MCP Time Server Node

# Python Stress Test Research Findings ## Date: 2025-01-28 ## Python Version: 3.11.2 (Debian Bookworm) ### Key Findings 1. **MCP Communication** - MCP servers communicate via JSON-RPC over stdin/stdout - Each request/response is a single line of JSON - Must flush stdin after writing - Responses come back on stdout, one per line 2. **Memory Measurement** - Use psutil.Process(pid).memory_info().rss for Resident Set Size - RSS is the most reliable metric for actual memory usage - Initial memory: ~143MB for our Node.js MCP server - Memory spike on first request: ~51MB increase (V8 warm-up) - Memory stabilizes and can drop below initial (GC working) 3. **Memory Leak Indicators** - Continuous growth without plateau = BAD - Initial spike then stabilization = NORMAL - Memory dropping after requests = GOOD (GC active) 4. **Process Management** - Use subprocess.Popen for full control - Set text=True for string I/O - Set bufsize=0 for unbuffered communication - Always terminate() and wait() for cleanup ### Verified Test Approach ```python # Start server server = subprocess.Popen( ['node', 'dist/index.js'], stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, bufsize=0 ) # Monitor memory proc = psutil.Process(server.pid) memory_mb = proc.memory_info().rss / 1024 / 1024 # Send MCP request request = {"jsonrpc": "2.0", "method": "tools/call", "id": 1, "params": {...}} server.stdin.write(json.dumps(request) + '\n') server.stdin.flush() # Read response response = json.loads(server.stdout.readline()) ``` ### Memory Patterns Observed 1. **Normal Pattern** (what we saw): ``` Initial: 143.28 MB After 1: 194.35 MB (+51.07 MB) - V8 warm-up After 2: 142.39 MB (-0.89 MB) - GC kicked in After 3: 142.39 MB (-0.89 MB) - Stable ``` 2. **Leak Pattern** (what to detect): ``` Initial: 143 MB After 1000: 200 MB After 2000: 250 MB After 3000: 300 MB (Continuous growth) ``` ### Next Steps 1. Create test that establishes baseline behavior 2. Create test that detects memory leaks 3. Create test runner that outputs context-efficient summaries