Skip to content

Event Replay - Professional

Kafka offsets provide partition-local progress; event-sourced stores such as EventStoreDB organize streams by aggregate; Elasticsearch aliases support atomic read-model swaps. At billion-event scale, source read bandwidth, sink write amplification, compaction, and live-tail convergence dominate. Dashboard events/s, bytes/s, lag derivative, error classes, checksum drift, and cutover readiness.

Best practices

  • Separate pure projection logic from side effects.
  • Version events and replay code for the full retention horizon.
  • Reserve capacity for live traffic and emergency rollback.
  • Prove convergence and semantic equivalence before cutover. 
    rebuild history -> follow live tail -> verify -> atomic swap

Test yourself

  1. How would you estimate catch-up time under continuing writes?
  2. What proof supports deletion of the old projection?
  3. How does log compaction change replay guarantees?

Further reading

  • Fowler, Event Sourcing.
  • Kafka log and offset documentation.
  • EventStoreDB projection documentation.