fluent-plugin-kafka, a plugin for Fluentd

A fluentd plugin to both consume and produce data for Apache Kafka.

Installation

Add this line to your application’s Gemfile:

gem 'fluent-plugin-kafka'

And then execute:

$ bundle

Or install it yourself as:

$ gem install fluent-plugin-kafka --no-document

If you want to use zookeeper related parameters, you also need to install zookeeper gem. zookeeper gem includes native extension, so development tools are needed, e.g. ruby-devel, gcc, make and etc.

Requirements

• Ruby 2.1 or later
• Input plugins work with kafka v0.9 or later
• Output plugins work with kafka v0.8 or later

Usage

Common parameters

SSL authentication

• ssl_ca_cert
• ssl_client_cert
• ssl_client_cert_key
• ssl_client_cert_key_password
• ssl_ca_certs_from_system

Set path to SSL related files. See Encryption and Authentication using SSL for more detail.

SASL authentication

with GSSAPI

• principal
• keytab

Set principal and path to keytab for SASL/GSSAPI authentication. See Authentication using SASL for more details.

with Plain/SCRAM

• username
• password
• scram_mechanism
• sasl_over_ssl

Set username, password, scram_mechanism and sasl_over_ssl for SASL/Plain or Scram authentication. See Authentication using SASL for more details.

Input plugin (@type ‘kafka’)

Consume events by single consumer.

<source>
  @type kafka

  brokers <broker1_host>:<broker1_port>,<broker2_host>:<broker2_port>,..
  topics <listening topics(separate with comma',')>
  format <input text type (text|json|ltsv|msgpack)> :default => json
  message_key <key (Optional, for text format only, default is message)>
  add_prefix <tag prefix (Optional)>
  add_suffix <tag suffix (Optional)>

  # Optionally, you can manage topic offset by using zookeeper
  offset_zookeeper    <zookeer node list (<zookeeper1_host>:<zookeeper1_port>,<zookeeper2_host>:<zookeeper2_port>,..)>
  offset_zk_root_node <offset path in zookeeper> default => '/fluent-plugin-kafka'

  # ruby-kafka consumer options
  max_bytes     (integer) :default => nil (Use default of ruby-kafka)
  max_wait_time (integer) :default => nil (Use default of ruby-kafka)
  min_bytes     (integer) :default => nil (Use default of ruby-kafka)
</source>

Supports a start of processing from the assigned offset for specific topics.

<source>
  @type kafka

  brokers <broker1_host>:<broker1_port>,<broker2_host>:<broker2_port>,..
  format <input text type (text|json|ltsv|msgpack)>
  <topic>
    topic     <listening topic>
    partition <listening partition: default=0>
    offset    <listening start offset: default=-1>
  </topic>
  <topic>
    topic     <listening topic>
    partition <listening partition: default=0>
    offset    <listening start offset: default=-1>
  </topic>
</source>

See also ruby-kafka README for more detailed documentation about ruby-kafka.

Consuming topic name is used for event tag. So when the target topic name is app_event, the tag is app_event. If you want to modify tag, use add_prefix or add_suffix parameters. With add_prefix kafka, the tag is kafka.app_event.

Input plugin (@type ‘kafka_group’, supports kafka group)

Consume events by kafka consumer group features..

<source>
  @type kafka_group

  brokers <broker1_host>:<broker1_port>,<broker2_host>:<broker2_port>,..
  consumer_group <consumer group name, must set>
  topics <listening topics(separate with comma',')>
  format <input text type (text|json|ltsv|msgpack)> :default => json
  message_key <key (Optional, for text format only, default is message)>
  kafka_message_key <key (Optional, If specified, set kafka's message key to this key)>
  add_headers <If true, add kafka's message headers to record>
  add_prefix <tag prefix (Optional)>
  add_suffix <tag suffix (Optional)>
  retry_emit_limit <Wait retry_emit_limit x 1s when BuffereQueueLimitError happens. The default is nil and it means waiting until BufferQueueLimitError is resolved>
  use_record_time (Deprecated. Use 'time_source record' instead.) <If true, replace event time with contents of 'time' field of fetched record>
  time_source <source for message timestamp (now|kafka|record)> :default => now
  time_format <string (Optional when use_record_time is used)>

