Pillar 05 • Arrow Flight SQL

Vectorized SIMD Analytics & Arrow Flight

Bridging 8KB columnar pages directly to Apache Arrow RecordBatch streams with SIMD pushdown at 87.36M rows/sec.

The Row-by-Row Serialization Tax

Traditional database wire protocols transfer query results row-by-row in text or tuple formats. When analytical frameworks (Python Polars, DuckDB, Pandas) scan millions of records, CPU time is bottlenecked by serialization overhead.

Zero-Copy Arrow Flight SQL (Port 8815)

Because UniDB stores relational data natively in 8KB columnar slotted pages, internal columnar arrays map directly into Apache Arrow RecordBatch streams and transmit over gRPC Arrow Flight at > 500 MB/sec with zero CPU copying.

Ingestion MethodWire FormatCPU OverheadThroughput
PostgreSQL Wire (Port 5432)Row TuplesHigh (Row-by-row parse)~45 MB/sec
HTTP JSON RESTText JSONHigh (Stringification)~28 MB/sec
UniDB Arrow Flight SQL (Port 8815)Arrow RecordBatch IPCZero (Direct Memory Map)> 500 MB/sec

Hardware-Level SIMD Pushdown

Analytical queries push aggregations directly into CPU vector registers using AVX-512 and AVX2 instructions, scanning and aggregating historical records at 87.36 Million rows/sec.