Limitations

These are the deliberate limits of Cohort, so you don't discover them in production.

Database

• PostgreSQL only. The SQL relies on RETURNING, = ANY(), FOR UPDATE ... SKIP LOCKED and advisory locks. There's no SQL Server or SQLite support, and none is planned.
• You own the schema. Cohort never creates or migrates its tables at runtime. See Schema & Migrations.

Model

• One retained entity maps to one independently owned table. Owned types, shared tables, entity and table splitting, and inheritance hierarchies are rejected at startup.
• Retained entities can't form a foreign-key cycle.
• Rows with a NULL anchor never age out of an ordinary sweep.
• Record IDs whose text form depends on session settings (timestamps, dates, intervals, floats, money, bytea) aren't supported.

Rules

• No built-in conditional, aliasing or caching rule providers. Your IRetentionRuleProvider is plain C#, so build whichever of those you need inside it.
• Anonymisation ships Null, EmptyString and FixedLiteral. Hashing, format-preserving and other schemes belong in an IAnonymiseValueFactory you own.

Delivery

• Row-handler OnAfterAsync is at-least-once, not exactly-once. Handlers must be idempotent.
• A handler that ignores its cancellation token can overlap its own retry.
• Audit observers are best effort, with no durable outbox. Integrations that must not miss an event should read the ledger tables or use CDC.

Scheduling

• Missed cron occurrences are skipped, not caught up.
• Replicas compare occurrence times using their own clocks, so keep them in sync.

Out of scope

• No source generator. Reflection is fine for a daily sweep.
• No OneTrust or Purview adapters.