SQLite is a relational database management system (RDBMS) that supports most of SQL.
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.
The main bottleneck in SQLite database performance is usually disk I/O. So, designing schemas that perform well requires understanding how SQLite stores data on disk.
At a very high level, a SQLite database is a forest of B-trees, some of which are B+-trees. The database file is an array of fixed-size pages, where each page stores a B-tree node. The page size can only be set when a database file is created, and impacts both SQL statement execution speed, and memory consumption.
The data in each table (usually called rows, records, or tuples) is stored in a separate B-tree. The data in each index (called entries, records or tuples) is also stored in a separate B-tree. So, each B-tree is associated with exactly one table. The Indexing section goes into further details.
Each B-tree node stores multiple tuples of values. The values and their encodings are described in the Value types section.
Tying everything together: The performance of a SQL statement is roughly the number of database pages touched (read / written) by the statement. These pages are nodes belonging to the B-trees associated with the tables mentioned in the statement. The number of pages touched when accessing a B-tree depends on the B-tree‘s depth. Each B-tree’s depth depends on its record count (number of records stored in it), and on its node width (how many records fit in a node).
SQLite stores values using 5 major types, which are summarized below.
NULL is a special type for the NULL
value.
INTEGER represents big-endian twos-complement integers. Boolean values (TRUE
and FALSE
) are represented as the integer values 1 and 0.
REAL represents IEEE 754-2008 64-bit floating point numbers.
TEXT represents strings (sequences of characters) encoded using a supported SQLite encoding. These values are sorted according to a collating sequence or a collating function.
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.
TINYINT
and BIGINT
, are treated as aliases for INTEGER
.DECIMAL
, FLOAT
, and DOUBLE PRECISION
are treated as aliases for REAL
.DECIMAL(5,2)
are ignored.CHAR
, CHARACTER VARYING
, VARCHAR
, and CLOB
, are treated as aliases for TEXT
.CHAR(255)
are ignored.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 very high level, SQLite compiles SQL statements (often called queries) into bytecode executed by a virtual machine called the VDBE, or the bytecode engine. A compiled statement 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.
Assuming effective use of cached statements, the performance of a SQL statement comes down to the query plan that SQLite generates for the statement. The query plan is the sequence of B-tree accesses used to execute the statement, which determines the number of B-tree pages touched.
The rest of this section summarizes the following SQLite documentation pages.
At a high level, a SQLite query plan is a sequence of nested loops, where each loop iterates over the data in a B-tree. Each loop can use the current record of the outer loops.
TODO: Complete this section. Cover joins, sorting, etc.
Ideally, the SQL schemas and statements used by Chrome features would be simple enough that the query plans would be obvious to the reader.
When this isn't the case, the fastest way to get the query plan is to load the schema in the SQLite shell, and use EXPLAIN QUERY PLAN
.
The following command builds a SQLite shell that uses Chrome's build of SQLite, and supports the EXPLAIN QUERY PLAN
command.
autoninja -C out/Default sqlite_dev_shell
Inside the SQLite shell, the .eqp on
directive automatically shows the results of EXPLAIN QUERY PLAN
for every SQL statement executed in the shell.
Query steps are the building blocks of SQLite query plans. Each query step is essentially a loop that iterates over the records in a B-tree. These loops differ in terms of how many B-tree pages they touch, and how many records they produce. This sub-section lists the types of steps implemented by SQLite.
Scans visit an entire (table or index) B-tree. For this reason, scans are almost never acceptable in Chrome. Most of our features don't have limits on the amount of stored data, so scans can result in an unbounded amount of I/O.
A table scan visits the entire table's B-tree.
A covering index scan visits an entire index B-tree, but doesn't access the associated table B-tree.
SQLite doesn't have any special optimization for COUNT(*)
queries. In other words, SQLite does not track subtree sizes in its B-tree nodes.
Reviewers sometimes emphasize performance issues by calling the scans full table scans and full index scans, where “full” references the fact that the number of B-tree pages accessed is proportional to the entire data set stored on disk.
TODO: Complete this section. Add examples in a way that doesn't make the section overly long.
Searches access a subset of a (table or index) B-tree nodes. Searches limit the amount of nodes they need to access based on query restrictions, such as terms in the WHERE
clause. Seeing a SEARCH
in a query plan is not a guarantee of performance. Searches can vary wildly in the amount of B-tree pages they need to access.
One of the fastest possible searches is a table search that performs exactly one B-tree lookup, and produces at most one record.
The other fastest possible search is a covering index search that also performs one lookup, and produces at most one record.
TODO: Complete this section. Add examples in a way that doesn't make the section overly long.
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
and 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 SQLite shell directive .eqp on
automatically issues EXPLAIN QUERY PLAN
for all future commands.
The following commands set up SQLite shells using Chrome's build of SQLite.
autoninja -C out/Default sqlite_shell sqlite_dev_shell
sqlite_shell
runs 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.sqlite_dev_shell
enables the EXPLAIN
and EXPLAIN QUERY PLAN
statements, 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.
SELECT
statements that retrieve more than one row should include an ORDER BY
clause to clarify the implicit ordering.
