Production hardening checklist

Pre-launch checklist for Livepeer applications: gateway selection, authentication, error handling, model selection, cost estimation, and monitoring.

Gateway selection

Environment	Gateway	Notes
Development / testing	dream-gateway.livepeer.cloud	Free, no API key required, unpredictable latency, no SLA
Production (managed)	Gateway provider with API key	Authenticated, rate-limited, provider-specific SLA
Production (self-hosted)	go-livepeer in broadcaster mode	Full control over pricing, routing, and orchestrator selection. Requires ETH deposit on Arbitrum One.

Switch before launch. Do not ship user-facing applications on the community gateway.

Authentication

• API key stored in environment variables or secrets manager, never in source code
• Backend API key used for server-side requests; CORS-enabled key for browser-side
• Key rotation schedule set (90-day recommended)
• .env files excluded from version control via .gitignore
• For self-hosted gateways: Ethereum keystore secured, TicketBroker deposit funded

Error handling

• 401 handler: log and surface a configuration error; do not retry (the key is wrong)
• 422 handler: log the full response body to identify the failing field; fix the request shape
• 503 handler: retry with exponential backoff; cold model load is expected behaviour, not a failure
• 429 handler: back off and retry after the rate limit window resets
• 500 handler: retry once; if persistent, check model_id, input dimensions, and file format
• Global timeout set on all requests (recommended: 300 seconds for cold model scenarios)
• SDK retry configuration enabled with exponential backoff

Model selection (AI applications)

Warm models respond immediately. Cold models take 30 seconds to 5 minutes to load. Use warm models for latency-sensitive paths:

Pipeline	Warm model
text-to-image	SG161222/RealVisXL_V4.0_Lightning
image-to-image	timbrooks/instruct-pix2pix
audio-to-text	openai/whisper-large-v3
image-to-text	Salesforce/blip-image-captioning-large
LLM	meta-llama/Meta-Llama-3.1-8B-Instruct

Custom model_id values cold-start on every request until an orchestrator warms them. Test custom models under expected load before launch.

Video applications

• Transcoding profiles defined for all target renditions (resolution, bitrate, FPS)
• Webhook endpoint configured and signature verification implemented (Livepeer-Signature header)
• Access control policy set on streams and assets (JWT or webhook playback policy)
• Recording enabled if VOD archival is required (record: true on stream creation)
• Player tested across target browsers (HLS fallback for non-WebRTC environments)

BYOC applications

• Container /health endpoint returns {"status": "ok"} under load
• Container registered in orchestrator’s aiModels.json with correct pipeline, model_id, and price_per_unit
• Container handles graceful shutdown on SIGTERM
• Container tested against local orchestrator before registering on mainnet
• GPU memory usage profiled under concurrent session load

Cost estimation

AI inference pricing is orchestrator-set and denominated in wei. Indicative rate for text-to-image: approximately $0.019 per megapixel of output.

Illustrative example for text-to-image at 1024x1024:

• 1024 x 1024 = 1,048,576 pixels = 1.05 megapixels
• At $0.019/MP: approximately$0.020 per image
• At 1,000 images/day: approximately $20/day

Actual rates vary by orchestrator and pipeline. Monitor real costs after the first week of production traffic to calibrate estimates.

Video transcoding is priced in wei per pixel across all output renditions.

Monitoring

• Log HTTP status codes for all gateway requests
• Track p50, p95, p99 latency per pipeline
• Alert on sustained 503 rate above 5% (warm model unavailability)
• Alert on sustained 5xx rate above 1% (inference failures)
• For self-hosted gateways: Prometheus metrics enabled (-metrics flag) and scraped
• TicketBroker deposit balance monitored (for self-hosted gateways)

Launch readiness

• Production gateway configured (not dream-gateway.livepeer.cloud)
• API key in secrets manager, not in code
• Retry policy implemented with exponential backoff
• All error types (401, 422, 503, 429, 500) handled
• Warm models used for latency-sensitive paths (AI applications)
• Cost projection completed
• Monitoring and alerting configured
• Incident response contact identified: Livepeer Discord #builders

See job debugging for the error diagnosis flow when production issues arise.