M3.8.3 — Compression Benchmarks & Tuning

Field	Value
Phase	M3.8 — Context optimization
Size	M — 1 day
Status	⬜ Not started
Depends	M3.8.1, M3.8.2
Blocks	M3.8.4

Goal

Measure compression performance across content types and validate that ratios meet targets without sacrificing quality.

Deliverables

1. Benchmark Suite

New module: crates/mem-core/src/optimizer/bench.rs (100 LOC)

pub fn benchmark_all_compressors() -> BenchmarkReport {
    // Real-world test fixtures:
    // - logs/npm-error.txt (5KB)
    // - logs/cargo-fail.txt (8KB)
    // - json/array-100.json (15KB)
    // - diff/patch-large.diff (10KB)
    // - text/prose-1000words.txt (6KB)
    
    // Measure per-compressor:
    // - compression ratio (%)
    // - time taken (µs)
    // - tokens before/after
}

Tests (4):

test_log_compression_meets_target (85-95%)
test_json_compression_meets_target (70-90%)
test_diff_compression_meets_target (60-80%)
test_text_compression_meets_target (30-50%)

2. Performance Profile

Command:

cargo test --release --lib optimizer::bench 2>&1 | grep "time:"

Expected output:

log compression: 89% ratio, 1.2ms
json compression: 78% ratio, 2.1ms
diff compression: 64% ratio, 1.5ms
text compression: 38% ratio, 1.8ms

3. Tuning Knobs

Document per-compressor parameters:

LogCompressor: error line threshold (currently: any line with "error", "failed", etc.)
JsonCrusher: importance budget (currently: 55%)
DiffCompressor: context lines kept (currently: 0)
TextCompressor: token retention ratio (currently: 40%)

Acceptance

All 4 compression targets met (measured >= target)
All benchmark tests passing
Performance < 3ms per chunk
Documentation of tuning parameters

1.8 KiB Raw Blame History