Configuration

pgbouncer.ini

Description

The configuration file is in “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.

Generic settings

logfile

Specifies the log file. For daemonization (-d), either this or syslog need 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_port

Which 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 to 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_mode

File system mode for Unix socket. Ignored for sockets in the abstract namespace. Not supported on Windows.

Default: 0777

unix_socket_group

Group name to use for Unix socket. Ignored for sockets in the abstract namespace. Not supported on Windows.

Default: not set

user

If 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_mode

Specifies 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_conn

Maximum 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 user name 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_size

How many server connections to allow per user/database pair. Can be overridden in the per-database configuration.

Default: 20

min_pool_size

Add more server connections to pool if below this number. Improves behavior when the normal load suddenly comes back after a period of total inactivity. The value is effectively capped at the pool size.

Only enforced for pools where at least one of the following is true:

  • the entry in the [database] section for the pool has a value set for the user key (aka forced user)
  • there is at least one client connected to the pool

Default: 0 (disabled)

reserve_pool_size

How many additional connections to allow to a pool (see reserve_pool_timeout). 0 disables.

Default: 0 (disabled)

reserve_pool_timeout

If a client has not been serviced in this time, use additional connections from the reserve pool. 0 disables. [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 PostgreSQL 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_robin

By 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_read_only, IntervalStyle)

Note: Most parameters cannot be tracked this way. The only parameters that can be tracked are ones that Postgres reports to the client. Postgres has an official list of parameters that it reports to the client. Postgres extensions can change this list though, they can add parameters themselves that they also report, and they can start reporting already existing parameters that Postgres does not report. Notably Citus 12.0+ causes Postgres to also report search_path.

The Postgres protocol allows specifying parameters 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 others parameters will raise an error. To allow others 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 parameters 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 track_extra_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 the docs for the [peers] section for more information. The maximum value that can be used for the peer_id is 16383.

Default: 0

disable_pqexec

Disable 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 config file. Changing it will make PgBouncer use another config file for next RELOAD / SIGHUP.

Default: file from command line

service_name

Used 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

max_prepared_statements

When this is set to a non-zero value PgBouncer tracks protocol-level named prepared statements related commands sent by the client in transaction and statement pooling mode. PgBouncer makes sure that any statement prepared by a client is available on the backing server connection. Even when the statement was originally prepared on another server connection.

PgBouncer internally examines all the queries that are sent by clients as a prepared statement, and gives each unique query string an internal name with the format PGBOUNCER_{unique_id}. If the same query string is prepared multiple times (possibly by different clients), then these queries share the same internal name. PgBouncer only prepares the statement on the actual PostgreSQL server using the internal name (so not the name provided by the client). PgBouncer keeps track of the name that the client gave to each prepared statement. It then rewrites each command that uses a prepared statement to by replacing the client side name with the the internal name (e.g. replacing my_prepared_statement with PGBOUNCER_123) before forwarding that command to the server. More importantly, if the prepared statement that the client wants to execute is not yet prepared on the server (e.g. because a different server is now assigned to the client than when the client prepared the statement), then PgBouncer transparently prepares the statement before executing it.

Note: This tracking and rewriting of prepared statement commands does not work for SQL-level prepared statement commands, so PREPARE, EXECUTE and DEALLOCATE are forwarded straight to Postgres. The exception to this rule are the DEALLOCATE ALL and DISCARD ALL commands, these do work as expected and will clear the prepared statements that PgBouncer tracked for the client that sends this command.

The actual value of this setting controls the number of prepared statements kept active in an LRU cache on a single server connection. When the setting is set to 0 prepared statement support for transaction and statement pooling is disabled. To get the best performance you should try to make sure that this setting is larger than the amount of commonly used prepared statements in your application. Keep in mind that the higher this value, the larger the memory footprint of each PgBouncer connection will be on your PostgreSQL server, because it will keep more queries prepared on those connections. It also increases the memory footprint of PgBouncer itself, because it now needs to keep track of query strings.

