pgbouncer — a Postgres Pro connection pooler
On Linux systems:
pgbouncer [ -d ] [ -R ] [ -v ] [ -u user ] pgbouncer.ini
pgbouncer -V | -h
On Windows:
pgbouncer [ -v ] [ -u user ] pgbouncer.ini
pgbouncer -V | -h
To use pgbouncer as a Windows service:
pgbouncer.exe --regservice pgbouncer.ini
pgbouncer.exe --unregservice pgbouncer.ini
pgbouncer is a Postgres Pro connection pooler. Any target application can be connected to pgbouncer as if it were a Postgres Pro server, and pgbouncer will create a connection to the actual server, or it will reuse one of its existing connections.
The aim of pgbouncer is to lower the performance impact of opening new connections to Postgres Pro.
In order not to compromise transaction semantics for connection pooling, pgbouncer supports several types of pooling when rotating connections:
Most polite method. When a client connects, a server connection will be assigned to it for the whole duration the client stays connected. When the client disconnects, the server connection will be put back into the pool. This is the default method.
A server connection is assigned to a client only during a transaction. When pgbouncer notices that transaction is over, the server connection will be put back into the pool.
Most aggressive method. The server connection will be put back into the pool immediately after a query completes. Multi-statement transactions are disallowed in this mode as they would break.
The administration interface of
pgbouncer consists of some new
SHOW commands available when connected to a
special “virtual” database
pgbouncer.
pgbouncer is provided with Postgres Pro Enterprise
as a separate pre-built package pgbouncer
(for the detailed installation instructions, see Chapter 17).
Basic setup and usage is as follows.
Create a pgbouncer.ini file. Details
in the pgbouncer(5) man page. Simple example:
[databases] template1 = host=localhost dbname=template1 auth_user=someuser [pgbouncer] listen_port = 6432 listen_addr = localhost auth_type = md5 auth_file = userlist.txt logfile = pgbouncer.log pidfile = pgbouncer.pid admin_users = someuser
Create a userlist.txt file that contains the users allowed in:
"someuser" "same_password_as_in_server"
Launch pgbouncer:
$ pgbouncer -d pgbouncer.ini
The above command does not work on Windows systems. Instead, pgbouncer must be launched as a service that first needs to be registered, as follows:
pgbouncer --regservice
Have your application (or the
psql client) connect to
pgbouncer instead of
directly to the Postgres Pro server:
$ psql -p 6432 -U someuser template1
Manage pgbouncer by
connecting to the special administration database
pgbouncer and issuing
SHOW HELP; to begin:
$ psql -p 6432 -U someuser pgbouncer pgbouncer=# SHOW HELP; NOTICE: Console usage DETAIL: SHOW [HELP|CONFIG|DATABASES|FDS|POOLS|CLIENTS|SERVERS|SOCKETS|LISTS|VERSION|...] SET key = arg RELOAD PAUSE SUSPEND RESUME SHUTDOWN [...]
If you made changes to the pgbouncer.ini file, you can reload
it with:
pgbouncer=# RELOAD;
-d, --daemon
Run in the background. Without it, the process will run in
the foreground. In daemon mode, setting pidfile
as well as logfile or syslog
is required. No log messages will be written to stderr
after going into the background.
Does not work on Windows, pgbouncer needs to run as service there.
-R, --reboot
This option is deprecated.
Instead of this option use a rolling restart with multiple
pgbouncer processes listening on the same port using so_reuseport instead.
Do an online restart. That means connecting to the running process, loading the open sockets from it, and then using them. If there is no active process, boot normally.
Works only if OS supports Unix sockets and the
unix_socket_dir is not disabled in configuration. Does not work on
Windows. Does not work with TLS connections, they
are dropped.
-u user, --user user
Switch to the given user on startup.
-v, --verbose
Increase verbosity. Can be used multiple times.
-q, --quiet
Be quiet: do not log to stderr.
This does not affect logging verbosity, only that
stderr is not to be used. For
use in init.d scripts.
-V, --version
Show version.
-h, --help
Show short help.
--regservice
Win32: Register to run as Windows service. The
service_name configuration
parameter value is used as the name to register under.
--unregservice
Win32: Unregister Windows service.
The console is available by connecting as normal to the database pgbouncer:
$ psql -p 6432 pgbouncer
Only users listed in the configuration parameters
admin_users or
stats_users are allowed to
log in to the console. (Except when auth_mode=any, then any user is
allowed in as a stats_user.)
Additionally, the user name
pgbouncer is allowed to log in
without password, if the login comes via the Unix socket and the
client has same Unix user uid as the running process.
The admin console currently only supports the simple query protocol. Some drivers use the extended query protocol for all commands; these drivers will not work for this.
The SHOW commands output
information. Each command is described below.
Shows statistics. In this and related commands, the total figures are
since process start, the averages are updated every stats_period.
database
Statistics are presented per database.
total_xact_count
Total number of SQL transactions pooled by pgbouncer.
total_query_count
Total number of SQL queries pooled by pgbouncer.
total_received
Total volume in bytes of network traffic received by pgbouncer.
total_sent
Total volume in bytes of network traffic sent by pgbouncer.
total_xact_time
Total number of microseconds spent by pgbouncer when connected to Postgres Pro in a transaction, either idle in transaction or executing queries.
total_query_time
Total number of microseconds spent by pgbouncer when actively connected to Postgres Pro, executing queries.
total_wait_time
Time spent by clients waiting for a server, in microseconds. Updated when a client connection is assigned a backend connection.
avg_xact_count
Average transactions per second in last stat period.
avg_query_count
Average queries per second in last stat period.
avg_recv
Average received (from clients) bytes per second.
avg_sent
Average sent (to clients) bytes per second.
avg_xact_time
Average transaction duration, in microseconds.
avg_query_time
Average query duration, in microseconds.
avg_wait_time
Average time spent by clients waiting for a server that were assigned a
backend connection within the current stats_period,
in microseconds (average per second within that period).
Subset of SHOW STATS showing the
total values (total_).
Subset of SHOW STATS showing the
average values (avg_).
