Scaling to 6,000 PDFs Per Hour
How I optimized a document generation system to process 6,000 PDFs per hour on a single Railway worker, a throughput number that typically requires 5-10 workers.
- High Scale
- Single Worker
- 99.7% Uptime
- Throughput
- 6,000/hr
- PDFs generated per hour
- P95 Latency
- <2s
- Response time at 95th percentile
- Infrastructure
- 1 Worker
- Single Railway instance
- Success Rate
- 99.7%
- Job completion rate
The Challenge
KamkmPDF started as a simple Next.js app with a cron job. It worked for 50 users, but when hundreds started uploading simultaneously, the queue backed up. Most teams would deploy 5-10 workers. I wanted to see how far a single optimized worker could go.
| Starting Point | Final Results | |
|---|---|---|
| Throughput | 200-300/hr | 6,000/hr |
| Generation time | 5-10s | <2s |
| Concurrent jobs | 3 before crash | 10 stable |
| AI keys | 1 (rate limits) | 11 with failover |
Key Optimizations
Four major changes that transformed performance
Go Worker Rewrite
Migrated from TypeScript to Go for better concurrency and memory efficiency using goroutines.
2x throughput increase
Cerebras AI Integration
Implemented priority-based AI model selection with Cerebras GPT-oss-120b as primary for high-speed inference.
5-10x faster than GPT-4
Database Optimization
Atomic job claiming with FOR UPDATE SKIP LOCKED, connection pooling, and strategic indexing.
Query time: 100ms → 5ms
Browser Pooling
Reusable Puppeteer instances instead of spawning new browsers per job.
10 concurrent PDFs, same memory
Architecture Highlights
Built for horizontal scaling from day one
Queue System
- PostgreSQL-backed queue with atomic job claiming
- Round-robin API key distribution for AI redundancy
- Multi-tier PDF fallback (Puppeteer → PDFKit → Static)
Worker Design
- Browser pool with automatic cleanup and monitoring
- Horizontal scaling ready — just add more workers
Technology Stack
Carefully chosen for maximum throughput
Frontend
- Next.js 15
- React 19
- Tailwind CSS
- tRPC
Worker
- Go
- Goroutines
- Railway
- Docker
Database
- PostgreSQL 14
- Drizzle ORM
- Connection Pooling
AI
- Cerebras GPT-oss-120b
- OpenRouter
- 11 API Keys
- Auto-failover
PDF Generation
- Puppeteer
- PDFKit
- Browser Pool
- Memory Management
Storage
- UploadThing v7
- Signed URLs
- File Metadata
Cost Efficiency
Enterprise throughput, startup budget
Infrastructure
- Single Railway worker
- ~$20/mo
- PostgreSQL database
- ~$15/mo
- No Redis or queues
- $0
AI Costs
- Cerebras GPT-oss-120b
- ~$0.002/PDF
- At 6,000 PDFs/hour
- ~$12/hr
- vs GPT-4 equivalent
- 5x savings