  # ruby-kafka consumer options
  max_bytes               (integer) :default => 1048576
  max_wait_time           (integer) :default => nil (Use default of ruby-kafka)
  min_bytes               (integer) :default => nil (Use default of ruby-kafka)
  offset_commit_interval  (integer) :default => nil (Use default of ruby-kafka)
  offset_commit_threshold (integer) :default => nil (Use default of ruby-kafka)
  fetcher_max_queue_size  (integer) :default => nil (Use default of ruby-kafka)
  refresh_topic_interval  (integer) :default => nil (Use default of ruby-kafka)
  start_from_beginning    (bool)    :default => true
</source>

See also ruby-kafka README for more detailed documentation about ruby-kafka options.

topics supports regex pattern since v0.13.1. If you want to use regex pattern, use /pattern/ like /foo.*/.

Consuming topic name is used for event tag. So when the target topic name is app_event, the tag is app_event. If you want to modify tag, use add_prefix or add_suffix parameter. With add_prefix kafka, the tag is kafka.app_event.

Input plugin (@type ‘rdkafka_group’, supports kafka consumer groups, uses rdkafka-ruby)

The in_rdkafka_group consumer was not yet tested under heavy production load. Use it at your own risk!

With the introduction of the rdkafka-ruby based input plugin we hope to support Kafka brokers above version 2.1 where we saw compatibility issues when using the ruby-kafka based @kafka_group input type. The rdkafka-ruby lib wraps the highly performant and production ready librdkafka C lib.

<source>
  @type rdkafka_group
  topics <listening topics(separate with comma',')>
  format <input text type (text|json|ltsv|msgpack)> :default => json
  message_key <key (Optional, for text format only, default is message)>
  kafka_message_key <key (Optional, If specified, set kafka's message key to this key)>
  add_headers <If true, add kafka's message headers to record>
  add_prefix <tag prefix (Optional)>
  add_suffix <tag suffix (Optional)>
  retry_emit_limit <Wait retry_emit_limit x 1s when BuffereQueueLimitError happens. The default is nil and it means waiting until BufferQueueLimitError is resolved>
  use_record_time (Deprecated. Use 'time_source record' instead.) <If true, replace event time with contents of 'time' field of fetched record>
  time_source <source for message timestamp (now|kafka|record)> :default => now
  time_format <string (Optional when use_record_time is used)>

  # The SSL and SASL parameters in "Common parameters" are mapped to librdkafka
  # properties. Keys given in kafka_configs take precedence over the mapped values.
  # ssl_ca_cert uses only its first entry, and ssl_client_cert_chain is not mapped.
  # When true, SASL authentication requires an SSL connection
  sasl_over_ssl (bool) :default => true

  # kafka consumer options
  max_wait_time_ms 500
  max_batch_size 10000
  kafka_configs {
    "bootstrap.servers": "brokers <broker1_host>:<broker1_port>,<broker2_host>:<broker2_port>",
    "group.id": "<consumer group name>"
  }
</source>

See also rdkafka-ruby and librdkafka for more detailed documentation about Kafka consumer options.

topics supports POSIX Extended Regular Expression pattern (not Ruby regex syntax) since v0.19.6. If you want to use regex pattern, use /pattern/ like /foo.*/.

Note: A caret (^) is automatically added to the beginning of the pattern.

Consuming topic name is used for event tag. So when the target topic name is app_event, the tag is app_event. If you want to modify tag, use add_prefix or add_suffix parameter. With add_prefix kafka, the tag is kafka.app_event.

Output plugin

This kafka2 plugin is for fluentd v1 or later. This plugin uses ruby-kafka producer for writing data. If ruby-kafka doesn’t fit your kafka environment, check rdkafka2 plugin instead. This will be out_kafka plugin in the future.

