How would you choose retention for an event stream, and when does compaction help?

In production each stream has an explicit recoverable window, owner, storage budget, and tested restore path. Alerts fire when oldest consumer lag approaches retention, leaving time to pause cleanup, restore from archive, or rebuild from an authoritative snapshot. I monitor segment growth, compaction lag, tombstone behavior, archival failures, and time required to replay at safe downstream throughput. Disaster exercises start a new consumer from a known checkpoint and prove that current code can process retained schema versions. Access controls and deletion policy account for personal or regulated data copied into long-lived events. The final design often combines broker retention for operational replay, periodic snapshots for fast rebuild, and object storage for longer evidence where policy permits.