Like SHOW STATS but aggregated
across all databases.
type
S, for server.
user
User name pgbouncer uses to connect to server.
database
Database name.
state
State of the pgbouncer server connection, one of
active, idle, used, tested,
new, active_cancel, or being_canceled.
addr
IP address of Postgres Pro server.
port
Port of Postgres Pro server.
local_addr
Connection start address on local machine.
local_port
Connection start port on local machine.
connect_time
When the connection was made.
request_time
When last request was issued.
wait
Not used for server connections.
wait_us
Not used for server connections.
close_needed
1 if the connection will be closed as soon as possible,
because a configuration file reload or DNS update changed the
connection information or RECONNECT was issued.
ptr
Address of internal object for this connection. Used as unique ID.
link
Address of client connection the server is paired with.
remote_pid
PID of backend server process. In case connection is made over Unix socket and OS supports getting process ID info, its OS PID. Otherwise it's extracted from cancel packet the server sent, which should be the PID in case the server is Postgres Pro, but it's a random number in case the server is another pgbouncer.
tls
A string with TLS connection information, or empty if not using TLS.
application_name
A string containing the application_name set on the linked
client connection, or empty if this is not set, or if there is no linked connection.
type
C, for client.
user
Client connected user.
database
Database name.
state
State of the client connection, one of
active, waiting,
active_cancel_req, or waiting_cancel_req.
addr
IP address of the client.
port
Source port of the client.
local_addr
Connection end address on local machine.
local_port
Connection end port on local machine.
connect_time
Timestamp of connect time.
request_time
Timestamp of latest client request.
wait
Current waiting time in seconds.
wait_us
Microsecond part of the current waiting time.
close_needed
Not used for clients.
ptr
Address of internal object for this connection. Used as unique ID.
link
Address of server connection the client is paired with.
remote_pid
Process ID, in case client connects over Unix socket and OS supports getting it.
tls
A string with TLS connection information, or empty if not using TLS.
application_name
A string containing the application_name set by the client
for this connection, or empty if this is not set.
A new pool entry is made for each couple of (database, user).
database
Database name.
user
User name.
cl_active
Client connections that are either linked to server connections or are idle with no queries waiting to be processed.
cl_waiting
Client connections that have sent queries but have not yet got a server connection.
cl_active_cancel_req
Client connections that have forwarded query cancellations to the server and are waiting for the server response.
cl_waiting_cancel_req
Client connections that have not forwarded query cancellations to the server yet.
sv_active
Server connections that are linked to a client.
sv_active_cancel
Server connections that are currently forwarding a cancel request.
sv_being_canceled
Servers that normally could become idle but are waiting to do so until all in-flight cancel requests have completed that were sent to cancel a query on this server.
sv_idle
Server connections that are unused and immediately usable for client queries.
sv_used
Server connections that have been idle for more than
server_check_delay, so they need server_check_query to
run on them before they can be used again.
sv_tested
Server connections that are currently running either
server_reset_query or server_check_query.
sv_login
Server connections currently in the process of logging in.
maxwait
How long the first (oldest) client in the queue has waited,
in seconds. If this starts increasing, then the current
pool of servers does not handle requests quickly enough.
The reason may be either an overloaded server or just too small
of a pool_size
setting.
maxwait_us
Microsecond part of the maximum waiting time.
pool_mode
The pooling mode in use.
A new peer_pool entry is made for each configured peer.
database
ID of the configured peer entry.
cl_active_cancel_req
Client connections that have forwarded query cancellations to the server and are waiting for the server response.
cl_waiting_cancel_req
Client connections that have not forwarded query cancellations to the server yet.
sv_active_cancel
Server connections that are currently forwarding a cancel request.
sv_login
Server connections currently in the process of logging in.
Show following internal information, in columns (not rows):
databases
Count of databases.
users
Count of users.
pools
Count of pools.
free_clients
Count of free clients.
used_clients
Count of used clients.
login_clients
Count of clients in
login state.
free_servers
Count of free servers.
used_servers
Count of used servers.
dns_names
Count of DNS names in the cache.
dns_zones
Count of DNS zones in the cache.
dns_queries
Count of in-flight DNS queries.
dns_pending
Not used.
name
The user name.
pool_mode
The user's override pool_mode, or NULL if the default
will be used instead.
name
Name of configured database entry.
host
Host pgbouncer connects to.
port
Port pgbouncer connects to.
database
Actual database name pgbouncer connects to.
force_user
When the user is part of the connection string, the connection between pgbouncer and Postgres Pro is forced to the given user, whatever the client user.
pool_size
Maximum number of server connections.
min_pool_size
Minimum number of server connections.
reserve_pool
Maximum number of additional connections for this database.
pool_mode
The database's override pool_mode, or NULL if the
default will be used instead.
max_connections
Maximum number of allowed connections for this database, as set by
max_db_connections, either globally or per database.
current_connections
Current number of connections for this database.
paused
1 if this database is currently paused, else 0.
disabled
1 if this database is currently disabled, else 0.
peer_id
ID of the configured peer entry.
host
Host pgbouncer connects to.
port
Port pgbouncer connects to.
pool_size
Maximum number of server connections that can be made to this peer.
Internal command — shows list of file descriptors (FDs) in use with internal state attached to them.
When the connected user has the user name pgbouncer,
connects through the Unix socket and has the same UID as the running
process, the actual FDs are passed over the connection. This
mechanism is used to do an online restart.
This does not work on Windows.
This command also blocks the internal event loop, so it should not be used while pgbouncer is in use.
fd
File descriptor numeric value.
task
One of pooler,
client or
server.
user
User of the connection using the FD.
database
Database of the connection using the FD.
addr
IP address of the connection using the FD,
unix if a Unix socket
is used.
port
Port used by the connection using the FD.
cancel
Cancel key for this connection.
link
File descriptor for corresponding server/client. NULL if idle.
Shows low-level information about sockets or only active sockets.
This includes the information shown under SHOW CLIENTS
and SHOW SERVERS as well as other more low-level information.
Show the current configuration settings, one per row, with the following columns:
key
Configuration variable name.
value
Configuration value.
default
Configuration default value.
changeable
Either yes or
no, shows if the
variable can be changed while running. If
no, the variable can
be changed only at boot-time. Use
SET to change a variable at run time.
Shows low-level information about the current sizes of various internal memory allocations. The information presented is subject to change.
Show host names in DNS cache.
hostname
Host name.
ttl
How many seconds until next lookup.
addrs
Comma separated list of addresses.
Show DNS zones in cache.
zonename
Zone name.
serial
Current serial.
count
Host names belonging to this zone.
