Heroku buildpack: pgbouncer

This is a Heroku buildpack that allows one to run pgbouncer in a dyno alongside application code. It is meant to be used in conjunction with other buildpacks.

The primary use of this buildpack is to allow for transaction pooling of PostgreSQL database connections among multiple workers in a dyno. For example, 10 unicorn workers would be able to share a single database connection, avoiding connection limits and Out Of Memory errors on the Postgres server.

FAQ

• Q: Why should I use transaction pooling?

• A: You have many workers per dyno that hold open idle Postgres connections and you want to reduce the number of unused connections. This is a slightly more complete answer from stackoverflow

• Q: Why shouldn't I use transaction pooling?

• A: If you need to use named prepared statements, advisory locks, listen/notify, or other features that operate on a session level. Please refer to PGBouncer's feature matrix for all transaction pooling caveats.

Disable Prepared Statements

With Rails 4.1, you can disable prepared statements by appending ?prepared_statements=false to the database's URI. Set the PGBOUNCER_PREPARED_STATEMENTS config var to false for the buildpack to do that for you.

Rails versions 4.0.0 - 4.0.3, reportedly can't disable prepared statements at all. Make sure your framework is up to date before troubleshooting prepared statements failures.

Rails 3.2 - 4.0 also requires an initializer to properly cast the prepared_statements configuration string as a boolean. This initializer is adapted from this commit. In file config/initializers/database_connection.rb insert the following:

require "active_record/connection_adapters/postgresql_adapter"

class ActiveRecord::ConnectionAdapters::PostgreSQLAdapter
  alias initialize_without_config_boolean_coercion initialize
  def initialize(connection, logger, connection_parameters, config)
    if config[:prepared_statements] == 'false'
      config = config.merge(prepared_statements: false)
    end
    initialize_without_config_boolean_coercion(connection, logger, connection_parameters, config)
  end
end

Usage

Example usage:

$ ls -a
Gemfile  Gemfile.lock  Procfile  config/  config.ru

$ heroku buildpacks:add heroku/pgbouncer
Buildpack added. Next release on pgbouncer-test-app will use heroku/pgbouncer.
Run `git push heroku master` to create a new release using this buildpack.

$ heroku buildpacks:add heroku/ruby
Buildpack added. Next release on pgbouncer-test-app will use:
  1. https://github.com/heroku/heroku-buildpack-pgbouncer
  2. https://github.com/heroku/heroku-buildpack-ruby
Run `git push heroku master` to create a new release using these buildpacks.

$ cat Procfile
web:    bin/start-pgbouncer bundle exec unicorn -p $PORT -c ./config/unicorn.rb -E $RACK_ENV
worker: bundle exec rake worker

$ git push heroku master
...
-----> Multipack app detected
-----> Fetching custom git buildpack... done
-----> pgbouncer app detected
       Using pgbouncer version: 1.7-heroku
-----> Fetching and vendoring pgbouncer into slug
-----> Moving the configuration generation script into app/bin
-----> Moving the start-pgbouncer script into app/bin
-----> pgbouncer done
-----> Fetching custom git buildpack... done
...

The buildpack will install and configure pgbouncer to connect to DATABASE_URL over a TLS connection, where available. Prepend bin/start-pgbouncer to any process in the Procfile to run pgbouncer alongside that process.

Multiple Databases

It is possible to connect to multiple databases through pgbouncer by setting PGBOUNCER_URLS to a list of config vars. Example:

$ heroku config:add PGBOUNCER_URLS="DATABASE_URL HEROKU_POSTGRESQL_ROSE_URL"
$ heroku run bash

~ $ env | grep 'HEROKU_POSTGRESQL_ROSE_URL\|DATABASE_URL'
HEROKU_POSTGRESQL_ROSE_URL=postgres://u9dih9htu2t3ll:[email protected]:5482/db6h3bkfuk5430
DATABASE_URL=postgres://uf2782hv7b3uqe:[email protected]:5622/deamhhcj6q0d31

~ $ bin/start-pgbouncer env # filtered for brevity
HEROKU_POSTGRESQL_ROSE_URL=postgres://u9dih9htu2t3ll:[email protected]:6000/db2
DATABASE_URL=postgres://uf2782hv7b3uqe:[email protected]:6000/db1

Follower Replica Databases

As of v0.3.2 of this buildpack, it is possible to use pgbouncer to connect to multiple databases that share a database name, such as a leader and follower. To use, add the follower's config var to PGBOUNCER_URLS as detailed in the Multiple Databases section.

