postgresql.driver provides a PG-API, postgresql.api, interface to a PostgreSQL server using PQ version 3.0 to facilitate communication. It makes use of the protocol’s extended features to provide binary datatype transmission and protocol level prepared statements for strongly typed parameters.

postgresql.driver currently supports PostgreSQL servers as far back as 8.0. Prior versions are not tested. While any version of PostgreSQL supporting version 3.0 of the PQ protocol should work, many features may not work due to absent functionality in the remote end.

For DB-API 2.0 users, the driver module is located at postgresql.driver.dbapi20. The DB-API 2.0 interface extends PG-API. All of the features discussed in this chapter are available on DB-API connections.


PostgreSQL versions 8.1 and earlier do not support standard conforming strings. In order to avoid subjective escape methods on connections, postgresql.driver.pq3 enables the standard_conforming_strings setting by default. Greater care must be taken when working versions that do not support standard strings. The majority of issues surrounding the interpolation of properly quoted literals can be easily avoided by using parameterized statements.

The following identifiers are regularly used as shorthands for significant interface elements:

postgresql.api.Connection, a database connection. Connections
postgresql.api.Statement, a prepared statement. Prepared Statements
postgresql.api.Cursor, a cursor; the results of a prepared statement. Cursors
postgresql.api.Connector, a connector. Connectors

Establishing a Connection

There are many ways to establish a postgresql.api.Connection to a PostgreSQL server using postgresql.driver. This section discusses those, connection creation, interfaces.

In the root package module, the open() function is provided for accessing databases using a locator string and optional connection keywords. The string taken by is a URL whose components make up the client parameters:

>>> db ="pq://localhost/postgres")

This will connect to the host, localhost and to the database named postgres via the pq protocol. open will inherit client parameters from the environment, so the user name given to the server will come from $PGUSER, or if that is unset, the result of getpass.getuser–the username of the user running the process. The user’s “pgpassfile” will even be referenced if no password is given:

>>> db ="pq://username:password@localhost/postgres")

In this case, the password is given, so ~/.pgpass would never be referenced. The user client parameter is also given, username, so $PGUSER or getpass.getuser will not be given to the server.

Settings can also be provided by the query portion of the URL:

>>> db ="pq://user@localhost/postgres?search_path=public&timezone=mst")

The above syntax ultimately passes the query as settings(see the description of the settings keyword in Connection Keywords). Driver parameters require a distinction. This distinction is made when the setting’s name is wrapped in square-brackets, ‘[‘ and ‘]’:

>>> db ="pq://user@localhost/postgres?[sslmode]=require&[connect_timeout]=5")

sslmode and connect_timeout are driver parameters. These are never sent to the server, but if they were not in square-brackets, they would be, and the driver would never identify them as driver parameters.

The general structure of a PQ-locator is:


Optionally, connection keyword arguments can be used to override anything given in the locator:

>>> db ="pq://user:secret@host", password = "thE_real_sekrat")

Or, if the locator is not desired, individual keywords can be used exclusively:

>>> db = = 'user', host = 'localhost', port = 6543)

In fact, all arguments to are optional as all arguments are keywords; iri is merely the first keyword argument taken by If the environment has all the necessary parameters for a successful connection, there is no need to pass anything to open:

>>> db =

For a complete list of keywords that can accept, see Connection Keywords. For more information about the environment variables, see PostgreSQL Environment Variables. For more information about the pgpassfile, see PostgreSQL Password File.

postgresql.driver.connect is a high-level interface to connection creation. It provides password resolution services and client parameter inheritance. For some applications, this is undesirable as such implicit inheritance may lead to failures due to unanticipated parameters being used. For those applications, use of is not recommended. Rather, postgresql.driver.connect should be used when explicit parameterization is desired by an application:

>>> import postgresql.driver as pg_driver
>>> db = pg_driver.connect(
...  user = 'usename',
...  password = 'secret',
...  host = 'localhost',
...  port = 5432
... )

This will create a connection to the server listening on port 5432 on the host localhost as the user usename with the password secret.


connect will not inherit parameters from the environment as libpq-based drivers do.

See Connection Keywords for a full list of acceptable keyword parameters and their meaning.


Connectors are the supporting objects used to instantiate a connection. They exist for the purpose of providing connections with the necessary abstractions for facilitating the client’s communication with the server, and to act as a container for the client parameters. The latter purpose is of primary interest to this section.

Each connection object is associated with its connector by the connector attribute on the connection. This provides the user with access to the parameters used to establish the connection in the first place, and the means to create another connection to the same server. The attributes on the connector should not be altered. If parameter changes are needed, a new connector should be created.

The attributes available on a connector are consistent with the names of the connection parameters described in Connection Keywords, so that list can be used as a reference to identify the information available on the connector.

Connectors fit into the category of “connection creation interfaces”, so connector instantiation normally takes the same parameters that the postgresql.driver.connect function takes.


Connector implementations are specific to the transport, so keyword arguments like host and port aren’t supported by the Unix connector.

The driver, postgresql.driver.default provides a set of connectors for making a connection:
Provides a getaddrinfo() abstraction for establishing a connection.
Connect to a single IPv4 addressed host.
Connect to a single IPv6 addressed host.
Connect to a single unix domain socket. Requires the unix keyword which must be an absolute path to the unix domain socket to connect to.

host is the usual connector used to establish a connection:

>>> C =
...  user = 'auser',
...  host = '',
...  port = 5432)
>>> # create
>>> db = C()
>>> # establish
>>> db.connect()

If a constant internet address is used, ip4 or ip6 can be used:

>>> C = postgresql.driver.default.ip4(user='auser', host='', port=5432)
>>> db = C()
>>> db.connect()

Additionally, db.connect() on db.__enter__() for with-statement support:

>>> with C() as db:
...  ...

Connectors are constant. They have no knowledge of PostgreSQL service files, environment variables or LDAP services, so changes made to those facilities will not be reflected in a connector’s configuration. If the latest information from any of these sources is needed, a new connector needs to be created as the credentials have changed.


host connectors use getaddrinfo(), so if DNS changes are made, new connections will use the latest information.

