> ## Documentation Index
> Fetch the complete documentation index at: https://docs.unpod.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Load Testing

> Perform load testing and stress testing.

## Getting Started

Load testing helps ensure your Unpod applications can handle concurrent user traffic and maintain optimal performance under various load conditions.

***

## Performance Targets and SLAs

The platform guarantees 99.99% uptime with automatic failover. SLA documentation

### Metric Commitment

| Metric                        | Commitment                      | Credit Policy              |
| ----------------------------- | ------------------------------- | -------------------------- |
| Uptime SLA                    | 99.90% available                | 0.5% credit per 0.1% below |
| End-to-End Latency p99        | less than 1500ms                | Included in E2E            |
| WebApp Service Latency        | less than 10ms internal routing | Included in E2E            |
| Vector Store Query p99        | less than 50ms                  | Included in E2E            |
| MongoDB Write Fire and Forget | less than 40ms                  | Included in E2E            |
| Data Purge Verification       | On-demand audit                 | Included Enterprise        |

### Baseline Performance Metrics

The following metrics represent measured performance under optimal baseline conditions single session, warm cache, optimal network:

| Component                 | Measured p50 | Measured p95 |
| ------------------------- | ------------ | ------------ |
| Platform Orchestration    | 8ms          | 12ms         |
| Speech-to-Text STT        | 0.5s         | 0.7s         |
| LLM Inference             | 0.8s         | 1.2s         |
| Text-to-Speech TTS        | 0.3s         | 0.5s         |
| End-to-End Voice Pipeline | 1.6s         | 2.4s         |

### Concurrent Load Test Results

Platform stability validated under concurrent session load:

| Test Scenario           | Concurrency | Success Rate | Avg Latency |
| ----------------------- | ----------- | ------------ | ----------- |
| Baseline Single Session | 1           | 100%         | 1.6s        |
| Low Concurrency         | 5           | 100%         | 1.65s       |
| Medium Concurrency      | 10          | 100%         | 1.7s        |
| High Concurrency        | 15          | 100%         | 1.7s        |

### Infrastructure Robustness

| Capability         | Status                            |
| ------------------ | --------------------------------- |
| Auto-scaling       | Horizontal pod scaling enabled    |
| Failover           | Multi-region redundancy           |
| Connection Pooling | Optimized for concurrent sessions |
| Rate Limiting      | Per-tenant throttling             |
| Observability      | Real-time latency monitoring      |
| Data Residency     | India region available            |

***

## Scalability Architecture

* Horizontal Scaling: Native HPA Horizontal Pod Autoscaler for all stateless components
* GPU Node Affinity: Dedicated GPU pools with NVIDIA A10G L4 for inference workloads
* Regional Infrastructure: Automatic routing through worldwide infrastructure for optimal latency
* Database Scaling: Postgres read replicas, MongoDB ReplicaSet with automatic failover
* SaaS Auto-scaling: Instant autoscale vs manual capacity planning for self-hosted

***

## Latency Optimization Techniques

* Streaming STT TTS: Real-time processing without full-file buffering
* Speculative Decoding: Parallel token generation for faster LLM responses
* Same Availability Zone: Co-located services to minimize network latency
* gRPC WebSocket: Low-overhead protocols for inter-service communication

***

## Notes

* End-to-End latency includes external service providers STT, LLM, TTS which contribute to variability under load.
* Platform orchestration layer maintains less than 15ms latency regardless of concurrent load.
* Performance optimizations for high-concurrency scenarios are actively being deployed.
* Custom SLA tiers available for enterprise customers with dedicated infrastructure.