Small. Fast. Reliable.
Choose any three.
Search for:
The Virtual Table Mechanism Of SQLite

1. Introduction

A virtual table is an object that is registered with an open SQLite database connection. From the perspective of an SQL statement, the virtual table object looks like any other table or view. But behind the scenes, queries and updates on a virtual table invoke callback methods of the virtual table object instead of reading and writing on the database file.

The virtual table mechanism allows an application to publish interfaces that are accessible from SQL statements as if they were tables. SQL statements can do almost anything to a virtual table that they can do to a real table, with the following exceptions:

Individual virtual table implementations might impose additional constraints. For example, some virtual implementations might provide read-only tables. Or some virtual table implementations might allow INSERT or DELETE but not UPDATE. Or some virtual table implementations might limit the kinds of UPDATEs that can be made.

A virtual table might represent an in-memory data structures. Or it might represent a view of data on disk that is not in the SQLite format. Or the application might compute the content of the virtual table on demand.

Here are some existing and postulated uses for virtual tables:

1.1. Usage

A virtual table is created using a CREATE VIRTUAL TABLE statement.

create-virtual-table-stmt:

syntax diagram create-virtual-table-stmt

The CREATE VIRTUAL TABLE statement creates a new table called table-name derived from the class class module-name. The module-name is the name that is registered for the virtual table by the sqlite3_create_module() interface.

CREATE VIRTUAL TABLE tablename USING modulename;

One can also provide comma-separated arguments to the module following the module name:

CREATE VIRTUAL TABLE tablename USING modulename(arg1, arg2, ...);

The format of the arguments to the module is very general. Each module-argument may contain keywords, string literals, identifiers, numbers, and punctuation. Each module-argument is passed as written (as text) into the constructor method of the virtual table implementation when the virtual table is created and that constructor is responsible for parsing and interpreting the arguments. The argument syntax is sufficiently general that a virtual table implementation can, if it wants to, interpret its arguments as column definitions in an ordinary CREATE TABLE statement. The implementation could also impose some other interpretation on the arguments.

Once a virtual table has been created, it can be used like any other table with the exceptions noted above and imposed by specific virtual table implementations. A virtual table is destroyed using the ordinary DROP TABLE syntax.

1.1.1. Temporary virtual tables

There is no "CREATE TEMP VIRTUAL TABLE" statement. To create a temporary virtual table, add the "temp" schema before the virtual table name.

CREATE VIRTUAL TABLE temp.tablename USING module(arg1, ...);

1.1.2. Eponymous virtual tables

Some virtual tables exist automatically in the "main" schema of every database connection in which their module is registered, even without a CREATE VIRTUAL TABLE statement. Such virtual tables are called "eponymous virtual tables". To use an eponymous virtual table, simply use the module name as if it were a table. Eponymous virtual tables exist in the "main" schema only, so they will not work if prefixed with a different schema name.

An example of an eponymous virtual table is the dbstat virtual table. To use the dbstat virtual table as an eponymous virtual table, simply query against the "dbstat" module name, as if it were an ordinary table. (Note that SQLite must be compiled with the SQLITE_ENABLE_DBSTAT_VTAB option to include the dbstat virtual table in the build.)

SELECT * FROM dbstat;

A virtual table is eponymous if its xCreate method is the exact same function as the xConnect method, or if the xCreate method is NULL. The xCreate method is called when a virtual table is first created using the CREATE VIRTUAL TABLE statement. The xConnect method whenever a database connection attaches to or reparses a schema. When these two methods are the same, that indicates that the virtual table has no persistent state that needs to be created and destroyed.

1.1.3. Eponymous-only virtual tables

If the xCreate method is NULL, then CREATE VIRTUAL TABLE statements are prohibited for that virtual table, and the virtual table is an "eponymous-only virtual table". Eponymous-only virtual tables are useful as table-valued functions.

Note that prior to version 3.9.0 (2015-10-14), SQLite did not check the xCreate method for NULL before invoking it. So if an eponymous-only virtual table is registered with SQLite version 3.8.11.1 (2015-07-29) or earlier and a CREATE VIRTUAL TABLE command is attempted against that virtual table module, a jump to a NULL pointer will occur, resulting in a crash.

1.2. Implementation

Several new C-level objects are used by the virtual table implementation:

typedef struct sqlite3_vtab sqlite3_vtab;
typedef struct sqlite3_index_info sqlite3_index_info;
typedef struct sqlite3_vtab_cursor sqlite3_vtab_cursor;
typedef struct sqlite3_module sqlite3_module;

The sqlite3_module structure defines a module object used to implement a virtual table. Think of a module as a class from which one can construct multiple virtual tables having similar properties. For example, one might have a module that provides read-only access to comma-separated-value (CSV) files on disk. That one module can then be used to create several virtual tables where each virtual table refers to a different CSV file.

The module structure contains methods that are invoked by SQLite to perform various actions on the virtual table such as creating new instances of a virtual table or destroying old ones, reading and writing data, searching for and deleting, updating, or inserting rows. The module structure is explained in more detail below.

