multi-process-workers.md

Multi Process Workers

This article describes how to use Fluentd's multi-process workers feature for high traffic. This feature launches two or more fluentd workers to utilize multiple CPU powers.

This feature can simply replace fluent-plugin-multiprocess.

How It Works

By default, one instance of fluentd launches a supervisor and a worker. A worker consists of input/filter/output plugins.

The multi-process workers feature launches multiple workers and use a separate process per worker. fluentd provides several features for multi-process workers.

Configuration

workers Parameter

<system> directive has workers parameter for specifying the number of workers:

<system>
  workers 4
</system>

With this configuration, fluentd launches four (4) workers.

<worker> directive

Some plugins do not work with multi-process workers feature automatically, e.g. in_tail. However, these plugins can be configured to run on specific workers with <worker N> directive. N is a zero-based worker index.

In the following example, the in_tail plugin will run only on worker 0 out of the 4 workers configured in the <system> directive:

<system>
  workers 4
</system>

# work on multi process workers. worker0 - worker3 run in_forward
<source>
  @type forward
</source>

# work on only worker 0. worker1 - worker3 don't run in_tail
<worker 0>
  <source>
    @type tail
  </source>
</worker>

# <worker 1>, <worker 2> or <worker 3> is also ok

With <worker> directive, non-multi-process-ready plugins can seamlessly be run along with multi-process-ready plugins.

<worker N-M> directive

As of Fluentd v1.4.0, <worker N-M> syntax has been introduced:

<system>
  workers 6
</system>

# work on worker 0 and worker 1
<worker 0-1>
  <source>
    @type forward
  </source>

  <filter test>
    @type record_transformer
    enable_ruby
    <record>
      worker_id ${ENV['SERVERENGINE_WORKER_ID']}
    </record>
  </filter>

  <match test>
    @type stdout
  </match>
</worker>

# work on worker 2 and worker 3
<worker 2-3>
  <source>
    @type tcp
    <parse>
      @type none
    </parse>
    tag test
  </source>

  <filter test>
    @type record_transformer
    enable_ruby
    <record>
      worker_id ${ENV['SERVERENGINE_WORKER_ID']}
    </record>
  </filter>

  <match test>
    @type stdout
  </match>
</worker>

# work on worker 4 and worker 5
<worker 4-5>
  <source>
    @type udp
    <parse>
      @type none
    </parse>
    tag test
  </source>

  <filter test>
    @type record_transformer
    enable_ruby
    <record>
      worker_id ${ENV['SERVERENGINE_WORKER_ID']}
    </record>
  </filter>

  <match test>
    @type stdout
  </match>
</worker>

With this directive, you can specify multiple workers per worker directive.

Operation

Each worker consumes memory and disk space separately. Take care while configuring buffer spaces and monitoring memory/disk consumption.

Multi-Process Workers and Plugins

Input Plugin

There are three (3) types of input plugins:

• feature supported and server helper based plugin
• feature supported and plain plugin
• feature unsupported

feature supported and server helper based plugin

Server plugin helper based plugin can share port between workers. For example, forward input plugin does not need multiple ports on multi process workers. forward input's port is shared among workers.

<system>
  workers 4
</system>

<source>
  @type forward
  port 24224 # 4 workers accept events on this port
</source>

feature supported and plain plugin

Non-server plugin helper based plugin set up socket/server in each worker. For example, monitor_agent needs multiple ports on multi-process workers. The port is assigned sequentially.

<system>
  workers 4
</system>

<source>
  @type monitor_agent
  port 25000 # worker0: 25000, worker1: 25001, ...
</source>

feature unsupported

Some plugins do not work on multi-process workers. For example, tail input does not work because in_tail cannot be implemented with multi process.

You can run these plugins with <worker N> directive. See "Configuration" section.

Output Plugin

By default, no additional changes are required but some plugins do need to specify the worker_id in the configuration. For example, file and S3 plugins store events into a specified path. The problem is if the plugins under multi-process workers flush events at the same time, the destination path is also the same which results in data loss. To avoid this problem, a worker_id or some random string can be configured.

# s3 plugin example

<match pattern>
  @type s3

  # Good
  path "logs/#{worker_id}/${tag}/%Y/%m/%d/"

  # Bad on multi process worker!
  path logs/${tag}/%Y/%m/%d/
</match>

See Configuration File article for embedded Ruby code feature.

FAQ

Fluentd cannot start with multi-process workers, why?

You may see following error in the fluentd logs:

2018-10-01 10:00:00 +0900 [error]: config error file="/path/to/fluentd.conf" error_class=Fluent::ConfigError error="Plugin 'tail' does not support multi workers configuration (Fluent::Plugin::TailInput)"

This means that the configured plugin does not support multi-process workers. All configured plugins must support multi-process workers. See "Multi-Process Worker and Plugins" section above.

If this article is incorrect or outdated, or omits critical information, please let us know. Fluentd is an open-source project under Cloud Native Computing Foundation (CNCF). All components are available under the Apache 2 License.