Documentation Source Text

      <a href="http://www.sqlite.org/cvstrac/timeline">
      http://www.sqlite.org/cvstrac/timeline</a>.</p>
    }
    hd_close_aux
    hd_enable_main 1
  }
}











chng {2009 March 31 (3.6.12)} {
<li>Fixed a bug that caused database corruption when an [incremental_vacuum] is
    rolled back in an in-memory database.  Ticket #3761.
<li>Added the [sqlite3_unlock_notify()] interface.
<li>Added the [reverse_unordered_selects pragma].
<li>The default page size on windows is automatically adjusted to match the

      <a href="http://www.sqlite.org/cvstrac/timeline">
      http://www.sqlite.org/cvstrac/timeline</a>.</p>
    }
    hd_close_aux
    hd_enable_main 1
  }
}

chng {2009 April 13 (3.6.13)} {
<li>Fix a bug in [version 3.6.12] that causes a segfault when running
    a count(*) on the sqlite_master table of an empty database. Ticket #3774.
<li>Fix a bug in [version 3.6.12] that causes a segfault that when 
    inserting into a table using a DEFAULT value where there is a 
    function as part of the DEFAULT value expression.  Ticket #3791.
<li>Fix data structure alignment issues on Sparc.  Ticket #3777.
<li>Other minor bug fixes.
}

chng {2009 March 31 (3.6.12)} {
<li>Fixed a bug that caused database corruption when an [incremental_vacuum] is
    rolled back in an in-memory database.  Ticket #3761.
<li>Added the [sqlite3_unlock_notify()] interface.
<li>Added the [reverse_unordered_selects pragma].
<li>The default page size on windows is automatically adjusted to match the

</td>
<td width="20"></td><td bgcolor="#80a796" width="1"></td><td width="20"></td>
<td valign="top">
<h3>Current Status</h3>

<p><ul>
<li><a href="releaselog/3_6_12.html">Version 3.6.12</a>
of SQLite is recommended for all new development.  Upgrading from
prior versions of SQLite is suggested for all users.</li>

</ul></p>

<h3>Common Links</h3>

<p><ul>
<li> <a href="features.html">Features</a> </li>
<li> <a href="faq.html">Frequently Asked Questions</a> </li>

</td>
<td width="20"></td><td bgcolor="#80a796" width="1"></td><td width="20"></td>
<td valign="top">
<h3>Current Status</h3>

<p><ul>
<li><a href="releaselog/3_6_13.html">Version 3.6.13</a>
of SQLite is recommended for all new development.  Upgrading from
version 3.6.12 is optional.  Upgrading from versions 3.6.11 and
earlier is recommended.</li>
</ul></p>

<h3>Common Links</h3>

<p><ul>
<li> <a href="features.html">Features</a> </li>
<li> <a href="faq.html">Frequently Asked Questions</a> </li>

in SQLite.  However, in many cases you can use an
[INSTEAD OF trigger] on the view to accomplish 
the same thing.  Views are removed 
with the [DROP VIEW] command.</p>

<tcl>
##############################################################################
Section {CREATE VIRTUAL TABLE} {createvtab} {{CREATE VIRTUAL_TABLE}}

BubbleDiagram create-virtual-table-stmt 1
</tcl>

<p>A virtual table is an interface to an external storage or computation
engine that appears to be a table but does not actually store information
in the database file.</p>

<p>In general, you can do anything with a virtual table that can be done
with an ordinary table, except that you cannot create indices or triggers on a
virtual table.  Some virtual table implementations might impose additional
restrictions.  For example, many virtual tables are read-only.
Virtual tables cannot be used in
[sqlite3_enable_shared_cache | shared cache mode].</p>

<p>The &lt;module-name&gt; is the name of an object that implements
the virtual table.  The &lt;module-name&gt; must be registered with
the SQLite database connection using
sqlite3_create_module()
prior to issuing the CREATE VIRTUAL TABLE statement.
The module takes zero or more comma-separated arguments.
The arguments can be just about any text as long as it has balanced
parentheses.  The argument syntax is sufficiently general that the
arguments can be made to appear as column definitions in a traditional
[CREATE TABLE] statement.  
SQLite passes the module arguments directly

in SQLite.  However, in many cases you can use an
[INSTEAD OF trigger] on the view to accomplish 
the same thing.  Views are removed 
with the [DROP VIEW] command.</p>

