pg-promise

PostgreSQL interface for Node.js

• About
• Support & Sponsorship
• Documentation
• Contributing
• Usage
  • Methods
  • Query Formatting
    • Index Variables
    • Named Parameters
      • Nested Named Parameters
  • Formatting Filters
    • SQL Names
      • Alias Filter
    • Raw Text
    • Open Values
    • JSON Filter
    • CSV Filter
  • Custom Type Formatting
    • Explicit CTF
    • Symbolic CTF
  • Query Files
  • Tasks
    • Conditional Tasks
  • Transactions
    • Nested Transactions
      • Limitations
    • Configurable Transactions
    • Conditional Transactions
  • Library de-initialization
• History

About

Built on top of node-postgres, this library adds the following:

• Automatic connections
• Automatic transactions
• Powerful query-formatting engine + query generation
• Declarative approach to handling query results
• Global events reporting for central handling
• Extensive support for external SQL files
• Support for all promise libraries

At its inception in 2015, this library was only adding promises to the base driver, hence the name pg-promise. And while the original name was kept, the library's functionality was vastly extended, with promises now being only its tiny part.

Support & Sponsorship

If you like this library, and want a continuous support, then please consider donating:

• Bitcoin: 1yki7MXMkuDw8qqe5icVdh1GJZSQSzKZp

Documentation

Chapter Usage below explains the basics you need to know, while the Official Documentation gets you started, and provides links to all other resources.

Contributing

Please read the Contribution Notes before opening any new issue or PR.

Usage

Once you have created a Database object, according to the steps in the Official Documentation, you get access to the methods documented below.

Methods

All query methods of the library are based off generic method query.

You should normally use only the derived, result-specific methods for executing queries, all of which are named according to how many rows of data the query is expected to return, so for each query you should pick the right method: none, one, oneOrNone, many, manyOrNone = any. Do not confuse the method name for the number of rows to be affected by the query, which is completely irrelevant.

By relying on the result-specific methods you protect your code from an unexpected number of data rows, to be automatically rejected (treated as errors).

There are also a few specific methods that you will often need:

• result, multi, multiResult - for verbose and/or multi-query results;
• map, each - for simpler/inline result pre-processing/re-mapping;
• func, proc - to simplify execution of SQL functions/procedures;
• stream - to access rows from a query via a read stream;
• connect, task, tx + txIf - for shared connections + automatic transactions, each exposing a connected protocol that has additional methods batch, page and sequence.

The protocol is fully customizable / extendable via event extend.

IMPORTANT:

The most important methods to understand from start are task and tx/txIf (see Tasks and Transactions). As documented for method query, it acquires and releases the connection, which makes it a poor choice for executing multiple queries at once. For this reason, Chaining Queries is a must-read, to avoid writing the code that misuses connections.

Learn by Example is a beginner's tutorial based on examples.

Query Formatting

This library comes with embedded query-formatting engine that offers high-performance value escaping, flexibility and extensibility. It is used by default with all query methods, unless you opt out of it entirely via option pgFormatting within Initialization Options.

All formatting methods used internally are available from the formatting namespace, so they can also be used directly when needed. The main method there is format, used by every query method to format the query.

The formatting syntax for variables is decided from the type of values passed in:

• Index Variables when values is an array or a single basic type;
• Named Parameters when values is an object (other than Array or null).

ATTENTION: Never use ES6 template strings or manual concatenation to generate queries, as both can easily result in broken queries! Only this library's formatting engine knows how to properly escape variable values for PostgreSQL.

Index Variables

The simplest (classic) formatting uses $1, $2, ... syntax to inject values into the query string, based on their index (from $1 to $100000) from the array of values:

await db.any('SELECT * FROM product WHERE price BETWEEN $1 AND $2', [1, 10])

The formatting engine also supports single-value parametrization for queries that use only variable $1:

await db.any('SELECT * FROM users WHERE name = $1', 'John')

This however works only for types number, bigint, string, boolean, Date and null, because types like Array and Object change the way parameters are interpreted. That's why passing in index variables within an array is advised as safer, to avoid ambiguities.

Named Parameters

When a query method is parameterized with values as an object, the formatting engine expects the query to use the Named Parameter syntax $*propName*, with * being any of the following open-close pairs: {}, (), <>, [], //.

