kohakuvault 0.3.0

SQLite-backed KV store (Rust+PyO3) for large media blobs with structured data support

KohakuVault

High-performance, SQLite-backed storage with dual interfaces: dict-like for blobs (key-value) and list-like for sequences (columnar). Rust core with Pythonic APIs.

Quick Start

pip install kohakuvault

KV Store - Dict-like interface for binary blobs (images, videos, documents):

from kohakuvault import KVault

vault = KVault("data.db")
vault["image:123"] = image_bytes
vault["video:456"] = video_bytes
data = vault["image:123"]

# Bulk writes with smart caching (NEW in v0.2.2!)
with vault.cache(64*1024*1024):
    for i in range(10000):
        vault[f"key:{i}"] = data
# Auto-flushes on exit!

Columnar Storage - List-like interface for typed sequences (timeseries, logs, events):

from kohakuvault import ColumnVault

cv = ColumnVault("data.db")

# Primitives
cv.create_column("temperatures", "f64")
temps = cv["temperatures"]
temps.extend([23.5, 24.1, 25.0])
print(temps[0])  # 23.5

# Structured data (NEW in v0.3.0!)
cv.create_column("users", "msgpack")
users = cv["users"]
users.append({"name": "Alice", "age": 30, "tags": ["vip"]})
print(users[0])  # {'name': 'Alice', 'age': 30, 'tags': ['vip']}

# Strings with encoding (NEW in v0.3.0!)
cv.create_column("messages", "str:utf8")
messages = cv["messages"]
messages.append("Hello, 世界!")
print(messages[0])  # 'Hello, 世界!'

DataPacker - Rust-based serialization (NEW in v0.3.0!):

from kohakuvault import DataPacker

# MessagePack for structured data
packer = DataPacker("msgpack")
packed = packer.pack({"user": "alice", "score": 95.5})
data = packer.unpack(packed, 0)

# Bulk operations
records = [{"id": i, "val": i*1.5} for i in range(1000)]
packed_all = packer.pack_many(records)  # Concatenated bytes

# Unpack with offsets (for variable-size)
offsets = [0, len(packer.pack(records[0]))]  # Calculate offsets
unpacked = packer.unpack_many(packed_all, offsets=offsets)

Features

• Dual interfaces: Dict for blobs (KVault), List for sequences (ColumnVault)
• Zero external dependencies: Single SQLite file, no services required
• Memory efficient: Stream multi-GB files, dynamic chunk growth
• Type-safe columnar: Fixed-size (i64, f64, bytes:N) and variable-size (bytes, str, msgpack, cbor)
• Rust performance: Native speed with Pythonic ergonomics
• Smart caching: Auto-flush context manager, daemon thread, capacity enforcement
• Structured data: Store dicts/lists directly with MessagePack/CBOR (NEW in v0.3.0!)
• DataPacker: Rust-based serialization with multi-encoding support (NEW in v0.3.0!)

Best Practices

Handling Many Large Binary Files

For thousands of large binaries (images, videos, documents), use a hybrid approach:

from kohakuvault import KVault, ColumnVault

kv = KVault("media.db")
cv = ColumnVault(kv)  # Share same database

# Store metadata in columnar (efficient for large lists)
cv.create_column("image_ids", "i64")
cv.create_column("image_names", "bytes")
cv.create_column("image_sizes", "i64")
cv.create_column("upload_times", "i64")

ids = cv["image_ids"]
names = cv["image_names"]
sizes = cv["image_sizes"]
times = cv["upload_times"]

# Store actual binaries in KV store
for img_id, img_data, img_name in image_stream:
    # Metadata in columnar (fast append, efficient iteration/filtering)
    ids.append(img_id)
    names.append(img_name)
    sizes.append(len(img_data))
    times.append(int(time.time()))

    # Binary data in KV (optimized for large blobs)
    kv[f"blob:{img_id}"] = img_data

# Query metadata without loading binaries
for i in range(len(ids)):
    if sizes[i] > 1024 * 1024:  # Find images > 1MB
        print(f"Large image: {names[i].decode()}")
        # Load binary only when needed
        data = kv[f"blob:{ids[i]}"]

Why this pattern?

• ✅ Columnar optimized for append-heavy metadata (millions of entries)
• ✅ KV optimized for large binary blobs (streaming, caching)
• ✅ Can query/filter metadata without loading binaries
• ✅ Both share same SQLite file (single-file deployment)
• ✅ Efficient iteration over metadata, lazy loading of binaries

Installation

pip install kohakuvault  # When published to PyPI
pip install .            # From source

Platform Support:

• ✅ Linux (x86_64)
• ✅ Windows (x86_64)
• ✅ macOS (Apple Silicon M1/M2/M3/M4 only - ARM64)
• ❌ macOS Intel (x86_64) - not supported

Development

Prerequisites: Python 3.10+, Rust (rustup.rs)

# Setup
git clone https://github.com/yourusername/kohakuvault.git
cd kohakuvault
python -m venv .venv && source .venv/bin/activate  # or .venv\Scripts\activate on Windows
pip install -e .[dev]
maturin develop  # Build Rust extension (once)

# Workflow
# - Edit Python files → changes live immediately
# - Edit Rust files → run `maturin develop` to rebuild

# Tools
pytest                  # Run tests
black src/kohakuvault   # Format Python
cargo fmt               # Format Rust
maturin build --release # Build production wheel