If you are using Octopus Replication to send reads to a replica, make sure to include the color url of your leader in the SLAVE_DISABLED_FOLLOWERS blacklist. Otherwise, Octopus will attempt to use your leader as a read-only replica, potentially doubling your connection count.

Tweak settings

Some settings are configurable through app config vars at runtime. Refer to the appropriate documentation for pgbouncer configurations to see what settings are right for you.

• PGBOUNCER_POOL_MODE Default is transaction
• PGBOUNCER_MAX_CLIENT_CONN Default is 100
• PGBOUNCER_DEFAULT_POOL_SIZE Default is 1
• PGBOUNCER_MIN_POOL_SIZE Default is 0
• PGBOUNCER_RESERVE_POOL_SIZE Default is 1
• PGBOUNCER_RESERVE_POOL_TIMEOUT Default is 5.0 seconds
• PGBOUNCER_SERVER_LIFETIME Default is 3600.0 seconds
• PGBOUNCER_SERVER_IDLE_TIMEOUT Default is 600.0 seconds
• PGBOUNCER_URLS should contain all config variables that will be overridden to connect to pgbouncer. For example, set this to AMAZON_RDS_URL to send RDS connections through pgbouncer. The default is DATABASE_URL.
• PGBOUNCER_LOG_CONNECTIONS Default is 1. If your app does not use persistent database connections, this may be noisy and should be set to 0.
• PGBOUNCER_LOG_DISCONNECTIONS Default is 1. If your app does not use persistent database connections, this may be noisy and should be set to 0.
• PGBOUNCER_LOG_POOLER_ERRORS Default is 1
• PGBOUNCER_STATS_PERIOD Default is 60
• PGBOUNCER_SERVER_RESET_QUERY Default is empty when pool mode is transaction, and "DISCARD ALL;" when session.
• PGBOUNCER_IGNORE_STARTUP_PARAMETERS Adds parameters to ignore when pgbouncer is starting. Some postgres libraries, like Go's pq, append this parameter, making it impossible to use this buildpack. Default is empty and the most common ignored parameter is extra_float_digits. Multiple parameters can be seperated via commas. Example: PGBOUNCER_IGNORE_STARTUP_PARAMETERS="extra_float_digits, some_other_param"
• PGBOUNCER_QUERY_WAIT_TIMEOUT Default is 120 seconds, helps when the server is down or the database rejects connections for any reason. If this is disabled, clients will be queued infinitely.
• PGBOUNCER_PKT_BUF Default is 4096.
• PGBOUNCER_MAX_PACKET_SIZE Default is 2147483647.
• PGBOUNCER_LISTEN_BACKLOG Default is 128.
• PGBOUNCER_SBUF_LOOPCNT Default is 5.
• PGBOUNCER_SUSPEND_TIMEOUT Default is 10.
• PGBOUNCER_TCP_DEFER_ACCEPT Default is 45.
• PGBOUNCER_TCP_KEEPALIVE Default is 1.
• PGBOUNCER_TCP_KEEPCNT Default is 9.
• PGBOUNCER_TCP_KEEPIDLE Default is 7200.
• PGBOUNCER_TCP_KEEPINTVL Default is 75.
• PGBOUNCER_STATS_USERNAME and PGBOUNCER_STATS_PASSWORD Set these to enable stats_users SHOW access to pgbouncer.

For more info, see CONTRIBUTING.md

Using the edge version of the buildpack

The heroku/pgbouncer buildpack points to the latest stable version of the buildpack published in the Buildpack Registry. To use the latest version of the buildpack (the code in this repository, run the following command:

$ heroku buildpacks:add https://github.com/heroku/heroku-buildpack-pgbouncer

Notes

Currently, the connection string parsing requires the connection string to be in a specific format:

postgres://<user>:<pass>@<host>:<port>/<database>

This corresponds to the regular expression ^postgres(?:ql)?:\/\/([^:]*):([^@]*)@(.*?):(.*?)\/(.*?)$. All components must be present in order for the buildpack to correctly parse the connection string.