The impact on PgBouncer memory usage is not that big though: - Each unique query is stored once in a global query cache. - Each client connection keeps a buffer that it uses to rewrite packets. This is, at most, 4 times the size of pkt_buf. This limit is often not reached though, it only happens when the queries in your prepared statements are between 2 and 4 times the size of pkt_buf.

So if you consider the following as an example scenario: - There are 1000 active clients - The clients prepare 200 unique queries - The average size of a query is 5kB - pkt_buf parameter is set to the default of 4096 (4kB)

Then, PgBouncer needs at most the following amount of memory to handle these prepared statements:

200 x 5kB + 1000 x 4 x 4kB = ~17MB of memory.

Tracking prepared statements does not only come with a memory cost, but also with increased CPU usage, because PgBouncer needs to inspect and rewrite the queries. Multiple PgBouncer instances can listen on the same port to use more than one core for processing, see the documentation for the so_reuseport option for details.

But of course there are also performance benefits to prepared statements. Just as when connecting to PostgreSQL directly, by preparing a query that is executed many times, it reduces the total amount of parsing and planning that needs to be done. The way that PgBouncer tracks prepared statements is especially beneficial to performance when multiple clients prepare the same queries. Because client connections automatically reuse a prepared statement on a server connection, even if it was prepared by another client. As an example, if you have a pool_size of 20 and you have 100 clients that all prepare the exact same query, then the query is prepared (and thus parsed) only 20 times on the PostgreSQL server.

The reuse of prepared statements has one downside. If the return or argument types of a prepared statement changes across executions then PostgreSQL currently throws an error such as:

ERROR:  cached plan must not change result type

You can avoid such errors by not having multiple clients that use the exact same query string in a prepared statement, but expecting different argument or result types. One of the most common ways of running into this issue is during a DDL migration where you add a new column or change a column type on an existing table. In those cases you can run RECONNECT on the PgBouncer admin console after doing the migration to force a re-prepare of the query and make the error go away.

Default: 0

Authentication settings

PgBouncer handles its own client authentication and has its own database of users. These settings control this.

auth_type

How to authenticate users.

cert
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.
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 the peer auth method, connections over TCP must use TLS.
pam
PAM 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. See section HBA file format below about details.

Default: not set

auth_ident_file

Identity map file to use when auth_type is hba and a user map will be defined. See section Ident map file format below about details.

Default: not set

auth_file

The name of the file to load user names and passwords from. See section Authentication file format below about details.

Most authentication types (see above) 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 the 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_query

Query 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_dbname

Database name in the [database] section to be used for authentication purposes. This option can be either global or overridden in the connection string if this parameter is specified.

Log settings

syslog

Toggles syslog on/off. On Windows, the event log is used instead.

Default: 0

syslog_ident

Under 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_connections

Log successful logins.

Default: 1

log_disconnections

Log disconnections with reasons.

Default: 1

log_pooler_errors

Log 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

Console access control

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

Connection sanity checks, timeouts

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 check is always run.

Default: 30.0

server_check_query

Simple 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_lifetime

The 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]

This can also be set per database in the [databases] section.

Default: 3600.0

server_idle_timeout

If a server connection has been idle more than this many seconds it will be closed. If 0 then this timeout is disabled. [seconds]

Default: 600.0

server_connect_timeout

If connection and login don’t finish in this amount of time, the connection will be closed. [seconds]

Default: 15.0

server_login_retry

If 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_timeout

If 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_ttl

How 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_period

Period 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.

Works only with c-ares backend (configure option --with-cares).

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.

Requires evdns (>= 2.0.3) or c-ares (>= 1.15.0) backend.

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)

TLS settings

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. The most common certificate file format usable by PgBouncer is pem.

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
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_file

Private key for PgBouncer to accept client connections.

Default: not set

client_tls_cert_file

Certificate for private key. Clients can validate it.

Default: not set

client_tls_ca_file

Root 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/fast/normal (these all use system wide OpenSSL defaults)
  • all (enables all ciphers, not recommended)

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: default

client_tls_ecdhcurve

Elliptic Curve name to use for ECDH key exchanges.