// We can use every supported variable syntax at the same time, if needed:
await db.none('INSERT INTO users(first_name, last_name, age) VALUES(${name.first}, $<name.last>, $/age/)', {
    name: {first: 'John', last: 'Dow'},
    age: 30
});

IMPORTANT: Never use the reserved ${} syntax inside ES6 template strings, as those have no knowledge of how to format values for PostgreSQL. Inside ES6 template strings you should only use one of the 4 alternatives - $(), $<>, $[] or $//. In general, you should either use the standard strings for SQL, or place SQL into external files - see Query Files.

Valid variable names are limited to the syntax of open-name JavaScript variables. And name this has special meaning - it refers to the formatting object itself (see below).

Keep in mind that while property values null and undefined are both formatted as null, an error is thrown when the property does not exist.

this reference

Property this refers to the formatting object itself, to be inserted as a JSON-formatted string.

await db.none('INSERT INTO documents(id, doc) VALUES(${id}, ${this})', {
    id: 123,
    body: 'some text'    
})
//=> INSERT INTO documents(id, doc) VALUES(123, '{"id":123,"body":"some text"}')

Nested Named Parameters

Named Parameters support property name nesting of any depth.

const obj = {
    one: {
        two: {
            three: {
                value1: 123,
                value2: a => {
                    // a = obj.one.two.three
                    return 'hello';
                },
                value3: function(a) {
                    // a = this = obj.one.two.three
                    return 'world';
                },
                value4: {
                    toPostgres: a => {
                        // Custom Type Formatting
                        // a = obj.one.two.three.value4
                        return a.text;
                    },
                    text: 'custom'
                }                
            }
        }
    }
};
await db.one('SELECT ${one.two.three.value1}', obj); //=> SELECT 123
await db.one('SELECT ${one.two.three.value2}', obj); //=> SELECT 'hello'
await db.one('SELECT ${one.two.three.value3}', obj); //=> SELECT 'world'
await db.one('SELECT ${one.two.three.value4}', obj); //=> SELECT 'custom'

The last name in the resolution can be anything, including:

• the actual value (basic JavaScript type)
• a function that returns:
  • the actual value
  • another function
  • a Custom Type Formatting object
• a Custom Type Formatting object that returns:
  • the actual value
  • another Custom Type Formatting object
  • a function

i.e. the resolution chain is infinitely flexible, and supports recursion without limits.

Please note, however, that nested parameters are not supported within the helpers namespace.

Formatting Filters

By default, all values are formatted according to their JavaScript type. Formatting filters (or modifiers), change that, so the value is formatted differently.

Filters use the same syntax for Index Variables and Named Parameters, following immediately the variable name:

await db.any('SELECT $1:name FROM $2:name', ['price', 'products'])
//=> SELECT "price" FROM "products"

await db.any('SELECT ${column:name} FROM ${table:name}', {
    column: 'price',
    table: 'products'    
});
//=> SELECT "price" FROM "products"

The following filters are supported:

• :name / ~ - SQL Names
  • :alias - Alias Filter
• :raw / ^ - Raw Text
• :value / # - Open Values
• :csv / :list - CSV Filter
• :json - JSON Filter

SQL Names

When a variable name ends with :name, or shorter syntax ~ (tilde), it represents an SQL name or identifier, to be escaped accordingly:

await db.query('INSERT INTO $1~($2~) VALUES(...)', ['Table Name', 'Column Name']);
//=> INSERT INTO "Table Name"("Column Name") VALUES(...)

await db.query('INSERT INTO $1:name($2:name) VALUES(...)', ['Table Name', 'Column Name']);
//=> INSERT INTO "Table Name"("Column Name") VALUES(...)

Typically, an SQL name variable is a text string, which must be at least 1 character long. However, pg-promise supports a variety of ways in which SQL names can be supplied:

• A string that contains only * (asterisks) is automatically recognized as all columns:

await db.query('SELECT $1:name FROM $2:name', ['*', 'table']);
//=> SELECT * FROM "table"

• An array of strings to represent column names:

await db.query('SELECT ${columns:name} FROM ${table:name}', {
    columns: ['column1', 'column2'],
    table: 'table'
});
//=> SELECT "column1","column2" FROM "table"

• Any object that's not an array gets its properties enumerated for column names:

const obj = {
    one: 1,
    two: 2
};

await db.query('SELECT $1:name FROM $2:name', [obj, 'table']);
//=> SELECT "one","two" FROM "table"

In addition, the syntax supports this to enumerate column names from the formatting object:

const obj = {
    one: 1,
    two: 2
};

await db.query('INSERT INTO table(${this:name}) VALUES(${this:csv})', obj);
//=> INSERT INTO table("one","two") VALUES(1, 2)

Relying on this type of formatting for sql names and identifiers, along with regular variable formatting protects your application from SQL injection.

