Report #53800
[architecture] Event sourcing read model corruption and inability to rebuild history after schema changes
Treat projections as ephemeral caches that can be dropped and rebuilt from the event store at any time. Never store business-critical state only in a projection. Implement upcasters \(schema migration functions\) that transform old event formats to new domain models during deserialization, allowing the event store schema to evolve without modifying historical records. Keep raw events append-only and immutable; never update or delete events in place.
Journey Context:
Common antipattern: Treating projections as the 'real' data and the event log as just an audit trail, or allowing ad-hoc updates to projected data. This destroys the fundamental guarantee of event sourcing: the ability to reconstruct any past state by replaying events. When a bug is found in projection logic \(e.g., interest calculation error\), you must be able to wipe the read model tables and rebuild them by re-reading all events from the beginning. If you've coupled write logic to projections or allowed ad-hoc updates, this becomes impossible. Schema evolution is the other hard part: if you change an event structure \(renaming a field\), old events in the store still have the old format. Without an upcaster \(a function that converts old event formats to new ones during deserialization\), replay will fail. Alternative of 'snapshotting' old events \(migrating the store in place\) breaks immutability and audit trails.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T20:47:53.100175+00:00— report_created — created