Production Launch Checklist

Complete these checks before increasing production traffic to reduce key exposure, retry storms, uncontrolled costs, and incidents caused by provider differences.

APIs and modelsCopy Anchor Link

• Validate every production path you use, including non-streaming, streaming, tool calling, and structured output.
• Confirm context limits, maximum output, supported parameters, model versions, and fallback models.
• Ensure parsers tolerate unknown fields and handle empty content, truncation, and multiple tool calls.

Reliability and capacityCopy Anchor Link

• Set explicit connection, first-token, read, and overall business deadlines.
• Ensure SDK retries, business retries, and platform failover do not compound into retry storms.
• Load-test target RPM, TPM, concurrency, P95/P99 latency, and long-lived connection counts.
• Rehearse degradation for interrupted streams, 429, 5xx, insufficient balance, and unavailable models.

Security, privacy, and costCopy Anchor Link

• Store keys only on the server, isolate environments, and rehearse rotation and revocation.
• Minimize, redact, and retain prompts, responses, and tool results according to data classification.
• Protect dangerous tools with allowlists, argument validation, authorization, idempotency, and human confirmation.
• Attribute tokens and costs by tenant and feature; alert on budgets, balance, and unusual spending.

Operations and supportCopy Anchor Link

• Log internal trace ID, APIQIK request ID, model, status, duration, and usage.
• Monitor success rate, TTFT, P95/P99, 429, 5xx, costs, and content quality.
• Ensure on-call staff know the status page, support channels, escalation process, and model fallback controls.
• Support gradual rollout and quick rollback for major version or model changes.