*** DRAFT ***

Editorial Note: This document was written in 2004 as a guide to programmers who were transitioning from SQLite2 to SQLite3. It is retained as part of the historical record of SQLite. Modern programmers should refer to more up-to-date documentation on SQLite is available elsewhere on this website.

SQLite Version 3 Overview

SQLite version 3.0 introduces important changes to the library, including:

This document is a quick introduction to the changes for SQLite 3.0 for users who are already familiar with SQLite version 2.8.

Naming Changes

SQLite version 2.8 will continue to be supported with bug fixes for the foreseeable future. In order to allow SQLite version 2.8 and SQLite version 3.0 to peacefully coexist, the names of key files and APIs in SQLite version 3.0 have been changed to include the character "3". For example, the include file used by C programs has been changed from "sqlite.h" to "sqlite3.h". And the name of the shell program used to interact with databases has been changed from "sqlite.exe" to "sqlite3.exe". With these changes, it is possible to have both SQLite 2.8 and SQLite 3.0 installed on the same system at the same time. And it is possible for the same C program to link against both SQLite 2.8 and SQLite 3.0 at the same time and to use both libraries at the same time.

New File Format

The format used by SQLite database files has been completely revised. The old version 2.1 format and the new 3.0 format are incompatible with one another. Version 2.8 of SQLite will not read a version 3.0 database files and version 3.0 of SQLite will not read a version 2.8 database file.

To convert an SQLite 2.8 database into an SQLite 3.0 database, have ready the command-line shells for both version 2.8 and 3.0. Then enter a command like the following:

sqlite OLD.DB .dump | sqlite3 NEW.DB

The new database file format uses B+trees for tables. In a B+tree, all data is stored in the leaves of the tree instead of in both the leaves and the intermediate branch nodes. The use of B+trees for tables allows for better scalability and the storage of larger data fields without the use of overflow pages. Traditional B-trees are still used for indices.

The new file format also supports variable pages sizes between 512 and 65536 bytes. The size of a page is stored in the file header so the same library can read databases with different pages sizes, in theory, though this feature has not yet been implemented in practice.

The new file format omits unused fields from its disk images. For example, indices use only the key part of a B-tree record and not the data. So for indices, the field that records the length of the data is omitted. Integer values such as the length of key and data are stored using a variable-length encoding so that only one or two bytes are required to store the most common cases but up to 64-bits of information can be encoded if needed. Integer and floating point data is stored on the disk in binary rather than being converted into ASCII as in SQLite version 2.8. These changes taken together result in database files that are typically 25% to 35% smaller than the equivalent files in SQLite version 2.8.

Details of the low-level B-tree format used in SQLite version 3.0 can be found in header comments to the btreeInt.h source file and in the file format documentation.

Manifest Typing and BLOB Support

SQLite version 2.8 will deal with data in various formats internally, but when writing to the disk or interacting through its API, SQLite 2.8 always converts data into ASCII text. SQLite 3.0, in contrast, exposes its internal data representations to the user and stores binary representations to disk when appropriate. The exposing of non-ASCII representations was added in order to support BLOBs.

SQLite version 2.8 had the feature that any type of data could be stored in any table column regardless of the declared type of that column. This feature is retained in version 3.0, though in a slightly modified form. Each table column will store any type of data, though columns have an affinity for the format of data defined by their declared datatype. When data is inserted into a column, that column will make an attempt to convert the data format into the column's declared type. All SQL database engines do this. The difference is that SQLite 3.0 will still store the data even if a format conversion is not possible.

For example, if you have a table column declared to be of type "INTEGER" and you try to insert a string, the column will look at the text string and see if it looks like a number. If the string does look like a number it is converted into a number and into an integer if the number does not have a fractional part, and stored that way. But if the string is not a well-formed number it is still stored as a string. A column with a type of "TEXT" tries to convert numbers into an ASCII-Text representation before storing them. But BLOBs are stored in TEXT columns as BLOBs because you cannot in general convert a BLOB into text.

In most other SQL database engines the datatype is associated with the table column that holds the data - with the data container. In SQLite 3.0, the datatype is associated with the data itself, not with its container. Paul Graham in his book ANSI Common Lisp calls this property "Manifest Typing". Other writers have other definitions for the term "manifest typing", so beware of confusion. But by whatever name, that is the datatype model supported by SQLite 3.0.