WITHOUT ROWID
tables, by the table's primary key.CREATE INDEX
statements should immediately follow the CREATE TABLE
statement for the indexed table.
Explicit CREATE UNIQUE INDEX
statements should be preferred to UNIQUE
constraints on CREATE TABLE
.
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 ?NNN
or :AAAA
, have few benefits in a codebase where the Bind
statements are right next to the queries, and are less known to readers.
SQL statements should be embedded in C++ as string literals. The char[]
type makes it possible for us to compute query length at compile time in the future. The static
and constexpr
qualifiers both ensure optimal code generation.
Do not execute multiple SQL statements (e.g., by calling Step()
or Run()
on 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 (INTEGER
, REAL
, TEXT
, BLOB
), so readers can avoid reasoning about SQLite's type affinity.
Columns that will store boolean values should have the INTEGER
type.
Columns that will store base::Time
values should have the INTEGER
type. Values should be serialized using sql::Statement::BindTime()
and deserialized using sql::Statement::ColumnTime()
.
Column types should not include information ignored by SQLite, such as numeric precision or scale specifiers, or string length specifiers.
Columns should have NOT NULL
constraints whenever possible. This saves maintainers from having to reason about the less intuitive cases of NULL
handling.
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.
Direct 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.
Access to virtual tables is disabled by default for SQLite databases opened with Chrome's sql::Database
infrastructure. This is intended to steer feature developers away from the discouraged feature.
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.
After WebSQL is removed from Chrome, we plan to disable SQLite's virtual table support using SQLITE_OMIT_VIRTUALTABLE.
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.
Foreign key constraints are not enforced by default on SQLite databases opened with Chrome's sql::Database
infrastructure. This is intended to steer feature developers away from the discouraged feature.
After WebSQL is removed from Chrome, we plan to disable SQLite's foreign key support using SQLITE_OMIT_FOREIGN_KEY.
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 DCHECK
.
After WebSQL is removed from Chrome, we plan to disable SQLite's CHECK constraint support using SQLITE_OMIT_CHECK.
SQL triggers should not be used.
Triggers significantly increase the difficulty of reviewing and maintaining Chrome features that use them.
Triggers are not executed on SQLite databases opened with Chrome's sql::Database
infrastructure. This is intended to steer feature developers away from the discouraged feature.
After WebSQL is removed from Chrome, we plan to disable SQLite's trigger support using SQLITE_OMIT_TRIGGER.
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.
Access to views is disabled by default for SQLite databases opened with Chrome's sql::Database
infrastructure. This is intended to steer feature developers away from the discouraged feature.
After WebSQL is removed from Chrome, we plan to disable SQLite's VIEW support using SQLITE_OMIT_VIEW.
String literals should always be single-quoted. That being said, string literals should be rare in Chrome code, because any user input must be injected using statement parameters and the Statement::Bind*()
methods.
Double-quoted string literals are non-standard SQL syntax. The SQLite authors currently consider this be a misfeature.
SQLite support for double-quoted string literals is disabled for databases opened with Chrome's sql::Database
infrastructure. This is intended to steer feature developers away from this discouraged feature.
After WebSQL is removed from Chrome, we plan to disable SQLite's support for double-quoted string literals using SQLITE_DQS=0.
Compound SELECT statements should not be used. Such statements should be broken down into simple SELECT statements, and the operators UNION
, UNION ALL
, INTERSECT
and 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).
After WebSQL is removed from Chrome, we plan to disable SQLite's compound SELECT support using SQLITE_OMIT_COMPOUND_SELECT.
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.
base::Time
functionality.printf()
and trim()
are best replaced by C++ code that uses the helpers in //base/strings/
.changes()
, last_insert_rowid()
, and total_changes()
, are best replaced by functionality in sql::Database
and sql::Statement
.sqlite_source_id()
and sqlite_version()
should not be necessary in Chrome code, and may suggest a problem in the feature's design.Math functions and Window functions are disabled in Chrome's SQLite build.
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.
We aim to disable SQLite features that should not be used in Chrome, subject to the constraint of keeping WebSQL's feature set stable. We currently disable all new SQLite features, to avoid expanding the attack surface exposed to WebSQL. This stance may change once WebSQL is removed from Chrome.
The following SQLite features have been disabled in Chrome.
Chrome features should prefer procotol buffers to JSON for on-disk (persistent) serialization of extensible structured data.
Chrome features should store the values used by indexes directly in their own columns, instead of relying on SQLite's JSON support.
SQLite's UPSERT implementation has been disabled in order to avoid increasing WebSQL's attack surface. UPSERT is disabled using the SQLITE_OMIT_UPSERT
macro, which is not currently included in the SQLite compile-time option list, but exists in the source code.
We currently think that the new UPSERT functionality is not essential to implementing Chrome features efficiently. An example where UPSERT is necessary for the success of a Chrome feature would likely get UPSERT enabled.
Window functions have been disabled primarily because they cause a significant binary size increase, which leads to a corresponding large increase in the attack surface exposed to WebSQL.
Window functions increase the difficulty of reviewing and maintaining the Chrome features that use them, because window functions add complexity to the mental model of query performance.
We currently think that this maintenance overhead of window functions exceeds any convenience and performance benefits (compared to simpler queries coordinated in C++).