<match app.**>
  @type kafka2

  brokers               <broker1_host>:<broker1_port>,<broker2_host>:<broker2_port>,.. # Set brokers directly

  # Kafka topic, placerholders are supported. Chunk keys are required in the Buffer section inorder for placeholders
  # to work.
  topic                 (string) :default => nil
  topic_key             (string) :default => 'topic'
  partition_key         (string) :default => 'partition'
  partition_key_key     (string) :default => 'partition_key'
  message_key_key       (string) :default => 'message_key'
  default_topic         (string) :default => nil
  default_partition_key (string) :default => nil
  record_key            (string) :default => nil
  default_message_key   (string) :default => nil
  exclude_topic_key     (bool)   :default => false
  exclude_partition_key (bool)   :default => false
  exclude_partition     (bool)   :default => false
  exclude_message_key   (bool)   :default => false
  get_kafka_client_log  (bool)   :default => false
  headers               (hash)   :default => {}
  headers_from_record   (hash)   :default => {}
  use_event_time        (bool)   :default => false
  use_default_for_unknown_topic (bool) :default => false
  discard_kafka_delivery_failed (bool) :default => false (No discard)
  partitioner_hash_function (enum) (crc32|murmur2) :default => 'crc32'
  share_producer        (bool)   :default => false
  idempotent            (bool)   :default => false

  # If you intend to rely on AWS IAM auth to MSK with long lived credentials
  # https://docs.aws.amazon.com/msk/latest/developerguide/iam-access-control.html
  #
  # For AWS STS support, see status in
  # - https://github.com/zendesk/ruby-kafka/issues/944
  # - https://github.com/zendesk/ruby-kafka/pull/951
  sasl_aws_msk_iam_access_key_id (string) :default => nil
  sasl_aws_msk_iam_secret_key_id (string) :default => nil
  sasl_aws_msk_iam_aws_region    (string) :default => nil

  <format>
    @type (json|ltsv|msgpack|attr:<record name>|<formatter name>) :default => json
  </format>

  # Optional. See https://docs.fluentd.org/v/1.0/configuration/inject-section
  <inject>
    tag_key tag
    time_key time
  </inject>

  # See fluentd document for buffer related parameters: https://docs.fluentd.org/v/1.0/configuration/buffer-section
  # Buffer chunk key should be same with topic_key. If value is not found in the record, default_topic is used.
  <buffer topic>
    flush_interval 10s
  </buffer>

  # ruby-kafka producer options
  idempotent        (bool)    :default => false
  sasl_over_ssl     (bool)    :default => true
  max_send_retries  (integer) :default => 1
  required_acks     (integer) :default => -1
  ack_timeout       (integer) :default => nil (Use default of ruby-kafka)
  compression_codec (string)  :default => nil (No compression. Depends on ruby-kafka: https://github.com/zendesk/ruby-kafka#compression)
</match>

The <formatter name> in <format> uses fluentd’s formatter plugins. See formatter article.

Note: Java based Kafka client uses murmur2 as partitioner function by default. If you want to use same partitioning behavior with fluent-plugin-kafka, change it to murmur2 instead of crc32. Note that for using murmur2 hash partitioner function, you must install digest-murmurhash gem.

ruby-kafka sometimes returns Kafka::DeliveryFailed error without good information. In this case, get_kafka_client_log is useful for identifying the error cause. ruby-kafka’s log is routed to fluentd log so you can see ruby-kafka’s log in fluentd logs.

Supports following ruby-kafka’s producer options.

• max_send_retries - default: 2 - Number of times to retry sending of messages to a leader.
• required_acks - default: -1 - The number of acks required per request. If you need flush performance, set lower value, e.g. 1, 2.
• ack_timeout - default: nil - How long the producer waits for acks. The unit is seconds.
• compression_codec - default: nil - The codec the producer uses to compress messages.
• max_send_limit_bytes - default: nil - Max byte size of one record after formatting. A larger record is dropped with a warning, and the other records in the chunk are sent. For example, if you set 1000000, a record larger than 1000000 bytes is dropped. This does not limit the size of a record batch, so it does not prevent MessageSizeTooLarge by itself. See FAQ.
• discard_kafka_delivery_failed - default: false - discard the record where Kafka::DeliveryFailed occurred

If you want to know about detail of monitoring, see also https://github.com/zendesk/ruby-kafka#monitoring

See also Kafka::Client for more detailed documentation about ruby-kafka.

