Kong Gateway is an open-source API gateway built on OpenResty/NGINX that actually works in production. Version 3.9.1 released in June 2025 includes serious AI capabilities, incremental configuration sync, and performance improvements that don't break existing setups.
The Reality Check
Kong's networking makes me want to throw my laptop when debugging hybrid mode, but it's the least terrible option for managing APIs at scale. Major enterprises use it for critical infrastructure - if it can handle enterprise traffic, it can probably handle your startup's 12 API calls per minute.
Architecture That Actually Makes Sense
Kong splits into data plane and control plane nodes in hybrid mode. The control plane handles configuration, the data plane processes traffic. This isn't marketing bullshit - it means you can configure policies centrally while keeping traffic processing distributed.
The data plane runs on port 8000 (proxy), the control plane on port 8001 (admin API). In hybrid deployments, data planes connect to control planes via mTLS on port 8005. The control plane config sync will fuck you over during rolling updates if you don't sequence them properly - ask me how I know.
Production Reality
Here's what actually matters: Kong can handle 50,000+ requests per second on decent hardware. That's not theoretical - it's measured with K6 load testing against real deployments.
The catch? Performance degrades with plugin complexity. Rate limiting adds ~2ms latency. OAuth validation adds ~5ms. JWT validation is surprisingly fast at ~1ms. The AI Gateway plugins (semantic caching, prompt guards) add 10-50ms depending on your LLM roundtrip.
The Plugin Ecosystem Problem
Kong has 300+ plugins. Half are useful, quarter are enterprise-only, and quarter are "why does this exist?" The good ones:
- Rate Limiting: Actually works, unlike AWS API Gateway's version
- OAuth 2.0: Proper implementation, not a toy
- JWT: Fast, handles rotation correctly
- Request/Response Transformer: Saves you from writing middleware
The problematic ones:
- CORS: Overly complex for what should be simple
- Prometheus: Resource-heavy, consider alternatives
- File Log: Will fill your disk, guarantee it
AI Gateway Capabilities (The New Hotness)
Kong AI Gateway launched in February 2024 and version 3.10 added automated RAG pipelines and PII sanitization. Version 3.11 includes prompt compression that reduces token costs by up to 5x.
Supports OpenAI, Azure OpenAI, AWS Bedrock, Anthropic Claude, Google Gemini, and Cohere. The semantic caching prevents redundant LLM calls - found this out at 2am when a routine config update took down half our API endpoints because cache keys weren't invalidating properly.
Database Requirements (Critical Decision)
Kong supports PostgreSQL and Cassandra for the database layer.
PostgreSQL: Always PostgreSQL. Cassandra sounds cool but the operational overhead isn't worth it unless you're Netflix-scale. Kong 3.x requires PostgreSQL 12+ and the migration scripts actually work.
DB-less mode: Use declarative YAML configuration instead of a database. Perfect for containerized deployments but limited plugin compatibility. Can't use rate limiting or OAuth plugins in DB-less mode - they need persistent storage.
I've run both in production. PostgreSQL is boring and reliable. Cassandra was exciting until the 3am pages about split-brain scenarios. For cloud deployments, check the database requirements and supported versions in the compatibility matrix.