Each virtual table instance is represented by an sqlite3_vtab structure. The sqlite3_vtab structure looks like this:

struct sqlite3_vtab {
  const sqlite3_module *pModule;
  int nRef;
  char *zErrMsg;
};

Virtual table implementations will normally subclass this structure to add additional private and implementation-specific fields. The nRef field is used internally by the SQLite core and should not be altered by the virtual table implementation. The virtual table implementation may pass error message text to the core by putting an error message string in zErrMsg. Space to hold this error message string must be obtained from an SQLite memory allocation function such as sqlite3_mprintf() or sqlite3_malloc(). Prior to assigning a new value to zErrMsg, the virtual table implementation must free any preexisting content of zErrMsg using sqlite3_free(). Failure to do this will result in a memory leak. The SQLite core will free and zero the content of zErrMsg when it delivers the error message text to the client application or when it destroys the virtual table. The virtual table implementation only needs to worry about freeing the zErrMsg content when it overwrites the content with a new, different error message.

The sqlite3_vtab_cursor structure represents a pointer to a specific row of a virtual table. This is what an sqlite3_vtab_cursor looks like:

struct sqlite3_vtab_cursor {
  sqlite3_vtab *pVtab;
};

Once again, practical implementations will likely subclass this structure to add additional private fields.

The sqlite3_index_info structure is used to pass information into and out of the xBestIndex method of the module that implements a virtual table.

Before a CREATE VIRTUAL TABLE statement can be run, the module specified in that statement must be registered with the database connection. This is accomplished using either of the sqlite3_create_module() or sqlite3_create_module_v2() interfaces:

int sqlite3_create_module(
  sqlite3 *db,               /* SQLite connection to register module with */
  const char *zName,         /* Name of the module */
  const sqlite3_module *,    /* Methods for the module */
  void *                     /* Client data for xCreate/xConnect */
);
int sqlite3_create_module_v2(
  sqlite3 *db,               /* SQLite connection to register module with */
  const char *zName,         /* Name of the module */
  const sqlite3_module *,    /* Methods for the module */
  void *,                    /* Client data for xCreate/xConnect */
  void(*xDestroy)(void*)     /* Client data destructor function */
);

The sqlite3_create_module() and sqlite3_create_module_v2() routines associates a module name with an sqlite3_module structure and a separate client data that is specific to each module. The only difference between the two create_module methods is that the _v2 method includes an extra parameter that specifies a destructor for client data pointer. The module structure is what defines the behavior of a virtual table. The module structure looks like this:

struct sqlite3_module {
  int iVersion;
  int (*xCreate)(sqlite3*, void *pAux,
               int argc, char **argv,
               sqlite3_vtab **ppVTab,
               char **pzErr);
  int (*xConnect)(sqlite3*, void *pAux,
               int argc, char **argv,
               sqlite3_vtab **ppVTab,
               char **pzErr);
  int (*xBestIndex)(sqlite3_vtab *pVTab, sqlite3_index_info*);
  int (*xDisconnect)(sqlite3_vtab *pVTab);
  int (*xDestroy)(sqlite3_vtab *pVTab);
  int (*xOpen)(sqlite3_vtab *pVTab, sqlite3_vtab_cursor **ppCursor);
  int (*xClose)(sqlite3_vtab_cursor*);
  int (*xFilter)(sqlite3_vtab_cursor*, int idxNum, const char *idxStr,
                int argc, sqlite3_value **argv);
  int (*xNext)(sqlite3_vtab_cursor*);
  int (*xEof)(sqlite3_vtab_cursor*);
  int (*xColumn)(sqlite3_vtab_cursor*, sqlite3_context*, int);
  int (*xRowid)(sqlite3_vtab_cursor*, sqlite_int64 *pRowid);
  int (*xUpdate)(sqlite3_vtab *, int, sqlite3_value **, sqlite_int64 *);
  int (*xBegin)(sqlite3_vtab *pVTab);
  int (*xSync)(sqlite3_vtab *pVTab);
  int (*xCommit)(sqlite3_vtab *pVTab);
  int (*xRollback)(sqlite3_vtab *pVTab);
  int (*xFindFunction)(sqlite3_vtab *pVtab, int nArg, const char *zName,
                     void (**pxFunc)(sqlite3_context*,int,sqlite3_value**),
                     void **ppArg);
  int (*Rename)(sqlite3_vtab *pVtab, const char *zNew);
  /* The methods above are in version 1 of the sqlite_module object. Those 
  ** below are for version 2 and greater. */
  int (*xSavepoint)(sqlite3_vtab *pVTab, int);
  int (*xRelease)(sqlite3_vtab *pVTab, int);
  int (*xRollbackTo)(sqlite3_vtab *pVTab, int);
};

The module structure defines all of the methods for each virtual table object. The module structure also contains the iVersion field which defines the particular edition of the module table structure. Currently, iVersion is always 1, but in future releases of SQLite the module structure definition might be extended with additional methods and in that case the iVersion value will be increased.

The rest of the module structure consists of methods used to implement various features of the virtual table. Details on what each of these methods do are provided in the sequel.

1.3. Virtual Tables And Shared Cache

Prior to SQLite version 3.6.17 (2009-08-10), the virtual table mechanism assumes that each database connection kept its own copy of the database schema. Hence, the virtual table mechanism could not be used in a database that has shared cache mode enabled. The sqlite3_create_module() interface would return an error if shared cache mode is enabled. That restriction was relaxed beginning with SQLite version 3.6.17.

1.4. Creating New Virtual Table Implementations

Follow these steps to create your own virtual table:

  1. Write all necessary methods.
  2. Create an instance of the sqlite3_module structure containing pointers to all the methods from step 1.
  3. Register your sqlite3_module structure using one of the sqlite3_create_module() or sqlite3_create_module_v2() interfaces.
  4. Run a CREATE VIRTUAL TABLE command that specifies the new module in the USING clause.

