Struct SoaBatch 

pub struct SoaBatch<'a> {
    pub commit_ts: Vec<u64>,
    pub txn_ids: Vec<u64>,
    pub keys: Vec<&'a [u8]>,
    pub value_handles: Vec<Option<ValueHandle<'a>>>,
    pub visibility: Vec<bool>,
    pub selection: Vec<usize>,
}

Structure of Arrays (SoA) batch for optimal SIMD visibility filtering

Why SoA?

AoS (Array of Structures) layout:

[key, value, commit_ts, txn_id], [key, value, commit_ts, txn_id], ...

• Poor cache utilization for visibility checks
• SIMD loads scatter data across cache lines

SoA (Structure of Arrays) layout:

commit_ts: [ts1, ts2, ts3, ts4, ...]  // Contiguous for SIMD
txn_ids:   [id1, id2, id3, id4, ...]  // Contiguous for SIMD
keys:      [k1, k2, k3, k4, ...]      // Only accessed for visible rows
values:    [v1, v2, v3, v4, ...]      // Late materialized

SIMD can process 4-8 timestamps per cycle with perfect cache utilization.

Late Materialization

Values are NOT copied into the batch. Instead, we store offsets/handles and only materialize values for rows that pass visibility filtering.

For scans where 90% of rows are filtered out, this saves ~90% of value copies.

Fields

commit_ts: Vec<u64>

Contiguous commit timestamps for SIMD visibility filtering

txn_ids: Vec<u64>

Contiguous transaction IDs for self-visibility checking

keys: Vec<&'a [u8]>

Key references (zero-copy)

value_handles: Vec<Option<ValueHandle<'a>>>

Value materializer handles (late binding) None = tombstone, Some = handle to materialize value

visibility: Vec<bool>

Pre-computed visibility mask (after SIMD filtering)

selection: Vec<usize>

Selection vector: indices of visible rows for late materialization

Implementations

impl<'a> SoaBatch<'a>

pub fn with_capacity(capacity: usize) -> Self

Create a new SoA batch with capacity

pub fn push( &mut self, key: &'a [u8], value: Option<&'a [u8]>, commit_ts: u64, txn_id: u64, )

Add an entry to the batch (SoA decomposition)

pub fn push_deferred( &mut self, key: &'a [u8], handle: Option<ValueHandle<'a>>, commit_ts: u64, txn_id: u64, )

Add with block handle (for disk-resident values)

pub fn len(&self) -> usize

Get batch size

pub fn is_empty(&self) -> bool

Check if empty

pub fn clear(&mut self)

Clear the batch for reuse

pub fn filter_visibility( &mut self, snapshot_ts: u64, current_txn_id: Option<u64>, )

Apply SIMD visibility filtering and build selection vector

This is the hot path - SIMD processes commit_ts array directly without touching keys/values until we know what’s visible.

pub fn visible_count(&self) -> usize

Get visible row count (after filtering)

pub fn iter_visible( &self, ) -> impl Iterator<Item = (&'a [u8], Option<&'a [u8]>)> + '_

Iterate over visible rows with late materialization

Values are only materialized for rows in the selection vector.

pub fn iter_visible_full( &self, ) -> impl Iterator<Item = (&'a [u8], Option<&'a [u8]>, u64, u64)> + '_

Iterate visible rows with full metadata

Trait Implementations

impl<'a> Debug for SoaBatch<'a>

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.