[Koshima]

Great question!

Our approach focuses on: - Fire-and-forget ingestion with in-memory queues so events don’t block product requests - Strict idempotency tokens tied to every event, enforced at the API layer - Lightweight retry logic that prevents double-counting but guarantees delivery under transient failures

Storage-wise, we’ve leaned on a mix of time-series DBs for raw events and pre-aggregated summaries for billing views.

Would love to swap notes on failure patterns or queue setups if you’ve dealt with similar scale.