The only really hard part is step 1. You might want to start with an existing virtual table implementation and modify it to suit your needs. There are several virtual table implementations in the SQLite source tree (for testing purposes). You might use one of those as a guide. Locate these test virtual table implementations by searching for "sqlite3_create_module".

You might also want to implement your new virtual table as a loadable extension.

2. Virtual Table Methods

2.1. The xCreate Method

int (*xCreate)(sqlite3 *db, void *pAux,
             int argc, char **argv,
             sqlite3_vtab **ppVTab,
             char **pzErr);

The xCreate method is called to create a new instance of a virtual table in response to a CREATE VIRTUAL TABLE statement. If the xCreate method is the same pointer as the xConnect method, then the virtual table is an eponymous virtual table. If the xCreate method is omitted (if it is a NULL pointer) then the virtual table is an eponymous-only virtual table.

The db parameter is a pointer to the SQLite database connection that is executing the CREATE VIRTUAL TABLE statement. The pAux argument is the copy of the client data pointer that was the fourth argument to the sqlite3_create_module() or sqlite3_create_module_v2() call that registered the virtual table module. The argv parameter is an array of argc pointers to null terminated strings. The first string, argv[0], is the name of the module being invoked. The module name is the name provided as the second argument to sqlite3_create_module() and as the argument to the USING clause of the CREATE VIRTUAL TABLE statement that is running. The second, argv[1], is the name of the database in which the new virtual table is being created. The database name is "main" for the primary database, or "temp" for TEMP database, or the name given at the end of the ATTACH statement for attached databases. The third element of the array, argv[2], is the name of the new virtual table, as specified following the TABLE keyword in the CREATE VIRTUAL TABLE statement. If present, the fourth and subsequent strings in the argv[] array report the arguments to the module name in the CREATE VIRTUAL TABLE statement.

The job of this method is to construct the new virtual table object (an sqlite3_vtab object) and return a pointer to it in *ppVTab.

As part of the task of creating a new sqlite3_vtab structure, this method must invoke sqlite3_declare_vtab() to tell the SQLite core about the columns and datatypes in the virtual table. The sqlite3_declare_vtab() API has the following prototype:

int sqlite3_declare_vtab(sqlite3 *db, const char *zCreateTable)

The first argument to sqlite3_declare_vtab() must be the same database connection pointer as the first parameter to this method. The second argument to sqlite3_declare_vtab() must a zero-terminated UTF-8 string that contains a well-formed CREATE TABLE statement that defines the columns in the virtual table and their data types. The name of the table in this CREATE TABLE statement is ignored, as are all constraints. Only the column names and datatypes matter. The CREATE TABLE statement string need not to be held in persistent memory. The string can be deallocated and/or reused as soon as the sqlite3_declare_vtab() routine returns.

The xCreate method need not initialize the pModule, nRef, and zErrMsg fields of the sqlite3_vtab object. The SQLite core will take care of that chore.

The xCreate should return SQLITE_OK if it is successful in creating the new virtual table, or SQLITE_ERROR if it is not successful. If not successful, the sqlite3_vtab structure must not be allocated. An error message may optionally be returned in *pzErr if unsuccessful. Space to hold the error message string must be allocated using an SQLite memory allocation function like sqlite3_malloc() or sqlite3_mprintf() as the SQLite core will attempt to free the space using sqlite3_free() after the error has been reported up to the application.

If the xCreate method is omitted (left as a NULL pointer) then the virtual table is an eponymous-only virtual table. New instances of the virtual table cannot be created using CREATE VIRTUAL TABLE and the virtual table can only be used via its module name. Note that SQLite versions prior to 3.9.0 do not understand eponymous-only virtual tables and will segfault if an attempt is made to CREATE VIRTUAL TABLE on an eponymous-only virtual table because the xCreate method was not checked for null.

If the xCreate method is the exact same pointer as the xConnect method, that indicates that the virtual table does not need to initialize backing store. Such a virtual table can be used as an eponymous virtual table or as a named virtual table using CREATE VIRTUAL TABLE or both.

2.1.1. Hidden columns in virtual tables

If a column datatype contains the special keyword "HIDDEN" (in any combination of upper and lower case letters) then that keyword it is omitted from the column datatype name and the column is marked as a hidden column internally. A hidden column differs from a normal column in three respects:

For example, if the following SQL is passed to sqlite3_declare_vtab():

CREATE TABLE x(a HIDDEN VARCHAR(12), b INTEGER, c INTEGER Hidden);

Then the virtual table would be created with two hidden columns, and with datatypes of "VARCHAR(12)" and "INTEGER".

An example use of hidden columns can be seen in the FTS3 virtual table implementation, where every FTS virtual table contains an FTS hidden column that is used to pass information from the virtual table into FTS auxiliary functions and to the FTS MATCH operator.

2.1.2. Table-valued functions

A virtual table that contains hidden columns can be used like a table-valued function in the FROM clause of a SELECT statement. The arguments to the table-valued function become constraints on the HIDDEN columns of the virtual table.