Show the pgbouncer version string.
Show the pgbouncer state settings. Current states are active, paused and suspended.
db]pgbouncer tries to disconnect from all servers. Disconnecting each server connection waits for that server connection to be released according to the server pool's pooling mode (in transaction pooling mode, the transaction must complete, in statement mode, the statement must complete, and in session pooling mode the client must disconnect). The command will not return before all server connections have been disconnected. To be used at the time of database restart.
If database name is given, only that database will be paused.
New client connections to a paused database will wait until
RESUME is called.
dbReject all new client connections on the given database.
db
Allow new client connections after a previous
DISABLE command.
dbClose each open server connection for the given database, or all databases, after it is released (according to the pooling mode), even if its lifetime is not up yet. New server connections can be made immediately and will connect as necessary according to the pool size settings.
This command is useful when the server connection setup has changed,
for example to perform a gradual switchover to a new server. It is
not necessary to run this command when the connection string in
pgbouncer.ini has been changed and reloaded
(see RELOAD) or when DNS resolution has changed,
because then the equivalent of this command will be run automatically.
This command is only necessary if something downstream of
pgbouncer routes the connections.
After this command is run, there could be an extended period where
some server connections go to an old destination and some server
connections go to a new destination. This is likely only sensible
when switching read-only traffic between read-only replicas, or when
switching between nodes of a multimaster replication setup. If all
connections need to be switched at the same time, PAUSE is
recommended instead. To close server connections without waiting (for
example, in emergency failover rather than gradual switchover
scenarios), also consider KILL.
dbImmediately drop all client and server connections on given database.
New client connections to a killed database will wait until
RESUME is called.
All socket buffers are flushed and pgbouncer stops listening for data on them. The command will not return before all buffers are empty. To be used at the time of pgbouncer online reboot.
New client connections to a suspended database will wait until
RESUME is called.
db]
Resume work from previous KILL,
PAUSE, or
SUSPEND command.
The pgbouncer process will exit.
The pgbouncer process will reload its configuration files and update
changeable settings. This includes the main configuration file as
well as the files specified by the settings auth_file and
auth_hba_file.
pgbouncer notices when a configuration file reload changes the connection parameters of a database definition. An existing server connection to the old destination will be closed when the server connection is next released (according to the pooling mode), and new server connections will immediately use the updated connection parameters.
db]
Wait until all server connections, either of the specified database or
of all databases, have cleared the close_needed state
(see the section called “SHOW SERVERS”). This can be called after
a RECONNECT or RELOAD to
wait until the respective configuration change has been fully
activated, for example in switchover scripts.
key = argChanges a configuration setting (see also the section called “SHOW CONFIG”). For example:
SET log_connections = 1; SET server_check_query = 'select 2';
(Note that this command is run on the pgbouncer
admin console and sets pgbouncer settings.
A SET command run on another database will be
passed to the Postgres Pro backend like any other SQL command.)
SIGHUP
Reload config. Same as issuing the command
RELOAD on the console.
SIGINT
Safe shutdown. Same as issuing
PAUSE and
SHUTDOWN on the console.
SIGTERM
Immediate shutdown. Same as issuing
SHUTDOWN on the console.
SIGUSR1
Same as issuing PAUSE on the console.
SIGUSR2
Same as issuing RESUME on the console.
From the libevent documentation:
It is possible to disable support for
epoll,kqueue,devpoll,poll, orselectby setting the environment variableEVENT_NOEPOLL,EVENT_NOKQUEUE,EVENT_NODEVPOLL,EVENT_NOPOLLorEVENT_NOSELECT, respectively.By setting the environment variable
EVENT_SHOW_METHOD,libeventdisplays the kernel notification method that it uses.
pgbouncer.ini Configuration File
The configuration file is in the .ini format. Section names are
between [ and ]. Lines starting with
; or # are taken as comments and ignored.
The characters ; and # are not recognized as special
when they appear later in the line.
logfile
Specifies the log file. For daemonization (-d),
either this or syslog has to be set.
The log file is kept open, so after rotation,
kill -HUP or on console
RELOAD; should be done. On Windows,
the service must be stopped and started.
Note that setting logfile does not by itself
turn off logging to stderr. Use the
command-line option -q or -d for that.
Default: not set
pidfile
Specifies the PID file. Without pidfile set,
daemonization (-d) is not allowed.
Default: not set
listen_addr
Specifies a list (comma-separated) of addresses where to listen for TCP
connections. You may also use * meaning
"listen on all addresses". When not set, only Unix
socket connections are accepted.
Addresses can be specified numerically (IPv4/IPv6) or by name.
Default: not set
listen_portWhich port to listen on. Applies to both TCP and Unix sockets.
Default: 6432
unix_socket_dir
Specifies the location for Unix sockets. Applies to both the listening
socket and server connections. If set to an empty string, Unix
sockets are disabled. A value that starts with @
specifies that a Unix socket in the abstract namespace
should be created (currently supported on Linux and Windows).
For online reboot (-R) to work,
a Unix socket needs to be configured, and it needs to be
in the file-system namespace.
Default: /tmp (empty on Windows)
unix_socket_modeFile system mode for Unix socket. Ignored for sockets in the abstract namespace. Not supported on Windows.
Default: 0777
unix_socket_groupGroup name to use for Unix socket. Ignored for sockets in the abstract namespace. Not supported on Windows.
Default: not set
userIf set, specifies the Unix user to change to after startup. Works only if pgbouncer is started as root or if it's already running as the given user.
Not supported on Windows.
Default: not set
pool_modeSpecifies when a server connection can be reused by other clients.
session
Server is released back to pool after client disconnects. Default.
transaction
Server is released back to pool after transaction finishes.
statement
Server is released back to pool after query finishes. Transactions spanning multiple statements are disallowed in this mode.
max_client_connMaximum number of client connections allowed.
When this setting is increased, then
the file descriptor limits in the operating system might also
have to be increased. Note that the number of file descriptors
potentially used is more than
max_client_conn. If each user connects under
its own username to the server, the theoretical maximum used is:
max_client_conn + (max pool_size * total databases * total users)
If a database user is specified in the connection string (all users connect under the same user name), the theoretical maximum is:
max_client_conn + (max pool_size * total databases)
The theoretical maximum should never be reached, unless somebody deliberately crafts a special load for it. Still, it means you should set the number of file descriptors to a safely high number.