Method as.name implements the formatting.

Alias Filter

An alias is a lighter/simpler version of SQL Names, which only supports a text string, and is used via the :alias filter.

For example, it will skip adding surrounding double quotes when the name is a same-case single word:

await db.any('SELECT full_name as $1:alias FROM $2:name', ['name', 'table']);
//=> SELECT full_name as name FROM "table"

For more details see method as.alias that implements the formatting.

Raw Text

When a variable name ends with :raw, or shorter syntax ^, the value is to be injected as raw text, without escaping.

Such variables cannot be null or undefined, because of the ambiguous meaning in this case, and those values will throw error Values null/undefined cannot be used as raw text.

const where = pgp.as.format('WHERE price BETWEEN $1 AND $2', [5, 10]); // pre-format WHERE condition
await db.any('SELECT * FROM products $1:raw', where);
//=> SELECT * FROM products WHERE price BETWEEN 5 AND 10

Special syntax this:raw / this^ is supported, to inject the formatting object as raw JSON string.

WARNING: This filter is unsafe, and should not be used for values that come from the client side, as it may result in SQL injection.

Open Values

When a variable name ends with :value, or shorter syntax #, it is escaped as usual, except when its type is a string, the trailing quotes are not added.

Open values are primarily to be able to compose complete LIKE/ILIKE dynamic statements in external SQL files, without having to generate them in the code.

i.e. you can either generate a filter like this in your code:

const name = 'John';
const filter = '%' + name + '%';

and then pass it in as a regular string variable, or you can pass in only name, and have your query use the open-value syntax to add the extra search logic:

SELECT * FROM table WHERE name LIKE '%$1:value%')

WARNING: This filter is unsafe, and should not be used for values that come from the client side, as it may result in SQL injection.

Method as.value implements the formatting.

JSON Filter

When a variable name ends with :json, explicit JSON formatting is applied to the value.

By default, any object that's not Date, Array, Buffer, null or Custom-Type (see Custom Type Formatting), is automatically formatted as JSON.

Method as.json implements the formatting.

CSV Filter

When a variable name ends with :csv or :list, it is formatted as a list of Comma-Separated Values, with each value formatted according to its JavaScript type.

Typically, you would use this for a value that's an array, though it works for single values also. See the examples below.

const ids = [1, 2, 3];
await db.any('SELECT * FROM table WHERE id IN ($1:csv)', [ids])
//=> SELECT * FROM table WHERE id IN (1,2,3)

const ids = [1, 2, 3];
await db.any('SELECT * FROM table WHERE id IN ($1:list)', [ids])
//=> SELECT * FROM table WHERE id IN (1,2,3)

Using automatic property enumeration:

const obj = {first: 123, second: 'text'};

await db.none('INSERT INTO table($1:name) VALUES($1:csv)', [obj])
//=> INSERT INTO table("first","second") VALUES(123,'text')

await db.none('INSERT INTO table(${this:name}) VALUES(${this:csv})', obj)
//=> INSERT INTO table("first","second") VALUES(123,'text')

const obj = {first: 123, second: 'text'};

await db.none('INSERT INTO table($1:name) VALUES($1:list)', [obj])
//=> INSERT INTO table("first","second") VALUES(123,'text')

await db.none('INSERT INTO table(${this:name}) VALUES(${this:list})', obj)
//=> INSERT INTO table("first","second") VALUES(123,'text')

Method as.csv implements the formatting.

Custom Type Formatting

The library supports dual syntax for CTF (Custom Type Formatting):

• Explicit CTF - extending the object/type directly, for ease of use, while changing its signature;
• Symbolic CTF - extending the object/type via Symbol properties, without changing its signature.

The library always first checks for the Symbolic CTF, and if no such syntax is used, only then it checks for the Explicit CTF.