<tcl>
##############################################################################
Section {CREATE VIRTUAL TABLE} {createvtab} {{CREATE VIRTUAL TABLE}}

BubbleDiagram create-virtual-table-stmt 1
</tcl>

<p>A [virtual table] is an interface to an external storage or computation
engine that appears to be a table but does not actually store information
in the database file.</p>

<p>In general, you can do anything with a [virtual table] that can be done
with an ordinary table, except that you cannot create indices or triggers on a
virtual table.  Some virtual table implementations might impose additional
restrictions.  For example, many virtual tables are read-only.
Virtual tables cannot be used in
[sqlite3_enable_shared_cache | shared cache mode].</p>

<p>The &lt;module-name&gt; is the name of an object that implements
the virtual table.  The &lt;module-name&gt; must be registered with
the SQLite database connection using
[sqlite3_create_module()] or [sqlite3_create_module_v2()]
prior to issuing the CREATE VIRTUAL TABLE statement.
The module takes zero or more comma-separated arguments.
The arguments can be just about any text as long as it has balanced
parentheses.  The argument syntax is sufficiently general that the
arguments can be made to appear as column definitions in a traditional
[CREATE TABLE] statement.  
SQLite passes the module arguments directly

  hd_puts "<h3>$date - $title</h3>"
  regsub -all "\n( *\n)+" $text "</p>\n\n<p>" txt
  regsub -all {[Tt]icket #(\d+)} $txt \
      {<a href="http://www.sqlite.org/cvstrac/tktview?tn=\1">\0</a>} txt
  hd_resolve "<p>$txt</p>"
  hd_puts "<hr width=\"50%\">"
}








newsitem {2009-Mar-31} {Version 3.6.12} {
  SQLite [version 3.6.12] fixes a database corruption bug.  If an
  [incremental_vacuum] is rolled back in an in-memory database, the
  database will often go corrupt.  This only happens for in-memory
  databases.  On-disk databases are unaffected.  And the corruption
  only appears if an incremental vacuum is rolled back.  Nevertheless,

  hd_puts "<h3>$date - $title</h3>"
  regsub -all "\n( *\n)+" $text "</p>\n\n<p>" txt
  regsub -all {[Tt]icket #(\d+)} $txt \
      {<a href="http://www.sqlite.org/cvstrac/tktview?tn=\1">\0</a>} txt
  hd_resolve "<p>$txt</p>"
  hd_puts "<hr width=\"50%\">"
}

newsitem {2009-Apr-14} {Version 3.6.13} {
  SQLite [version 3.6.13] fixes several minor issues that appeared
  in previous version, including Ticket #3774, #3791, and #3777.
  This is a bug-fix release only.   There are no new features or
  enhancements.  
}

newsitem {2009-Mar-31} {Version 3.6.12} {
  SQLite [version 3.6.12] fixes a database corruption bug.  If an
  [incremental_vacuum] is rolled back in an in-memory database, the
  database will often go corrupt.  This only happens for in-memory
  databases.  On-disk databases are unaffected.  And the corruption
  only appears if an incremental vacuum is rolled back.  Nevertheless,

<title>The Virtual Table Mechanism Of SQLite</title>

<h1 align="center">The Virtual Table Mechanism Of SQLite</h1>
<tcl>hd_keywords {virtual table} {virtual tables}</tcl>

<h2>1.0 Introduction</h2>

<p>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 from and updates to a virtual table
invoke callback methods on the virtual table object instead of
reading and writing to the database file.

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

<p>
<ul>
<li> One cannot create a trigger on a virtual table.
<li> One cannot create additional indices on a virtual table. 
     (Virtual tables can have indices but that must be built into
     the virtual table implementation.  Indices cannot be added
     separately using CREATE INDEX statements.)
<li> One cannot run ALTER TABLE commands against a virtual table.
<li> Virtual tables cannot be used in a database that makes use of
     the [shared cache] feature. 
</ul>

<p>Particular 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.

<p>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.

<p>Here are some postulated uses for virtual tables:

<ul>
<li> A full-text search interface
<li> Spatial indices using R-Trees
<li> Access to the filesystem of the host computer
<li> Enabling SQL manipulation of data in statistics packages like R
</ul>

<h3>1.1 Usage</h3>

<p>A virtual table is created using using a CREATE VIRTUAL TABLE statement.
This statement creates a table with a particular name and associates the
table with a "module".