Usage

Basic Operations

vault = KVault("media.db")

# Dict-like interface
vault["key"] = b"value"
data = vault["key"]
del vault["key"]
if "key" in vault: ...

# Safe retrieval
data = vault.get("key", default=b"")

# Iteration
for key in vault:
    print(f"{key}: {len(vault[key])} bytes")

Streaming Large Files

vault = KVault("media.db", chunk_size=1024*1024)  # 1 MiB chunks

# Stream from file → vault
with open("large_video.mp4", "rb") as f:
    vault.put_file("video:789", f)

# Stream from vault → file
with open("output.mp4", "wb") as f:
    vault.get_to_file("video:789", f)

Bulk Operations with Caching

Recommended: Use context manager for automatic flush

vault = KVault("media.db")

# Safest: Context manager auto-flushes
with vault.cache(cap_bytes=64*1024*1024):
    for i in range(1000):
        vault[f"item:{i}"] = data
# Auto-flushed here, guaranteed!

# Long-running: Daemon thread auto-flushes every 5 seconds
vault.enable_cache(cap_bytes=64*1024*1024, flush_interval=5.0)
while True:
    vault["sensor_data"] = read_sensor()
# Daemon flushes automatically

# Manual control (backward compatible)
vault.enable_cache(cap_bytes=64*1024*1024)
for i in range(1000):
    vault[f"item:{i}"] = data
vault.flush_cache()  # Manual flush
vault.disable_cache()  # Auto-flushes before disabling

Configuration

vault = KVault(
    path="media.db",
    chunk_size=2*1024*1024,   # Streaming chunk size
    retries=10,                # Retry attempts for busy DB
    enable_wal=True,           # Write-Ahead Logging
    cache_kb=20000,            # SQLite cache size
)

Columnar Storage (NEW!)

List-like interface for typed sequences (timeseries, logs, events):

from kohakuvault import ColumnVault

cv = ColumnVault("data.db")

# Fixed-size types: i64, f64, bytes:N
cv.create_column("sensor_temps", "f64")
cv.create_column("timestamps", "i64")
cv.create_column("hashes", "bytes:32")  # 32-byte fixed

temps = cv["sensor_temps"]
temps.append(23.5)
temps.extend([24.1, 25.0, 25.3])
print(temps[0], temps[-1], len(temps))  # 23.5, 25.3, 4

# Variable-size bytes (for strings, JSON, etc.)
cv.create_column("log_messages", "bytes")  # No size = variable!
logs = cv["log_messages"]
logs.append(b"Short message")
logs.append(b"This is a much longer log entry with details...")
print(logs[0])  # Exact bytes, no padding

# Iterate
for temp in temps:
    print(temp)

Why columnar?

• Append-heavy workloads (O(1) amortized, like Python list)
• Typed data (int/float/bytes with type safety)
• Efficient iteration and random access
• Dynamic chunk growth (128KB → 16MB, exponential like std::vector)
• Cross-chunk element support (byte-based addressing)
• Minimal memory overhead (incremental BLOB I/O)

See docs/COLUMNAR_GUIDE.md and examples/columnar_demo.py for complete guide.

API Reference

Constructor

KVault(path, chunk_size=1048576, retries=4, backoff_base=0.02,
       table="kvault", enable_wal=True, page_size=4096,
       mmap_size=268435456, cache_kb=20000)

Methods

Storage

• put(key, value) - Store bytes
• put_file(key, reader, size=None, chunk_size=None) - Stream from file-like
• get(key, default=None) - Retrieve bytes
• get_to_file(key, writer, chunk_size=None) - Stream to file-like
• delete(key) - Remove key
• exists(key) - Check existence

Caching

• enable_cache(cap_bytes, flush_threshold) - Enable write-back cache
• disable_cache() - Disable and flush cache
• flush_cache() - Commit cached writes, returns count

Maintenance

• optimize() - VACUUM database
• close() - Flush and close

Dict Interface: vault[key], del vault[key], key in vault, len(vault), vault.keys(), vault.values(), vault.items(), etc.

Exceptions: KohakuVaultError, NotFound, DatabaseBusy, InvalidArgument, IoError

Architecture

Python wrapper (src/kohakuvault/proxy.py)
    ↓ PyO3 bindings
Rust core (src/kvault-rust/lib.rs)
    ↓ rusqlite
SQLite database (bundled)

Why hybrid? Rust handles SQLite operations safely and efficiently. Python provides the ergonomic dict-like interface.

Contributing

# Setup
git checkout -b feature-name
# Make changes
black src/kohakuvault && cargo fmt  # Format
pytest                               # Test
git commit && git push
# Open PR

Releasing

GitHub Actions automatically builds wheels and publishes to PyPI when you push a tag:

# 1. Update version in pyproject.toml and Cargo.toml
# 2. Commit changes
git add pyproject.toml Cargo.toml
git commit -m "Bump version to 0.1.0"

# 3. Create and push tag
git tag v0.1.0
git push origin main --tags

# 4. GitHub Actions will:
#    - Build wheels for all platforms
#    - Create GitHub Release with wheels attached
#    - Publish to PyPI (with skip-existing for safety)

What happens:

• Wheels are built for Linux, Windows, macOS (Apple Silicon)
• All wheels are uploaded to the GitHub Release (downloadable)
• Wheels are published to PyPI
• If some wheels already exist on PyPI, they're skipped (no error)

License

Apache 2.0 - see LICENSE