FAST.
LOCK-FREE.
DURABLE.
The world's most unique in-memory database. Engineered with $O(\log n)$ efficiency and the custom Nexus Protocol for sub-microsecond mastery.
git clone https://github.com/AmitNilajkar/VeloDB
BUILT DIFFERENT.
VeloDB bypasses traditional database bottlenecks with architectural innovations designed for modern many-core hardware.
Lock-Free B-Tree
Immutable slabs and atomic root swaps enable wait-free reads. No mutexes. No bottlenecks. Linearizable consistency without the cost.
Nexus Protocol
Custom binary wire protocol optimized for high-bandwidth, event-driven data transport. 10x throughput compared to standard JSON over TCP.
SAI Indexing
Succinct Adaptive Indexing via Elias-Fano encoding. $O(1)$ rank/select logic for high-speed range scans with zero footprint.
Checksummed WAL
Enterprise-grade Write-Ahead Log. Each entry is CRC32-C validated. Safe truncation ensures zero corruption on recovery.
Python Bindings
Seamless high-performance C++ core exposed via a Pybind11 API. Build enterprise data pipelines with the ease of Python.
Zero-Copy Processing
Direct mapping to memory structures. Eliminates the CPU tax of parsing, decoding, and intermediate representation.
DEEP DIVE: COW B-TREE
-
Path-Copying ensures old data is never mutated, providing immutable snapshots for zero-lock thread safety.
-
Structural growth is handled via atomic 64-bit root swaps, maintaining linearizability under extreme parallel stress.
-
Cache-line alignment in custom SlabArenas minimizes L3 misses, pushing read latencies to the hardware floor.
PHD-LEVEL SPECIFICATION
VeloDB utilizes Zipfian Key Distribution management coupled with a Count-Min Sketch for probabilistic write-amplification control (PWAC). This ensures that heavy write cycles do not degrade read snapshots.
The Nexus Protocol utilizes the Proactor pattern on high-concurrency event loops, achieving near-perfect core utilization on many-core Windows systems via native IOCP.
Storage density is optimized via Elias-Fano Elias-Fano encoding, transforming range lookups into bit-vector rank/select operations. This reduces the theoretical search complexity while maintaining a constant space footprint.
Atomic Linearizability is guaranteed via hardware-level fences, ensuring that every version of the database remains perfectly consistent across NUMA boundaries.
THE FUTURE IS
LOCK-FREE.
"Software is either fast or wrong. We chose fast."
— AMIT NILAJKAR