Allowed values: none (DH is disabled), auto (256-bit ECDH), curve name

Default: auto

client_tls_dheparams

DHE 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 PostgreSQL servers. The default mode is prefer.

disable
Plain TCP. TLS is not even requested from the server.
allow
FIXME: if server rejects plain, try TLS?
prefer
TLS connection is always requested first from PostgreSQL. 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_file

Root certificate file to validate PostgreSQL server certificates.

Default: not set

server_tls_key_file

Private key for PgBouncer to authenticate against PostgreSQL server.

Default: not set

server_tls_cert_file

Certificate for private key. PostgreSQL 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/fast/normal (these all use system wide OpenSSL defaults)
  • all (enables all ciphers, not recommended)

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: default

Dangerous timeouts

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_timeout

Maximum 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. 0 disables. 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_timeout

Maximum 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_timeout

Client 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_timeout

If a client has been in “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

Low-level network settings

pkt_buf

Internal 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_size

Maximum size for PostgreSQL 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

Backlog argument for listen(2). 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.)

The behavior in detail depends on the operating system kernel. As of this writing, this setting has the desired effect on (sufficiently recent versions of) Linux, DragonFlyBSD, and FreeBSD. (On FreeBSD, it applies the socket option SO_REUSEPORT_LB instead.) Some other operating systems support the socket option but it won’t have the desired effect: It will allow multiple processes to bind to the same port but only one of them will get the connections. See your operating system’s setsockopt() documentation for details.

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 look at docs for the peer_id configuration option and the peers configuration section. There’s also an example that uses peering and so_reuseport in the example section of these docs.

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_buffer

Default: not set

tcp_keepalive

Turns 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_keepcnt

Default: not set

tcp_keepidle

Default: not set

tcp_keepintvl

Default: 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

Section [databases]

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 characters need to be quoted with standard SQL identifier quoting: double quotes, with “” for a single instance of a double quote.

The database name “pgbouncer” is reserved for the admin console and cannot be used as a key here.

“*” acts as a 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 an entry (and no other overriding entries)

* = host=foo

then a connection to PgBouncer specifying a database “bar” will effectively behave as if an entry

bar = host=foo dbname=bar

exists (taking advantage of the default for dbname being the client-side database name; see below).

Such automatically created database entries are cleaned up if they stay idle longer than the time specified by the autodb_idle_timeout parameter.

dbname

Destination 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

port

Default: 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 will be used for the user specified above. Dynamic forms of password discovery such as auth_query are not currently supported.

auth_user

Override of the global auth_user setting, if specified.

auth_query

Override of the global auth_query setting, if specified. The entire SQL statement needs to be enclosed in single quotes.

auth_dbname

Override of the global auth_dbname 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.

Only enforced if at least one of the following is true:

  • this entry in the [database] section has a value set for the user key (aka forced user)
  • there is at least one client connected to the pool

reserve_pool

Set additional connections for this database. If not set, reserve_pool_size is used.

connect_query

Query 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_connections

Configure a database-wide maximum (i.e. all pools within the database will not have more than this many server connections).

server_lifetime

Configure the server_lifetime per database. If not set the database will fall back to the instance wide configured value for server_lifetime

client_encoding

Ask specific client_encoding from server.

datestyle

Ask specific datestyle from server.

timezone

Ask specific timezone from server.

Section [users]

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 key=value pairs of configuration settings specific for this user. Example:

user1 = pool_mode=session

Only a few settings are available here.

Note that when auth_file is configured, if a user is defined in this section but not listed in auth_file, pgBouncer will attempt to use auth_query to find a password for that user if auth_user is set. If auth_user is not set, pgBouncer will pretend the user exists and fail to return “no such user” messages to the client, but neither will it accept any provided password.

pool_size

Set the maximum size of pools for all connections from this user. If not set, the database or default_pool_size is used.

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_connections

Configure a maximum for the user (i.e. all pools with the user will not have more than this many server connections).

Section [peers]

The section [peers] 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 cancellations 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. A more in-depth explanation is provided in this recording of a conference talk.

