*** DRAFT ***
The Session Extension

1. Introduction

The session extension provide a mechanism for recording changes to some or all of the rowid tables in an SQLite database, and packaging those changes into a "changeset" or "patchset" file that can later be used to apply the same set of changes to another database with the same schema and compatible starting data. A "changeset" may also be inverted and used to "undo" a session.

This document is an introduction to the session extension. The details of the interface are in the separate Session Extension C-language Interface document.

1.1. Typical Use Case

Suppose SQLite is used as the application file format for a particular design application. Two users, Alice and Bob, each start with a baseline design that is about a gigabyte in size. They work all day, in parallel, each making their own customizations and tweaks to the design. At the end of the day, they would like to merge their changes together into a single unified design.

The session extension facilitates this by recording all changes to both Alice's and Bob's databases and writing those changes into changeset or patchset files. At the end of the day, Alice can send her changeset to Bob and Bob can "apply" it to his database. The result (assuming there are no conflicts) is that Bob's database then contains both his changes and Alice's changes. Likewise, Bob can send a changeset of his work over to Alice and she can apply his changes to her database.

In other words, the session extension provides a facility for SQLite database files that is similar to the unix patch utility program, or to the "merge" capabilities of version control systems such as Fossil, Git, or Mercurial.

1.2. Obtaining the Session Extension

Since version 3.13.0 (2016-05-18), the session extension has been included in the SQLite amalgamation source distribution. By default, the session extension is disabled. To enable it, build with the following compiler switches:

-DSQLITE_ENABLE_SESSION -DSQLITE_ENABLE_PREUPDATE_HOOK

Or, if using the autoconf build system, pass the --enable-session option to the configure script.

1.3. Limitations

2. Concepts

2.1. Changesets and Patchsets

The sessions module revolves around creating and manipulating changesets. A changeset is a blob of data that encodes a series of changes to a database. Each change in a changeset is one of the following:

A single changeset may contain changes that apply to more than one database table. For each table that the changeset includes at least one change for, it also encodes the following data:

Changesets may only be applied to databases that contain tables matching the above three criteria as stored in the changeset.

A patchset is similar to a changeset. It is slightly more compact than a changeset, but provides more limited conflict detection and resolution options (see the next section for details). The differences between a patchset and a changeset are that:

2.2. Conflicts

When a changeset or patchset is applied to a database, an attempt is made to insert a new row for each INSERT change, remove a row for each DELETE change and modify a row for each UPDATE change. If the target database is in the same state as the original database that the changeset was recorded on, this is a simple matter. However, if the contents of the target database is not in exactly this state, conflicts can occur when applying the changeset or patchset.

When processing an INSERT change, the following conflicts can occur:

When processing a DELETE change, the following conflicts may be detected:

When processing an UPDATE change, the following conflicts may be detected:

Depending on the type of conflict, a sessions application has a variety of configurable options for dealing with conflicts, ranging from omitting the conflicting change, aborting the entire changeset application or applying the change despite the conflict. For details, refer to the documentation for the sqlite3changeset_apply() API.

2.3. Changeset Construction

After a session object has been configured, it begins monitoring for changes to its configured tables. However, it does not record an entire change each time a row within the database is modified. Instead, it records just the PRIMARY KEY fields for each inserted row, and just the PRIMARY KEY and all original row values for any updated or deleted rows. If a row is modified more than once by a single session, no new information is recorded.

The other information required to create a changeset or patchset is read from the database file when sqlite3session_changeset() or sqlite3session_patchset() is called. Specifically,

One implication of the above is that if a change is made and then unmade within a single session (for example if a row is inserted and then deleted again), the sessions module does not report any change at all. Or if a row is updated multiple times within the same session, all updates are coalesced into a single update within any changeset or patchset blob.

3. Using The Session Extension

This section provides examples that demonstrate how to use the sessions extension.

3.1. Capturing a Changeset

The example code below demonstrates the steps involved in capturing a changeset while executing SQL commands. In summary:

  1. A session object (type sqlite3_session*) is created by making a call to the sqlite3session_create() API function.

    A single session object monitors changes made to a single database (i.e. "main", "temp" or an attached database) via a single sqlite3* database handle.

