Virtual Tables

Virtual Tables let a developer provide an underlying table implementations, while still presenting a normal SQL interface to the user. The person writing SQL doesn’t need to know or care that some of the tables come from elsewhere.

Some examples of how you might use this:

  • Translating to/from information stored in other formats

  • Accessing the data remotely (eg you could make a table that backends into the cloud)

  • Dynamic information (eg currently running processes, files and directories, objects in your program)

  • There are other examples on the SQLite page


You’ll find initial development a lot quicker by using apsw.ext.make_virtual_module() which lets you export a Python function as a virtual table, being able to provide positional and keyword arguments as part of your query.

See the example.

To write a virtual table, you need to have 3 types of object. A module providing the module, a virtual table, and a cursor that moves through a table.

IndexInfo class

class apsw.IndexInfo

IndexInfo represents the sqlite3_index_info and associated methods used in the VTTable.BestIndexObject() method.

Naming is identical to the C structure rather than Pythonic. You can access members directly while needing to use get/set methods for array members.

You will get ValueError if you use the object outside of an BestIndex method.

apsw.ext.index_info_to_dict() provides a convenient representation of this object as a dict.

IndexInfo.colUsed: set[int]

(Read-only) Columns used by the statement. Note that a set is returned, not the underlying integer.

IndexInfo.distinct: int

(Read-only) How the query planner would like output ordered if the query is using group by or distinct.

Calls: sqlite3_vtab_distinct

IndexInfo.estimatedCost: float

Estimated cost of using this index

IndexInfo.estimatedRows: int

Estimated number of rows returned

IndexInfo.get_aConstraintUsage_argvIndex(which: int) int

Returns argvIndex for aConstraintUsage[which]

IndexInfo.get_aConstraintUsage_in(which: int) bool

Returns True if the constraint is in - eg column in (3, 7, 9)

Calls: sqlite3_vtab_in

IndexInfo.get_aConstraintUsage_omit(which: int) bool

Returns omit for aConstraintUsage[which]

IndexInfo.get_aConstraint_collation(which: int) str

Returns collation name for aConstraint[which]

Calls: sqlite3_vtab_collation

IndexInfo.get_aConstraint_iColumn(which: int) int

Returns iColumn for aConstraint[which]

IndexInfo.get_aConstraint_op(which: int) int

Returns op for aConstraint[which]

IndexInfo.get_aConstraint_rhs(which: int) SQLiteValue

Returns right hand side value if known, else None.

Calls: sqlite3_vtab_rhs_value

IndexInfo.get_aConstraint_usable(which: int) bool

Returns usable for aConstraint[which]

IndexInfo.get_aOrderBy_desc(which: int) bool

Returns desc for aOrderBy[which]

IndexInfo.get_aOrderBy_iColumn(which: int) int

Returns iColumn for aOrderBy[which]

IndexInfo.idxFlags: int


IndexInfo.idxNum: int

Number used to identify the index

IndexInfo.idxStr: str | None

Name used to identify the index

IndexInfo.nConstraint: int

(Read-only) Number of constraint entries

IndexInfo.nOrderBy: int

(Read-only) Number of order by entries

IndexInfo.orderByConsumed: bool

True if index output is already ordered

IndexInfo.set_aConstraintUsage_argvIndex(which: int, argvIndex: int) None

Sets argvIndex for aConstraintUsage[which]

IndexInfo.set_aConstraintUsage_in(which: int, filter_all: bool) None

If which is an in constraint, and filter_all is True then your VTCursor.Filter() method will have all of the values at once.

Calls: sqlite3_vtab_in

IndexInfo.set_aConstraintUsage_omit(which: int, omit: bool) None

Sets omit for aConstraintUsage[which]

VTModule class

class apsw.VTModule


There is no actual VTModule class - it is shown this way for documentation convenience and is present as a typing protocol.

A module instance is used to create the virtual tables. Once you have a module object, you register it with a connection by calling Connection.create_module():

# make an instance

# register the vtable on connection con
con.create_module("modulename", mymod)

# tell SQLite about the table
con.execute("create VIRTUAL table tablename USING modulename('arg1', 2)")

The create step is to tell SQLite about the existence of the table. Any number of tables referring to the same module can be made this way.

VTModule.Connect(connection: Connection, modulename: str, databasename: str, tablename: str, *args: tuple[SQLiteValue, ...]) tuple[str, VTTable]

The parameters and return are identical to Create(). This method is called when there are additional references to the table. Create() will be called the first time and Connect() after that.