Search for ulimit in your favorite shell man
page. Note: ulimit does not apply in a
Windows environment.
Default: 100
default_pool_sizeHow many server connections to allow per user/database pair. Can be overridden in the per-database configuration.
Default: 20
min_pool_sizeAdd more server connections to pool if below this number. Improves the behavior when the normal load suddenly comes back after a period of total inactivity. The value is effectively capped at the pool size.
Default: 0 (disabled)
reserve_pool_size
How many additional connections to allow to a pool (see
reserve_pool_timeout). The 0 value disables this parameter.
Default: 0 (disabled)
reserve_pool_timeoutIf a client has not been serviced in this time, pgbouncer enables use of additional connections from the reserve pool. The 0 value disables this parameter. [seconds]
Default: 5.0
max_db_connections
Do not allow more than this many server connections per database
(regardless of user). This considers the pgbouncer
database that the client has connected to, not the
Postgres Pro database of the outgoing connection.
This can also be set per database in the [databases] section.
Note that when you hit the limit, closing a client connection to one pool will not immediately allow a server connection to be established for another pool, because the server connection for the first pool is still open. Once the server connection closes (due to idle timeout), a new server connection will immediately be opened for the waiting pool.
Default: 0 (unlimited)
max_user_connections
Do not allow more than this many server connections per user
(regardless of database). This considers the pgbouncer
user that is associated with a pool, which is either the user specified for the
server connection or in absence of that the user the client has connected as.
This can also be set per user in the [users] section.
Note that when you hit the limit, closing a client connection to one pool will not immediately allow a server connection to be established for another pool, because the server connection for the first pool is still open. Once the server connection closes (due to idle timeout), a new server connection will immediately be opened for the waiting pool.
Default: 0 (unlimited)
server_round_robinBy default, pgbouncer reuses server connections in LIFO (last-in, first-out) manner, so that few connections get the most load. This gives best performance if you have a single server serving a database. But if there is a round-robin system behind a database address (TCP, DNS, or host list), then it is better if pgbouncer also uses connections in that manner, thus achieving uniform load.
Default: 0
track_extra_parameters
By default, pgbouncer tracks client_encoding,
datestyle, timezone,
standard_conforming_strings and application_name
parameters per client. To allow other parameters to be tracked, they can be
specified here, so that pgbouncer knows that they should be
maintained in the client variable cache and restored in the server whenever the client
becomes active.
If you need to specify multiple values, use a comma-separated list (e.g.
default_transaction_readonly, IntervalStyle)
Most parameters cannot be tracked this way. The only parameters that can be
tracked are ones that Postgres Pro reports to the client.
Postgres Pro has
an official list of parameters that it reports to the client.
Postgres Pro extensions can change this list though,
they can add parameters themselves that they also report,
and they can start reporting already existing paremeters that Postgres Pro
does not report.
Notably Citus 12.0+ causes PostgreSQL to also report search_path.
The postgres protocol allows specifying parameter settings, both directly as a
parameter in the startup packet, or inside the options
startup packet. Parameters specified using both of these methods are
supported by track_extra_parameters. However, it's not possible to include
options itself in track_extra_parameters, only the parameters contained in
options.
Default: IntervalStyle
ignore_startup_parameters
By default, pgbouncer allows only parameters it can keep track
of in startup packets: client_encoding,
datestyle, timezone and
standard_conforming_strings.
All other parameters will raise an error. To allow other parameters, they can be specified here, so that pgbouncer knows that they are handled by the admin and it can ignore them.
If you need to specify multiple values, use a comma-separated list (e.g.
options,extra_float_digits).
The postgres protocol allows specifying parameter settings,
both directly as a parameter in the startup packet, or inside the
options
startup packet. Parameters specified using both of these methods are
supported by ignore_startup_parameters. It's even possible to
include options itself in ignore_startup_parameters,
which results in any unknown parameters contained inside options
to be ignored.
Default: empty
peer_id
The peer ID used to identify this pgbouncer process
in a group of pgbouncer processes that are peered together.
The peer_id value should be unique within
a group of peered pgbouncer processes. When set to 0,
pgbouncer peering is disabled. See also
[peers] Section for more information. The
maximum value that can be used for the peer_id is 16383.
Default: 0
disable_pqexecDisable the Simple Query protocol (PQexec). Unlike the Extended Query protocol, Simple Query allows multiple queries in one packet, which allows some classes of SQL-injection attacks. Disabling it can improve security. Obviously, this means only clients that exclusively use the Extended Query protocol will stay working.
Default: 0
application_name_add_host
Add the client host address and port to the application name
setting set on connection start. This helps in identifying the
source of bad queries, etc. This logic applies only at the start of
a connection. If application_name is later changed with SET,
pgbouncer does not change it again.
Default: 0
conffile
Show location of current configuration file. Changing it will make
pgbouncer use another configuration file for next
RELOAD / SIGHUP.
Default: file from command line
service_nameUsed on win32 service registration.
Default: pgbouncer
job_name
Alias for service_name.
stats_period
Sets how often the averages shown in various
SHOW commands are updated and how often
aggregated statistics are written to the log
(but see log_stats). [seconds]
Default: 60
pgbouncer handles its own client authentication and has its own database of users. These settings control this.
auth_typeHow to authenticate users.
cert
The client must connect over TLS connection with a valid client
certificate. The user name is then taken from the CommonName field from the
certificate.
md5
Use MD5-based password check. This is the default authentication
method. auth_file may contain both
MD5-encrypted and plain-text passwords. If md5
is configured and a user has a SCRAM secret, then SCRAM
authentication is used automatically instead.
scram-sha-256
Use password check with SCRAM-SHA-256.
auth_file has to contain
SCRAM secrets or plain-text passwords. Note that SCRAM secrets
can only be used for verifying the password of a client but not
for logging into a server. To be able to use SCRAM on server
connections, use plain-text passwords.
plain
The clear-text password is sent over the wire. Deprecated.
trust
No authentication is done. The user name must still exist in
auth_file.
any
Like the trust method, but the user name
given is ignored. Requires that all databases are
configured to log in as a specific user. Additionally, the
console database allows any user to log in as admin.
hba
The actual authentication type is loaded from auth_hba_file. This
allows different authentication methods for different access
paths, for example: connections over Unix socket use
peer authentication method, connections over TCP
must use TLS.
pam
Pluggable Authentication Modules (PAM) method is used to authenticate users,
auth_file is ignored. This method is not compatible with databases using the
auth_user option. The service name reported to PAM is
pgbouncer. pam is not supported in the
HBA configuration file.
auth_hba_file
HBA configuration file to use when auth_type is
hba.