Connection Keywords

The following is a list of keywords accepted by connection creation interfaces:

The user to connect as.
The user’s password.
The name of the database to connect to. (PostgreSQL defaults it to user)
The hostname or IP address to connect to.
The port on the host to connect to.
The unix domain socket to connect to. Exclusive with host and port. Expects a string containing the absolute path to the unix domain socket to connect to.
A dictionary or key-value pair sequence stating the parameters to give to the database. These settings are included in the startup packet, and should be used carefully as when an invalid setting is given, it will cause the connection to fail.
Amount of time to wait for a connection to be made. (in seconds)
Hint given to the driver to properly encode password data and some information in the startup packet. This should only be used in cases where connections cannot be made due to authentication failures that occur while using known-correct credentials.
Don’t allow SSL connections.
Try without SSL first, but if that doesn’t work, try with.
Try SSL first, then without.
Require an SSL connection.
Certificate file path given to ssl.wrap_socket.
Key file path given to ssl.wrap_socket.
Root certificate file path given to ssl.wrap_socket
Revocation list file path. [Currently not checked.]

Connections and postgresql.driver.connect provide the means to establish a connection. Connections provide a postgresql.api.Database interface to a PostgreSQL server; specifically, a postgresql.api.Connection.

Connections are one-time objects. Once, it is closed or lost, it can longer be used to interact with the database provided by the server. If further use of the server is desired, a new connection must be established.


Cannot connect failures, exceptions raised on connect(), are also terminal.

In cases where operations are performed on a closed connection, a postgresql.exceptions.ConnectionDoesNotExistError will be raised.

Database Interface Points

After a connection is established:

>>> import postgresql
>>> db =

The methods and properties on the connection object are ready for use:

Create a postgresql.api.Statement object for querying the database. This provides an “SQL statement template” that can be executed multiple times. See Prepared Statements for more information.
Create a postgresql.api.StoredProcedure object referring to a stored procedure on the database. The returned object will provide a collections.Callable interface to the stored procedure on the server. See Stored Procedures for more information.
Create a postgresql.api.Statement object from an existing statement identifier. This is used in cases where the statement was prepared on the server. See Prepared Statements for more information.
Create a postgresql.api.Cursor object from an existing cursor identifier. This is used in cases where the cursor was declared on the server. See Cursors for more information., source)
Execute a DO statement on the server using the specified language. DO statements are available on PostgreSQL 9.0 and greater. Executing this method on servers that do not support DO statements will likely cause a SyntaxError.
Run a block of SQL on the server. This method returns None unless an error occurs. If errors occur, the processing of the statements will stop and the error will be raised.
Connection.xact(isolation = None, mode = None)
The postgresql.api.Transaction constructor for creating transactions. This method creates a transaction reference. The transaction will not be started until it’s instructed to do so. See Transactions for more information.
A property providing a collections.MutableMapping interface to the database’s SQL settings. See Settings for more information.
Create a new connection object based on the same factors that were used to create db. The new connection returned will already be connected.
By default, the msghook attribute does not exist. If set to a callable, any message that occurs during an operation of the database or an operation of a database derived object will be given to the callable. See the Database Messages section for more information.
Start listening for asynchronous notifications in the specified channels. Sends a batch of LISTEN statements to the server.
Stop listening for asynchronous notifications in the specified channels. Sends a batch of UNLISTEN statements to the server.
Return an iterator producing the channel names that are currently being listened to.
Connection.notify(*channels, **channel_and_payload)

NOTIFY the channels with the given payload. Sends a batch of NOTIFY statements to the server.

Equivalent to issuing “NOTIFY <channel>” or “NOTIFY <channel>, <payload>” for each item in channels and channel_and_payload. All NOTIFYs issued will occur in the same transaction, regardless of auto-commit.

The items in channels can either be a string or a tuple. If a string, no payload is given, but if an item is a builtins.tuple, the second item in the pair will be given as the payload, and the first as the channel. channels offers a means to issue NOTIFYs in guaranteed order:

>>> db.notify('channel1', ('different_channel', 'payload'))

In the above, NOTIFY "channel1"; will be issued first, followed by NOTIFY "different_channel", 'payload';.

The items in channel_and_payload are all payloaded NOTIFYs where the keys are the channels and the values are the payloads. Order is undefined:

>>> db.notify(channel_name = 'payload_data')

channels and channels_and_payload can be used together. In such cases all NOTIFY statements generated from channels_and_payload will follow those in channels.

Connection.iternotifies(timeout = None)
Return an iterator to the NOTIFYs received on the connection. The iterator will yield notification triples consisting of (channel, payload, pid). While iterating, the connection should not be used in other threads. The optional timeout can be used to enable “idle” events in which None objects will be yielded by the iterator. See Notification Management for details.

When a connection is established, certain pieces of information are collected from the backend. The following are the attributes set on the connection object after the connection is made:

The version string of the server; the result of SELECT version().
A sys.version_info form of the server_version setting. eg. (8, 1, 2, 'final', 0).
None if no security. 'ssl' if SSL is enabled.
The process-id of the backend process.
When backend was started. datetime.datetime instance.
The address of the client that the backend is communicating with.
The port of the client that the backend is communicating with.
Method to get the file descriptor number of the connection’s socket. This method will return None if the socket object does not have a fileno. Under normal circumstances, it will return an int.

The backend_start, client_address, and client_port are collected from pg_stat_activity. If this information is unavailable, the attributes will be None.

Prepared Statements

Prepared statements are the primary entry point for initiating an operation on the database. Prepared statement objects represent a request that will, likely, be sent to the database at some point in the future. A statement is a single SQL command.

The prepare entry point on the connection provides the standard method for creating a postgersql.api.Statement instance bound to the connection(db) from an SQL statement string:

>>> ps = db.prepare("SELECT 1")
>>> ps()

Statement objects may also be created from a statement identifier using the statement_from_id method on the connection. When this method is used, the statement must have already been prepared or an error will be raised.

>>> db.execute("PREPARE a_statement_id AS SELECT 1;")
>>> ps = db.statement_from_id('a_statement_id')
>>> ps()

When a statement is executed, it binds any given parameters to a new cursor and the entire result-set is returned.