  2. The session object is configured with a set of tables to monitor changes on.

    By default a session object does not monitor changes on any database table. Before it does so it must be configured. There are three ways to configure the set of tables to monitor changes on:

    • By explicitly specifying tables using one call to sqlite3session_attach() for each table, or
    • By specifying that all tables in the database should be monitored for changes using a call to sqlite3session_attach() with a NULL argument, or
    • By configuring a callback to be invoked the first time each table is written to that indicates to the session module whether or not changes on the table should be monitored.

    The example code below uses the second of the methods enumerated above - it monitors for changes on all database tables.

  3. Changes are made to the database by executing SQL statements. The session object records these changes.

  4. A changeset blob is extracted from the session object using a call to sqlite3session_changeset() (or, if using patchsets, a call to the sqlite3session_patchset() function).

  5. The session object is deleted using a call to the sqlite3session_delete() API function.

    It is not necessary to delete a session object after extracting a changeset or patchset from it. It can be left attached to the database handle and will continue monitoring for changes on the configured tables as before. However, if sqlite3session_changeset() or sqlite3session_patchset() is called a second time on a session object, the changeset or patchset will contain all changes that have taken place on the connection since the session was created. In other words, a session object is not reset or zeroed by a call to sqlite3session_changeset() or sqlite3session_patchset().

/*
** Argument zSql points to a buffer containing an SQL script to execute 
** against the database handle passed as the first argument. As well as
** executing the SQL script, this function collects a changeset recording
** all changes made to the "main" database file. Assuming no error occurs,
** output variables (*ppChangeset) and (*pnChangeset) are set to point
** to a buffer containing the changeset and the size of the changeset in
** bytes before returning SQLITE_OK. In this case it is the responsibility
** of the caller to eventually free the changeset blob by passing it to
** the sqlite3_free function.
**
** Or, if an error does occur, return an SQLite error code. The final
** value of (*pChangeset) and (*pnChangeset) are undefined in this case.
*/
int sql_exec_changeset(
  sqlite3 *db,                  /* Database handle */
  const char *zSql,             /* SQL script to execute */
  int *pnChangeset,             /* OUT: Size of changeset blob in bytes */
  void **ppChangeset            /* OUT: Pointer to changeset blob */
){
  sqlite3_session *pSession = 0;
  int rc;

  /* Create a new session object */
  rc = sqlite3session_create(db, "main", &pSession);

  /* Configure the session object to record changes to all tables */
  if( rc==SQLITE_OK ) rc = sqlite3session_attach(pSession, NULL);

  /* Execute the SQL script */
  if( rc==SQLITE_OK ) rc = sqlite3_exec(db, zSql, 0, 0, 0);

  /* Collect the changeset */
  if( rc==SQLITE_OK ){
    rc = sqlite3session_changeset(pSession, pnChangeset, ppChangeset);
  }

  /* Delete the session object */
  sqlite3session_delete(pSession);

  return rc;
}

3.2. Applying a Changeset to a Database

Applying a changeset to a database is simpler than capturing a changeset. Usually, a single call to sqlite3changeset_apply(), as depicted in the example code below, suffices.

In cases where it is complicated, the complications in applying a changeset lie in conflict resolution. Refer to the API documentation linked above for details.

/*
** Conflict handler callback used by apply_changeset(). See below.
*/
static int xConflict(void *pCtx, int eConflict, sqlite3_changset_iter *pIter){
  int ret = (int)pCtx;
  return ret;
}

/*
** Apply the changeset contained in blob pChangeset, size nChangeset bytes,
** to the main database of the database handle passed as the first argument.
** Return SQLITE_OK if successful, or an SQLite error code if an error
** occurs.
**
** If parameter bIgnoreConflicts is true, then any conflicting changes 
** within the changeset are simply ignored. Or, if bIgnoreConflicts is
** false, then this call fails with an SQLTIE_ABORT error if a changeset
** conflict is encountered.
*/
int apply_changeset(
  sqlite3 *db,                  /* Database handle */
  int bIgnoreConflicts,         /* True to ignore conflicting changes */
  int nChangeset,               /* Size of changeset in bytes */
  void *pChangeset              /* Pointer to changeset blob */
){
  return sqlite3changeset_apply(
      db, 
      nChangeset, pChangeset, 
      0, xConflict, 
      (void*)bIgnoreConflicts
  );
}

3.3. Inspecting the Contents of a Changeset

The example code below demonstrates the techniques used to iterate through and extract the data related to all changes in a changeset. To summarize:

  1. The sqlite3changeset_start() API is called to create and initialize an iterator to iterate through the contents of a changeset. Initially, the iterator points to no element at all.

