SQLite is architected as a library that can be embedded in another application, such as Chrome. SQLite runs in the application's process, and shares its memory and other resources. This is similar to embedded databases like LevelDB and BerkeleyDB. By contrast, most popular RDMBSes, like PostgreSQL and MySQL, are structured as standalone server processes that accept queries from client processes.
TODO: Explain the process model and locking
TODO: Explain Chrome decisions -- exclusive locking, full per-feature isolation (separate databases and page caches)
The section summarizes aspects of SQLite that are relevant to schema and query design, and may be surprising to readers with prior experience in other popular SQL database systems, such as PostgreSQL and MySQL.
SQLite stores data using 5 major types, which are summarized below.
NULL is a special type for the
INTEGER represents big-endian twos-complement integers. Boolean values (
FALSE) are represented as the integer values 1 and 0.
REAL represents IEEE 754-2008 64-bit floating point numbers.
BLOB represents sequences of bytes that are opaque to SQLite. These values are sorted using the bitwise binary comparison offered by memcmp.
SQLite stores index keys and row values (records / tuples) using a tightly packed format that makes heavy use of varints and variable-length fields. The column types have almost no influence on the encoding of values. This has the following consequences.
BIGINT, are treated as aliases for
DOUBLE PRECISIONare treated as aliases for
CLOB, are treated as aliases for
SQLite uses clever heuristics, called type affinity, to map SQL column types such as
VARCHAR to the major types above.
Chrome database schemas should avoid type affinity, and should not include any information ignored by SQLite.
SQLite uses B-trees to store both table and index data.
The exclusive use of B-trees reduces the amount of schema design decisions. Notable examples:
There is no equivalent to PostgreSQL's index types. In particular, since there are no hashed indexes, the design does not need to consider whether the index only needs to support equality queries, as opposed to greater/smaller than comparisons.
There is no equivalent to PostgreSQL's table access methods. Each table is clustered by a primary key index, which is implicitly stored in the table's B-tree.
By default, table rows (records / tuples) are stored in a B-tree keyed by rowid, an automatically assigned 64-bit integer key. Effectively, these tables are clustered by rowid, which acts as an implicit primary key. Opting out of this SQLite-specific default requires appending
WITHOUT ROWID to the
CREATE TABLE instruction.
SQLite's B-tree page format has optimized special cases for tables clustered by rowid. This makes rowid the most efficient surrogate key implementation in SQLite. To make this optimization easier to use, any column that is a primary key and has an
INTEGER type is considered an alias for rowid.
Each SQLite index is stored in a B-tree. Each index entry is stored as a B-tree node whose key is made up of the record‘s index key column values, followed by the record’s primary key column values.
WITHOUT ROWID table indexes can include primary key columns without additional storage costs. This is because indexes for
WITHOUT ROWID tables enjoy a space optimization where columns in both the primary key and the index key are not stored twice in B-tree nodes.
At a high level, SQLite compiles SQL queries into bytecode executed by a virtual machine called the VDBE, or the bytecode engine. A compiled query can be executed multiple times, amortizing the costs of query parsing and planning. Chrome's SQLite abstraction layer makes it easy to use compiled queries.
The following SQLite documentation pages cover the query planner and optimizer.
TODO: Present a simplified model that's sufficient for most database design.
The following pieces of advice usually come up in code reviews.
The SQLite shell offers quick feedback for converging on valid SQL statement syntax, and avoiding SQLite features that are disabled in Chrome. In addition, the
EXPLAIN QUERY PLAN statements show the results of SQLite's query planner and optimizer, which are very helpful for reasoning about the performance of complex queries.
The following commands set up SQLite shells using Chrome's build of SQLite.
autoninja -C out/Default sqlite_shell sqlite_dev_shell
sqlite_shellruns the SQLite build that we ship in Chrome. It offers the ground truth on whether a SQL statement can be used in Chrome code or not.
EXPLAIN QUERY PLANstatements, as well as a few features used by Perfetto's analysis tools.
SQLite queries are usually embedded as string literals in C++ code. The advice here has the following goals.
Easy to read queries. The best defense against subtle bugs is making the queries very easy to read, so that any bugs become obvious at code review time. SQL string literals don't benefit from our code analysis infrastructure, so the only lines of defense against bugs are testing and code review.
Simplify crash debugging. We will always have a low volume of non-actionable crash reports, because Chrome runs on billions of devices, some of which have faulty RAM or processors.
No unnecessary performance overheads. The C++ optimizer doesn't understand SQL query literals, so the queries end up as written in the Chrome binary. Extra characters cost binary size, as well as CPU time (which turns into battery usage) during query parsing.
Match the embedding language (C++) style guide. This reduces the mental context switch overhead for folks who write and/or review C++ code that contains SQL.
Format statements like so.
static constexpr char kOriginInfoSql = // clang-format off "CREATE TABLE origin_infos(" "origin TEXT NOT NULL," "last_modified INTEGER NOT NULL," "secure INTEGER NOT NULL)"; // clang-format on static constexpr char kInsertSql = // clang-format off "INSERT INTO infos(origin,last_modified,secure) " "VALUES (?,?,?)"; // clang-format on static constexpr char kSelectSql = // clang-format off "SELECT origin,last_modified,secure FROM origins " "WHERE last_modified > ? " "ORDER BY last_modified"; // clang-format on
SQLite keywords should use ALL CAPS. This makes SQL query literals easier to distinguish and search for.
Identifiers, such as table and row names, should use snake_case.
Identifiers, keywords, and parameter placeholders (
?) should be separated by exactly one character. Separators may be spaces (
), commas (
,), or parentheses (
Statement-ending semicolons (
;) are omitted.