This plugin supports compression codec “snappy” also. Install snappy module before you use snappy compression.

$ gem install snappy --no-document

snappy gem uses native extension, so you need to install several packages before. On Ubuntu, need development packages and snappy library.

$ sudo apt-get install build-essential autoconf automake libtool libsnappy-dev

On CentOS 7 installation is also necessary.

$ sudo yum install gcc autoconf automake libtool snappy-devel

This plugin supports compression codec “lz4” also. Install extlz4 module before you use lz4 compression.

$ gem install extlz4 --no-document

This plugin supports compression codec “zstd” also. Install zstd-ruby module before you use zstd compression.

$ gem install zstd-ruby --no-document

Load balancing

Messages will be assigned a partition at random as default by ruby-kafka, but messages with the same partition key will always be assigned to the same partition by setting default_partition_key in config file. If key name partition_key_key exists in a message, this plugin set the value of partition_key_key as key.

default_partition_key	partition_key_key	behavior
Not set	Not exists	All messages are assigned a partition at random
Set	Not exists	All messages are assigned to the specific partition
Not set	Exists	Messages which have partition_key_key record are assigned to the specific partition, others are assigned a partition at random
Set	Exists	Messages which have partition_key_key record are assigned to the specific partition with partition_key_key, others are assigned to the specific partition with default_parition_key

If key name message_key_key exists in a message, this plugin publishes the value of message_key_key to kafka and can be read by consumers. Same message key will be assigned to all messages by setting default_message_key in config file. If message_key_key exists and if partition_key_key is not set explicitly, messsage_key_key will be used for partitioning.

Headers

It is possible to set headers on Kafka messages. This only works for kafka2 and rdkafka2 output plugin.

The format is like key1:value1,key2:value2. For example:

<match app.**>
  @type kafka2
  [...]
  headers some_header_name:some_header_value
<match>

You may set header values based on a value of a fluentd record field. For example, imagine a fluentd record like:

{"source": { "ip": "127.0.0.1" }, "payload": "hello world" }

And the following fluentd config:

<match app.**>
  @type kafka2
  [...]
  headers_from_record source_ip:$.source.ip
<match>

The Kafka message will have a header of source_ip=12.7.0.0.1.

The configuration format is jsonpath. It is descibed in https://docs.fluentd.org/plugin-helper-overview/api-plugin-helper-record_accessor

Excluding fields

Fields can be excluded from output data. Only works for kafka2 and rdkafka2 output plugin.

Fields must be specified using an array of dot notation $., for example:

<match app.**>
  @type kafka2
  [...]
  exclude_fields $.source.ip,$.HTTP_FOO
<match>

This config can be used to remove fields used on another configs.

For example, $.source.ip can be extracted with config headers_from_record and excluded from message payload.

Using this config to remove unused fields is discouraged. A filter plugin can be used for this purpose.

Send only a sub field as a message payload

If record_key is provided, the plugin sends only a sub field given by that key. The configuration format is jsonpath.

e.g. When the following configuration and the incoming record are given:

configuration:

<match **>
  @type kafka2
  [...]
  record_key '$.data'
</match>

record:

{
    "specversion" : "1.0",
    "type" : "com.example.someevent",
    "id" : "C234-1234-1234",
    "time" : "2018-04-05T17:31:00Z",
    "datacontenttype" : "application/json",
    "data" : {
        "appinfoA" : "abc",
        "appinfoB" : 123,
        "appinfoC" : true
    },
    ...
}

only the data field will be serialized by the formatter and sent to Kafka. The toplevel data key will be removed.

Buffered output plugin

This plugin uses ruby-kafka producer for writing data. This plugin is for v0.12. If you use v1, see kafka2. Support of fluentd v0.12 has ended. kafka_buffered will be an alias of kafka2 and will be removed in the future.

