Pithos – Roadmap & Architectural Extensions¶
Architectural Roadmap & Next Steps for the Model-Isomorphic Vector Database
1. Completed Core Milestones (v2.0 – v2.2)¶
| Feature Area | Status | Key Innovations |
|---|---|---|
| Universal Container Format | Released (v2.0) | Schema-agnostic .pithos single-file container with embedded Apache Arrow IPC partition directory. |
| 4x8-Bit MIH Prefix Routing | Released (v2.1) | Direct-mapped CSR inverted table pruning 98.5% of search space in \(O(1)\) sub-microsecond time. |
| Precision Sidecar Reranking | Released (v2.0) | Asymmetric Query LUTs decoding FP8 (E4M3), NVFP4 (E2M1), FP16 with Monotonic Early Distance Cutoff. |
| LSM DeltaBuffer & WAL | Released (v2.0) | Lock-free real-time insertions, tombstone soft-deletes, write-ahead log, and zero-cost binary compactions. |
| Zero-Copy NumPy FFI | Released (v2.1) | Direct memory view returns (search_numpy(), batch_search_numpy()) bypassing Python wrapper conversions. |
| CUDA GPU Acceleration | Released (v2.0) | Fused batch Hamming kernels, Fast Walsh-Hadamard transform, and multi-family resonant voting. |
| Real Foundation Model Recall | Released (v2.2) | Validated \(>92.3\%\) Recall@10 on real Matryoshka models (nomic-embed-text, bge-base, text-embedding-3-small). |
| Disruptor 4.0.0 & VarHandles | Released (v2.2) | Replaced sun.misc.Unsafe with java.lang.invoke.VarHandle, 0 JVM deprecation warnings on Java 25+. |
| Zero-Allocation BitSets | Released (v2.2) | ThreadLocal<long[]> scratch buffers with 0 heap bytes allocated during search queries. |
| Zero-Overhead Security | Released (v2.2) | Defensive C-API pointer guards, Superblock/TOC bounds verification, and path traversal sanitization. |
2. Forward Roadmap & Future Directions¶
2.1 FPGA Co-Design via PCIe DMA Streaming¶
- Objective: Stream off-heap binary tier blocks directly from NVMe / Host RAM to FPGA accelerator cards (e.g. AMD Xilinx Alveo U280) via PCIe Gen5 DMA.
- Architecture: Host CPU performs asynchronous query preconditioning; FPGA runs massive pipelined Hamming popcount reduction units over 100+ parallel channels.
- Direct Virtual Addressing: Built-in
vdb_get_tier_address()provides contiguous off-heap addresses for zero-copy DMA engine setup.
2.2 GPUDirect Storage (GDS) & NVMe-oF Distributed Search¶
- Objective: Bypass Host CPU and DRAM entirely during massive multi-billion vector batch scans by streaming
.pithosfiles directly from NVMe into GPU High Bandwidth Memory (HBM3e). - Architecture: Leverage cuFile APIs and NVMe over Fabrics (NVMe-oF) for cluster-wide partition retrieval.
2.3 Distributed Sharding & Raft Cluster Consensus¶
- Objective: Partition multi-billion scale datasets across horizontal nodes with sub-millisecond scatter-gather query aggregation.
- Architecture: Consistent hashing over 16-bit Walsh-Hadamard prefixes, routing queries directly to target cluster shards.
2.4 Autonomous In-Storage Compaction Engine¶
- Objective: Background merge workers running on smart NVMe computational storage drives to compact WAL DeltaBuffers without consuming host CPU cycles.