The advise is to create caches, generated data and other heavyweight processing on Create() calls and then find and reuse that on the subsequent Connect() calls.

The corresponding call is VTTable.Disconnect(). If you have a simple virtual table implementation, then just set Connect() to be the same as Create():

class MyModule:

     def Create(self, connection, modulename, databasename, tablename, *args):
         # do lots of hard work


SQLite xConnect reference

VTModule.Create(connection: Connection, modulename: str, databasename: str, tablename: str, *args: tuple[SQLiteValue, ...]) tuple[str, VTTable]

Called when a table is first created on a connection.

  • connection – An instance of Connection

  • modulename – The string name under which the module was registered

  • databasename – The name of the database. main, temp, the name in ATTACH

  • tablename – Name of the table the user wants to create.

  • args – Any arguments that were specified in the create virtual table statement.


A list of two items. The first is a SQL create table statement. The columns are parsed so that SQLite knows what columns and declared types exist for the table. The second item is an object that implements the table methods.

The corresponding call is VTTable.Destroy().

SQLite xCreate reference

VTModule.ShadowName(table_suffix: str) bool

This method is called to check if table_suffix is a shadow name

The default implementation always returns False.

If a virtual table is created using this module named example and then a real table is created named example_content, this would be called with a table_suffix of content

SQLite xShadowName reference

VTTable class

class apsw.VTTable


There is no actual VTTable class - it is shown this way for documentation convenience and is present as a typing protocol.

The VTTable object contains knowledge of the indices, makes cursors and can perform transactions.

A virtual table is structured as a series of rows, each of which has the same number of columns. The value in a column must be one of the 5 supported types, but the type can be different between rows for the same column. The virtual table routines identify the columns by number, starting at zero.

Each row has a unique 64 bit integer rowid with the Cursor routines operating on this number, as well as some of the Table routines such as UpdateChangeRow.

It is possible to not have a rowid

VTTable.Begin() None

This function is used as part of transactions. You do not have to provide the method.

SQLite xBegin reference

VTTable.BestIndex(constraints: Sequence[tuple[int, int]], orderbys: Sequence[tuple[int, int]]) Any

This is a complex method. To get going initially, just return None and you will be fine. You should also consider using BestIndexObject() instead.

Implementing this method reduces the number of rows scanned in your table to satisfy queries, but only if you have an index or index like mechanism available.


The implementation of this method differs slightly from the SQLite documentation for the C API. You are not passed “unusable” constraints. The argv/constraintarg positions are not off by one. In the C api, you have to return position 1 to get something passed to VTCursor.Filter() in position 0. With the APSW implementation, you return position 0 to get Filter arg 0, position 1 to get Filter arg 1 etc.

The purpose of this method is to ask if you have the ability to determine if a row meets certain constraints that doesn’t involve visiting every row. An example constraint is price > 74.99. In a traditional SQL database, queries with constraints can be speeded up with indices. If you return None, then SQLite will visit every row in your table and evaluate the constraints itself. Your index choice returned from BestIndex will also be passed to the Filter() method on your cursor object. Note that SQLite may call this method multiple times trying to find the most efficient way of answering a complex query.


You will be passed the constraints as a sequence of tuples containing two items. The first item is the column number and the second item is the operation.

Example query: select * from foo where price > 74.99 and quantity<=10 and customer='Acme Widgets'

If customer is column 0, price column 2 and quantity column 5 then the constraints will be:


Note that you do not get the value of the constraint (ie “Acme Widgets”, 74.99 and 10 in this example).

If you do have any suitable indices then you return a sequence the same length as constraints with the members mapping to the constraints in order. Each can be one of None, an integer or a tuple of an integer and a boolean. Conceptually SQLite is giving you a list of constraints and you are returning a list of the same length describing how you could satisfy each one.

Each list item returned corresponding to a constraint is one of:


This means you have no index for that constraint. SQLite will have to iterate over every row for it.


This is the argument number for the constraintargs being passed into the Filter() function of your cursor (the values “Acme Widgets”, 74.99 and 10 in the example).

(integer, boolean)

By default SQLite will check what you return. For example if you said that you had an index on price and so would only return rows greater than 74.99, then SQLite will still check that each row you returned is greater than 74.99. If the boolean is True then SQLite will not double check, while False retains the default double checking.