Additional information about datatypes in SQLite version 3.0 is available separately.

Support for UTF-8 and UTF-16

The new API for SQLite 3.0 contains routines that accept text as both UTF-8 and UTF-16 in the native byte order of the host machine. Each database file manages text as either UTF-8, UTF-16BE (big-endian), or UTF-16LE (little-endian). Internally and in the disk file, the same text representation is used everywhere. If the text representation specified by the database file (in the file header) does not match the text representation required by the interface routines, then text is converted on-the-fly. Constantly converting text from one representation to another can be computationally expensive, so it is suggested that programmers choose a single representation and stick with it throughout their application.

In the current implementation of SQLite, the SQL parser only works with UTF-8 text. So if you supply UTF-16 text it will be converted. This is just an implementation issue and there is nothing to prevent future versions of SQLite from parsing UTF-16 encoded SQL natively.

When creating new user-defined SQL functions and collating sequences, each function or collating sequence can specify if it works with UTF-8, UTF-16be, or UTF-16le. Separate implementations can be registered for each encoding. If an SQL function or collating sequence is required but a version for the current text encoding is not available, then the text is automatically converted. As before, this conversion takes computation time, so programmers are advised to pick a single encoding and stick with it in order to minimize the amount of unnecessary format juggling.

SQLite is not particular about the text it receives and is more than happy to process text strings that are not normalized or even well-formed UTF-8 or UTF-16. Thus, programmers who want to store IS08859 data can do so using the UTF-8 interfaces. As long as no attempts are made to use a UTF-16 collating sequence or SQL function, the byte sequence of the text will not be modified in any way.

User-defined Collating Sequences

A collating sequence is just a defined order for text. When SQLite 3.0 sorts (or uses a comparison operator like "<" or ">=") the sort order is first determined by the data type.

Collating sequences are used for comparing two text strings. The collating sequence does not change the ordering of NULLs, numbers, or BLOBs, only text.

A collating sequence is implemented as a function that takes the two strings being compared as inputs and returns negative, zero, or positive if the first string is less than, equal to, or greater than the second. SQLite 3.0 comes with a single built-in collating sequence named "BINARY" which is implemented using the memcmp() routine from the standard C library. The BINARY collating sequence works well for English text. For other languages or locales, alternative collating sequences may be preferred.

The decision of which collating sequence to use is controlled by the COLLATE clause in SQL. A COLLATE clause can occur on a table definition, to define a default collating sequence to a table column, or on field of an index, or in the ORDER BY clause of a SELECT statement. Planned enhancements to SQLite are to include standard CAST() syntax to allow the collating sequence of an expression to be defined.

64-bit ROWIDs

Every row of a table has a unique rowid. If the table defines a column with the type "INTEGER PRIMARY KEY" then that column becomes an alias for the rowid. But with or without an INTEGER PRIMARY KEY column, every row still has a rowid.

In SQLite version 3.0, the rowid is a 64-bit signed integer. This is an expansion of SQLite version 2.8 which only permitted rowids of 32-bits.

To minimize storage space, the 64-bit rowid is stored as a variable length integer. Rowids between 0 and 127 use only a single byte. Rowids between 0 and 16383 use just 2 bytes. Up to 2097152 uses three bytes. And so forth. Negative rowids are allowed but they always use nine bytes of storage and so their use is discouraged. When rowids are generated automatically by SQLite, they will always be non-negative.

Improved Concurrency

SQLite version 2.8 allowed multiple simultaneous readers or a single writer but not both. SQLite version 3.0 allows one process to begin writing the database while other processes continue to read. The writer must still obtain an exclusive lock on the database for a brief interval in order to commit its changes, but the exclusive lock is no longer required for the entire write operation. A more detailed report on the locking behavior of SQLite version 3.0 is available separately.

A limited form of table-level locking is now also available in SQLite. If each table is stored in a separate database file, those separate files can be attached to the main database (using the ATTACH command) and the combined databases will function as one. But locks will only be acquired on individual files as needed. So if you redefine "database" to mean two or more database files, then it is entirely possible for two processes to be writing to the same database at the same time. To further support this capability, commits of transactions involving two or more ATTACHed database are now atomic.

Credits

SQLite version 3.0 is made possible in part by AOL developers supporting and embracing great Open-Source Software.

*** DRAFT ***