Soft deletes versus archive tables

What it is

Soft delete: a deleted_at timestamptz (or is_deleted boolean) column; deleting sets it, and every read adds WHERE deleted_at IS NULL. Archive: the row is removed from the live table and inserted into an archive table with the same columns plus archived_at, archived_by and a reason, in one transaction (a data-modifying CTE with DELETE ... RETURNING feeding an INSERT does both in one statement). A third design, a history table filled by trigger for every change, records deletion as one event among updates.

Why it matters

Deletion is a business event ("account closed", "draft discarded"), and the two designs answer different questions afterwards. Soft deletes make undelete trivial and keep foreign keys intact, but every query, unique constraint, count and join must remember the filter; one forgotten predicate shows deleted data to users. Archive tables keep the live schema honest but turn restore into a copy back and complicate references from other tables.

How to apply

• Soft delete when deleted rows are read often (undo within days, "show deleted" views, rows still referenced elsewhere). Expose live rows through a view (CREATE VIEW customers AS SELECT ... FROM customers_all WHERE deleted_at IS NULL) or a row-level security policy, so the filter exists in one place.
• Turn unique constraints into partial unique indexes: the PostgreSQL documentation describes a unique index with a WHERE predicate as enforcing uniqueness only among rows that satisfy the predicate, so CREATE UNIQUE INDEX ... ON users (email) WHERE deleted_at IS NULL frees an address when the account is deleted.
• Archive when deleted rows are rarely read (retention, occasional investigation). Decide what happens to child rows: archive them in the same transaction or forbid deletion while children exist.
• Define a purge in both designs: soft-deleted rows past the retention period are physically deleted; archive tables are partitioned or pruned by date.
• A marker is not removal: where records must actually disappear, plan physical deletion or anonymisation.

Pitfalls

Soft-deleted rows still count against foreign keys and non-partial unique indexes. ORM-level global filters are bypassed by raw SQL and reporting tools. Indexes on a soft-deleting table grow with dead history unless they are partial. Archive tables drift from the live schema when migrations forget them; a test that inserts a live row into the archive table catches this.