Statements created using prepare() will leverage garbage collection in order to automatically close statements that are no longer referenced. However, statements created from pre-existing identifiers, statement_from_id, must be explicitly closed if the statement is to be discarded.

Statement objects are one-time objects. Once closed, they can no longer be used.

Statement Interface Points

Prepared statements can be executed just like functions:

>>> ps = db.prepare("SELECT 'hello, world!'")
>>> ps()
[('hello, world!',)]

The default execution method, __call__, produces the entire result set. It is the simplest form of statement execution. Statement objects can be executed in different ways to accommodate for the larger results or random access(scrollable cursors).

Prepared statement objects have a few execution methods:

As shown before, statement objects can be invoked like a function to get the statement’s results.
Return a iterator to all the rows produced by the statement. This method will stream rows on demand, so it is ideal for situations where each individual row in a large result-set must be processed.

Convenience interface that executes the rows() method without arguments. This enables the following syntax:

>>> for table_name, in db.prepare("SELECT table_name FROM information_schema.tables"):
...  print(table_name)
Return a iterator to the first column produced by the statement. This method will stream values on demand, and should only be used with statements that have a single column; otherwise, bandwidth will ultimately be wasted as the other columns will be dropped. This execution method cannot be used with COPY statements.

For simple statements, cursor objects are unnecessary. Consider the data contained in c from above, ‘hello world!’. To get at this data directly from the __call__(...) method, it looks something like:

>>> ps = db.prepare("SELECT 'hello, world!'")
>>> ps()[0][0]
'hello, world!'

To simplify access to simple data, the first method will simply return the “first” of the result set:

>>> ps.first()
'hello, world!'
The first value.
When the result set consists of a single column, first() will return that column in the first row.
The first row.
When the result set consists of multiple columns, first() will return that first row.
The first, and only, row count.

When DML–for instance, an INSERT-statement–is executed, first() will return the row count returned by the statement as an integer.


DML that returns row data, RETURNING, will not return a row count.

The result set created by the statement determines what is actually returned. Naturally, a statement used with first() should be crafted with these rules in mind.

This access point is designed for situations where rows are being streamed out quickly. It is a method that returns a collections.Iterator that produces sequences of rows. This is the most efficient way to get rows from the database. The rows in the sequences are builtins.tuple objects.
Create a scrollable cursor with hold. This returns a postgresql.api.Cursor ready for accessing random rows in the result-set. Applications that use the database to support paging can use this method to manage the view.
Close the statement inhibiting further use.
Given an iterable producing parameters, execute the statement for each iteration. Always returns None.

Given an iterable of iterables producing parameters, execute the statement for each parameter produced. However, send the all execution commands with the corresponding parameters of each chunk before reading any results. Always returns None. This access point is designed to be used in conjunction with Statement.chunks() for transferring rows from one connection to another with great efficiency:

>>> dst.prepare(...).load_chunks(src.prepare(...).chunks())
Create a new statement object based on the same factors that were used to create ps.
By default, the msghook attribute does not exist. If set to a callable, any message that occurs during an operation of the statement or an operation of a statement derived object will be given to the callable. See the Database Messages section for more information.

In order to provide the appropriate type transformations, the driver must acquire metadata about the statement’s parameters and results. This data is published via the following properties on the statement object:

A sequence of SQL type names specifying the types of the parameters used in the statement.
A sequence of SQL type names specifying the types of the columns produced by the statement. None if the statement does not return row-data.
A sequence of PostgreSQL type Oid’s specifying the types of the parameters used in the statement.
A sequence of PostgreSQL type Oid’s specifying the types of the columns produced by the statement. None if the statement does not return row-data.
A sequence of Python types that the statement expects.
A sequence of Python types that the statement will produce.
A sequence of str objects specifying the names of the columns produced by the statement. None if the statement does not return row-data.

The indexes of the parameter sequences correspond to the parameter’s identifier, N+1: sql_parameter_types[0] -> '$1'.

>>> ps = db.prepare("SELECT $1::integer AS intname, $2::varchar AS chardata")
>>> ps.sql_parameter_types
>>> ps.sql_column_types
>>> ps.column_names
>>> ps.column_types
(<class 'int'>, <class 'str'>)

Parameterized Statements

Statements can take parameters. Using statement parameters is the recommended way to interrogate the database when variable information is needed to formulate a complete request. In order to do this, the statement must be defined using PostgreSQL’s positional parameter notation. $1, $2, $3, etc:

>>> ps = db.prepare("SELECT $1")
>>> ps('hello, world!')[0][0]
'hello, world!'

PostgreSQL determines the type of the parameter based on the context of the parameter’s identifier:

>>> ps = db.prepare(
...  "SELECT * FROM information_schema.tables WHERE table_name = $1 LIMIT $2"
... )
>>> ps("tables", 1)
[('postgres', 'information_schema', 'tables', 'VIEW', None, None, None, None, None, 'NO', 'NO', None)]

Parameter $1 in the above statement will take on the type of the table_name column and $2 will take on the type required by the LIMIT clause(text and int8).

However, parameters can be forced to a specific type using explicit casts:

>>> ps = db.prepare("SELECT $1::integer")
>>> ps.first(-400)

Parameters are typed. PostgreSQL servers provide the driver with the type information about a positional parameter, and the serialization routine will raise an exception if the given object is inappropriate. The Python types expected by the driver for a given SQL-or-PostgreSQL type are listed in Type Support.

This usage of types is not always convenient. Notably, the datetime module does not provide a friendly way for a user to express intervals, dates, or times. There is a likely inclination to forego these parameter type requirements.

In such cases, explicit casts can be made to work-around the type requirements:

>>> ps = db.prepare("SELECT $1::text::date")
>>> ps.first('yesterday'), 3, 11)

The parameter, $1, is given to the database as a string, which is then promptly cast into a date. Of course, without the explicit cast as text, the outcome would be different:

>>> ps = db.prepare("SELECT $1::date")
>>> ps.first('yesterday')

The function that processes the parameter expects a object, and the given str object does not provide the necessary interfaces for the conversion, so the driver raises a postgresql.exceptions.ParameterError from the original conversion exception.

