Module batch_wal 

Batched WAL with Vectored I/O

This module implements batched WAL writing using vectored I/O (writev) to minimize syscall overhead and improve write throughput.

Problem Analysis

Current WAL writes one entry per syscall:

• Syscall overhead: ~200-400 cycles
• Context switch potential: ~1000 cycles
• For 100K rows × 4 cols: 400K syscalls

Solution

Batch WAL entries into single vectored write:

• Single syscall per batch (up to 1000 entries)
• Vectored I/O (writev) eliminates intermediate copies

Math

Syscall amortization for N=100K entries, k=4 columns, B=1000 batch size:

T_unbatched = N × k × S = 100K × 4 × 300 = 120M cycles T_batched = ⌈N × k / B⌉ × S = ⌈400K / 1000⌉ × 300 = 120K cycles

Speedup: 1000× (syscall portion only)

Performance

Expected throughput: 10-20× improvement for bulk inserts

Structs

BatchAccumulator

Batch entry accumulator for a single transaction

BatchedWalReader

Batch reader for recovery

BatchedWalStats

Statistics for batched WAL writer

BatchedWalWriter

Batched WAL writer with vectored I/O

ConcurrentBatchedWal

Thread-safe batched WAL writer

Constants

DEFAULT_MAX_BATCH_BYTES

Default maximum batch bytes (64KB)

DEFAULT_MAX_BATCH_SIZE

Default maximum batch size (number of entries)