Explicit CTF

Any value/object that implements function toPostgres is treated as a custom-formatting type. The function is then called to get the actual value, passing it the object via this context, and plus as a single parameter (in case toPostgres is an ES6 arrow function):

const obj = {
    toPostgres(self) {
        // self = this = obj
        
        // return a value that needs proper escaping
    }
}

Function toPostgres can return anything, including another object with its own toPostgres function, i.e. nested custom types are supported.

The value returned from toPostgres is escaped according to its JavaScript type, unless the object also contains property rawType set to a truthy value, in which case the returned value is considered pre-formatted, and thus injected directly, as Raw Text:

const obj = {
    toPostgres(self) {
        // self = this = obj
        
        // return a pre-formatted value that does not need escaping
    },
    rawType: true // use result from toPostgres directly, as Raw Text
}

Example below implements a class that auto-formats ST_MakePoint from coordinates:

class STPoint {
    constructor(x, y) {
        this.x = x;
        this.y = y;
        this.rawType = true; // no escaping, because we return pre-formatted SQL
    }
    
    toPostgres(self) {
        return pgp.as.format('ST_MakePoint($1, $2)', [this.x, this.y]);
    }
}

And a classic syntax for such a class is even simpler:

function STPoint(x, y){
    this.rawType = true; // no escaping, because we return pre-formatted SQL
    this.toPostgres = () => pgp.as.format('ST_MakePoint($1, $2)', [x, y]);
}

With this class you can use new STPoint(12, 34) as a formatting value that will be injected correctly.

You can also use CTF to override any standard type:

Date.prototype.toPostgres = a => a.getTime();

Symbolic CTF

The only difference from Explicit CTF is that we set toPostgres and rawType as ES6 Symbol properties, defined in the ctf namespace:

const {toPostgres, rawType} = pgp.as.ctf; // Global CTF symbols

const obj = {
    [toPostgres](self) {
        // self = this = obj
        
        // return a pre-formatted value that does not need escaping
    },
    [rawType]: true // use result from toPostgres directly, as Raw Text
};

As CTF symbols are global, you can also configure objects independently of this library:

const ctf = {
    toPostgres: Symbol.for('ctf.toPostgres'),
    rawType: Symbol.for('ctf.rawType')
};

Other than that, it works exactly as the Explicit CTF, but without changing the object's signature.

If you do not know what it means, read the ES6 Symbol API and its use for unique property names. But in short, Symbol properties are not enumerated via for(name in obj), i.e. they are not generally visible within JavaScript, only through specific API Object.getOwnPropertySymbols.

Query Files

Use of external SQL files (via QueryFile) offers many advantages:

• Much cleaner JavaScript code, with all SQL kept in external files;
• Much easier to write large and well-formatted SQL, with many comments and whole revisions;
• Changes in external SQL can be automatically re-loaded (option debug), without restarting the app;
• Pre-formatting SQL upon loading (option params), automating two-step SQL formatting;
• Parsing and minifying SQL (options minify + compress), for early error detection and compact queries.

const {join: joinPath} = require('path');

// Helper for linking to external query files:
function sql(file) {
    const fullPath = joinPath(__dirname, file);
    return new pgp.QueryFile(fullPath, {minify: true});
}

// Create a QueryFile globally, once per file:
const sqlFindUser = sql('./sql/findUser.sql');

db.one(sqlFindUser, {id: 123})
    .then(user => {
        console.log(user);
    })
    .catch(error => {
        if (error instanceof pgp.errors.QueryFileError) {
            // => the error is related to our QueryFile
        }
    });

File findUser.sql:

/*
    multi-line comments are supported
*/
SELECT name, dob -- single-line comments are supported
FROM Users
WHERE id = ${id}

Every query method of the library can accept type QueryFile as its query parameter. Type QueryFile never throws any error, leaving it for query methods to gracefully reject with QueryFileError.

Use of Named Parameters within external SQL files is recommended over the Index Variables, because it makes the SQL much easier to read and understand, and because it also allows Nested Named Parameters, so variables in a large and complex SQL file can be grouped in namespaces for even easier visual separation.

