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.
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.
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
$ 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
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:firstname.lastname@example.org:5482/db6h3bkfuk5430 DATABASE_URL=postgres://uf2782hv7b3uqe:email@example.com:5622/deamhhcj6q0d31 ~ $ bin/start-pgbouncer env # filtered for brevity HEROKU_POSTGRESQL_ROSE_URL=postgres://u9dih9htu2t3ll:firstname.lastname@example.org:6000/db2 DATABASE_URL=postgres://uf2782hv7b3uqe:email@example.com:6000/db1
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.
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_MODEDefault is transaction
PGBOUNCER_MAX_CLIENT_CONNDefault is 100
PGBOUNCER_DEFAULT_POOL_SIZEDefault is 1
PGBOUNCER_MIN_POOL_SIZEDefault is 0
PGBOUNCER_RESERVE_POOL_SIZEDefault is 1
PGBOUNCER_RESERVE_POOL_TIMEOUTDefault is 5.0 seconds
PGBOUNCER_SERVER_LIFETIMEDefault is 3600.0 seconds
PGBOUNCER_SERVER_IDLE_TIMEOUTDefault is 600.0 seconds
PGBOUNCER_URLSshould contain all config variables that will be overridden to connect to pgbouncer. For example, set this to
AMAZON_RDS_URLto send RDS connections through pgbouncer. The default is
PGBOUNCER_CONNECTION_RETRYDefault is no
PGBOUNCER_LOG_CONNECTIONSDefault is 1. If your app does not use persistent database connections, this may be noisy and should be set to 0.
PGBOUNCER_LOG_DISCONNECTIONSDefault is 1. If your app does not use persistent database connections, this may be noisy and should be set to 0.
PGBOUNCER_LOG_POOLER_ERRORSDefault is 1
PGBOUNCER_STATS_PERIODDefault is 60
PGBOUNCER_SERVER_RESET_QUERYDefault is empty when pool mode is transaction, and "DISCARD ALL;" when session.
PGBOUNCER_IGNORE_STARTUP_PARAMETERSAdds 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_CONNECT_QUERYThe query to be executed on newly created connections. This is useful for setting things like
PGBOUNCER_CONNECT_QUERY="set statement_timeout to 10000"
PGBOUNCER_QUERY_WAIT_TIMEOUTDefault 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.
For more info, see CONTRIBUTING.md
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
Currently, the connection string parsing requires the connection string to be in a specific format:
This corresponds to the regular expression
^postgres(?:ql)?:\/\/([^:]*):([^@]*)@(.*?):(.*?)\/(.*?)$. All components must be present in order for the buildpack to correctly parse the connection string.
Copy the snippet above into CLI.