Default: not set
auth_fileThe name of the file to load user names and passwords from. See the section called “Authentication File Format” for details.
Most authentication types (see auth_type)
require that either auth_file or
auth_user be set; otherwise there would be no users defined.
Default: not set
auth_user
If auth_user is set, then any user not specified
in auth_file will be queried through the
auth_query query from pg_shadow
in the database, using auth_user. The password of
auth_user will be taken from auth_file.
(If auth_user does not require a password,
then it does not need to be defined in auth_file.)
Direct access to pg_shadow requires admin rights. It's
preferable to use a non-superuser that calls a SECURITY DEFINER
function instead.
Default: not set
auth_queryQuery to load user's password from database.
Direct access to pg_shadow requires admin rights. It's
preferable to use a non-superuser that calls a SECURITY DEFINER
function instead.
Note that the query is run inside the target database. So if a function is used, it needs to be installed into each database.
Default:
SELECT usename, passwd FROM pg_shadow WHERE usename=$1
auth_dbnameDatabase name in the [databases] section to be used for authentication purposes. This option can be either global or overriden in the connection string if this parameter is specified.
syslogToggles syslog on/off. On Windows, the event log is used instead.
Default: 0
syslog_identUnder what name to send logs to syslog.
Default: pgbouncer (program name)
syslog_facility
Under what facility to send logs to syslog. Possibilities:
auth, authpriv,
daemon, user,
local0-7.
Default: daemon
log_connectionsLog successful logins.
Default: 1
log_disconnectionsLog disconnections with reasons.
Default: 1
log_pooler_errorsLog error messages the pooler sends to clients.
Default: 1
log_stats
Write aggregated statistics into the log, every
stats_period. This can be disabled if
external monitoring tools are used to grab the same data
from SHOW commands.
Default: 1
verbose
Increase verbosity. Mirrors the -v switch on the command
line. For example, using -v -v on the command line is
the same as verbose=2.
Default: 0
admin_users
Comma-separated list of database users that are allowed to
connect and run all commands on the console. Ignored when auth_type
is any, in which case any user name is allowed
in as admin.
Default: empty
stats_users
Comma-separated list of database users that are allowed to
connect and run read-only queries on the console. That means all
SHOW commands except SHOW FDS.
Default: empty
server_reset_query
Query sent to server on connection release, before making it
available to other clients. At that moment no transaction is in
progress, so the value should not include ABORT or
ROLLBACK.
The query is supposed to clean any changes made to the database
session so that the next client gets the connection in a well-defined
state. The default is DISCARD ALL, which cleans
everything, but that leaves the next client no pre-cached state. It
can be made lighter, e.g. DEALLOCATE ALL to
just drop prepared statements, if the application does not break
when some state is kept around.
When transaction pooling is used, the server_reset_query is not
used, because in that mode, clients must not use any session-based features, since each
transaction ends up in a different connection and thus gets a
different session state.
Default: DISCARD ALL
server_reset_query_always
Whether server_reset_query should be run in all pooling modes.
When this setting is off (default), the server_reset_query will
be run only in pools that are in sessions-pooling mode.
Connections in transaction-pooling mode should not have any need
for a reset query.
This setting is for working around broken setups that run applications that use session features over a transaction-pooled pgbouncer. It changes non-deterministic breakage to deterministic breakage: clients always lose their state after each transaction.
Default: 0
server_check_delay
How long to keep released connections available for immediate
re-use, without running server_check_query
on it. If 0 then the query is always run.
Default: 30.0
server_check_querySimple do-nothing query to check if the server connection is alive.
If an empty string, then sanity checking is disabled.
Default: select 1
server_fast_close
Disconnect a server in session pooling mode immediately or after the
end of the current transaction if it is in close_needed mode (set by
RECONNECT, RELOAD that changes connection settings, or DNS
change), rather than waiting for the session end. In statement or
transaction pooling mode, this has no effect since that is the default
behavior there.
If because of this setting a server connection is closed before the end of the client session, the client connection is also closed. This ensures that the client notices that the session has been interrupted.
This setting makes connection configuration changes take effect sooner if session pooling and long-running sessions are used. The downside is that client sessions are liable to be interrupted by a configuration change, so client applications will need logic to reconnect and reestablish session state. But note that no transactions will be lost, because running transactions are not interrupted, only idle sessions.
Default: 0
server_lifetimeThe pooler will close an unused (not currently linked to any client connection) server connection that has been connected longer than this. Setting it to 0 means the connection is to be used only once, then closed. [seconds]
Default: 3600.0
server_idle_timeoutIf a server connection has been idle more than this many seconds it will be closed. If 0 then timeout is disabled. [seconds]
Default: 600.0
server_connect_timeoutIf connection and login don't finish in this amount of time, the connection will be closed. [seconds]
Default: 15.0
server_login_retryIf login to the server failed, because of failure to connect or from authentication, the pooler waits this much before retrying to connect. During the waiting interval, new clients trying to connect to the failing server will get an error immediately without another connection attempt. [seconds]
The purpose of this behavior is that clients don't unnecessarily queue
up waiting for a server connection to become available if the server
is not working. However, it also means that if a server is
momentarily failing, for example during a restart or if the
configuration was erroneous, then it will take at least this long
until the pooler will consider connecting to it again. Planned events
such as restarts should normally be managed using the PAUSE
command to avoid this.
Default: 15.0
client_login_timeout
If a client connects but does not manage to log in in this amount
of time, it will be disconnected. Mainly needed to avoid dead
connections stalling SUSPEND and thus online restart. [seconds]
Default: 60.0
autodb_idle_timeoutIf the automatically created (via "*") database pools have been unused this many seconds, they are freed. The negative aspect of that is that their statistics are also forgotten. [seconds]
Default: 3600.0
dns_max_ttlHow long DNS lookups can be cached. The actual DNS TTL is ignored. [seconds]
Default: 15.0
dns_nxdomain_ttl
How long DNS errors and NXDOMAIN
DNS lookups can be cached. [seconds]
Default: 15.0
dns_zone_check_periodPeriod to check if a zone serial has changed.
pgbouncer can collect DNS zones from host names (everything after first dot) and then periodically check if the zone serial changes. If it notices changes, all host names under that zone are looked up again. If any host IP changes, its connections are invalidated.