<blockquote><pre>
   CREATE VIRTUAL TABLE tablename USING modulename;
</pre></blockquote>

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

<blockquote><pre>
   CREATE VIRTUAL TABLE tablename USING modulename(arg1, arg2, ...);
</pre></blockquote>

<p>The format of the arguments to the module is very general. Each argument 
can consist of keywords, string literals, identifiers, numbers, and 
punctuation. The arguments are passed as written (as text) into the
constructor metod 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.

<p>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.

<h2>Implementation</h2>

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

<blockquote><pre>
  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;
</pre></blockquote>

<p>The sqlite3_module structure defines a module object used to implement
a virtual table. Think of a module as a class from which you 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.

<p>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.

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

<blockquote><pre>
  struct sqlite3_vtab {
    const sqlite3_module *pModule;
    int nRef;
    char *zErrMsg;
  };
</pre></blockquote>

<p>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 can pass error message text to the core by putting 
an error message string obtained from [sqlite3_mprintf()] in zErrMsg.
Prior to assigning a new value to zErrMsg, the virtual table 
implementation should free any preexisting content of zErrMsg using 
[sqlite3_free()]. Failure to do this might 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.

<p>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:

<blockquote><pre>
  struct sqlite3_vtab_cursor {
    sqlite3_vtab *pVtab;
  };
</pre></blockquote>

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

<p>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.

<p>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:

<blockquote><pre>
  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 */
  );
</pre></blockquote>

<p>The sqlite3_create_module() 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:

<blockquote><pre>  
  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);
  };
</pre></blockquote>

<p>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.

<p>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.

<h3>1.2 Virtual Tables And Shared Cache</h3>

<p>The virtual table mechanism assumes that each database connection keeps 
its own copy of the database schema. Hence, the virtual table mechanism
cannot be used in a database that has [shared cache] enabled. 
The [sqlite3_create_module()] interface will return an error if 
the [shared cache] feature is enabled.

<h3>1.3 Creating New Virtual Table Implementations</h3>

<p>Follow these steps to create your own virtual table:

<p>
<ol>
<li> Write all necessary methods.
<li> Create an instance of the [sqlite3_module] structure containing pointers
     to all the methods from step 1.
<li> Register your [sqlite3_module] structure using the
     [sqlite3_create_module()] API.
<li> Run a [CREATE VIRTUAL TABLE] command that specifies your module in 
     the USING clause. 
</ol>

<p>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".

<p>You might also want to implement your new virtual table as a 
[sqlite3_load_extension | loadable extension].

<h2>2.0 Virtual Table Methods</h2>

<h3>2.1 The xCreate Method</h3>

<blockquote><pre>
  int (*xCreate)(sqlite3 *db, void *pAux,
               int argc, char **argv,
               sqlite3_vtab **ppVTab,
               char **pzErr);
</pre></blockquote>

<p>This method is called to create a new instance of a virtual table 
in response to a [CREATE VIRTUAL TABLE] statement. 
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 second, argv[1], is the name of the database in which the new virtual table is being created. The third element of the array, argv[2], 
is the name of the new virtual table. 
If present, the fourth and subsquent strings in the argv[] array report 
the arguments to the module name in the [CREATE VIRTUAL TABLE] statement.

<p>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.

<p>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:

<blockquote><pre>
    int sqlite3_declare_vtab(sqlite3 *db, const char *zCreateTable)
</pre></blockquote>

<p>The first argument to [sqlite3_declare_vtab()] is the pointer to the 
sqlite [database connection].  The first argument must be the same
database connection pointer that was passed into the xCreate method.
The second argument to [sqlite3_declare_vtab()] is 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.

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

<p>
<ul>
<li> Hidden columns are not listed in the dataset returned by 
     "[PRAGMA table_info]",
<li> Hidden columns are not included in the expansion of a "*"
     expression in the result set of a [SELECT], and
<li> Hidden columns are not included in the implicit column-list 
     used by an [INSERT] statement that lacks an explicit column-list. 
</ul>

<p>For example, if the following SQL is passed to [sqlite3_declare_vtab()]:

<blockquote><pre>
   CREATE TABLE x(a HIDDEN VARCHAR(12), b INTEGER, c INTEGER Hidden);
</pre></blockquote>

<p>Then the virtual table would be created with two hidden columns, declaration types "VARCHAR(12)" and "INTEGER".

<p>The xCreate method 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, no [sqlite3_vtab] structure should be allocated. 
An error message may optionally be returned in *pzErr if unsuccessful.
The text of the error message should be obtained from [sqlite3_mprintf()].

