Last updated November 21, 2022

By design, connections to Postgres databases are persistent to reduce the performance impact of having to re-establish a connection for every invocation. While this increases the performance of your application it also requires properly establishing the connection, especially in forked environments.

Forked environments

If you’re using a framework or library that uses forked processes, connections to Postgres (and any other resources) should be established after the fork completes. This ensures that each forked process has its own connection and avoids several of the most common connection errors such as no connection to the server and SSL SYSCALL error: EOF detected and SSL error: decryption failed or bad record mac.

Connection instructions for several common frameworks and libraries are included here.

Database connection pools

When using a fork-based server of any kind, each fork receives its own database connection pool. For this reason, most applications using fork-based web servers receive the best performance when the application-level (for example, Rails, Sequel, or Django) connection pool isn’t used, or used with a pool size of one.

Overuse of database connections can result in lower overall performance, “out of memory” errors reported by the database server, and the database server’s refusal to accept connections from additional clients (connection limit reached).

Applications that are exempt from this advice tend to have been purposefully written to take advantage of or require multiple simultaneous transactions within the context of a single HTTP request.

Ruby on Rails

Important Postgres reconnection bugs have been fixed in ActiveRecord 3.2.9. Previous releases (3.1, 3.0, 2.x) haven’t received these enhancements. If you’re using ActiveRecord 3.2, upgrading to version 3.2.9 or later is recommended.

For previous versions, exiting the process when PGError is propagated from the application is recommended since Heroku will automatically restart crashed dynos.

Unicorn server

Unicorn is a Rack HTTP server that uses forked process to handle incoming requests. Specify the before and after fork behavior in the unicorn.conf configuration file.

In a Rails app or an app using ActiveRecord, add the following before_fork and after_fork blocks in unicorn.conf:

before_fork do |server, worker|

  Signal.trap 'TERM' do
    puts 'Unicorn master intercepting TERM and sending myself QUIT instead'
    Process.kill 'QUIT', Process.pid
  end

  defined?(ActiveRecord::Base) and
    ActiveRecord::Base.connection.disconnect!
end

after_fork do |server, worker|

  Signal.trap 'TERM' do
    puts 'Unicorn worker intercepting TERM and doing nothing. Wait for master to sent QUIT'
  end

  defined?(ActiveRecord::Base) and
    ActiveRecord::Base.establish_connection(
      Rails.application.config.database_configuration[Rails.env]
    )

end

Resque Ruby queuing

Resque uses forking to create new worker processes. The main process connection should be disconnected before forking (to avoid consuming unnecessary resources) while worker connections should be established after the fork occurs.

You can specify this behavior by cleaning up and re-establishing connections in an initializer:

  Resque.before_fork do
    defined?(ActiveRecord::Base) and
      ActiveRecord::Base.connection.disconnect!
  end

  Resque.after_fork do
    defined?(ActiveRecord::Base) and
      ActiveRecord::Base.establish_connection
  end

Sidekiq

Sidekiq uses threads to handle many jobs at the same time in the same process. To prevent sharing of connections you must configure the Sidekiq server to correctly establish its own connection.

You can specify this behavior by cleaning up and re-establishing connections in a Sidekiq initializer:

See Concurrency and Database Connections for Background Workers

Disabling New Relic EXPLAIN

The current implementation of New Relic auto-EXPLAIN can cause one extra database connection to be used per fork. For high-volume applications that cannot tolerate the extra connection, it may be worthwhile to disable the automatic EXPLAIN feature of New Relic. For instructions about how to do this, see the New Relic documentation in Manually changing your configuration.