- File doc/shared_schema.md — part of check-in [9a78d89c] at 2019-02-20 17:36:10 on branch reuse-schema — Add test and fixes for SQLITE_OPEN_SHARED_SCHEMA mode. (user: dan size: 3885)
Shared-Schema Mode Notes
This branch contains a patch to allow SQLite connections to share schemas between database connections within the same process in order to save memory. Schemas may be shared between multiple databases attached to the same or distinct connection handles.
To activate shared-schemas, a database connection must be opened using the sqlite3_open_v2() API with the SQLITE_OPEN_SHARED_SCHEMA flag specified. The main database and any attached databases will then share an in-memory Schema object with any other database opened within the process for which:
- the contents of the sqlite_master table, including all object names, SQL statements and root pages are identical, and
- have the same values for the schema-cookie.
Temp databases (those populated with "CREATE TEMP TABLE" and similar statements) never share schemas.
Connections opened with the SQLITE_OPEN_SHARED_SCHEMA flag specified may not modify any database schema except that belonging to the temp database in anyway. This includes creating or dropping database objects, vacuuming the database, or running ANALYZE when the sqlite_stat tables do not exist.
For SQLITE_OPEN_SHARED_SCHEMA connections, the SQLITE_DBSTATUS_SCHEMA_USED sqlite3_db_status() verb distributes the memory used for a shared schema object evenly between all database connections that share it.
A single Schema object is never used by more than one database simultaneously, regardless of whether or not those databases are attached to the same or different database handles. Instead, a pool of schema objects is maintained for each unique sqlite_master-contents/schema-cookie combination opened within the process. Each time database schemas are required by a connection, for example as part of an sqlite3_prepare*(), sqlite3_blob_open() or sqlite3_blob_open() call, it obtains the minimum number of schemas required from the various schema-pools, returning them at the end of the call. This means that a single schema-pool only ever contains more than one copy of the schema if:
- Two threads require schemas from the same pool at the same time, or
- A single sqlite3_prepare*() call requires schemas for two or more attached databases that use the same schema-pool.
The size of a schema-pool never shrinks. Each schema pool always maintains a number of schema objects equal to the highwater mark of schema objects simultaneously required by clients.
This approach is preferred to allowing multiple databases to use the same Schema object simultaneously for three reasons:
- The Schema object is not completely read-only. For example, the Index.zIdxAff string is allocated lazily.
- Throughout the statement compiler, SQLite uses variables like Table.pSchema and Index.pSchema with the sqlite3SchemaToIndex() routine in order to determine which attached database a Table or Index object resides in. This mechanism does not work if the same Schema may be used by two or more attached databases.
- It may be easier to modify this approach in order to allow SQLITE_OPEN_SHARED_SCHEMA connections to modify database schemas, should that be required.
SQLITE_OPEN_SHARED_SCHEMA connections do not store their virtual-table handles in the Table.pVTable list of each table. This would not work, as (a) there is no guarantee that a connection will be assigned the same Schema object each time it requests one from a schema-pool and (b) a single Schema (and therefore Table) object may correspond to tables in two or more databases attached to a single connection. Instead, all virtual-table handles associated with a single database are stored in a linked-list headed at Db.pVTable.