<match app.**>
  @type kafka_buffered

  # Brokers: you can choose either brokers or zookeeper. If you are not familiar with zookeeper, use brokers parameters.
  brokers             <broker1_host>:<broker1_port>,<broker2_host>:<broker2_port>,.. # Set brokers directly
  zookeeper           <zookeeper_host>:<zookeeper_port> # Set brokers via Zookeeper
  zookeeper_path      <broker path in zookeeper> :default => /brokers/ids # Set path in zookeeper for kafka

  topic_key             (string) :default => 'topic'
  partition_key         (string) :default => 'partition'
  partition_key_key     (string) :default => 'partition_key'
  message_key_key       (string) :default => 'message_key'
  default_topic         (string) :default => nil
  default_partition_key (string) :default => nil
  default_message_key   (string) :default => nil
  exclude_topic_key     (bool)   :default => false
  exclude_partition_key (bool)   :default => false
  exclude_partition     (bool)   :default => false
  exclude_message_key   (bool)   :default => false
  output_data_type      (json|ltsv|msgpack|attr:<record name>|<formatter name>) :default => json
  output_include_tag    (bool) :default => false
  output_include_time   (bool) :default => false
  exclude_topic_key     (bool) :default => false
  exclude_partition_key (bool) :default => false
  get_kafka_client_log  (bool) :default => false
  use_event_time        (bool) :default => false
  partitioner_hash_function (enum) (crc32|murmur2) :default => 'crc32'

  # See fluentd document for buffer related parameters: https://docs.fluentd.org/v/0.12/buffer

  # ruby-kafka producer options
  idempotent                   (bool)    :default => false
  sasl_over_ssl                (bool)    :default => true
  max_send_retries             (integer) :default => 1
  required_acks                (integer) :default => -1
  ack_timeout                  (integer) :default => nil (Use default of ruby-kafka)
  compression_codec            (string)  :default => nil (No compression. Depends on ruby-kafka: https://github.com/zendesk/ruby-kafka#compression)
  kafka_agg_max_bytes          (integer) :default => 4096
  kafka_agg_max_messages       (integer) :default => nil (No limit)
  max_send_limit_bytes         (integer) :default => nil (No drop)
  discard_kafka_delivery_failed   (bool) :default => false (No discard)
  monitoring_list              (array)   :default => []
</match>

kafka_buffered supports the following ruby-kafka parameters:

• max_send_retries - default: 2 - Number of times to retry sending of messages to a leader.
• required_acks - default: -1 - The number of acks required per request. If you need flush performance, set lower value, e.g. 1, 2.
• ack_timeout - default: nil - How long the producer waits for acks. The unit is seconds.
• compression_codec - default: nil - The codec the producer uses to compress messages.
• max_send_limit_bytes - default: nil - Max byte size of one record after formatting. A larger record is dropped with a warning, and the other records in the chunk are sent. For example, if you set 1000000, a record larger than 1000000 bytes is dropped. This does not limit the size of a record batch, so it does not prevent MessageSizeTooLarge by itself. See FAQ.
• discard_kafka_delivery_failed - default: false - discard the record where Kafka::DeliveryFailed occurred
• monitoring_list - default: [] - library to be used to monitor. statsd and datadog are supported

kafka_buffered has two additional parameters:

• kafka_agg_max_bytes - default: 4096 - Maximum value of total message size to be included in one batch transmission.
• kafka_agg_max_messages - default: nil - Maximum number of messages to include in one batch transmission.

Note: Java based Kafka client uses murmur2 as partitioner function by default. If you want to use same partitioning behavior with fluent-plugin-kafka, change it to murmur2 instead of crc32. Note that for using murmur2 hash partitioner function, you must install digest-murmurhash gem.

Non-buffered output plugin

This plugin uses ruby-kafka producer for writing data. For performance and reliability concerns, use kafka_bufferd output instead. This is mainly for testing.

