diff --git a/doc/source/admin/index.rst b/doc/source/admin/index.rst
index 287e9d8fb56f..93b4e6a55408 100644
--- a/doc/source/admin/index.rst
+++ b/doc/source/admin/index.rst
@@ -230,3 +230,4 @@ Once you are running nova, the following information is extremely useful.
    node-down
    hw-machine-type
    hw-emulation-architecture
+   soft-delete-shadow-tables
diff --git a/doc/source/admin/soft-delete-shadow-tables.rst b/doc/source/admin/soft-delete-shadow-tables.rst
new file mode 100644
index 000000000000..126279c4d062
--- /dev/null
+++ b/doc/source/admin/soft-delete-shadow-tables.rst
@@ -0,0 +1,62 @@
+=============================
+Soft Delete and Shadow Tables
+=============================
+
+Nova has two unrelated features which are called ``soft delete``:
+
+Soft delete instances that can be restored
+------------------------------------------
+
+After an instance delete request, the actual delete is
+delayed by a configurable amount of time (config option
+:oslo.config:option:`reclaim_instance_interval`).  During the delay,
+the instance is marked to be in state ``SOFT_DELETED`` and can be
+restored (:command:`openstack server restore`) by an admin in order to
+gracefully handle human mistakes. If the instance is not restored during
+the configured delay, a periodic job actually deletes the instance.
+
+This feature is optional and by default off.
+
+See also:
+
+- "Delete, Restore" in `API Guide: Server Concepts
+  <https://docs.openstack.org/api-guide/compute/server_concepts.html#server-actions>`_
+- config reference: :oslo.config:option:`reclaim_instance_interval`
+
+Soft delete database rows to shadow tables
+------------------------------------------
+
+At an actual instance delete, no DB record is deleted. Instead the
+records are marked as deleted (for example ``instances.deleted``
+in Nova cell databases). This preserves historic information
+for debugging and audit uses. But it also leads to accumulation
+of data in Nova cell DB tables, which may have an effect on
+Nova DB performance as documented in `DB prune deleted rows
+<https://docs.openstack.org/nova/latest/admin/upgrades.html#concepts>`_.
+
+The records marked as deleted can be cleaned up in multiple stages.
+First you can move them to so-called shadow tables (tables with prefix
+``shadow_`` in Nova cell databases).  This is called *archiving the
+deleted rows*.  Nova does not query shadow tables, therefore data moved
+to the shadow tables no longer affect DB performance. However storage
+space is still consumed.  Then you can actually delete the information
+from the shadow tables.  This is called *DB purge*.
+
+These operations can be performed by nova-manage:
+
+- https://docs.openstack.org/nova/latest/cli/nova-manage.html#db-archive-deleted-rows
+- https://docs.openstack.org/nova/latest/cli/nova-manage.html#db-purge
+
+This feature is not optional. Every long-running deployment should
+regularly archive and purge the deleted rows. For example via a cron
+job to regularly call :program:`nova-manage db archive_deleted_rows` and
+:program:`nova-manage db purge`.  The tradeoffs between data retention,
+DB performance and storage needs should be considered.
+
+In the Mitaka release there was an agreement between Nova developers that
+it's not desirable to provide shadow tables for every table in the Nova
+database, `documented in a spec
+<https://specs.openstack.org/openstack/nova-specs/specs/mitaka/implemented/no-more-soft-delete.html>`_.
+
+Therefore not all information about an instance is preserved in the shadow
+tables. Since then new shadow tables are not introduced.