For example, the "generate_series" extension (located in the ext/misc/series.c file in the source tree) implements an eponymous virtual table with the following schema:

CREATE TABLE generate_series(
  value,
  start HIDDEN,
  stop HIDDEN,
  step HIDDEN
);

The sqlite3_module.xBestIndex method in the implementation of this table checks for equality constraints against the HIDDEN columns, and uses those as input parameters to determine the range of integer "value" outputs to generate. Reasonable defaults are used for any unconstrained columns. For example, to list all integers between 5 and 50:

SELECT value FROM generate_series(5,50);

The previous query is equivalent to the following:

SELECT value FROM generate_series WHERE start=5 AND stop=50;

Arguments on the virtual table name are matched to hidden columns in order. The number of arguments can be less than the number of hidden columns, in which case the latter hidden columns are unconstrained. However, an error results if there are more arguments than there are hidden columns in the virtual table.

2.1.3. WITHOUT ROWID Virtual Tables

Beginning with SQLite version 3.14.0 (2016-08-08), the CREATE TABLE statement that is passed into sqlite3_declare_vtab() may contain a WITHOUT ROWID clause. This is useful for cases where the virtual table rows cannot easily be mapped into unique integers. A CREATE TABLE statement that includes WITHOUT ROWID must define one or more columns as the PRIMARY KEY. Every column of the PRIMARY KEY must individually be NOT NULL and all columns for each row must be collectively unique.

Note that SQLite does not enforce the PRIMARY KEY for a WITHOUT ROWID virtual table. Enforcement is the responsibility of the underlying virtual table implementation. But SQLite does assume that the PRIMARY KEY constraint is valid - that the identified columns really are UNIQUE and NOT NULL - and it uses that assumption to optimize queries against the virtual table.

The rowid column is not accessible on a WITHOUT ROWID virtual table (of course). Furthermore, since the xUpdate method depends on having a valid rowid, the xUpdate method must be NULL for a WITHOUT ROWID virtual table. That in turn means that WITHOUT ROWID virtual tables must be read-only.

2.2. The xConnect Method

int (*xConnect)(sqlite3*, void *pAux,
             int argc, char **argv,
             sqlite3_vtab **ppVTab,
             char **pzErr);

The xConnect method is very similar to xCreate. It has the same parameters and constructs a new sqlite3_vtab structure just like xCreate. And it must also call sqlite3_declare_vtab() like xCreate.

The difference is that xConnect is called to establish a new connection to an existing virtual table whereas xCreate is called to create a new virtual table from scratch.

The xCreate and xConnect methods are only different when the virtual table has some kind of backing store that must be initialized the first time the virtual table is created. The xCreate method creates and initializes the backing store. The xConnect method just connects to an existing backing store. When xCreate and xConnect are the same, the table is an eponymous virtual table.

As an example, consider a virtual table implementation that provides read-only access to existing comma-separated-value (CSV) files on disk. There is no backing store that needs to be created or initialized for such a virtual table (since the CSV files already exist on disk) so the xCreate and xConnect methods will be identical for that module.

Another example is a virtual table that implements a full-text index. The xCreate method must create and initialize data structures to hold the dictionary and posting lists for that index. The xConnect method, on the other hand, only has to locate and use an existing dictionary and posting lists that were created by a prior xCreate call.

The xConnect method must return SQLITE_OK if it is successful in creating the new virtual table, or SQLITE_ERROR if it is not successful. If not successful, the sqlite3_vtab structure must not be allocated. An error message may optionally be returned in *pzErr if unsuccessful. Space to hold the error message string must be allocated using an SQLite memory allocation function like sqlite3_malloc() or sqlite3_mprintf() as the SQLite core will attempt to free the space using sqlite3_free() after the error has been reported up to the application.

The xConnect method is required for every virtual table implementation, though the xCreate and xConnect pointers of the sqlite3_module object may point to the same function if the virtual table does not need to initialize backing store.

2.3. The xBestIndex Method

SQLite uses the xBestIndex method of a virtual table module to determine the best way to access the virtual table. The xBestIndex method has a prototype like this:

int (*xBestIndex)(sqlite3_vtab *pVTab, sqlite3_index_info*);

The SQLite core communicates with the xBestIndex method by filling in certain fields of the sqlite3_index_info structure and passing a pointer to that structure into xBestIndex as the second parameter. The xBestIndex method fills out other fields of this structure which forms the reply. The sqlite3_index_info structure looks like this:

struct sqlite3_index_info {
  /* Inputs */
  const int nConstraint;     /* Number of entries in aConstraint */
  const struct sqlite3_index_constraint {
     int iColumn;              /* Column constrained.  -1 for ROWID */
     unsigned char op;         /* Constraint operator */
     unsigned char usable;     /* True if this constraint is usable */
     int iTermOffset;          /* Used internally - xBestIndex should ignore */
  } *const aConstraint;      /* Table of WHERE clause constraints */
  const int nOrderBy;        /* Number of terms in the ORDER BY clause */
  const struct sqlite3_index_orderby {
     int iColumn;              /* Column number */
     unsigned char desc;       /* True for DESC.  False for ASC. */
  } *const aOrderBy;         /* The ORDER BY clause */