Inserting and DML

Loading data into the database is facilitated by prepared statements. In these examples, a table definition is necessary for a complete illustration:

>>> db.execute(
...     """
... CREATE TABLE employee (
...     employee_name text,
...     employee_salary numeric,
...     employee_dob date,
...     employee_hire_date date
... );
...     """
... )

Create an INSERT statement using prepare:

>>> mkemp = db.prepare("INSERT INTO employee VALUES ($1, $2, $3, $4)")

And add “Mr. Johnson” to the table:

>>> import datetime
>>> r = mkemp(
...     "John Johnson",
...     "92000",
..., 12, 10),
..., 4, 23)
... )
>>> print(r[0])
>>> print(r[1])

The execution of DML will return a tuple. This tuple contains the completed command name and the associated row count.

Using the call interface is fine for making a single insert, but when multiple records need to be inserted, it’s not the most efficient means to load data. For multiple records, the ps.load_rows([...]) provides an efficient way to load large quantities of structured data:

>>> from datetime import date
>>> mkemp.load_rows([
...  ("Jack Johnson", "85000", date(1962, 11, 23), date(1990, 3, 5)),
...  ("Debra McGuffer", "52000", date(1973, 3, 4), date(2002, 1, 14)),
...  ("Barbara Smith", "86000", date(1965, 2, 24), date(2005, 7, 19)),
... ])

While small, the above illustrates the ps.load_rows() method taking an iterable of tuples that provides parameters for the each execution of the statement.

load_rows is also used to support COPY ... FROM STDIN statements:

>>> copy_emps_in = db.prepare("COPY employee FROM STDIN")
>>> copy_emps_in.load_rows([
...  b'Emp Name1\t72000\t1970-2-01\t1980-10-22\n',
...  b'Emp Name2\t62000\t1968-9-11\t1985-11-1\n',
...  b'Emp Name3\t62000\t1968-9-11\t1985-11-1\n',
... ])

Copy data goes in as bytes and come out as bytes regardless of the type of COPY taking place. It is the user’s obligation to make sure the row-data is in the appropriate encoding.

COPY Statements

postgresql.driver transparently supports PostgreSQL’s COPY command. To the user, COPY will act exactly like other statements that produce tuples; COPY tuples, however, are bytes objects. The only distinction in usability is that the COPY should be completed before other actions take place on the connection–this is important when a COPY is invoked via rows() or chunks().

In situations where other actions are invoked during a COPY TO STDOUT, the entire result set of the COPY will be read. However, no error will be raised so long as there is enough memory available, so it is very desirable to avoid doing other actions on the connection while a COPY is active.

In situations where other actions are invoked during a COPY FROM STDIN, a COPY failure error will occur. The driver manages the connection state in such a way that will purposefully cause the error as the COPY was inappropriately interrupted. This not usually a problem as load_rows(...) and load_chunks(...) methods must complete the COPY command before returning.

Copy data is always transferred using bytes objects. Even in cases where the COPY is not in BINARY mode. Any needed encoding transformations must be made the caller. This is done to avoid any unnecessary overhead by default:

>>> ps = db.prepare("COPY (SELECT i FROM generate_series(0, 99) AS g(i)) TO STDOUT")
>>> r = ps()
>>> len(r)
>>> r[0]
>>> r[-1]

Of course, invoking a statement that way will read the entire result-set into memory, which is not usually desirable for COPY. Using the chunks(...) iterator is the fastest way to move data:

>>> ci = ps.chunks()
>>> import sys
>>> for rowset in ps.chunks():
...  sys.stdout.buffer.writelines(rowset)
<lots of data>

COPY FROM STDIN commands are supported via postgresql.api.Statement.load_rows. Each invocation to load_rows is a single invocation of COPY. load_rows takes an iterable of COPY lines to send to the server:

>>> db.execute("""
... CREATE TABLE sample_copy (
...     sc_number int,
...     sc_text text
... );
... """)
>>> copyin = db.prepare('COPY sample_copy FROM STDIN')
>>> copyin.load_rows([
...     b'123\tone twenty three\n',
...     b'350\ttree fitty\n',
... ])

For direct connection-to-connection COPY, use of load_chunks(...) is recommended as it will provide the most efficient transfer method:

>>> copyout = src.prepare('COPY atable TO STDOUT')
>>> copyin = dst.prepare('COPY atable FROM STDIN')
>>> copyin.load_chunks(copyout.chunks())

Specifically, each chunk of row data produced by chunks() will be written in full by load_chunks() before getting another chunk to write.


When a prepared statement is declared, ps.declare(...), a postgresql.api.Cursor is created and returned for random access to the rows in the result set. Direct use of cursors is primarily useful for applications that need to implement paging. For situations that need to iterate over the result set, the ps.rows(...) or ps.chunks(...) execution methods should be used.

Cursors can also be created directly from cursor_id‘s using the cursor_from_id method on connection objects:

>>> db.execute('DECLARE the_cursor_id CURSOR WITH HOLD FOR SELECT 1;')
>>> c = db.cursor_from_id('the_cursor_id')
>>> c.close()


If the cursor that needs to be opened is going to be treated as an iterator, then a FETCH-statement should be prepared instead using cursor_from_id.

Like statements created from an identifier, cursors created from an identifier must be explicitly closed in order to destroy the object on the server. Likewise, cursors created from statement invocations will be automatically released when they are no longer referenced.


PG-API cursors are a direct interface to single result-set SQL cursors. This is in contrast with DB-API cursors, which have interfaces for dealing with multiple result-sets. There is no execute method on PG-API cursors.

Cursor Interface Points

For cursors that return row data, these interfaces are provided for accessing those results: = None, direction = None)

This method name is borrowed from file objects, and are semantically similar. However, this being a cursor, rows are returned instead of bytes or characters. When the number of rows returned is less then the quantity requested, it means that the cursor has been exhausted in the configured direction. The direction argument can be either 'FORWARD' or True to FETCH FORWARD, or 'BACKWARD' or False to FETCH BACKWARD.