SQL statements are stored in variables typed
static constexpr char, or in string literals passed directly to methods.
INSERT statements should list all the table columns by name, in the same order as the corresponding
CREATE TABLE statements.
SELECT statements should list the desired table columns by name, in the same order as the corresponding
CREATE TABLE statements.
SELECT * is strongly discouraged, at least until we have schema checks on database opens.
CREATE UNIQUE INDEX statements should be preferred to
UNIQUE constraints on
Values must either be embedded in the SQL statement string literal, or bound using parameters.
Parameter placeholders should always use the
? syntax. Alternative syntaxes, such as
:AAAA, have few benefits in a codebase where the
Bind statements are right next to the queries, and are less known to readers.
Do not execute multiple SQL statements (e.g., by calling
sql::Statement) on the same C++ line. It‘s difficult to get more than line numbers from crash reports’ stack traces.
Identifiers (table / index / column names and aliases) must not be current SQLite keywords. Identifiers may not start with the
sqlite_ prefix, to avoid conflicting with the name of a SQLite internal schema object.
Column types should only be one of the the SQLite storage types (
BLOB), so readers can avoid reasoning about SQLite's type affinity.
Columns that will store boolean values should have the
Columns that will store
base::Time values should have the
INTEGER type. Values should be serialized using
sql::Statement::BindTime() and deserialized using
Column types should not include information ignored by SQLite, such as numeric precision or scale specifiers, or string length specifiers.
NOT NULL constraints must be explicitly stated in column definitions that include
PRIMARY KEY specifiers. For historical reasons, SQLite allows NULL primary keys in most cases. When a table‘s primary key is composed of multiple columns, each column’s definition should have a
NOT NULL constraint.
Columns should avoid
DEFAULT values. Columns that have
NOT NULL constraints and lack a
DEFAULT value are easier to review and maintain, as SQLite takes over the burden of checking that
INSERT statements aren't missing these columns.
Surrogate primary keys should use the column type
INTEGER PRIMARY KEY, to take advantage of SQLite's rowid optimizations.
AUTOINCREMENT should only be used where primary key reuse would be unacceptable.
SQLite exposes a vast array of functionality via SQL statements. The following features are not a good match for SQL statements used by Chrome feature code.
PRAGMA statements should never be used directly. Chrome's SQLite abstraction layer should be modified to support the desired effects instead.
PRAGMA use limits our ability to customize and secure our SQLite build.
PRAGMA statements may turn on code paths with less testing / fuzzing coverage. Furthermore, some
PRAGMA statements invalidate previously compiled queries, reducing the efficiency of Chrome's compiled query cache.
CREATE VIRTUAL TABLE statements should not be used. The desired functionality should be implemented in C++, and access storage using standard SQL statements.
Virtual tables are SQLite's module system. SQL statements on virtual tables are essentially running arbitrary code, which makes them very difficult to reason about and maintain. Furthermore, the virtual table implementations don't receive the same level of fuzzing coverage as the SQLite core.
Chrome's SQLite build has virtual table functionality reduced to the minimum needed to support FTS3 in WebSQL, and an internal feature. SQLite's run-time loading mechanism is disabled, and most built-in virtual tables are disabled as well.
SQL foreign key constraints should not be used. All data validation should be performed using explicit
SELECT statements (generally wrapped as helper methods) inside transactions. Cascading deletions should be performed using explicit
DELETE statements inside transactions.
Chrome features cannot rely on foreign key enforcement, due to the possibility of data corruption. Furthermore, foreign key constraints make it more difficult to reason about system behavior (Chrome feature code + SQLite) when the database gets corrupted. Foreign key constraints also make it more difficult to reason about query performance.
SQL CHECK constraints should not be used, for the same reasons as foreign key constraints. The equivalent checks should be performed in C++, typically using
SQL triggers should not be used.
Triggers significantly increase the difficulty of reviewing and maintaining Chrome features that use them.
SQL Common Table Expressions (CTEs) should not be used. Chrome's SQL schemas and queries should be simple enough that the factoring afforded by ordinary CTEs is not necessary. Recursive CTEs should be implemented in C++.
Common Table Expressions do not open up any query optimizations that would not be available otherwise, and make it more difficult to review / analyze queries.
SQL views, managed by the
CREATE VIEW statement and the
DROP VIEW statement, should not be used. Chrome's SQL schemas and queries should be simple enough that the factoring afforded by views is not necessary.
Views are syntactic sugar, and do not open up any new SQL capabilities. SQL statements on views are more difficult to understand and maintain, because of the extra layer of indirection.
Compound SELECT statements should not be used. Such statements should be broken down into simple SELECT statements, and the operators
EXCEPT should be implemented in C++.
A single compound SELECT statement is more difficult to review and properly unit-test than the equivalent collection of simple SELECT statements. Furthermore, the compound SELECT statement operators can be implemented more efficiently in C++ than in SQLite's bytecode interpreter (VDBE).
SQLite's built-in functions should be only be used in SQL statements where they unlock significant performance improvements. Chrome features should store data in a format that leaves the most room for query optimizations, and perform any necessary transformations after reading / before writing the data.
trim()are best replaced by C++ code that uses the helpers in
total_changes(), are best replaced by functionality in
sqlite_version()should not be necessary in Chrome code, and may suggest a problem in the feature's design.
ATTACH DATABASE statements should not be used. Each Chrome feature should store its data in a single database. Chrome code should not assume that transactions across multiple databases are atomic.
We plan to remove all existing
ATTACH DATABASE use from Chrome.