<match app.**>
  @type kafka

  # Brokers: you can choose either brokers or zookeeper.
  brokers        <broker1_host>:<broker1_port>,<broker2_host>:<broker2_port>,.. # Set brokers directly
  zookeeper      <zookeeper_host>:<zookeeper_port> # Set brokers via Zookeeper
  zookeeper_path <broker path in zookeeper> :default => /brokers/ids # Set path in zookeeper for kafka

  default_topic         (string) :default => nil
  default_partition_key (string) :default => nil
  default_message_key   (string) :default => nil
  output_data_type      (json|ltsv|msgpack|attr:<record name>|<formatter name>) :default => json
  output_include_tag    (bool) :default => false
  output_include_time   (bool) :default => false
  exclude_topic_key     (bool) :default => false
  exclude_partition_key (bool) :default => false
  partitioner_hash_function (enum) (crc32|murmur2) :default => 'crc32'

  # ruby-kafka producer options
  max_send_retries    (integer) :default => 1
  required_acks       (integer) :default => -1
  ack_timeout         (integer) :default => nil (Use default of ruby-kafka)
  compression_codec   (string)  :default => nil (No compression. Depends on ruby-kafka: https://github.com/zendesk/ruby-kafka#compression)
  max_buffer_size     (integer) :default => nil (Use default of ruby-kafka)
  max_buffer_bytesize (integer) :default => nil (Use default of ruby-kafka)
</match>

This plugin also supports ruby-kafka related parameters. See Buffered output plugin section.

Note: Java based Kafka client uses murmur2 as partitioner function by default. If you want to use same partitioning behavior with fluent-plugin-kafka, change it to murmur2 instead of crc32. Note that for using murmur2 hash partitioner function, you must install digest-murmurhash gem.

rdkafka based output plugin

This plugin uses rdkafka instead of ruby-kafka for kafka client. You need to install rdkafka gem.

# rdkafka is C extension library. Need to install development tools like ruby-devel, gcc and etc
# for v0.12 or later
$ gem install rdkafka --no-document
# for v0.11 or earlier
$ gem install rdkafka -v 0.6.0 --no-document

rdkafka2 is for fluentd v1.0 or later.

<match app.**>
  @type rdkafka2

  brokers <broker1_host>:<broker1_port>,<broker2_host>:<broker2_port>,.. # Set brokers directly

  topic_key             (string) :default => 'topic'
  default_topic         (string) :default => nil
  partition_key         (string) :default => 'partition'
  partition_key_key     (string) :default => 'partition_key'
  message_key_key       (string) :default => 'message_key'
  use_default_for_unknown_topic           (bool) :default => false
  use_default_for_unknown_partition_error (bool) :default => false
  default_partition_key (string) :default => nil
  default_message_key   (string) :default => nil
  exclude_topic_key     (bool) :default => false
  exclude_partition_key (bool) :default => false
  discard_kafka_delivery_failed (bool) :default => false (No discard)
  discard_kafka_delivery_failed_regex (regexp) :default => nil (No discard)
  use_event_time        (bool) :default => false
  # When true, SASL authentication requires an SSL connection
  sasl_over_ssl         (bool) :default => true

  # same with kafka2
  headers               (hash) :default => {}
  headers_from_record   (hash) :default => {}
  record_key            (string) :default => nil

  <format>
    @type (json|ltsv|msgpack|attr:<record name>|<formatter name>) :default => json
  </format>

  # Optional. See https://docs.fluentd.org/v/1.0/configuration/inject-section
  <inject>
    tag_key tag
    time_key time
  </inject>

  # See fluentd document for buffer section parameters: https://docs.fluentd.org/v/1.0/configuration/buffer-section
  # Buffer chunk key should be same with topic_key. If value is not found in the record, default_topic is used.
  <buffer topic>
    flush_interval 10s
  </buffer>

  # You can set any rdkafka configuration via this parameter: https://github.com/edenhill/librdkafka/blob/master/CONFIGURATION.md
  rdkafka_options {
    "log_level" : 7
  }