Tasks

A task represents a shared connection for executing multiple queries:

db.task(t => {
    // execute a chain of queries against the task context, and return the result:
    return t.one('SELECT count(*) FROM events WHERE id = $1', 123, a => +a.count)
        .then(count => {
            if(count > 0) {
                return t.any('SELECT * FROM log WHERE event_id = $1', 123)
                    .then(logs => {
                        return {count, logs};
                    })
            }
            return {count};
        });    
})
    .then(data => {
        // success, data = either {count} or {count, logs}
    })
    .catch(error => {
        // failed    
    });

Tasks provide a shared connection context for its callback function, to be released when finished, and they must be used whenever executing more than one query at a time. See also Chaining Queries to understand the importance of using tasks.

You can optionally tag tasks (see Tags), and use ES7 async syntax:

db.task(async t => {
    const count = await t.one('SELECT count(*) FROM events WHERE id = $1', 123, a => +a.count);
    if(count > 0) {
        const logs = await t.any('SELECT * FROM log WHERE event_id = $1', 123);
        return {count, logs};
    }
    return {count};
})
    .then(data => {
        // success, data = either {count} or {count, logs}
    })
    .catch(error => {
        // failed    
    });

db.task('get-event-logs', async t => {
    const count = await t.one('SELECT count(*) FROM events WHERE id = $1', 123, a => +a.count);
    if(count > 0) {
        const logs = await t.any('SELECT * FROM log WHERE event_id = $1', 123);
        return {count, logs};
    }
    return {count};
})
    .then(data => {
        // success, data = either {count} or {count, logs}
    })
    .catch(error => {
        // failed    
    });

Conditional Tasks

Method taskIf creates a new task only when required, according to the condition.

The default condition is to start a new task only when necessary, such as on the top level.

db.taskIf(t1 => {
    // new task has started, as the top level doesn't have one
    return t1.taskIf(t2 => {
        // Task t1 is being used, according to the default condition
        // t2 = t1
    });
})

db.taskIf({cnd: false}, t1 => {
    // new task is created, i.e. option cnd is ignored here,
    // because the task is required on the top level
    return t1.taskIf({cnd: true}, t2 => {
        // new task created, because we specified that we want one;
        // t2 != t1
    });
})

const cnd = c => {
    // c.ctx - task/tx context (not available on the top level)
    // default condition: return !c.ctx;
    return someValue;
};

db.taskIf({cnd}, t1 => {
    // new task is always created, because it is required on the top level
    return t1.taskIf({cnd}, t2 => {
        // if someValue is truthy, a new task is created (t2 != t1);
        // otherwise, we continue with the containing task (t2 = t1).
    });
})

Transactions

Transaction method tx is like task, which also executes BEGIN + COMMIT/ROLLBACK:

db.tx(t => {
    // creating a sequence of transaction queries:
    const q1 = t.none('UPDATE users SET active = $1 WHERE id = $2', [true, 123]);
    const q2 = t.one('INSERT INTO audit(entity, id) VALUES($1, $2) RETURNING id', ['users', 123]);

    // returning a promise that determines a successful transaction:
    return t.batch([q1, q2]); // all of the queries are to be resolved;
})
    .then(data => {
        // success, COMMIT was executed
    })
    .catch(error => {
        // failure, ROLLBACK was executed
    });

If the callback function returns a rejected promise or throws an error, the method will automatically execute ROLLBACK at the end. In all other cases the transaction will be automatically closed by COMMIT.

The same as tasks, transactions support Tags and ES7 async:

db.tx(async t => {
    await t.none('UPDATE users SET active = $1 WHERE id = $2', [true, 123]);
    await t.one('INSERT INTO audit(entity, id) VALUES($1, $2) RETURNING id', ['users', 123]);
})
    .then(data => {
        // success, COMMIT was executed
    })
    .catch(error => {
        // failure, ROLLBACK was executed
    });

db.tx('update-user', async t => {
    await t.none('UPDATE users SET active = $1 WHERE id = $2', [true, 123]);
    await t.one('INSERT INTO audit(entity, id) VALUES($1, $2) RETURNING id', ['users', 123]);
})
    .then(data => {
        // success, COMMIT was executed
    })
    .catch(error => {
        // failure, ROLLBACK was executed
    });

Nested Transactions

Nested transactions automatically share the connection between all levels. This library sets no limitation as to the depth (nesting levels) of transactions supported.

db.tx(t => {
    const queries = [
        t.none('DROP TABLE users;'),
        t.none('CREATE TABLE users(id SERIAL NOT NULL, name TEXT NOT NULL)')
    ];
    for (let i = 1; i <= 100; i++) {
        queries.push(t.none('INSERT INTO users(name) VALUES($1)', 'name-' + i));
    }
    queries.push(
        t.tx(t1 => {
            return t1.tx(t2 => {
                return t2.one('SELECT count(*) FROM users');
            });
        }));
    return t.batch(queries);
})
    .then(data => {
        // success
    })
    .catch(error => {
        // failure
    });

If you want to avoid automatic occurrence of nested transactions, see Conditional Transactions.

