Summary

Load testing with 4000 Locust users reveals that 271 endpoints still use Depends(get_db) which holds database sessions for the entire request lifecycle. This causes:

1. OOM kills - Gateway workers exceed 8GB container limit under load
2. Idle-in-transaction buildup - 120+ connections stuck for 60-290 seconds
3. Memory pressure - Held sessions + pending responses consume RAM

Evidence

From profiling RCA (todo/claude-rca-2026-01-22-profiling.md):

Memory cgroup out of memory: Killed process 1172024 (mcpgateway work)
total-vm:7470296kB, anon-rss:1102856kB

Stuck queries in pg_stat_activity:

• `SELECT tools.*` - stuck 64-221 seconds
• `SELECT email_teams.*` - stuck 69-215 seconds

Endpoints to Fix (Priority Order)

High Priority (main.py - 52 endpoints)

```bash
grep -n "Depends(get_db)" mcpgateway/main.py | wc -l

52 endpoints

```

Key endpoints:

• `/tools`, `/tools/{id}` - CRUD operations
• `/servers`, `/servers/{id}` - CRUD operations
• `/resources`, `/resources/{id}` - CRUD operations
• `/prompts`, `/prompts/{id}` - CRUD operations
• `/health`, `/ready` - Health checks

Medium Priority (admin.py)

• `/admin/` - Heavy HTML template rendering (5-7s response times)
• `/admin/tools/partial` - HTMX partials
• `/admin/resources/partial` - HTMX partials
• `/admin/prompts/partial` - HTMX partials

Lower Priority (routers/)

• `mcpgateway/routers/tokens.py` - 10 endpoints
• `mcpgateway/routers/sso.py` - 9 endpoints
• `mcpgateway/routers/oauth_router.py` - 7 endpoints
• Other routers

Pattern to Apply

Replace:
```python
@router.get("/tools")
async def list_tools(db: Session = Depends(get_db)):
tools = tool_service.list_tools(db)
return tools
```

With:
```python
@router.get("/tools")
async def list_tools():
with fresh_db_session() as db:
tools = tool_service.list_tools(db)
return [t.model_dump() for t in tools] # Serialize inside context
```

Acceptance Criteria

• All 271 `Depends(get_db)` usages replaced with `fresh_db_session()`
• Unit tests pass
• Load test (4000 users) runs for 10+ minutes without OOM kills
• No `idle in transaction` connections older than 60 seconds

Related Issues

• [BUG][PERFORMANCE][DB]: Endpoint handlers hold DB sessions during slow MCP backend calls #2323 - Database session lifetime performance issues
• [BUG][PERFORMANCE]: TokenScopingMiddleware causes connection pool exhaustion under load #2330 - TokenScopingMiddleware connection pool exhaustion

References

• `todo/claude-rca-2026-01-22-profiling.md` - Full profiling report
• `todo/rca-part-2.md` - Initial RCA
• `docs/docs/development/profiling.md` - Profiling guide