  2. The first call to sqlite3changeset_next() on the iterator moves it to point to the first change in the changeset (or to EOF, if the changeset is completely empty). sqlite3changeset_next() returns SQLITE_ROW if it moves the iterator to point to a valid entry, SQLITE_DONE if it moves the iterator to EOF, or an SQLite error code if an error occurs.

  3. If the iterator points to a valid entry, the sqlite3changeset_op() API may be used to determine the type of change (INSERT, UPDATE or DELETE) that the iterator points to. Additionally, the same API can be used to obtain the name of the table the change applies to and its expected number of columns and primary key columns.

  4. If the iterator points to a valid INSERT or UPDATE entry, the sqlite3changeset_new() API may be used to obtain the new.* values within the change payload.

  5. If the iterator points to a valid DELETE or UPDATE entry, the sqlite3changeset_old() API may be used to obtain the old.* values within the change payload.

  6. An iterator is deleted using a call to the sqlite3changeset_finalize() API. If an error occured while iterating, an SQLite error code is returned (even if the same error code has already been returned by sqlite3changeset_next()). Or, if no error has occurred, SQLITE_OK is returned.

/*
** Print the contents of the changeset to stdout.
*/
static int print_changeset(void *pChangeset, int nChangeset){
  int rc;
  sqlite3_changeset_iter *pIter = 0;

  /* Create an iterator to iterate through the changeset */
  rc = sqlite3changeset_start(&pIter, nChangeset, pChangeset);
  if( rc!=SQLITE_OK ) return rc;

  /* This loop runs once for each change in the changeset */
  while( SQLITE_ROW==sqlite3changeset_next(pIter) ){
    const char *zTab;           /* Table change applies to */
    int nCol;                   /* Number of columns in table zTab */
    int op;                     /* SQLITE_INSERT, UPDATE or DELETE */
    sqlite3_value *pVal;

    /* Print the type of operation and the table it is on */
    rc = sqlite3changeset_op(pIter, &zTab, &nCol, &op, 0);
    if( rc!=SQLITE_OK ) goto exit_print_changeset;
    printf("%s on table %s\n",
      op==SQLITE_INSERT?"INSERT" : op==SQLITE_UPDATE?"UPDATE" : "DELETE",
      zTab
    );

    /* If this is an UPDATE or DELETE, print the old.* values */
    if( op==SQLITE_UPDATE || op==SQLITE_DELETE ){
      printf("Old values:");
      for(i=0; i<nCol; i++){
        rc = sqlite3changeset_old(pIter, i, &pVal);
        if( rc!=SQLITE_OK ) goto exit_print_changeset;
        printf(" %s", pVal ? sqlite3_value_text(pVal) : "-");
      }
      printf("\n");
    }

    /* If this is an UPDATE or INSERT, print the new.* values */
    if( op==SQLITE_UPDATE || op==SQLITE_INSERT ){
      printf("New values:");
      for(i=0; i<nCol; i++){
        rc = sqlite3changeset_new(pIter, i, &pVal);
        if( rc!=SQLITE_OK ) goto exit_print_changeset;
        printf(" %s", pVal ? sqlite3_value_text(pVal) : "-");
      }
      printf("\n");
    }
  }

  /* Clean up the changeset and return an error code (or SQLITE_OK) */
 exit_print_changeset:
  rc2 = sqlite3changeset_finalize(pIter);
  if( rc==SQLITE_OK ) rc = rc2;
  return rc;
}

4. Extended Functionality

Most applications will only use the session module functionality described in the previous section. However, the following additional functionality is available for the use and manipulation of changeset and patchset blobs:

*** DRAFT ***