Default: 0.0 (disabled)
resolv_conf
The location of a custom resolv.conf file. This is to allow
specifying custom DNS servers and perhaps other name resolution
options, independent of the global operating system configuration.
The parsing of the file is done by the DNS backend library, not pgbouncer, so see the library's documentation for details on allowed syntax and directives.
Default: empty (use operating system defaults)
client_tls_sslmode
TLS mode to use for connections from clients. TLS connections
are disabled by default. When enabled, client_tls_key_file and
client_tls_cert_file must be also configured to set up the key and
certificate pgbouncer uses to accept client connections.
disable
Plain TCP. If client requests TLS, it's ignored. Default.
allow
If client requests TLS, it is used. If not, plain TCP is used. If the client presents a client certificate, it is not validated.
prefer
Same as allow.
require
The client must use TLS. If not, the client connection is rejected. If the client presents a client certificate, it is not validated.
verify-ca
Client must use TLS with valid client certificate.
verify-full
Same as verify-ca.
client_tls_key_filePrivate key for pgbouncer to accept client connections.
Default: not set
client_tls_cert_fileCertificate for private key. Clients can validate it.
Default: not set
client_tls_ca_fileRoot certificate file to validate client certificates.
Default: not set
client_tls_protocols
Which TLS protocol versions are allowed. Allowed values:
tlsv1.0, tlsv1.1,
tlsv1.2, tlsv1.3. Shortcuts: all
(tlsv1.0,tlsv1.1,tlsv1.2,tlsv1.3), secure (tlsv1.2,tlsv1.3),
legacy (all).
Default: secure
client_tls_ciphers
Allowed TLS ciphers, in OpenSSL syntax. Shortcuts:
default/secure,
compat/legacy,
insecure/all, normal,
fast.
Only connections using TLS version 1.2 and lower are affected. There is currently no setting that controls the cipher choices used by TLS version 1.3 connections.
Default: fast
client_tls_ecdhcurveElliptic Curve name to use for ECDH key exchanges.
Allowed values: none (DH is disabled),
auto (256-bit ECDH), curve name.
Default: auto
client_tls_dheparamsDHE key exchange type.
Allowed values: none (DH is disabled),
auto (2048-bit DH), legacy
(1024-bit DH).
Default: auto
server_tls_sslmode
TLS mode to use for connections to Postgres Pro servers.
The default mode is prefer.
disable
Plain TCP. TLS is not even requested from the server.
prefer
TLS connection is always requested first from Postgres Pro. If refused, the connection will be established over plain TCP. Server certificate is not validated. Default.
require
Connection must go over TLS. If server rejects it, plain TCP is not attempted. Server certificate is not validated.
verify-ca
Connection must go over TLS and server certificate must be
valid according to server_tls_ca_file. Server host name is
not checked against certificate.
verify-full
Connection must go over TLS and server certificate must be
valid according to server_tls_ca_file. Server host name
must match certificate information.
server_tls_ca_fileRoot certificate file to validate Postgres Pro server certificates.
Default: not set
server_tls_key_filePrivate key for pgbouncer to authenticate against Postgres Pro server.
Default: not set
server_tls_cert_fileCertificate for private key. Postgres Pro server can validate it.
Default: not set
server_tls_protocols
Which TLS protocol versions are allowed. Allowed values:
tlsv1.0, tlsv1.1,
tlsv1.2, tlsv1.3. Shortcuts: all
(tlsv1.0,tlsv1.1,tlsv1.2,tlsv1.3), secure (tlsv1.2,tlsv1.3),
legacy (all).
Default: secure
server_tls_ciphers
Allowed TLS ciphers, in OpenSSL syntax. Shortcuts:
default/secure,
compat/legacy,
insecure/all, normal,
fast.
Only connections using TLS version 1.2 and lower are affected. There is currently no setting that controls the cipher choices used by TLS version 1.3 connections.
Default: fast
Setting the following timeouts can cause unexpected errors.
query_timeout
Queries running longer than that are canceled. This should be
used only with a slightly smaller server-side statement_timeout,
to apply only for network problems. [seconds]
Default: 0.0 (disabled)
query_wait_timeoutMaximum time queries are allowed to spend waiting for execution. If the query is not assigned to a server during that time, the client is disconnected. The 0 value disables this parameter. If this is disabled, clients will be queued indefinitely. [seconds]
This setting is used to prevent unresponsive servers from grabbing up connections. It also helps when the server is down or rejects connections for any reason.
Default: 120.0
cancel_wait_timeoutMaximum time cancellation requests are allowed to spend waiting for execution. If the cancel request is not assigned to a server during that time, the client is disconnected. 0 disables. If this is disabled, cancel requests will be queued indefinitely. [seconds]
This setting is used to prevent a client locking up when a cancel cannot be forwarded due to the server being down.
Default: 10.0
client_idle_timeoutClient connections idling longer than this many seconds are closed. This should be larger than the client-side connection lifetime settings, and only used for network problems. [seconds]
Default: 0.0 (disabled)
idle_transaction_timeoutIf a client has been in the “idle in transaction” state longer, it will be disconnected. [seconds]
Default: 0.0 (disabled)
suspend_timeout
How long to wait for buffer flush during SUSPEND or
reboot (-R). A connection is dropped if the flush does
not succeed. [seconds]
Default: 10
pkt_bufInternal buffer size for packets. Affects size of TCP packets sent and general memory usage. Actual libpq packets can be larger than this, so no need to set it large.
Default: 4096
max_packet_sizeMaximum size for Postgres Pro packets that pgbouncer allows through. One packet is either one query or one result set row. The full result set can be larger.
Default: 2147483647
listen_backlog
The value of the backlog argument for
listen(). Determines how many new
unanswered connection attempts are kept in the queue. When the queue is
full, further new connections are dropped.
Default: 128
sbuf_loopcnt
How many times to process data on one connection, before
proceeding. Without this limit, one connection with a big
result set can stall pgbouncer for a long time. One loop
processes one pkt_buf amount of data. 0 means no limit.