  /* Outputs */
  struct sqlite3_index_constraint_usage {
    int argvIndex;           /* if >0, constraint is part of argv to xFilter */
    unsigned char omit;      /* Do not code a test for this constraint */
  } *const aConstraintUsage;
  int idxNum;                /* Number used to identify the index */
  char *idxStr;              /* String, possibly obtained from sqlite3_malloc */
  int needToFreeIdxStr;      /* Free idxStr using sqlite3_free() if true */
  int orderByConsumed;       /* True if output is already ordered */
  double estimatedCost;      /* Estimated cost of using this index */
  /* Fields below are only available in SQLite 3.8.2 and later */
  sqlite3_int64 estimatedRows;    /* Estimated number of rows returned */
  /* Fields below are only available in SQLite 3.9.0 and later */
  int idxFlags;              /* Mask of SQLITE_INDEX_SCAN_* flags */
  /* Fields below are only available in SQLite 3.10.0 and later */
  sqlite3_uint64 colUsed;    /* Input: Mask of columns used by statement */
};

Note the warnings on the "estimatedRows", "idxFlags", and colUsed fields. These fields were added with SQLite versions 3.8.2, 3.9.0, and 3.10.0, respectively. Any extension that reads or writes these fields must first check that the version of the SQLite library in use is greater than or equal to appropriate version - perhaps comparing the value returned from sqlite3_libversion_number() against constants 3008002, 3009000, and/or 3010000. The result of attempting to access these fields in an sqlite3_index_info structure created by an older version of SQLite are undefined.

In addition, there are some defined constants:

#define SQLITE_INDEX_CONSTRAINT_EQ      2
#define SQLITE_INDEX_CONSTRAINT_GT      4
#define SQLITE_INDEX_CONSTRAINT_LE      8
#define SQLITE_INDEX_CONSTRAINT_LT     16
#define SQLITE_INDEX_CONSTRAINT_GE     32
#define SQLITE_INDEX_CONSTRAINT_MATCH  64
#define SQLITE_INDEX_CONSTRAINT_LIKE   65     /* 3.10.0 and later only */
#define SQLITE_INDEX_CONSTRAINT_GLOB   66     /* 3.10.0 and later only */
#define SQLITE_INDEX_CONSTRAINT_REGEXP 67     /* 3.10.0 and later only */
#define SQLITE_INDEX_SCAN_UNIQUE        1     /* Scan visits at most 1 row */

The SQLite core calls the xBestIndex method when it is compiling a query that involves a virtual table. In other words, SQLite calls this method when it is running sqlite3_prepare() or the equivalent. By calling this method, the SQLite core is saying to the virtual table that it needs to access some subset of the rows in the virtual table and it wants to know the most efficient way to do that access. The xBestIndex method replies with information that the SQLite core can then use to conduct an efficient search of the virtual table.

While compiling a single SQL query, the SQLite core might call xBestIndex multiple times with different settings in sqlite3_index_info. The SQLite core will then select the combination that appears to give the best performance.

Before calling this method, the SQLite core initializes an instance of the sqlite3_index_info structure with information about the query that it is currently trying to process. This information derives mainly from the WHERE clause and ORDER BY or GROUP BY clauses of the query, but also from any ON or USING clauses if the query is a join. The information that the SQLite core provides to the xBestIndex method is held in the part of the structure that is marked as "Inputs". The "Outputs" section is initialized to zero.

The information in the sqlite3_index_info structure is ephemeral and may be overwritten or deallocated as soon as the xBestIndex method returns. If the xBestIndex method needs to remember any part of the sqlite3_index_info structure, it should make a copy. Care must be take to store the copy in a place where it will be deallocated, such as in the idxStr field with needToFreeIdxStr set to 1.

Note that xBestIndex will always be called before xFilter, since the idxNum and idxStr outputs from xBestIndex are required inputs to xFilter. However, there is no guarantee that xFilter will be called following a successful xBestIndex.

The xBestIndex method is required for every virtual table implementation.

2.3.1. Inputs

The main thing that the SQLite core is trying to communicate to the virtual table is the constraints that are available to limit the number of rows that need to be searched. The aConstraint[] array contains one entry for each constraint. There will be exactly nConstraint entries in that array.

Each constraint will correspond to a term in the WHERE clause or in a USING or ON clause that is of the form

column OP EXPR

Where "column" is a column in the virtual table, OP is an operator like "=" or "<", and EXPR is an arbitrary expression. So, for example, if the WHERE clause contained a term like this:

a = 5

Then one of the constraints would be on the "a" column with operator "=" and an expression of "5". Constraints need not have a literal representation of the WHERE clause. The query optimizer might make transformations to the WHERE clause in order to extract as many constraints as it can. So, for example, if the WHERE clause contained something like this:

x BETWEEN 10 AND 100 AND 999>y

The query optimizer might translate this into three separate constraints:

x >= 10
x <= 100
y < 999

For each constraint, the aConstraint[].iColumn field indicates which column appears on the left-hand side of the constraint. The first column of the virtual table is column 0. The rowid of the virtual table is column -1. The aConstraint[].op field indicates which operator is used. The SQLITE_INDEX_CONSTRAINT_* constants map integer constants into operator values. Columns occur in the order they were defined by the call to sqlite3_declare_vtab() in the xCreate or xConnect method. Hidden columns are counted when determining the column index.

The aConstraint[] array contains information about all constraints that apply to the virtual table. But some of the constraints might not be usable because of the way tables are ordered in a join. The xBestIndex method must therefore only consider constraints that have an aConstraint[].usable flag which is true.