Limitations

It is important to know that PostgreSQL does not support full/atomic nested transactions, it only supports savepoints inside top-level transactions, to allow partial rollbacks.

Postgres uses BEGIN with COMMIT / ROLLBACK for top-level transactions, and SAVEPOINT name with RELEASE / ROLLBACK TO name for inner save-points.

This library automatically executes all such transaction and savepoint commands, with unique savepoint names, based on the transaction level.

Configurable Transactions

TransactionMode type can extend your BEGIN command with transaction configuration:

const {TransactionMode, isolationLevel} = pgp.txMode;
 
// Create a reusable transaction mode (serializable + read-only + deferrable):
const mode = new TransactionMode({
    tiLevel: isolationLevel.serializable,
    readOnly: true,
    deferrable: true
});

db.tx({mode}, t => {
    // do transaction queries here
})
    .then(() => {
        // success;
    })
    .catch(error => {
        // failure    
    });

Instead of the default BEGIN, such transaction will open with the following command:

BEGIN ISOLATION LEVEL SERIALIZABLE READ ONLY DEFERRABLE

Transaction Mode is set via option mode, preceding the callback function. See methods tx and txIf.

This is the most efficient and best-performing way of configuring transactions. In combination with Transaction Snapshots you can make the most out of transactions in terms of performance and concurrency.

Conditional Transactions

Method txIf executes a transaction / tx when a specified condition is met, or else it executes a task.

When no condition is specified, the default is to start a transaction, if currently not in one, or else it starts a task. It is useful when you want to avoid Nested Transactions - savepoints.

db.txIf(t => {
    // transaction is started, as the top level doesn't have one
    return t.txIf(t2 => {
        // a task is started, because there is a parent transaction        
    });
})

db.txIf({cnd: someValue}, t => {
    // if condition is truthy, a transaction is started
    return t.txIf(t2 => {
        // a task is started, if the parent is a transaction
        // a transaction is started, if the parent is a task
    });
})

const cnd = c => {
    // c.ctx - task/transaction context (not available on the top level)
    // default condition: return !c.ctx || !c.ctx.inTransaction;
    return someValue;
};

db.txIf({cnd}, t => {
    // if condition is truthy, a transaction is started
    return t.txIf(t2 => {
        // a task is started, if the parent is a transaction
        // a transaction is started, if the parent is a task
    });
})

Library de-initialization

This library manages all database connections via the connection pool, which internally caches them.

Connections in the cache expire due to inactivity after idleTimeoutMillis number of milliseconds, which you can set only when creating the Database object.

While there is a single open connection in the pool, the process cannot terminate by itself, only via process.exit(). If you want the process to finish by itself, without waiting for all connections in the pool to expire, you need to force the pool to shut down all the connections it holds:

db.$pool.end(); // shuts down the connection pool associated with the Database object

For example, if you are using the Bluebird library, you can chain the last promise in the process like this:

.finally(db.$pool.end);

IMPORTANT: Note that if your app is an HTTP service, or generally an application that does not feature any exit point, then you should not do any de-initialization at all. It is only if your app is a run-through process/utility, then you might want to use it, so the process ends without delays.

In applications that either use multiple databases or execute a multi-pool strategy for balanced query loads, you would end up with multiple Database objects, each with its own connection pool. In this scenario, in order to exit the process normally, at a particular point, you can call pgp.end to shut down all connection pools at once:

pgp.end(); // shuts down all connection pools created in the process

or promise-chained to the last query block in the process:

.finally(pgp.end);

Once you have shut down the pool associated with your Database object, you can no longer use the object, and any of its query methods will be rejecting with Error = Connection pool of the database object has been destroyed.

See the relevant API: pgp.end, Database.$pool

History

Although this project formally maintains a CHANGELOG, for a short list of the top-level changes, for detailed changes between versions you should see the corresponding release notes.