Default: 5
so_reuseport
Specifies whether to set the socket option SO_REUSEPORT
on TCP listening sockets. On some operating systems,
this allows running multiple pgbouncer instances
on the same host listening on the same port and having the kernel distribute
the connections automatically. This option is a way to get
pgbouncer to use more CPU cores.
(pgbouncer is single-threaded and uses
one CPU core per instance.)
This setting has the desired effect on Linux. On systems that don't support the socket option at all, turning this setting on will result in an error.
Each pgbouncer instance on the same host
needs different settings for at least unix_socket_dir
and pidfile, as well as logfile
if that is used. Also note that if you make use of this option, you can no
longer connect to a specific pgbouncer instance via TCP/IP, which
might have implications for monitoring and metrics collection.
To make sure query cancellations keep working, you should set up
pgbouncer peering between the different
pgbouncer processes. For details see
the peer_id
configuration option and the
[peers] configuration section.
There's also an example that uses peering and so_reuseport
in the Examples section.
Default: 0
tcp_defer_accept
Sets the TCP_DEFER_ACCEPT socket option; see man 7 tcp
for details. (This is a Boolean option: 1 means enabled. The actual
value set if enabled is currently hardcoded to 45 seconds.)
This is currently only supported on Linux.
Default: 1 on Linux, otherwise 0
tcp_socket_bufferDefault: not set
tcp_keepaliveTurns on basic keepalive with OS defaults.
On Linux, the system defaults are tcp_keepidle=7200, tcp_keepintvl=75, tcp_keepcnt=9. They are probably similar on other operating systems.
Default: 1
tcp_keepcntDefault: not set
tcp_keepidleDefault: not set
tcp_keepintvlDefault: not set
tcp_user_timeout
Sets the TCP_USER_TIMEOUT socket option.
This specifies the maximum amount of time in milliseconds that
transmitted data may remain unacknowledged before the TCP
connection is forcibly closed. If set to 0, then operating
system's default is used.
This is currently only supported on Linux.
Default: 0
The section [databases] defines the names of the databases that
clients of pgbouncer can connect to and specifies where those
connections will be routed. The section contains key=value lines like
dbname = connection string
where the key will be taken as a database name and the value as a connection string, consisting of key=value pairs of connection parameters, described below (similar to libpq, but the actual libpq is not used and the set of available features is different).
Example:
foodb = host=host1.example.com port=5432 bardb = host=localhost dbname=bazdb
The database name can contain characters _0-9A-Za-z
without quoting. Names that contain other chars need to be quoted
with standard SQL ident quoting: double quotes where ""
is taken as single quote.
The database name pgbouncer is reserved for the admin console and
cannot be used as a key here.
* acts as fallback database: if the exact name does
not exist, its value is taken as connection string for the requested
database. For example, if there is the following entry (and no other overriding entries):
* = host=foo
In this case, a connection to pgbouncer
specifying a database bar will effectively behave as if the following entry
exists (taking advantage of the default for dbname being the
client-side database name):
bar = host=foo dbname=bar
Such automatically created database entries are cleaned
up if they stay idle longer than the time specified by the
autodb_idle_timeout parameter.
dbnameDestination database name.
Default: same as client-side database name
host
Host name or IP address to connect to. Host names are resolved
at connection time, the result is cached per dns_max_ttl parameter.
When a host name's resolution changes, existing server connections are
automatically closed when they are released (according to the pooling
mode), and new server connections immediately use the new resolution.
If DNS returns several results, they are used in a round-robin
manner.
If the value begins with /, then a Unix socket in the file-system
namespace is used. If the value begins with @, then a Unix socket
in the abstract namespace is used.
A comma-separated list of host names or addresses can be specified.
In that case, connections are made in a round-robin manner. (If a
host list contains host names that in turn resolve via DNS
to multiple addresses, the round-robin systems operate independently. This is an
implementation dependency that is subject to change.) Note that in a
list, all hosts must be available at all times: there are no
mechanisms to skip unreachable hosts or to select only available hosts
from a list or similar. (This is different from what a host list in
libpq means.) Also note that this only affects
how the destinations of new connections are chosen. See also the setting
server_round_robin for how clients are assigned
to already established server connections.
Examples:
host=localhost host=127.0.0.1 host=2001:0db8:85a3:0000:0000:8a2e:0370:7334 host=/var/run/postgresql host=192.168.0.1,192.168.0.2,192.168.0.3
Default: not set, meaning to use a Unix socket
portDefault: 5432
user
If user= is set, all connections to the
destination database will be done with the specified user,
meaning that there will be only one pool for this database.
Otherwise pgbouncer logs into the destination database with the client user name, meaning that there will be one pool per user.
password
If no password is specified here, the password from the
auth_file or auth_query will be used.
auth_user
Override of the global auth_user setting, if specified.
pool_size
Set the maximum size of pools for this database. If not set, the
default_pool_size is used.
min_pool_size
Set the minimum pool size for this database. If not set, the global
min_pool_size is used.
reserve_pool
Set additional connections for this database.
If not set, reserve_pool_size is used.
connect_queryQuery to be executed after a connection is established, but before allowing the connection to be used by any clients. If the query raises errors, they are logged but ignored otherwise.
pool_mode
Set the pool mode specific to this database. If not set, the
default pool_mode is used.
max_db_connectionsConfigure a database-wide maximum (i.e. all pools within the database will not have more than this many server connections).
client_encoding
Ask specific client_encoding from server.
datestyle
Ask specific datestyle from server.
timezone
Ask specific timezone from
server.
This section contains key=value lines like
user1 = settings
where the key will be taken as a user name and the value as a list of configuration settings specific for this user.
Example:
user1 = pool_mode=session
Only a few settings are available here.
pool_mode
Set the pool mode to be used for all connections from this user.
If not set, the database or default pool_mode is used.
max_user_connectionsConfigure a maximum for the user (i.e. all pools with the user will not have more than this many server connections).
This section defines the peers that pgbouncer can forward cancellation requests to and where those cancellation requests will be routed.
pgbouncer processes can be peered together in a group
by defining a peer_id
value and a [peers] section in the configs of all the
pgbouncer processes.
These pgbouncer processes can then forward cancellation
requests to the process that it originated from. This is needed to make cancellations
work when multiple pgbouncer processes (possibly on different
servers) are behind the same TCP load balancer. Cancellation requests are sent over different TCP
connections than the query they are cancelling, so a TCP load balancer might
send the cancellation request connection to a different process than the one
that it was meant for. By peering them these cancellation requests eventually
end up at the right process.