In addition to WHERE clause constraints, the SQLite core also tells the xBestIndex method about the ORDER BY clause. (In an aggregate query, the SQLite core might put in GROUP BY clause information in place of the ORDER BY clause information, but this fact should not make any difference to the xBestIndex method.) If all terms of the ORDER BY clause are columns in the virtual table, then nOrderBy will be the number of terms in the ORDER BY clause and the aOrderBy[] array will identify the column for each term in the order by clause and whether or not that column is ASC or DESC.

In SQLite version 3.10.0 (2016-01-06) and later, the colUsed field is available to indicate which fields of the virtual table are actually used by the statement being prepared. If the lowest bit of colUsed is set, that means that the first column is used. The second lowest bit corresponds to the second column. And so forth. If the most significant bit of colUsed is set, that means that one or more columns other than the first 63 columns are used. If column usage information is needed by the xFilter method, then the required bits must be encoded into either the idxNum or idxStr output fields.

2.3.2. Outputs

Given all of the information above, the job of the xBestIndex method it to figure out the best way to search the virtual table.

The xBestIndex method fills the idxNum and idxStr fields with information that communicates an indexing strategy to the xFilter method. The information in idxNum and idxStr is arbitrary as far as the SQLite core is concerned. The SQLite core just copies the information through to the xFilter method. Any desired meaning can be assigned to idxNum and idxStr as long as xBestIndex and xFilter agree on what that meaning is.

The idxStr value may be a string obtained from an SQLite memory allocation function such as sqlite3_mprintf(). If this is the case, then the needToFreeIdxStr flag must be set to true so that the SQLite core will know to call sqlite3_free() on that string when it has finished with it, and thus avoid a memory leak.

If the virtual table will output rows in the order specified by the ORDER BY clause, then the orderByConsumed flag may be set to true. If the output is not automatically in the correct order then orderByConsumed must be left in its default false setting. This will indicate to the SQLite core that it will need to do a separate sorting pass over the data after it comes out of the virtual table.

The estimatedCost field should be set to the estimated number of disk access operations required to execute this query against the virtual table. The SQLite core will often call xBestIndex multiple times with different constraints, obtain multiple cost estimates, then choose the query plan that gives the lowest estimate.

If the current version of SQLite is 3.8.2 or greater, the estimatedRows field may be set to an estimate of the number of rows returned by the proposed query plan. If this value is not explicitly set, the default estimate of 25 rows is used.

If the current version of SQLite is 3.9.0 or greater, the idxFlags field may be set to SQLITE_INDEX_SCAN_UNIQUE to indicate that the virtual table will return only zero or one rows given the input constraints. Additional bits of the idxFlags field might be understood in later versions of SQLite.

The aConstraintUsage[] array contains one element for each of the nConstraint constraints in the inputs section of the sqlite3_index_info structure. The aConstraintUsage[] array is used by xBestIndex to tell the core how it is using the constraints.

The xBestIndex method may set aConstraintUsage[].argvIndex entries to values greater than zero. Exactly one entry should be set to 1, another to 2, another to 3, and so forth up to as many or as few as the xBestIndex method wants. The EXPR of the corresponding constraints will then be passed in as the argv[] parameters to xFilter.

For example, if the aConstraint[3].argvIndex is set to 1, then when xFilter is called, the argv[0] passed to xFilter will have the EXPR value of the aConstraint[3] constraint.

By default, the SQLite core double checks all constraints on each row of the virtual table that it receives. If such a check is redundant, the xBestFilter method can suppress that double-check by setting aConstraintUsage[].omit.

2.4. The xDisconnect Method

int (*xDisconnect)(sqlite3_vtab *pVTab);

This method releases a connection to a virtual table. Only the sqlite3_vtab object is destroyed. The virtual table is not destroyed and any backing store associated with the virtual table persists. This method undoes the work of xConnect.

This method is a destructor for a connection to the virtual table. Contrast this method with xDestroy. The xDestroy is a destructor for the entire virtual table.

The xDisconnect method is required for every virtual table implementation, though it is acceptable for the xDisconnect and xDestroy methods to be the same function if that makes sense for the particular virtual table.

2.5. The xDestroy Method

int (*xDestroy)(sqlite3_vtab *pVTab);

This method releases a connection to a virtual table, just like the xDisconnect method, and it also destroys the underlying table implementation. This method undoes the work of xCreate.

The xDisconnect method is called whenever a database connection that uses a virtual table is closed. The xDestroy method is only called when a DROP TABLE statement is executed against the virtual table.

The xDestroy method is required for every virtual table implementation, though it is acceptable for the xDisconnect and xDestroy methods to be the same function if that makes sense for the particular virtual table.

2.6. The xOpen Method

int (*xOpen)(sqlite3_vtab *pVTab, sqlite3_vtab_cursor **ppCursor);

The xOpen method creates a new cursor used for accessing (read and/or writing) a virtual table. A successful invocation of this method will allocate the memory for the sqlite3_vtab_cursor (or a subclass), initialize the new object, and make *ppCursor point to the new object. The successful call then returns SQLITE_OK.

For every successful call to this method, the SQLite core will later invoke the xClose method to destroy the allocated cursor.

The xOpen method need not initialize the pVtab field of the sqlite3_vtab_cursor structure. The SQLite core will take care of that chore automatically.

A virtual table implementation must be able to support an arbitrary number of simultaneously open cursors.

When initially opened, the cursor is in an undefined state. The SQLite core will invoke the xFilter method on the cursor prior to any attempt to position or read from the cursor.

The xOpen method is required for every virtual table implementation.