<h3>2.2 The xConnect Method</h3>

<blockquote><pre>
  int (*xConnect)(sqlite3*, void *pAux,
               int argc, char **argv,
               sqlite3_vtab **ppVTab,
               char **pzErr);
</pre></blockquote>

<p>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.

<p>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.

<p>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.

<p>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.

<p>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.

<p>The xConnect method 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, no [sqlite3_vtab] structure should be 
allocated. An error message may optionally be returned in *pzErr if 
unsuccessful. 
The text of the error message should be obtained from [sqlite3_mprintf()].

<h3>2.3 The xBestIndex Method</h3>

<p>The xBestIndex method has a prototype like this:

<blockquote><pre>
  int (*xBestIndex)(sqlite3_vtab *pVTab, sqlite3_index_info*);
</pre></blockquote>

<p>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:

<blockquote><pre>
  struct sqlite3_index_info {
    /* Inputs */
    const int nConstraint;     /* Number of entries in aConstraint */
    const struct sqlite3_index_constraint {
       int iColumn;              /* Column on left-hand side of constraint */
       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 */
  };
</pre></blockquote>

<p>In addition, there are some defined constants:

<blockquote><pre>
  #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
</pre></blockquote>

<p>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()]. 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.

<p>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.

<h4>2.3.1 Inputs</h4>

<p>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.

<p>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.

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

<blockquote>
     column  OP  EXPR
</blockquote>

<p>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:

<blockquote><pre>
     a = 5
</pre></blockquote>

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

<blockquote><pre>
     x BETWEEN 10 AND 100 AND 999>y
</pre></blockquote>

<p>The query optimizer would translate this into three separate constraints:

<blockquote><pre>
     x >= 10
     x <= 100
     y < 999
</pre></blockquote>

<p>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.

<p>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 should therefore only consider constraints 
that have a aConstraint[].usable flag which is true.

<p>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.

<h4>2.3.2 Outputs</h4>

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

<p>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.

<p>The idxStr value can be a string obtained from [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.

<p>If the virtual table will output rows in the order specified by 
the ORDER BY clause, then the orderByConsumed flag should be set to 
true. If the output is not automatically in the correct order 
then orderByConsumed should 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.

<p>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.

<p>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.

<p>The xBestIndex method may set aConstraintUsage[].argvIndex 
entries to values greater than one. 
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.

<p>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.

<p>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 check by 
setting aConstraintUsage[].omit.

<h3>2.4 The xDisconnect Method</h3>

<blockquote><pre>
  int (*xDisconnect)(sqlite3_vtab *pVTab);
</pre></blockquote>

<p>This method releases a connection to a virtual table. 
The virtual table is not destroyed and any backing store 
associated with the virtual table persists. 
This method is the opposite of xConnect.

<h3>2.5 The xDestroy Method</h3>

<blockquote><pre>
  int (*xDestroy)(sqlite3_vtab *pVTab);
</pre></blockquote>

<p>This method releases a connection to a virtual table, just like 
the xDisconnect method, and it also destroys the underlying 
table implementation. This method is the opposite of xCreate.

<p>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.

<h3>2.6 The xOpen Method</h3>

<blockquote><pre>
  int (*xOpen)(sqlite3_vtab *pVTab, sqlite3_vtab_cursor **ppCursor);
</pre></blockquote>

<p>The xOpen method creates a new cursor used for reading information
out of 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].

<p>The cursor is not immediately usable. It must first be positioned 
using xFilter. Then elements of a row of the virtual table can be 
accessed using xColumn and xRowid.

<p>The xClose method is used to close the cursor.

<h3>2.7 The xClose Method</h3>

<blockquote><pre>
  int (*xClose)(sqlite3_vtab_cursor*);
</pre></blockquote>

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

<p>This method must release all resources allocated by the
corresponding xOpen call. The routine will not be called again if it
returns an error.

<h3>2.8 The xEof Method</h3>

<blockquote><pre>
  int (*xEof)(sqlite3_vtab_cursor*);
</pre></blockquote>

<p>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.

<h3>2.9 The xFilter Method</h3>

<blockquote><pre>
  int (*xFilter)(sqlite3_vtab_cursor*, int idxNum, const char *idxStr,
                int argc, sqlite3_value **argv);
</pre></blockquote>

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

<p>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.