Like, seek(), the direction property on the cursor object effects this method.[, whence = 0])
When the cursor is scrollable, this seek interface can be used to move the position of the cursor. See Scrollable Cursors for more information.
This fetches the next row in the cursor object. Cursors support the iterator protocol. While equivalent to[0], StopIteration is raised if the returned sequence is empty. (__next__())
For cursors opened using cursor_from_id(), this method must be called in order to CLOSE the cursor. For cursors created by invoking a prepared statement, this is not necessary as the garbage collection interface will take the appropriate steps.
Create a new cursor object based on the same factors that were used to create c.
By default, the msghook attribute does not exist. If set to a callable, any message that occurs during an operation of the cursor will be given to the callable. See the Database Messages section for more information.

Cursors have some additional configuration properties that may be modified during the use of the cursor:

A value of True, the default, will cause read to fetch forwards, whereas a value of False will cause it to fetch backwards. 'BACKWARD' and 'FORWARD' can be used instead of False and True.

Cursors normally share metadata with the statements that create them, so it is usually unnecessary for referencing the cursor’s column descriptions directly. However, when a cursor is opened from an identifier, the cursor interface must collect the metadata itself. These attributes provide the metadata in absence of a statement object:

A sequence of SQL type names specifying the types of the columns produced by the cursor. None if the cursor does not return row-data.
A sequence of PostgreSQL type Oid’s specifying the types of the columns produced by the cursor. None if the cursor does not return row-data.
A sequence of Python types that the cursor will produce.
A sequence of str objects specifying the names of the columns produced by the cursor. None if the cursor does not return row-data.
The statement that was executed that created the cursor. None if unknown–db.cursor_from_id().

Scrollable Cursors

Scrollable cursors are supported for applications that need to implement paging. When statements are invoked via the declare(...) method, the returned cursor is scrollable.


Scrollable cursors never pre-fetch in order to provide guaranteed positioning.

The cursor interface supports scrolling using the seek method. Like read, it is semantically similar to a file object’s seek().

seek takes two arguments: position and whence:

The position to scroll to. The meaning of this is determined by whence.

How to use the position: absolute, relative, or absolute from end:

absolute: 'ABSOLUTE' or 0 (default)
seek to the absolute position in the cursor relative to the beginning of the cursor.
relative: 'RELATIVE' or 1
seek to the relative position. Negative position‘s will cause a MOVE backwards, while positive position‘s will MOVE forwards.
from end: 'FROM_END' or 2
seek to the end of the cursor and then MOVE backwards by the given position.

The whence keyword argument allows for either numeric and textual specifications.

Scrolling through employees:

>>> emps_by_age = db.prepare("""
...     employee_name, employee_salary, employee_dob, employee_hire_date,
...     EXTRACT(years FROM AGE(employee_dob)) AS age
... ORDER BY age ASC
... """)
>>> c = emps_by_age.declare()
>>> # seek to the end, ``2`` works as well.
>>>, 'FROM_END')
>>> # scroll back one, ``1`` works as well.
>>>, 'RELATIVE')
>>> # and back to the beginning again

Additionally, scrollable cursors support backward fetches by specifying the direction keyword argument:

>>>, 2)
>>>, 'BACKWARD')

Cursor Direction

The direction property on the cursor states the default direction for read and seek operations. Normally, the direction is True, 'FORWARD'. When the property is set to 'BACKWARD' or False, the read method will fetch backward by default, and seek operations will be inverted to simulate a reversely ordered cursor. The following example illustrates the effect:

>>> reverse_c = db.prepare('SELECT i FROM generate_series(99, 0, -1) AS g(i)').declare()
>>> c = db.prepare('SELECT i FROM generate_series(0, 99) AS g(i)').declare()
>>> reverse_c.direction = 'BACKWARD'
>>> ==

Furthermore, when the cursor is configured to read backwards, specifying 'BACKWARD' for read’s direction argument will ultimately cause a forward fetch. This potentially confusing facet of direction configuration is implemented in order to create an appropriate symmetry in functionality. The cursors in the above example contain the same rows, but are ultimately in reverse order. The backward direction property is designed so that the effect of any read or seek operation on those cursors is the same:

>>> ==
>>>, 'BACKWARD') ==, 'BACKWARD')

And for relative seeks:

>>>, 1)
>>>, 1)
>>>, 'BACKWARD') ==, 'BACKWARD')


Rows received from PostgreSQL are instantiated into postgresql.types.Row objects. Rows are both a sequence and a mapping. Items accessed with an int are seen as indexes and other objects are seen as keys:

>>> row = db.prepare("SELECT 't'::text AS col0, 2::int4 AS col1").first()
>>> row
('t', 2)
>>> row[0]
>>> row["col0"]

However, this extra functionality is not free. The cost of instantiating postgresql.types.Row objects is quite measurable, so the chunks() execution method will produce builtins.tuple objects for cases where performance is critical.


Attributes aren’t used to provide access to values due to potential conflicts with existing method and property names.

Row Interface Points

Rows implement the collections.Mapping and collections.Sequence interfaces.

An iterable producing the column names. Order is not guaranteed. See the column_names property to get an ordered sequence.
Iterable to the values in the row.
Row.get(key_or_index[, default=None])
Get the item in the row. If the key doesn’t exist or the index is out of range, return the default.
Iterable of key-value pairs. Ordered by index.
Iterable to the values in index order.
value in Row
Whether or not the value exists in the row. (__contains__)
If key_or_index is an integer, return the value at that index. If the index is out of range, raise an IndexError. Otherwise, return the value associated with column name. If the given key, key_or_index, does not exist, raise a KeyError.
Return the index associated with the given key.
Return the key associated with the given index.
Row.transform(*args, **kw)
Create a new row object of the same length, with the same keys, but with new values produced by applying the given callables to the corresponding items. Callables given as args will be associated with values by their index and callables given as keywords will be associated with values by their key, column name.

While the mapping interfaces will provide most of the needed information, some additional properties are provided for consistency with statement and cursor objects.


Property providing an ordered sequence of column names. The index corresponds to the row value-index that the name refers to.