2.7. The xClose Method

int (*xClose)(sqlite3_vtab_cursor*);

The xClose method closes a cursor previously opened by xOpen. The SQLite core will always call xClose once for each cursor opened using xOpen.

This method must release all resources allocated by the corresponding xOpen call. The routine will not be called again even if it returns an error. The SQLite core will not use the sqlite3_vtab_cursor again after it has been closed.

The xClose method is required for every virtual table implementation.

2.8. The xEof Method

int (*xEof)(sqlite3_vtab_cursor*);

The xEof method must return false (zero) if the specified cursor currently points to a valid row of data, or true (non-zero) otherwise. This method is called by the SQL engine immediately after each xFilter and xNext invocation.

The xEof method is required for every virtual table implementation.

2.9. The xFilter Method

int (*xFilter)(sqlite3_vtab_cursor*, int idxNum, const char *idxStr,
              int argc, sqlite3_value **argv);

This method begins a search of a virtual table. The first argument is a cursor opened by xOpen. The next two arguments define a particular search index previously chosen by xBestIndex. The specific meanings of idxNum and idxStr are unimportant as long as xFilter and xBestIndex agree on what that meaning is.

The xBestIndex function may have requested the values of certain expressions using the aConstraintUsage[].argvIndex values of the sqlite3_index_info structure. Those values are passed to xFilter using the argc and argv parameters.

If the virtual table contains one or more rows that match the search criteria, then the cursor must be left point at the first row. Subsequent calls to xEof must return false (zero). If there are no rows match, then the cursor must be left in a state that will cause the xEof to return true (non-zero). The SQLite engine will use the xColumn and xRowid methods to access that row content. The xNext method will be used to advance to the next row.

This method must return SQLITE_OK if successful, or an sqlite error code if an error occurs.

The xFilter method is required for every virtual table implementation.

2.10. The xNext Method

int (*xNext)(sqlite3_vtab_cursor*);

The xNext method advances a virtual table cursor to the next row of a result set initiated by xFilter. If the cursor is already pointing at the last row when this routine is called, then the cursor no longer points to valid data and a subsequent call to the xEof method must return true (non-zero). If the cursor is successfully advanced to another row of content, then subsequent calls to xEof must return false (zero).

This method must return SQLITE_OK if successful, or an sqlite error code if an error occurs.

The xNext method is required for every virtual table implementation.

2.11. The xColumn Method

int (*xColumn)(sqlite3_vtab_cursor*, sqlite3_context*, int N);

The SQLite core invokes this method in order to find the value for the N-th column of the current row. N is zero-based so the first column is numbered 0. The xColumn method may return its result back to SQLite using one of the following interface:

If the xColumn method implementation calls none of the functions above, then the value of the column defaults to an SQL NULL.

To raise an error, the xColumn method should use one of the result_text() methods to set the error message text, then return an appropriate error code. The xColumn method must return SQLITE_OK on success.

The xColumn method is required for every virtual table implementation.

2.12. The xRowid Method

int (*xRowid)(sqlite3_vtab_cursor *pCur, sqlite_int64 *pRowid);

A successful invocation of this method will cause *pRowid to be filled with the rowid of row that the virtual table cursor pCur is currently pointing at. This method returns SQLITE_OK on success. It returns an appropriate error code on failure.

The xRowid method is required for every virtual table implementation.

2.13. The xUpdate Method

int (*xUpdate)(
  sqlite3_vtab *pVTab,
  int argc,
  sqlite3_value **argv,
  sqlite_int64 *pRowid
);

All changes to a virtual table are made using the xUpdate method. This one method can be used to insert, delete, or update.

The argc parameter specifies the number of entries in the argv array. The value of argc will be 1 for a pure delete operation or N+2 for an insert or replace or update where N is the number of columns in the table. In the previous sentence, N includes any hidden columns.

Every argv entry will have a non-NULL value in C but may contain the SQL value NULL. In other words, it is always true that argv[i]!=0 for i between 0 and argc-1. However, it might be the case that sqlite3_value_type(argv[i])==SQLITE_NULL.

The argv[0] parameter is the rowid of a row in the virtual table to be deleted. If argv[0] is an SQL NULL, then no deletion occurs.

The argv[1] parameter is the rowid of a new row to be inserted into the virtual table. If argv[1] is an SQL NULL, then the implementation must choose a rowid for the newly inserted row. Subsequent argv[] entries contain values of the columns of the virtual table, in the order that the columns were declared. The number of columns will match the table declaration that the xConnect or xCreate method made using the sqlite3_declare_vtab() call. All hidden columns are included.

When doing an insert without a rowid (argc>1, argv[1] is an SQL NULL), the implementation must set *pRowid to the rowid of the newly inserted row; this will become the value returned by the sqlite3_last_insert_rowid() function. Setting this value in all the other cases is a harmless no-op; the SQLite engine ignores the *pRowid return value if argc==1 or argv[1] is not an SQL NULL.

