaps, durable after flush()

Recovery: On open, reads all metadata entries and rebuilds in-memory structures. Deleted regions are identified by zeroed metadata.