Fold-RS

Fold-RS is an LSM-tree-based storage engine implemented in Rust, built with durability and crash recovery as its first priorities. It iscurrently synchronous and single-threaded.

Project status: actively developed, pre-1.0. Core write/read/recovery paths are implemented and tested; compaction and background operation are not yet implemented. See Status below.

Features

Implemented today:

• Write-ahead log (WAL) with per-record CRC32 validation and synchronous (sync_all()-per-write) durability
• WAL generations / rotation (wal.NNNNNNNN.log)
• Crash-safe WAL recovery, including safe handling of a truncated final record and replay across multiple WAL generations
• Ordered in-memory Memtable (BTreeMap) with tombstone support and live size accounting
• Immutable SSTables, published atomically via temp-file-then-rename (NNNNNNNN.sst.tmp → NNNNNNNN.sst)
• mmap-based SSTable reads (via memmap2)
• Sparse SSTable index for O(log B) block lookup
• Bloom filters per SSTable (false positives allowed, false negatives are a correctness bug)
• Newest-to-oldest multi-SSTable reads with correct tombstone shadowing
• Automatic Memtable flush at a configurable-in-code size threshold
• Orphaned .sst.tmp cleanup on startup
• A read-amplification benchmark demonstrating the cost of growing SSTable count in the absence of compaction

Not yet implemented (see ARCHITECTURE.md for design intent):

• Compaction
• Background flushing / background compaction
• A manifest / version set
• A block cache
• Concurrent writers, transactions, snapshot isolation

Documentation

Document	Contents
ARCHITECTURE.md	Component design, write/delete/flush paths, WAL rotation, recovery, durability guarantee, read path, tombstone semantics, compaction as future work, limitations
FORMAT.md	Exact on-disk byte layouts: WAL record format, SSTable file layout, data block/index/Bloom format, the 32-byte footer, temp-file protocol
RESULTS.md	Benchmark results: WAL append latency, read-amplification measurements, SSTable size verification, multi-flush test

Project Layout

src/
  storage_engine.rs   # StorageEngine: put/get/delete, flush, recovery, orchestration
  memtable.rs          # In-memory sorted Memtable (BTreeMap<Bytes, Entry>)
  error.rs             # StorageError / Result

  wal/
    format.rs          # WAL record encode/decode, CRC32
    writer.rs           # WalWriter (append + sync_all)
    reader.rs           # WalReader (sequential replay, truncation-safe)

  sstable/
    writer.rs           # SstableWriter (blocks, index, bloom, footer)
    reader.rs           # SstableReader (mmap-backed lookups)
    footer.rs            # 32-byte footer format + magic
    mod.rs

  bloom.rs             # Bloom filter implementation

  bin/
    crash_writer.rs     # Standalone writer used by crash-safety integration tests

benches/
  wal_append.rs         # WAL append latency by value size
  recovery.rs            # Recovery-path benchmark
  read_amplification.rs # get() latency vs. SSTable count
  memtable.rs             # Memtable operation benchmark
  crc32.rs                # CRC32 benchmark

tests/
  recovery_test.rs
  sstable_reader_test.rs
  sstable_writer_test.rs

Usage

Fold-RS is currently a library crate (fold-rs), used programmatically via StorageEngine:

use fold_rs::storage_engine::StorageEngine;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let mut engine = StorageEngine::open("/path/to/data-dir")?;

    engine.put(b"key1", b"value1")?;
    engine.put(b"key2", b"value2")?;

    assert_eq!(engine.get(b"key1")?.as_deref(), Some(&b"value1"[..]));

    engine.delete(b"key1")?;
    assert_eq!(engine.get(b"key1")?, None);

    Ok(())
}

StorageEngine's methods return fold_rs::error::Result<T> (an alias for Result<T, fold_rs::error::StorageError>). StorageError implements the standard std::error::Error trait via thiserror, so ? works directly against Box<dyn std::error::Error> as shown above — you don't need to name fold_rs::error::Result explicitly unless you want to match on a specific StorageError variant (Io, KeyTooLarge, ValueTooLarge).

• StorageEngine::open(dir) creates the data directory if needed, cleans up any orphaned .sst.tmp files, and replays WAL generations to rebuild in-memory state — see ARCHITECTURE.md §8.
• Every put/delete is durable on return (sync_all() on the WAL before the call succeeds) — see ARCHITECTURE.md §9.
• get() returns Ok(Some(value)), Ok(None) (key absent or deleted), or an Err on I/O failure.

Building and Testing

# Build the library
cargo build

# Run the test suite
cargo test

Running Benchmarks

Most benchmarks run with the standard cargo bench:

cargo bench --bench wal_append
cargo bench --bench recovery
cargo bench --bench memtable
cargo bench --bench crc32

The read-amplification benchmark relies on the test-only flush_for_test() helper, which is gated behind the bench-utils Cargo feature and must be enabled explicitly:

cargo bench --features bench-utils --bench read_amplification

⚠️ cargo bench -- --bench-utils is not equivalent — that passes --bench-utils as an argument to the benchmark binary instead of enabling the Cargo feature, and will not work.

See RESULTS.md for recorded results from these benchmarks.

Status

Fold-RS's current priorities, in order, are:

1. Correctness — write/read/delete semantics, tombstone shadowing, and crash recovery are covered by unit and integration tests, including a repeated (20-iteration) crash-safety test built around the crash_writer binary.
2. Durability — every write is fsync'd before being acknowledged.
3. Crash recovery — WAL replay, orphaned temp-file cleanup, and truncated-record handling are all explicitly tested.
4. Explicit, understandable systems-level behavior — favoring a simple, inspectable implementation over premature optimization.

The engine does not yet perform compaction, so SSTable count — and therefore read amplification — grows monotonically with the amount of data written and flushed. This is a known, measured, and intentional current limitation; see RESULTS.md §5 for the measured effect and ARCHITECTURE.md §14 for the planned design. Compaction is the next major subsystem planned for this project.

For the full list of current limitations and planned future work, see ARCHITECTURE.md §16–17.

License

MIT License

Copyright (c) 2026 Ayush Verma

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.