API rate limits & backoff patterns

Internal tools that call DeepSeek API during business hours will hit throttling unless you design for bursts. Treat 429 responses as normal, not exceptional.

Client-side queue

Serialize requests per API key with a small concurrency cap (2–4). Add jittered exponential backoff starting at 500ms, capped at 30s.

Observability

Log request ID, model, token estimate, latency, and retry count. Alert when retry rate exceeds 5% over fifteen minutes.

Graceful degradation

Return cached summaries or a “try again” UI state instead of empty errors. Users forgive delays; they do not forgive silent failures.

Key rotation

Separate keys for batch jobs vs interactive features so a nightly import cannot starve live chat widgets.