Rust Performance Safety Patterns

Ask: "Fetch tokio documentation using Context7"

serde (/serde-rs/serde)

Serialization framework patterns

Ask: "Fetch serde documentation using Context7"

Zero-Copy Patterns

Core Principle

Avoid copying data unnecessarily. Use Bytes for shared buffer references.

See zero-copy-patterns.md for detailed patterns from existing codebase.

Buffer Sharing with Bytes

use bytes::Bytes;

// Share buffer without copying
fn parse_partition(buffer: Bytes, offset: usize, len: usize) -> Result<Partition> {
    // Slice creates new Bytes pointing to same underlying buffer
    let partition_data = buffer.slice(offset..offset + len);
    
    // Pass slices to child parsers
    let header = parse_header(partition_data.slice(0..10))?;
    let rows = parse_rows(partition_data.slice(10..))?;
    
    Ok(Partition { header, rows })
}

Avoiding Unnecessary Clones

// ❌ BAD: Copies data
fn parse_text(data: &[u8]) -> Result<String> {
    let bytes = data.to_vec();  // COPY 1
    String::from_utf8(bytes)    // COPY 2 (if validation needed)
}

// ✅ GOOD: Minimal copying
fn parse_text(data: Bytes) -> Result<String> {
    // Only copy if UTF-8 validation requires it
    let s = std::str::from_utf8(&data)?;
    Ok(s.to_string())  // Single copy only when needed
}

// ✅ BETTER: Keep as Bytes if possible
fn parse_blob(data: Bytes) -> Result<Bytes> {
    // No copy at all
    Ok(data)
}

Lifetime Management

Borrowing vs Owning

// Struct with borrowed data (careful with lifetimes)
struct Row<'a> {
    key: &'a [u8],
    values: Vec<&'a [u8]>,
}

// Struct with owned data (simpler, but copies)
struct RowOwned {
    key: Bytes,      // Shared ownership, no copy
    values: Vec<Bytes>,
}

Lifetime Elision

// Explicit lifetimes
fn parse_row<'a>(data: &'a [u8]) -> Result<Row<'a>> { ... }

// Elided (compiler infers)
fn parse_row(data: &[u8]) -> Result<Row> { ... }

Common Lifetime Patterns

// Pattern 1: Return borrowed data
fn find_cell<'a>(row: &'a Row, column: &str) -> Option<&'a [u8]> {
    row.cells.get(column).map(|c| c.value.as_ref())
}

// Pattern 2: Return owned data (use Bytes for zero-copy)
fn find_cell_owned(row: &Row, column: &str) -> Option<Bytes> {
    row.cells.get(column).map(|c| c.value.clone())  // Bytes::clone is cheap
}

Async Patterns

Async File I/O

use tokio::fs::File;
use tokio::io::AsyncReadExt;

async fn read_sstable(path: &Path) -> Result<Bytes> {
    let mut file = File::open(path).await?;
    let mut buffer = Vec::new();
    file.read_to_end(&mut buffer).await?;
    Ok(Bytes::from(buffer))
}

Async Decompression

use tokio::task;

async fn decompress_chunk(compressed: Bytes) -> Result<Bytes> {
    // CPU-intensive work in blocking task
    task::spawn_blocking(move || {
        let decompressed = lz4::block::decompress(&compressed, None)?;
        Ok(Bytes::from(decompressed))
    }).await?
}

Async Iteration

use futures::stream::{Stream, StreamExt};

async fn parse_rows<S>(row_stream: S) -> Result<Vec<Row>>
where
    S: Stream<Item = Result<Bytes>>,
{
    let mut rows = Vec::new();
    tokio::pin!(row_stream);
    
    while let Some(row_data) = row_stream.next().await {
        let row = parse_row(row_data?)?;
        rows.push(row);
    }
    
    Ok(rows)
}

Memory Management

PRD Target: <128MB

Track memory usage for large SSTables:

// Don't hold entire SSTable in memory
struct SstableReader {
    file: File,
    index: Vec<IndexEntry>,  // Keep index in memory
    cache: LruCache<u64, Bytes>,  // Cache hot blocks
}

// Read only what's needed
async fn read_partition(&mut self, offset: u64) -> Result<Partition> {
    // Check cache first
    if let Some(block) = self.cache.get(&offset) {
        return parse_partition(block.clone(), 0, block.len());
    }
    
    // Read minimal block
    let block = self.read_block(offset).await?;
    self.cache.put(offset, block.clone());
    parse_partition(block, 0, block.len())
}