Each call to xUpdate will fall into one of cases shown below. Not that references to argv[i] mean the SQL value held within the argv[i] object, not the argv[i] object itself.

argc = 1

The single row with rowid equal to argv[0] is deleted. No insert occurs.

argc > 1
argv[0] = NULL

A new row is inserted with a rowid argv[1] and column values in argv[2] and following. If argv[1] is an SQL NULL, the a new unique rowid is generated automatically.

argc > 1
argv[0] ≠ NULL
argv[0] = argv[1]

The row with rowid argv[0] is updated with new values in argv[2] and following parameters.

argc > 1
argv[0] ≠ NULL
argv[0] ≠ argv[1]

The row with rowid argv[0] is updated with rowid argv[1] and new values in argv[2] and following parameters. This will occur when an SQL statement updates a rowid, as in the statement:

UPDATE table SET rowid=rowid+1 WHERE ...;

The xUpdate method must return SQLITE_OK if and only if it is successful. If a failure occurs, the xUpdate must return an appropriate error code. On a failure, the pVTab->zErrMsg element may optionally be replaced with error message text stored in memory allocated from SQLite using functions such as sqlite3_mprintf() or sqlite3_malloc().

If the xUpdate method violates some constraint of the virtual table (including, but not limited to, attempting to store a value of the wrong datatype, attempting to store a value that is too large or too small, or attempting to change a read-only value) then the xUpdate must fail with an appropriate error code.

There might be one or more sqlite3_vtab_cursor objects open and in use on the virtual table instance and perhaps even on the row of the virtual table when the xUpdate method is invoked. The implementation of xUpdate must be prepared for attempts to delete or modify rows of the table out from other existing cursors. If the virtual table cannot accommodate such changes, the xUpdate method must return an error code.

The xUpdate method is optional. If the xUpdate pointer in the sqlite3_module for a virtual table is a NULL pointer, then the virtual table is read-only.

2.14. The xFindFunction Method

int (*xFindFunction)(
  sqlite3_vtab *pVtab,
  int nArg,
  const char *zName,
  void (**pxFunc)(sqlite3_context*,int,sqlite3_value**),
  void **ppArg
);

This method is called during sqlite3_prepare() to give the virtual table implementation an opportunity to overload functions. This method may be set to NULL in which case no overloading occurs.

When a function uses a column from a virtual table as its first argument, this method is called to see if the virtual table would like to overload the function. The first three parameters are inputs: the virtual table, the number of arguments to the function, and the name of the function. If no overloading is desired, this method returns 0. To overload the function, this method writes the new function implementation into *pxFunc and writes user data into *ppArg and returns 1.

Note that infix functions (LIKE, GLOB, REGEXP, and MATCH) reverse the order of their arguments. So "like(A,B)" is equivalent to "B like A". For the form "B like A" the B term is considered the first argument to the function. But for "like(A,B)" the A term is considered the first argument.

The function pointer returned by this routine must be valid for the lifetime of the sqlite3_vtab object given in the first parameter.

2.15. The xBegin Method

int (*xBegin)(sqlite3_vtab *pVTab);

This method begins a transaction on a virtual table. This is method is optional. The xBegin pointer of sqlite3_module may be NULL.

This method is always followed by one call to either the xCommit or xRollback method. Virtual table transactions do not nest, so the xBegin method will not be invoked more than once on a single virtual table without an intervening call to either xCommit or xRollback. Multiple calls to other methods can and likely will occur in between the xBegin and the corresponding xCommit or xRollback.

2.16. The xSync Method

int (*xSync)(sqlite3_vtab *pVTab);

This method signals the start of a two-phase commit on a virtual table. This is method is optional. The xSync pointer of sqlite3_module may be NULL.

This method is only invoked after call to the xBegin method and prior to an xCommit or xRollback. In order to implement two-phase commit, the xSync method on all virtual tables is invoked prior to invoking the xCommit method on any virtual table. If any of the xSync methods fail, the entire transaction is rolled back.

2.17. The xCommit Method

int (*xCommit)(sqlite3_vtab *pVTab);

This method causes a virtual table transaction to commit. This is method is optional. The xCommit pointer of sqlite3_module may be NULL.

A call to this method always follows a prior call to xBegin and xSync.

2.18. The xRollback Method

int (*xRollback)(sqlite3_vtab *pVTab);

This method causes a virtual table transaction to rollback. This is method is optional. The xRollback pointer of sqlite3_module may be NULL.

A call to this method always follows a prior call to xBegin.

2.19. The xRename Method

int (*xRename)(sqlite3_vtab *pVtab, const char *zNew);

This method provides notification that the virtual table implementation that the virtual table will be given a new name. If this method returns SQLITE_OK then SQLite renames the table. If this method returns an error code then the renaming is prevented.

The xRename method is required for every virtual table implementation.

2.20. The xSavepoint, xRelease, and xRollbackTo Methods

int (*xSavepoint)(sqlite3_vtab *pVtab, int);
int (*xRelease)(sqlite3_vtab *pVtab, int);
int (*xRollbackTo)(sqlite3_vtab *pVtab, int);

These methods provide the virtual table implementation an opportunity to implement nested transactions. They are always optional and will only be called in SQLite version 3.7.7 (2011-06-23) and later.

When xSavepoint(X,N) is invoked, that is a signal to the virtual table X that it should save its current state as savepoint N. A subsequent call to xRollbackTo(X,R) means that the state of the virtual table should return to what it was when xSavepoint(X,R) was last called. The call to xRollbackTo(X,R) will invalidate all savepoints with N>R; none of the invalided savepoints will be rolled back or released without first being reinitialized by a call to xSavepoint(). A call to xRelease(X,M) invalidates all savepoints where N>=M.

None of the xSavepoint(), xRelease(), or xRollbackTo() methods will ever be called except in between calls to xBegin() and either xCommit() or xRollback().