>>> row[row.column_names[i]] == row[i]

Row Transformations

After a row is returned, sometimes the data in the row is not in the desired format. Further processing is needed if the row object is to going to be given to another piece of code which requires an object of differring consistency.

The transform method on row objects provides a means to create a new row object consisting of the old row’s items, but with certain columns transformed using the given callables:

>>> row = db.prepare("""
...   'XX9301423'::text AS product_code,
...   2::int4 AS quantity,
...   '4.92'::numeric AS total
... """).first()
>>> row
('XX9301423', 2, Decimal("4.92"))
>>> row.transform(quantity = str)
('XX9301423', '2', Decimal("4.92"))

transform supports both positional and keyword arguments in order to assign the callable for a column’s transformation:

>>> from operator import methodcaller
>>> row.transform(methodcaller('strip', 'XX'))
('9301423', 2, Decimal("4.92"))

Of course, more than one column can be transformed:

>>> stripxx = methodcaller('strip', 'XX')
>>> row.transform(stripxx, str, str)
('9301423', '2', '4.92')

None can also be used to indicate no transformation:

>>> row.transform(None, str, str)
('XX9301423', '2', '4.92')

More advanced usage can make use of lambdas for compound transformations in a single pass of the row:

>>> strip_and_int = lambda x: int(stripxx(x))
>>> row.transform(strip_and_int)
(9301423, 2, Decimal("4.92"))

Transformations will be, more often than not, applied against rows as opposed to a row. Using operator.methodcaller with map provides the necessary functionality to create simple iterables producing transformed row sequences:

>>> import decimal
>>> apply_tax = lambda x: (x * decimal.Decimal("0.1")) + x
>>> transform_row = methodcaller('transform', strip_and_int, None, apply_tax)
>>> r = map(transform_row, [row])
>>> list(r)
[(9301423, 2, Decimal('5.412'))]

And finally, functools.partial can be used to create a simple callable:

>>> from functools import partial
>>> transform_rows = partial(map, transform_row)
>>> list(transform_rows([row]))
[(9301423, 2, Decimal('5.412'))]


Queries in py-postgresql are single use prepared statements. They exist primarily for syntactic convenience, but they also allow the driver to recognize the short lifetime of the statement.

Single use statements are supported using the query property on connection objects, postgresql.api.Connection.query. The statement object is not available when using queries as the results, or handle to the results, are directly returned.

Queries have access to all execution methods:

  • Connection.query(sql, *parameters)
  • Connection.query.rows(sql, *parameters)
  • Connection.query.column(sql, *parameters)
  • Connection.query.first(sql, *parameters)
  • Connection.query.chunks(sql, *parameters)
  • Connection.query.declare(sql, *parameters)
  • Connection.query.load_rows(sql, collections.Iterable(parameters))
  • Connection.query.load_chunks(collections.Iterable(collections.Iterable(parameters)))

In cases where a sequence of one-shot queries needs to be performed, it may be important to avoid unnecessary repeat attribute resolution from the connection object as the query property is an interface object created on access. Caching the target execution methods is recommended:

qrows = db.query.rows
l = []
for x in my_queries:

The characteristic of Each execution method is discussed in the prior Prepared Statements section.

Stored Procedures

The proc method on postgresql.api.Database objects provides a means to create a reference to a stored procedure on the remote database. postgresql.api.StoredProcedure objects are used to represent the referenced SQL routine.

This provides a direct interface to functions stored on the database. It leverages knowledge of the parameters and results of the function in order to provide the user with a natural interface to the procedure:

>>> func = db.proc('version()')
>>> func()
'PostgreSQL 8.3.6 on ...'

Stored Procedure Interface Points

It’s more-or-less a function, so there’s only one interface point:

func(*args, **kw) (__call__)

Stored procedure objects are callable, executing a procedure will return an object of suitable representation for a given procedure’s type signature.

If it returns a single object, it will return the single object produced by the procedure.

If it’s a set returning function, it will return an iterable to the values produced by the procedure.

In cases of set returning function with multiple OUT-parameters, a cursor will be returned.

Stored Procedure Type Support

Stored procedures support most types of functions. “Function Types” being set returning functions, multiple-OUT parameters, and simple single-object returns.

Set-returning functions, SRFs return a sequence:

>>> generate_series = db.proc('generate_series(int,int)')
>>> gs = generate_series(1, 20)
>>> gs
<generator object <genexpr>>
>>> next(gs)
>>> list(gs)
[2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20]

For functions like generate_series(), the driver is able to identify that the return is a sequence of solitary integer objects, so the result of the function is just that, a sequence of integers.

Functions returning composite types are recognized, and return row objects:

>>> db.execute("""
... CREATE FUNCTION composite(OUT i int, OUT t text)
... $body$
...  SELECT 900::int AS i, 'sample text'::text AS t;
... $body$;
... """)
>>> composite = db.proc('composite()')
>>> r = composite()
>>> r
(900, 'sample text')
>>> r['i']
>>> r['t']
'sample text'

Functions returning a set of composites are recognized, and the result is a postgresql.api.Cursor object whose column names are consistent with the names of the OUT parameters:

>>> db.execute("""
... CREATE FUNCTION srfcomposite(out i int, out t text)
... $body$
...  SELECT 900::int AS i, 'sample text'::text AS t
...  SELECT 450::int AS i, 'more sample text'::text AS t
... $body$;
... """)
>>> srfcomposite = db.proc('srfcomposite()')
>>> r = srfcomposite()
>>> next(r)
(900, 'sample text')
>>> v = next(r)
>>> v['i'], v['t']
(450, 'more sample text')


Transactions are managed by creating an object corresponding to a transaction started on the server. A transaction is a transaction block, a savepoint, or a prepared transaction. The xact(...) method on the connection object provides the standard method for creating a postgresql.api.Transaction object to manage a transaction on the connection.

The creation of a transaction object does not start the transaction. Rather, the transaction must be explicitly started using the start() method on the transaction object. Usually, transactions should be managed with the context manager interfaces:

>>> with db.xact():
...  ...