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

Note 1: 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 of these docs. 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 config management easy. Because it allows using the exact same [peers] section for multiple configs.

Note 2: Cross-version peering is supported as long as all peers are on the same side of the v1.21.0 version boundary. In v1.21.0 some breaking changes were made in how we encode the cancellation tokens that made them incompatible with the ones created by earlier versions.

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

port

Default: 6432

pool_size

Set the maximum number of cancel requests that can be in flight to the peer at the same time. It’s quite normal for cancel requests to arrive in bursts, e.g. when the backing Postgres server slow or down. So it’s important for pool_size to not be so low that it cannot handle these bursts.

If not set, the default_pool_size is used.

Include directive

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 file name is not an absolute path, it is taken as relative to the current working directory.

Authentication file format

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 2 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.

PostgreSQL MD5-hashed password format:

"md5" + md5(password + username)

So user admin with password 1234 will have MD5-hashed password md545f2603610af569b6155c45067268c6b.

PostgreSQL SCRAM secret format:

SCRAM-SHA-256$<iterations>:<salt>$<storedkey>:<serverkey>

See the PostgreSQL documentation and RFC 5803 for details on this.

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).

Limitations

If the password is stored in plain text, it can be used for any password-based authentication used in the backend server; plain text, MD5 or SCRAM (see https://www.postgresql.org/docs/current/auth-password.html for details).

MD5-hashed passwords can be used if backend server uses MD5 authentication (or specific users have MD5-hashed passwords).

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 PostgreSQL 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.

Note on managed servers

If the backend server is configured to use SCRAM password authentication PgBouncer cannot successfully authenticate if it does not know either a) user password in plain text or b) corresponding SCRAM secret.

Some cloud providers (i.e. AWS RDS) prohibit access to PostgreSQL sensitive system tables for fetching passwords. Even for the most privileged user (i.e. member of rds_superuser) the select * from pg_authid; returns the ERROR: permission denied for table pg_authid. That is a known behaviour (blog).

Therefore, fetching an existing SCRAM secret once it has been stored in a managed server is impossible which makes it hard to configure PgBouncer to use the same SCRAM secret. Nevertheless, SCRAM secret can still be configured and used on both sides using the following trick:

Generate SCRAM secret for arbitrary password with a tool that is capable of printing out the secret. For example psql --echo-hidden and the command \password prints out the SCRAM secret to the console before sending it over to the server.

$ psql --echo-hidden <connection_string>
postgres=# \password <role_name>
Enter new password for user "<role_name>": 
Enter it again: 
********* QUERY **********
ALTER USER <role_name> PASSWORD 'SCRAM-SHA-256$<iterations>:<salt>$<storedkey>:<serverkey>'
**************************

Note down the SCRAM secret from the QUERY and set it in PgBouncer’s userlist.txt.

If you used a tool other than psql --echo-hidden then you need to set the SCRAM secret also in the server (you can use alter role <role_name> password '<scram_secret>' for that).

HBA file format

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 PostgreSQL pg_hba.conf file (see https://www.postgresql.org/docs/current/auth-pg-hba-conf.html).

  • Supported record types: local, host, hostssl, hostnossl.
  • Database field: Supports all, replication, sameuser, @file, multiple names. Not supported: samerole, samegroup.
  • User name field: Supports all, @file, multiple names. Not supported: +groupname.
  • Address field: Supports all, IPv4, IPv6. Not supported: samehost, samenet, 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 supported when auth_type is cert or peer.

Ident map file format

The location of the ident map file is specified by the setting auth_ident_file. It is only loaded if auth_type is set to hba.

The file format is a simplified variation of the PostgreSQL ident map file (see https://www.postgresql.org/docs/current/auth-username-maps.html).

  • Supported lines are only of the form map-name system-username database-username.
  • There is no support for including file/directory.
  • System-username field: Not supported: regular expressions.
  • Database-username field: Supports all or a single postgres user name. Not supported: +groupname, regular expressions.

Examples

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

See also

pgbouncer(1) - man page for general usage, console commands

https://www.pgbouncer.org/