Example query: select * from foo where price > 74.99 and quantity<=10 and customer=='Acme Widgets'. customer is column 0, price column 2 and quantity column 5. You can index on customer equality and price.

Constraints (in)

Constraints used (out)


When your Filter method in the cursor is called, constraintarg[0] will be “Acme Widgets” (customer constraint value) and constraintarg[1] will be 74.99 (price constraint value). You can also return an index number (integer) and index string to use (SQLite attaches no significance to these values - they are passed as is to your VTCursor.Filter() method as a way for the BestIndex method to let the Filter() method know which of your indices or similar mechanism to use.


The second argument to BestIndex is a sequence of orderbys because the query requested the results in a certain order. If your data is already in that order then SQLite can give the results back as is. If not, then SQLite will have to sort the results first.

Example query: select * from foo order by price desc, quantity asc

Price is column 2, quantity column 5 so orderbys will be:

(2, True),  # True means descending, False is ascending
(5, False)


You should return up to 5 items. Items not present in the return have a default value.

0: constraints used (default None)

This must either be None or a sequence the same length as constraints passed in. Each item should be as specified above saying if that constraint is used, and if so which constraintarg to make the value be in your VTCursor.Filter() function.

1: index number (default zero)

This value is passed as is to VTCursor.Filter()

2: index string (default None)

This value is passed as is to VTCursor.Filter()

3: orderby consumed (default False)

Return True if your output will be in exactly the same order as the orderbys passed in

4: estimated cost (default a huge number)

Approximately how many disk operations are needed to provide the results. SQLite uses the cost to optimise queries. For example if the query includes A or B and A has 2,000 operations and B has 100 then it is best to evaluate B before A.

A complete example

Query is select * from foo where price>74.99 and quantity<=10 and customer=="Acme Widgets" order by price desc, quantity asc. Customer is column 0, price column 2 and quantity column 5. You can index on customer equality and price.

BestIndex(constraints, orderbys)

constraints= ( (2, apsw.SQLITE_INDEX_CONSTRAINT_GT),
               (5, apsw.SQLITE_INDEX_CONSTRAINT_LE),
               (0, apsw.SQLITE_INDEX_CONSTRAINT_EQ)  )

orderbys= ( (2, True), (5, False) )

# You return

( (1, None, 0),   # constraints used
  27,             # index number
  "idx_pr_cust",  # index name
  False,          # results are not in orderbys order
  1000            # about 1000 disk operations to access index

# Your Cursor.Filter method will be called with:

27,              # index number you returned
"idx_pr_cust",   # index name you returned
"Acme Widgets",  # constraintarg[0] - customer
74.99            # constraintarg[1] - price

SQLite xBestIndex reference

VTTable.BestIndexObject(index_info: IndexInfo) bool

This method is called instead of BestIndex() if use_bestindex_object was True in the call to Connection.create_module().

Use the IndexInfo to tell SQLite about your indexes, and extract other information.

Return True to indicate all is well. If you return False or there is an error, then SQLITE_CONSTRAINT is returned to SQLite.

SQLite xBestIndex reference

VTTable.Commit() None

This function is used as part of transactions. You do not have to provide the method.

SQLite xCommit reference

VTTable.Destroy() None

The opposite of VTModule.Create(). This method is called when the table is no longer used. Note that you must always release resources even if you intend to return an error, as it will not be called again on error.

SQLite xDestroy reference

VTTable.Disconnect() None

The opposite of VTModule.Connect(). This method is called when a reference to a virtual table is no longer used, but VTTable.Destroy() will be called when the table is no longer used.

SQLite xDisconnect reference

VTTable.FindFunction(name: str, nargs: int) None | Callable | tuple[int, Callable]

Called to find if the virtual table has its own implementation of a particular scalar function. You do not have to provide this method.

  • name – The function name

  • nargs – How many arguments the function takes

Return None if you don’t have the function. Zero is then returned to SQLite.

Return a callable if you have one. One is then returned to SQLite with the function.

Return a sequence of int, callable. The int is returned to SQLite with the function. This is useful for SQLITE_INDEX_CONSTRAINT_FUNCTION returns.

It isn’t possible to tell SQLite about exceptions in this function, so an unraisable exception is used.

SQLite xFindFunction reference

VTTable.Integrity(schema: str, name: str, is_quick: int) str | None

If present, check the integrity of the virtual table.


