Perfmaxx json2xml

BENCHMARKS.md

@@ -5,9 +5,10 @@ Comprehensive performance comparison between all json2xml implementations.
 ## Test Environment
 
 - **Machine**: Apple Silicon (arm64)
-- **OS**: macOS 26.4.1 (Darwin 25.4.0)
+- **OS**: macOS 26.5 (Darwin 25.5.0)
 - **Python**: 3.14.4
-- **Date**: April 24, 2026
+- **Date**: May 27, 2026
+- **CLI tools**: `json2xml-go` and `json2xml-zig` from `/Users/vinitkumar/.local/bin`
 
 To make new runs comparable, record the same fields for your machine before
 publishing results:
@@ -33,65 +34,72 @@ which json2xml-go json2xml-zig 2>/dev/null || true
 | Size | Description | Bytes |
 |------|-------------|-------|
 | Small | Simple object `{"name": "John", "age": 30, "city": "New York"}` | 47 |
-| Medium | 10 generated records with nested structures | 3,212 |
+| Medium | 10 generated records with nested structures | 3,215 |
 | bigexample.json | Real-world patent data | 2,018 |
-| Large | 100 generated records with nested structures | 32,207 |
-| Very Large | 1,000 generated records with nested structures | 323,148 |
+| Large | 100 generated records with nested structures | 32,206 |
+| Very Large | 1,000 generated records with nested structures | 323,131 |
 
 ## Results
 
 ### Performance Summary
 
 | Test Case | Python | Rust | Go | Zig |
 |-----------|--------|------|-----|-----|
-| Small (47B) | 31.49µs | 0.55µs | 4.09ms | 2.02ms |
-| Medium (3.2KB) | 1.69ms | 16.15µs | 4.07ms | 2.09ms |
-| bigexample (2KB) | 819.86µs | 6.44µs | 4.37ms | 2.11ms |
-| Large (32KB) | 17.97ms | 168.21µs | 4.10ms | 2.42ms |
-| Very Large (323KB) | 183.33ms | 1.42ms | 4.20ms | 5.12ms |
+| Small (47B) | 3.19µs | 0.86µs | 6.05ms | 3.08ms |
+| Medium (3.2KB) | 214.83µs | 18.41µs | 5.85ms | 3.12ms |
+| bigexample (2KB) | 91.20µs | 7.32µs | 5.76ms | 3.08ms |
+| Large (32KB) | 2.07ms | 175.46µs | 5.89ms | 3.73ms |
+| Very Large (323KB) | 21.20ms | 1.48ms | 6.82ms | 7.82ms |
 
 ### Speedup vs Pure Python
 
 | Test Case | Rust | Go | Zig |
 |-----------|------|-----|-----|
-| Small (47B) | **56.8x** | 0.0x* | 0.0x* |
-| Medium (3.2KB) | **105.0x** | 0.4x* | 0.8x* |
-| bigexample (2KB) | **127.2x** | 0.2x* | 0.4x* |
-| Large (32KB) | **106.8x** | 4.4x | **7.4x** |
-| Very Large (323KB) | **129.0x** | **43.6x** | **35.8x** |
+| Small (47B) | **3.7x** | 0.0x* | 0.0x* |
+| Medium (3.2KB) | **11.7x** | 0.0x* | 0.1x* |
+| bigexample (2KB) | **12.5x** | 0.0x* | 0.0x* |
+| Large (32KB) | **11.8x** | 0.4x* | 0.6x* |
+| Very Large (323KB) | **14.4x** | **3.1x** | **2.7x** |
 
-*CLI tools have process spawn overhead (~2-4ms) which dominates for small inputs.
+*CLI tools have process spawn overhead (~3-6ms) which dominates for small inputs.
 
 ## Key Observations
 
 ### 1. Rust Extension is the Best Choice for Python Users 🦀
 
 The Rust extension (json2xml-rs) provides:
-- **~57-129x faster** conversion than pure Python in this run
+- **~4-14x faster** conversion than the optimized pure Python path in this run
 - **Zero process overhead** - called directly from Python
 - **Automatic fallback** - pure Python used if Rust is unavailable or a feature requires it
 - **Easy install**: `pip install json2xml[fast]`
 
-### 2. Go Excels for Very Large CLI Workloads 🚀
+### 2. Python Optimizations Changed the Relative Speedups
+
+Recent pure Python improvements substantially reduced conversion time:
+- Small inputs dropped to microsecond-level library-call timings
+- Medium and large inputs are roughly an order of magnitude faster than the April 2026 run
+- Rust remains the fastest library backend, but its relative speedup is lower because the baseline improved
+
+### 3. Go Excels for Very Large CLI Workloads 🚀
 
 For very large inputs (323KB+):
-- **43.6x faster** than Python
-- But ~4ms startup overhead hurts small file performance
+- **3.1x faster** than optimized Python in this run
+- But ~6ms startup overhead hurts small and medium file performance
 - Best for batch processing or large file conversions from shell scripts
 
-### 3. Zig is Highly Competitive for CLI Use ⚡
+### 4. Zig is Highly Competitive for CLI Use ⚡
 
 In this run:
-- **35.8x faster** than Python for very large files
-- **7.4x faster** for large files (32KB)
-- Faster startup than Go (~2ms vs ~4ms)
+- **2.7x faster** than optimized Python for very large files
+- Slower than optimized Python for the 32KB library-call benchmark once process startup is included
+- Faster startup than Go (~3ms vs ~6ms)
 - Best balance of startup time and throughput for mixed CLI workloads
 
-### 4. Process Spawn Overhead Matters
+### 5. Process Spawn Overhead Matters
 
 CLI tools (Go, Zig) have process spawn overhead:
-- Go: ~4ms startup overhead
-- Zig: ~2ms startup overhead
+- Go: ~6ms startup overhead
+- Zig: ~3ms startup overhead
 - Dominates for small inputs (makes them appear slower than Python)
 - Negligible for large inputs where actual work dominates
 - Rust extension avoids this entirely by being a native Python module
@@ -100,9 +108,9 @@ CLI tools (Go, Zig) have process spawn overhead:
 
 | Use Case | Recommended | Why |
 |----------|-------------|-----|
-| Python library calls | **Rust** (`pip install json2xml[fast]`) | 57-129x faster, no process overhead |
-| Small files via CLI | **Zig** (json2xml-zig) | Fastest startup among native CLIs (~2ms) |
-| Large files via CLI | **Go** or **Zig** | Both excellent; Zig wins at 32KB, Go wins at 323KB in this run |
+| Python library calls | **Rust** (`pip install json2xml[fast]`) | 4-14x faster, no process overhead |
+| Small files via CLI | **Zig** (json2xml-zig) | Fastest startup among native CLIs (~3ms) |
+| Large files via CLI | **Go** or **Zig** | Both excellent; Go wins at 323KB in this run |
 | Batch processing | **Go** or **Rust** | Both excellent depending on shell vs Python integration |
 | Pure Python required | **Python** (json2xml) | Always available |