The transaction in the above example is opened, started, by the __enter__ method invoked by the with-statement’s usage. It will be subsequently committed or rolled-back depending on the exception state and the error state of the connection when __exit__ is called.

Using the with-statement syntax for managing transactions is strongly recommended. By using the transaction’s context manager, it allows for Python exceptions to be properly treated as fatal to the transaction as when an uncaught exception of any kind occurs within the block, it is unlikely that the state of the transaction can be trusted. Additionally, the __exit__ method provides a safe-guard against invalid commits. This can occur if a database error is inappropriately caught within a block without being raised.

The context manager interfaces are higher level interfaces to the explicit instruction methods provided by postgresql.api.Transaction objects.

Transaction Configuration

Keyword arguments given to xact() provide the means for configuring the properties of the transaction. Only three points of configuration are available:


The isolation level of the transaction. This must be a string. It will be interpolated directly into the START TRANSACTION statement. Normally, ‘SERIALIZABLE’ or ‘READ COMMITTED’:

>>> with db.xact('SERIALIZABLE'):
...  ...
A string, ‘READ ONLY’ or ‘READ WRITE’. States the mutability of stored information in the database. Like isolation, this is interpolated directly into the START TRANSACTION string.

The specification of any of these transaction properties imply that the transaction is a block. Savepoints do not take configuration, so if a transaction identified as a block is started while another block is running, an exception will be raised.

Transaction Interface Points

The methods available on transaction objects manage the state of the transaction and relay any necessary instructions to the remote server in order to reflect that change of state.

>>> x = db.xact(...)
Start the transaction.
Commit the transaction.
Abort the transaction.

These methods are primarily provided for applications that manage transactions in a way that cannot be formed around single, sequential blocks of code. Generally, using these methods require additional work to be performed by the code that is managing the transaction. If usage of these direct, instructional methods is necessary, it is important to note that if the database is in an error state when a transaction block’s commit() is executed, an implicit rollback will occur. The transaction object will simply follow instructions and issue the COMMIT statement, and it will succeed without exception.

Error Control

Handling database errors inside transaction CMs is generally discouraged as any database operation that occurs within a failed transaction is an error itself. It is important to trap any recoverable database errors outside of the scope of the transaction’s context manager:

>>> try:
...  with db.xact():
...   ...
... except postgresql.exceptions.UniqueError:
...  pass

In cases where the database is in an error state, but the context exits without an exception, a postgresql.exceptions.InFailedTransactionError is raised by the driver:

>>> with db.xact():
...  try:
...   ...
...  except postgresql.exceptions.UniqueError:
...   pass
Traceback (most recent call last):
postgresql.exceptions.InFailedTransactionError: invalid block exit detected
CODE: 25P02

Normally, if a COMMIT is issued on a failed transaction, the command implies a ROLLBACK without error. This is a very undesirable result for the CM’s exit as it may allow for code to be ran that presumes the transaction was committed. The driver intervenes here and raises the postgresql.exceptions.InFailedTransactionError to safe-guard against such cases. This effect is consistent with savepoint releases that occur during an error state. The distinction between the two cases is made using the source property on the raised exception.


SQL’s SHOW and SET provides a means to configure runtime parameters on the database(“GUC”s). In order to save the user some grief, a collections.MutableMapping interface is provided to simplify configuration.

The settings attribute on the connection provides the interface extension.

The standard dictionary interface is supported:

>>> db.settings['search_path'] = "$user,public"

And update(...) is better performing for multiple sets:

>>> db.settings.update({
...  'search_path' : "$user,public",
...  'default_statistics_target' : "1000"
... })


The transaction_isolation setting cannot be set using the settings mapping. Internally, settings uses set_config, which cannot adjust that particular setting.

Settings Interface Points

Manipulation and interrogation of the connection’s settings is achieved by using the standard collections.MutableMapping interfaces.

Get the value of a single setting.
Connection.settings[k] = v
Set the value of a single setting.
Connection.settings.update([(k1,v2), (k2,v2), ..., (kn,vn)])
Set multiple settings using a sequence of key-value pairs.
Connection.settings.update({k1 : v1, k2 : v2, ..., kn : vn})
Set multiple settings using a dictionary or mapping object.
Connection.settings.getset([k1, k2, ..., kn])
Get a set of a settings. This is the most efficient way to get multiple settings as it uses a single request.
Get all available setting names.
Get all setting values.
Get a sequence of key-value pairs corresponding to all settings on the database.

Settings Management

postgresql.api.Settings objects can create context managers when called. This gives the user with the ability to specify sections of code that are to be ran with certain settings. The settings’ context manager takes full advantage of keyword arguments in order to configure the context manager:

>>> with db.settings(search_path = 'local,public', timezone = 'mst'):
...  ...

postgresql.api.Settings objects are callable; the return is a context manager configured with the given keyword arguments representing the settings to use for the block of code that is about to be executed.

When the block exits, the settings will be restored to the values that they had before the block entered.

Type Support

The driver supports a large number of PostgreSQL types at the binary level. Most types are converted to standard Python types. The remaining types are usually PostgreSQL specific types that are converted into objects whose class is defined in postgresql.types.

When a conversion function is not available for a particular type, the driver will use the string format of the type and instantiate a str object for the data. It will also expect str data when parameter of a type without a conversion function is bound.


Generally, these standard types are provided for convenience. If conversions into these datatypes are not desired, it is recommended that explicit casts into text are made in statement string.

Python types used to represent PostgreSQL types.
PostgreSQL Types Python Types SQL Types
postgresql.types.INT2OID int smallint
postgresql.types.INT4OID int integer
postgresql.types.INT8OID int bigint
postgresql.types.FLOAT4OID float float
postgresql.types.FLOAT8OID float double
postgresql.types.VARCHAROID str varchar
postgresql.types.BPCHAROID str char
postgresql.types.XMLOID xml.etree (cElementTree) xml
postgresql.types.DATEOID date
postgresql.types.TIMESTAMPOID datetime.datetime timestamp
postgresql.types.TIMESTAMPTZOID datetime.datetime (tzinfo) timestamptz
postgresql.types.TIMEOID datetime.time time
postgresql.types.TIMETZOID datetime.time timetz
postgresql.types.INTERVALOID datetime.timedelta interval
postgresql.types.NUMERICOID decimal.Decimal numeric
postgresql.types.BYTEAOID bytes bytea
postgresql.types.TEXTOID str text
<contrib_hstore> dict hstore

The mapping in the above table normally goes both ways. So when a parameter is passed to a statement, the type should be consistent with the corresponding Python type. However, many times, for convenience, the object will be passed through the type’s constructor, so it is not always necessary.