Streaming Instead of Buffering

// ❌ BAD: Buffer everything
async fn process_sstable(path: &Path) -> Result<Vec<Row>> {
    let data = tokio::fs::read(path).await?;  // Load entire file
    parse_all_rows(&data)
}

// ✅ GOOD: Stream rows
async fn process_sstable(path: &Path) -> Result<()> {
    let reader = SstableReader::open(path).await?;
    
    while let Some(row) = reader.next_row().await? {
        process_row(row)?;
        // Row dropped here, memory freed
    }
    
    Ok(())
}

Error Handling

Result Propagation

use thiserror::Error;

#[derive(Error, Debug)]
enum ParseError {
    #[error("Not enough bytes: need {need}, have {have}")]
    NotEnoughBytes { need: usize, have: usize },
    
    #[error("Invalid UTF-8: {0}")]
    InvalidUtf8(#[from] std::str::Utf8Error),
    
    #[error("Compression error: {0}")]
    Compression(String),
}

// Use ? operator for clean propagation
fn parse_row(data: &[u8]) -> Result<Row, ParseError> {
    let flags = data.get(0).ok_or(ParseError::NotEnoughBytes { 
        need: 1, 
        have: data.len() 
    })?;
    
    let text = std::str::from_utf8(&data[1..])?;  // Auto-converts Utf8Error
    
    Ok(Row { flags: *flags, text: text.to_string() })
}

Safe Unsafe Code

When Unsafe is Necessary

// Reading fixed-size integers from buffer
fn read_u32_be(data: &[u8]) -> u32 {
    // Safe version (bounds check)
    u32::from_be_bytes([data[0], data[1], data[2], data[3]])
    
    // Unsafe version (skip bounds check if you're certain)
    unsafe {
        u32::from_be_bytes(*(data.as_ptr() as *const [u8; 4]))
    }
}

Safety Documentation

/// # Safety
/// 
/// `data` must be at least 4 bytes long, and properly aligned.
/// Caller must ensure this invariant.
unsafe fn read_u32_unchecked(data: &[u8]) -> u32 {
    debug_assert!(data.len() >= 4);
    u32::from_be_bytes(*(data.as_ptr() as *const [u8; 4]))
}

Prefer Safe Alternatives

// ✅ BEST: Safe with slice pattern matching
fn read_u32_safe(data: &[u8]) -> Option<u32> {
    match data {
        [a, b, c, d, ..] => Some(u32::from_be_bytes([*a, *b, *c, *d])),
        _ => None,
    }
}

Performance Profiling

Cargo Flamegraph

cargo install flamegraph
cargo flamegraph --bin cqlite -- parse large-file.db

Criterion Benchmarks

use criterion::{criterion_group, criterion_main, Criterion};

fn parse_row_benchmark(c: &mut Criterion) {
    let data = generate_test_row();
    
    c.bench_function("parse_row", |b| {
        b.iter(|| parse_row(&data))
    });
}

criterion_group!(benches, parse_row_benchmark);
criterion_main!(benches);

Memory Profiling

cargo install cargo-instruments
cargo instruments -t Allocations --bin cqlite -- parse large-file.db

PRD Alignment

Supports Milestone M1 (Core Reading Library):

• Zero-copy deserialization
• Memory target: <128MB for large files
• Type-safe parsing

Supports Milestone M6 (Performance Validation):

• Parse 1GB files in <10 seconds
• Sub-millisecond partition lookups

Common Patterns from Codebase

See zero-copy-patterns.md for patterns extracted from:

• v5_compressed_legacy.rs (1997 lines)
• Bytes usage
• Async decompression
• Buffer management

Anti-Patterns to Avoid

1. Unnecessary Allocations

❌ Vec::new() then push in loop with unknown size ✅ Vec::with_capacity(known_size)

2. Clone Everything

❌ .clone() on every data structure ✅ Use & references or Bytes for shared ownership

3. Blocking in Async

❌ CPU-intensive work in async fn ✅ tokio::task::spawn_blocking for CPU work

4. Ignoring Capacity

❌ String::new() then many push_str calls ✅ String::with_capacity(estimated_size)

Next Steps

When optimizing performance:

1. Profile first (don't guess)
2. Use flamegraph to find hotspots
3. Check allocations with Instruments/heaptrack
4. Benchmark changes with Criterion
5. Validate memory usage stays <128MB

References

• zero-copy-patterns.md - Patterns from codebase
• Context7: /tokio-rs/bytes, /tokio-rs/tokio, /serde-rs/serde
• Rust Performance Book: https://nnethercote.github.io/perf-book/