Heroku buildpack: pgbouncer

This is a Heroku buildpack that allows one to run pgbouncer and stunnel in a dyno alongside application code. It is meant to be used in conjunction with other buildpacks as part of a multi-buildpack.

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.

It uses stunnel and pgbouncer.

FAQ

Q: Why should I use transaction pooling?
A: You have many workers per dyno that hold open idle Postgres connections and 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 diable 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
.buildpacks  Gemfile  Gemfile.lock  Procfile  config/  config.ru

$ heroku config:add BUILDPACK_URL=https://github.com/ddollar/heroku-buildpack-multi.git

$ cat .buildpacks
https://github.com/gregburek/heroku-buildpack-pgbouncer.git#v0.3.2
https://github.com/heroku/heroku-buildpack-ruby.git

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

$ git push heroku master
...
-----> Fetching custom git buildpack... done
-----> Multipack app detected
=====> Downloading Buildpack: https://github.com/gregburek/heroku-buildpack-pgbouncer.git
=====> Detected Framework: pgbouncer-stunnel
       Using pgbouncer version: 1.5.4
       Using stunnel version: 5.02
       Using stack version: cedar
-----> Fetching and vendoring pgbouncer into slug
-----> Fetching and vendoring stunnel into slug
-----> Moving the configuration generation script into app/bin
-----> Moving the start-pgbouncer-stunnel script into app/bin
-----> pgbouncer/stunnel done
=====> Downloading Buildpack: https://github.com/heroku/heroku-buildpack-ruby.git
=====> Detected Framework: Ruby/Rack
-----> Using Ruby version: ruby-1.9.3
-----> Installing dependencies using Bundler version 1.3.2
...

The buildpack will install and configure pgbouncer and stunnel to connect to DATABASE_URL over a SSL connection. Prepend bin/start-pgbouncer-stunnel to any process in the Procfile to run pgbouncer and stunnel 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:password@ec2-107-20-228-134.compute-1.amazonaws.com:5482/db6h3bkfuk5430
DATABASE_URL=postgres://uf2782hv7b3uqe:password@ec2-50-19-210-113.compute-1.amazonaws.com:5622/deamhhcj6q0d31

~ $ bin/start-pgbouncer-stunnel env # filtered for brevity
HEROKU_POSTGRESQL_ROSE_URL=postgres://u9dih9htu2t3ll:password@127.0.0.1:6000/db2
DATABASE_URL=postgres://uf2782hv7b3uqe:password@127.0.0.1:6000/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 and stunnel 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_RESERVE_POOL_SIZE Default is 1
PGBOUNCER_RESERVE_POOL_TIMEOUT Default is 5.0 seconds
PGBOUNCER_URLS Default is DATABASE_URL
PGBOUNCER_CONNECTION_RETRY Default is no
PGBOUNCER_LOG_CONNECTIONS Default is yes
PGBOUNCER_LOG_DISCONNECTIONS Default is yes
PGBOUNCER_LOG_POOLER_ERRORS Default is yes
PGBOUNCER_STATS_PERIOD Default is 60
PGBOUNCER_SERVER_RESET_QUERY Default is empty when pool mode is transaction, and "DISCARD ALL;" when session.

For more info, see CONTRIBUTING.md

michaeldiscala / heroku-buildpack-pgbouncer