Arrays of PostgreSQL types are supported with near transparency. For simple arrays, arbitrary iterables can just be given as a statement’s parameter and the array’s constructor will consume the objects produced by the iterator into a postgresql.types.Array instance. However, in situations where the array has multiple dimensions, list objects are used to delimit the boundaries of the array.

>>> ps = db.prepare("select $1::int[]")
>>> ps.first([(1,2), (2,3)])

In the above case, it is apparent that this array is supposed to have two dimensions. However, this is not the case for other types:

>>> ps = db.prepare("select $1::point[]")
>>> ps.first([(1,2), (2,3)])
postgresql.types.Array([postgresql.types.point((1.0, 2.0)), postgresql.types.point((2.0, 3.0))])

Lists are used to provide the necessary boundary information:

>>> ps = db.prepare("select $1::int[]")
>>> ps.first([[1,2],[2,3]])

The above is the appropriate way to define the array from the original example.


The root-iterable object given as an array parameter does not need to be a list-type as it’s assumed to be made up of elements.


Composites are supported using postgresql.types.Row objects to represent the data. When a composite is referenced for the first time, the driver queries the database for information about the columns that make up the type. This information is then used to create the necessary I/O routines for packing and unpacking the parameters and columns of that type:

>>> db.execute("CREATE TYPE ctest AS (i int, t text, n numeric);")
>>> ps = db.prepare("SELECT $1::ctest")
>>> i = (100, 'text', "100.02013")
>>> r = ps.first(i)
>>> r["t"]
>>> r["n"]

Or if use of a dictionary is desired:

>>> r = ps.first({'t' : 'just-the-text'})
>>> r
(None, 'just-the-text', None)

When a dictionary is given to construct the row, absent values are filled with None.

Database Messages

By default, py-postgresql gives detailed reports of messages emitted by the database. Often, the verbosity is excessive due to single target processes or existing application infrastructure for tracing the sources of various events. Normally, this verbosity is not a significant problem as the driver defaults the client_min_messages setting to 'WARNING' by default.

However, if NOTICE or INFO messages are needed, finer grained control over message propagation may be desired, py-postgresql’s object relationship model provides a common protocol for controlling message propagation and, ultimately, display.

The msghook attribute on elements–for instance, Statements, Connections, and Connectors–is absent by default. However, when present on an object that contributed the cause of a message event, it will be invoked with the Message, postgresql.message.Message, object as its sole parameter. The attribute of the object that is closest to the event is checked first, if present it will be called. If the msghook() call returns a True value(specficially, bool(x) is True), the message will not be propagated any further. However, if a False value–notably, None–is returned, the next element is checked until the list is exhausted and the message is given to postgresql.sys.msghook. The normal list of elements is as follows:

Output → Statement → Connection → Connector → Driver → postgresql.sys

Where Output can be a postgresql.api.Cursor object produced by declare(...) or an implicit output management object used internally by Statement.__call__() and other statement execution methods. Setting the msghook attribute on postgresql.api.Statement gives very fine control over raised messages. Consider filtering the notice message on create table statements that implicitly create indexes:

>>> db =
>>> db.settings['client_min_messages'] = 'NOTICE'
>>> ct_this = db.prepare('CREATE TEMP TABLE "this" (i int PRIMARY KEY)')
>>> ct_that = db.prepare('CREATE TEMP TABLE "that" (i int PRIMARY KEY)')
>>> def filter_notices(msg):
...  if msg.details['severity'] == 'NOTICE':
...   return True
>>> ct_that()
NOTICE: CREATE TABLE / PRIMARY KEY will create implicit index "that_pkey" for table "that"
>>> ct_this.msghook = filter_notices
>>> ct_this()

The above illustrates the quality of an installed msghook that simply inhibits further propagation of messages with a severity of ‘NOTICE’–but, only notices coming from objects derived from the ct_this postgresql.api.Statement object.

Subsequently, if the filter is installed on the connection’s msghook:

>>> db =
>>> db.settings['client_min_messages'] = 'NOTICE'
>>> ct_this = db.prepare('CREATE TEMP TABLE "this" (i int PRIMARY KEY)')
>>> ct_that = db.prepare('CREATE TEMP TABLE "that" (i int PRIMARY KEY)')
>>> def filter_notices(msg):
...  if msg.details['severity'] == 'NOTICE':
...   return True
>>> db.msghook = filter_notices
>>> ct_that()
>>> ct_this()

Any message with 'NOTICE' severity coming from the connection, db, will be suffocated by the filter_notices function. However, if a msghook is installed on either of those statements, it would be possible for display to occur depending on the implementation of the hook installed on the statement objects.

Message Metadata

PostgreSQL messages, postgresql.message.Message, are primarily described in three parts: the SQL-state code, the main message string, and a mapping containing the details. The follow attributes are available on message objects:

The primary message string.
The SQL-state code associated with a given message.
The origins of the message. Normally, 'SERVER' or 'CLIENT'.
A terse, textual representation of 'file', 'line', and 'function' provided by the associated details.

A mapping providing extended information about a message. This mapping object can contain the following keys:

Any of 'DEBUG', 'INFO', 'NOTICE', 'WARNING', 'ERROR', 'FATAL', or 'PANIC'; the latter three are usually associated with a postgresql.exceptions.Error instance.
The CONTEXT portion of the message.
The DETAIL portion of the message.
The HINT portion of the message.
A number identifying the position in the statement string that caused a parse error.
The name of the file that emitted the message. (normally server information)
The name of the function that emitted the message. (normally server information)
The line of the file that emitted the message. (normally server information)