Respect service limits
Check documented request, payload, and rate limits, including whether they apply per account, resource, or time window. Leave capacity for interactive traffic and other clients using the same service.
Group operations that share a purpose and can be retried safely. Avoid placing unrelated destructive actions in one batch if a partial failure would leave an unclear result.
Handle partial success
Inspect per-item results when the service provides them; an overall success response may still contain rejected items. Track stable item identifiers so failed entries can be retried without resubmitting completed work.
Measure latency, throttling, and error rates while adjusting batch size. Add bounded backoff for rate-limit responses and pause when errors indicate a persistent configuration or permission issue.