None if there are no problems, else a string to be used as an error message. The string is returned to the pragma as is, so it is recommended that you include the database and table name to clarify what database and table the message is referring to.

SQLite xIntegrity reference

VTTable.Open() VTCursor

Returns a cursor object.

SQLite xOpen reference

VTTable.Release(level: int) None

Release nested transactions back to level.

If you do not provide this method then the call succeeds (matching SQLite behaviour when no callback is provided).

SQLite xRelease reference

VTTable.Rename(newname: str) None

Notification that the table will be given a new name. If you return without raising an exception, then SQLite renames the table (you don’t have to do anything). If you raise an exception then the renaming is prevented. You do not have to provide this method.

SQLite xRename reference

VTTable.Rollback() None

This function is used as part of transactions. You do not have to provide the method.

SQLite xRollbackTo reference

VTTable.Savepoint(level: int) None

Set nested transaction to level.

If you do not provide this method then the call succeeds (matching SQLite behaviour when no callback is provided).

SQLite xSavepoint reference

VTTable.Sync() None

This function is used as part of transactions. You do not have to provide the method.

SQLite xSync reference

VTTable.UpdateChangeRow(row: int, newrowid: int, fields: tuple[SQLiteValue, ...]) None

Change an existing row. You may also need to change the rowid - for example if the query was UPDATE table SET rowid=rowid+100 WHERE ...

  • row – The existing 64 bit integer rowid

  • newrowid – If not the same as row then also change the rowid to this.

  • fields – A tuple of values the same length and order as columns in your table

SQLite xUpdate reference

VTTable.UpdateDeleteRow(rowid: int) None

Delete the row with the specified rowid.


rowid – 64 bit integer

SQLite xUpdate reference

VTTable.UpdateInsertRow(rowid: int | None, fields: tuple[SQLiteValue, ...]) int | None

Insert a row with the specified rowid.

  • rowidNone if you should choose the rowid yourself, else a 64 bit integer

  • fields – A tuple of values the same length and order as columns in your table


If rowid was None then return the id you assigned to the row. If rowid was not None then the return value is ignored.

SQLite xUpdate reference

VTCursor class

class apsw.VTCursor


There is no actual VTCursor class - it is shown this way for documentation convenience and is present as a typing protocol.

The VTCursor object is used for iterating over a table. There may be many cursors simultaneously so each one needs to keep track of where in the table it is.

VTCursor.Close() None

This is the destructor for the cursor. Note that you must cleanup. The method will not be called again if you raise an exception.

SQLite xClose reference

VTCursor.Column(number: int) SQLiteValue

Requests the value of the specified column number of the current row. If number is -1 then return the rowid.


Must be one one of the 5 supported types

SQLite xColumn reference

VTCursor.ColumnNoChange(number: int) SQLiteValue

VTTable.UpdateChangeRow() is going to be called which includes values for all columns. However this column is not going to be changed in that update.

If you return apsw.no_change then VTTable.UpdateChangeRow() will have apsw.no_change for this column. If you return anything else then it will have that value - as though VTCursor.Column() had been called.

This method will only be called if use_no_change was True in the call to Connection.create_module().

SQLite xColumn reference

Calls: sqlite3_vtab_nochange

VTCursor.Eof() bool

Called to ask if we are at the end of the table. It is called after each call to Filter and Next.


False if the cursor is at a valid row of data, else True


This method can only return True or False to SQLite. If you have an exception in the method or provide a non-boolean return then True (no more data) will be returned to SQLite.

SQLite xEof reference

VTCursor.Filter(indexnum: int, indexname: str, constraintargs: tuple | None) None

This method is always called first to initialize an iteration to the first row of the table. The arguments come from the BestIndex() or BestIndexObject() with constraintargs being a tuple of the constraints you requested. If you always return None in BestIndex then indexnum will be zero, indexstring will be None and constraintargs will be empty).

If you had an in constraint and set IndexInfo.set_aConstraintUsage_in() then that value will be a set.

SQLite xFilter reference

VTCursor.Next() None

Move the cursor to the next row. Do not have an exception if there is no next row. Instead return False when Eof() is subsequently called.

If you said you had indices in your VTTable.BestIndex() return, and they were selected for use as provided in the parameters to Filter() then you should move to the next appropriate indexed and constrained row.

SQLite xNext reference

VTCursor.Rowid() int

Return the current rowid.

SQLite xRowid reference