The section contains key=value lines like
peer_id = connection string
where the key will be taken as a peer_id and the value as a connection string,
consisting of key=value pairs of connection parameters, described below (similar
to libpq, but the actual libpq is not
used and the set of available features is different).
Example:
1 = host=host1.example.com 2 = host=/tmp/pgbouncer-2 port=5555
For peering to work, the peer_id of each pgbouncer
process in the group
must be unique within the peered group. And the [peers] section should
contain entries for each of those peer IDs. An example can be found in the
Examples section. It is allowed, but not necessary, for the
[peers] section to contain the peer_id of the
pgbouncer that the config is
for. Such an entry will be ignored, but it is allowed to make config management easier.
Because it allows using the exact same [peers] section for multiple
configs.
host
Host name or IP address to connect to. Host names are resolved at connection
time, the result is cached per dns_max_ttl parameter. If DNS returns several
results, they are used in a round-robin manner. But in general it's not
recommended to use a hostname that resolves to multiple IPs, because then the
cancel request might still be forwarded to the wrong node and it would need to
be forwarded again (which is only allowed up to three times).
If the value begins with /, then a Unix socket in the file-system namespace is
used. If the value begins with @, then a Unix socket in the abstract
namespace is used.
Examples:
host=localhost host=127.0.0.1 host=2001:0db8:85a3:0000:0000:8a2e:0370:7334 host=/var/run/pgbouncer-1
portDefault: 6432
pool_size
If not set, the default_pool_size is used.
The pgbouncer configuration file can contain include directives, which specify another configuration file to read and process. This allows splitting the configuration file into physically separate parts. The include directives look like this:
%include filename
If the filename is not an absolute path, it is taken as relative to the current working directory.
This section describes the format of the file specified by the
auth_file setting. It is a text file in the following format:
"username1" "password" ... "username2" "md5abcdef012342345" ... "username2" "SCRAM-SHA-256$iterations:salt$storedkey:serverkey"
There should be at least two fields, surrounded by double quotes. The first field is the user name and the second is either a plain-text, a MD5-hashed password, or a SCRAM secret. pgbouncer ignores the rest of the line. Double quotes in a field value can be escaped by writing two double quotes.
Postgres Pro MD5-hashed password format:
"md5" + md5(password + username)
So user admin with password 1234 will have MD5-hashed password
md545f2603610af569b6155c45067268c6b.
Postgres Pro SCRAM secret format:
SCRAM-SHA-256$iterations:salt$storedkey:serverkey
The passwords or secrets stored in the authentication file serve two purposes. First, they are used to verify the passwords of incoming client connections, if a password-based authentication method is configured. Second, they are used as the passwords for outgoing connections to the backend server, if the backend server requires password-based authentication (unless the password is specified directly in the database's connection string). The latter works if the password is stored in plain text or MD5-hashed. SCRAM secrets can only be used for logging into a server if the client authentication also uses SCRAM, the pgbouncer database definition does not specify a user name, and the SCRAM secrets are identical in pgbouncer and the Postgres Pro server (same salt and iterations, not merely the same password). This is due to an inherent security property of SCRAM: the stored SCRAM secret cannot by itself be used for deriving login credentials.
The authentication file can be written by hand, but it's also useful
to generate it from some other list of users and passwords. See
./etc/mkauth.py for a sample script to generate the authentication
file from the pg_shadow system table.
Alternatively, use auth_query instead of
auth_file to avoid having to maintain a separate authentication file.
The location of the HBA file is specified by the setting
auth_hba_file. It is only used
if auth_type is set to hba.
The file follows the format of the Postgres Pro pg_hba.conf
file described in Section 20.1.
Supported record types: local, host, hostssl, hostnossl.
Database field: Supports all, sameuser,
@file, multiple names.
Not supported: replication, samerole, samegroup.
User name field: Supports all, @file, multiple names. Not
supported: +groupname.
Address field: Supports IPv4, IPv6.
Not supported: DNS names, domain prefixes.
Auth-method field: Only methods supported by
pgbouncer's auth_type
are supported, plus peer and reject, but except any
and pam, which only work globally. User name map (map=) parameter
is not supported.
Small example configuration:
[databases] template1 = host=localhost dbname=template1 auth_user=someuser [pgbouncer] pool_mode = session listen_port = 6432 listen_addr = localhost auth_type = md5 auth_file = users.txt logfile = pgbouncer.log pidfile = pgbouncer.pid admin_users = someuser stats_users = stat_collector
Database examples:
[databases] ; foodb over Unix socket foodb = ; redirect bardb to bazdb on localhost bardb = host=localhost dbname=bazdb ; access to destination database will go with single user forcedb = host=localhost port=300 user=baz password=foo client_encoding=UNICODE datestyle=ISO
Example of a secure function for auth_query:
CREATE OR REPLACE FUNCTION pgbouncer.user_lookup(in i_username text, out uname text, out phash text)
RETURNS record AS $$
BEGIN
SELECT usename, passwd FROM pg_catalog.pg_shadow
WHERE usename = i_username INTO uname, phash;
RETURN;
END;$$ LANGUAGE plpgsql SECURITY DEFINER;
REVOKE ALL ON FUNCTION pgbouncer.user_lookup(text) FROM public, pgbouncer;
GRANT EXECUTE ON FUNCTION pgbouncer.user_lookup(text) TO pgbouncer;
Example configs for 2 peered pgbouncer processes to create a multi-core
pgbouncer setup using so_reuseport.
The config for the first process:
[databases] postgres = host=localhost dbname=postgres [peers] 1 = host=/tmp/pgbouncer1 2 = host=/tmp/pgbouncer2 [pgbouncer] listen_addr=127.0.0.1 auth_file=auth_file.conf so_reuseport=1 unix_socket_dir=/tmp/pgbouncer1 peer_id=1
The config for the second process:
[databases] postgres = host=localhost dbname=postgres [peers] 1 = host=/tmp/pgbouncer1 2 = host=/tmp/pgbouncer2 [pgbouncer] listen_addr=127.0.0.1 auth_file=auth_file.conf so_reuseport=1 ; only unix_socket_dir and peer_id are different unix_socket_dir=/tmp/pgbouncer2 peer_id=2