<p>If the filtering constraints configured by idxNum and idxStr do 
not match any rows of the virtual table (i.e. no rows of data will 
be returned), then a subsequent call to the xEof method of the same 
table should return non-zero. Otherwise, xEof should return zero 
and the cursor should be left pointing at a row of the virtual table. 
The xColumn and xRowid methods can be used to access that row. 
The xNext method can be used to advance to the next row.

<p>This method should return SQLITE_OK if successful, or an sqlite 
error code if an error occurs.

<h3>2.10 The xNext Method</h3>

<blockquote><pre>
  int (*xNext)(sqlite3_vtab_cursor*);
</pre></blockquote>

<p>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 should return non-zero. 
Otherwise, the xEof method returns zero.

<p>This method should return SQLITE_OK if successful, or an sqlite 
error code if an error occurs.

<h3>2.11 The xColumn Method</h3>

<blockquote><pre>
  int (*xColumn)(sqlite3_vtab_cursor*, sqlite3_context*, int N);
</pre></blockquote>

<p>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 uses one of the 
[sqlite3_result_blob | sqlite3_result_*() APIs] 
to return the result. 
This method can invoke [sqlite3_result_error()] to raise an exception, 
if desired.

<h3>2.12 The xRowid Method</h3>

<blockquote><pre>
  int (*xRowid)(sqlite3_vtab_cursor*, sqlite_int64 *pRowid);
</pre></blockquote>

<p>This method writes the 64-bit rowid of the current row into *pRowid.

<h3>2.13 The xUpdate Method</h3>

<blockquote><pre>
  int (*xUpdate)(
    sqlite3_vtab *pVTab,
    int argc,
    sqlite3_value **argv,
    sqlite_int64 *pRowid
  );
</pre></blockquote>

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

<p>The argc parameter specifies the number of entries in the argv array. 
Every argv entry will have a non-NULL value in C (but may contain the 
SQL value NULL).

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

<p>The argv[1] parameter is the rowid of a new row to be inserted 
into the virtual table. If argv[1] is 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.

<p>When doing an insert without a rowid (argc>1, argv[1]==NULL), the 
implementation must set *pRowid to the id 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 has no effect,
and cannot be used to change the rowid of an existing row or if you 
don't like what was in argv[1].

<p>Each call to xUpdate will fall into one of the following cases:

<blockquote>
<dl>
<dt><b>argc == 1</b>
<dd><p>The single row argv[0] is deleted; no insert occurs

<dt><b>argc > 1 && argv[0] == NULL</b>
<dd><p>A new row is inserted with rowid argv[1]

<dt><b>argc > 1 && argv[0] != NULL && argv[0] == argv[1]</b>
<dd><p>The row with rowid argv[0] is updated with new values 
       in argv[2] and following parameters

<dt><b>argc > 1 && argv[0] != NULL && argv[0] != argv[1]</b>
<dd><p> 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:
<blockquote>
   [UPDATE] table SET rowid=rowid+1 WHERE ...; 
</blockquote>
</dl>
</blockquote>

<p>It is currently undocumented what effect changes should have on 
existing cursors.

<h3>2.14 The xFindFunction Method</h3>

<blockquote><pre>
  int (*xFindFunction)(
    sqlite3_vtab *pVtab,
    int nArg,
    const char *zName,
    void (**pxFunc)(sqlite3_context*,int,sqlite3_value**),
    void **ppArg
  );
</pre></blockquote>

<p>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.

<p>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.

<p>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.

<p>It is currently undocumented how often this function is called or 
when you can free any resources allocated in order to respond - #2095

<h3>2.15 The xBegin Method</h3>

<blockquote><pre>
  int (*xBegin)(sqlite3_vtab *pVTab);
</pre></blockquote>

<i>TBD...</i>

<h3>2.16 The xSync Method</h3>

<blockquote><pre>
  int (*xSync)(sqlite3_vtab *pVTab);
</pre></blockquote>

<i>TBD...</i>

<h3>2.17 The xCommit Method</h3>

<blockquote><pre>
  int (*xCommit)(sqlite3_vtab *pVTab);
</pre></blockquote>

<i>TBD...</i>

<h3>2.18 The xRollback Method</h3>

<blockquote><pre>
  int (*xRollback)(sqlite3_vtab *pVTab);
</pre></blockquote>

<i>TBD...</i>

<h3>2.19 The xRename Method</h3>

<blockquote><pre>
  int (*xRename)(sqlite3_vtab *pVtab, const char *zNew);
</pre></blockquote>

<p>This method provides sotification 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 then the renaming is prevented.