  # rdkafka2 specific parameters

  # share kafka producer between flush threads. This is mainly for reducing kafka operations like kerberos
  share_producer (bool) :default => false
  # Timeout for polling message wait. If 0, no wait.
  rdkafka_delivery_handle_poll_timeout (integer) :default => 30
  # Max byte size of one record after formatting. Larger records are dropped.
  # This does not limit the size of a record batch. Default is no limit
  max_send_limit_bytes (integer) :default => nil
  # The maximum number of enqueueing bytes per second. It can reduce the
  # load of both Fluentd and Kafka when excessive messages are attempted
  # to send. Default is no limit.
  max_enqueue_bytes_per_second (integer) :default => nil
  unrecoverable_error_codes (array) :default => ["topic_authorization_failed", "msg_size_too_large"]

</match>

rdkafka2 supports discard_kafka_delivery_failed_regex parameter:

• discard_kafka_delivery_failed_regex - default: nil - discard the record where the Kafka::DeliveryFailed occurred and the emitted message matches the given regex pattern, such as /unknown_topic/.

rdkafka2 supports unrecoverable_error_codes parameter:

• unrecoverable_error_codes - default: ["topic_authorization_failed", "msg_size_too_large"] - librdkafka error codes to treat as unrecoverable errors. When a listed code is returned, Fluent::UnrecoverableError is raised, so Fluentd hands the whole chunk to the secondary output or to the backup directory instead of retrying it until retry_timeout expires. Both the produce call and the delivery report are covered, the latter only while rdkafka_delivery_handle_poll_timeout is not 0. Set the parameter to an empty value keeping retrying every error, as this plugin did before v0.19.x.
• A code is spelled the way Rdkafka::RdkafkaError#code reports it, that is the librdkafka name lowercased with the leading underscore removed, so RD_KAFKA_RESP_ERR__UNKNOWN_PARTITION is written as unknown_partition. A name that matches no error code is accepted and never matches anything.
• discard_kafka_delivery_failed and discard_kafka_delivery_failed_regex take precedence over this parameter. The regex is matched against the librdkafka message, such as Broker: Message size too large (msg_size_too_large).

If you use v0.12, use rdkafka instead.

<match kafka.**>
  @type rdkafka

  default_topic kafka
  flush_interval 1s
  output_data_type json

  rdkafka_options {
    "log_level" : 7
  }
</match>

FAQ

Why fluent-plugin-kafka can’t send data to our kafka cluster?

We got lots of similar questions. Almost cases, this problem happens by version mismatch between ruby-kafka and kafka cluster. See ruby-kafka README for more details: https://github.com/zendesk/ruby-kafka#compatibility

To avoid the problem, there are 2 approaches:

• Upgrade your kafka cluster to latest version. This is better because recent version is faster and robust.
• Downgrade ruby-kafka/fluent-plugin-kafka to work with your older kafka.

Why does Fluentd fail with Kafka::MessageSizeTooLarge (or msg_size_too_large)?

Kafka rejects a record batch that is larger than the size limit of the topic. The limit is max.message.bytes of the topic if it is set, and message.max.bytes of the broker otherwise. The topic value can be larger or smaller than the broker value. Below, “the topic limit” means this value. If one output writes to several topics, use the smallest one.

kafka2 sends all records of a buffer chunk at once, and the records for the same partition go into one record batch. In the worst case, the whole chunk becomes one record batch. To make the batches smaller, lower chunk_limit_size in <buffer>.

chunk_limit_size is the size of a chunk in the buffer, where records are kept as msgpack. The records sent to Kafka can be larger, because of <format>, <inject> and the overhead of each Kafka record. Start with about half of the topic limit, and lower it if the error continues. If <buffer> has compress gzip or compress zstd, chunk_limit_size limits the compressed size, and a chunk can hold several times more data. In that case, divide it further by the compression ratio of your data. compression_codec works the other way. Kafka checks the size of a record batch after compression, so compressed batches leave more room. How much depends on how well your data compresses.

max_send_limit_bytes does not make the batch smaller. It checks only the formatted value of each record, and drops the records larger than this setting. The message key, headers and the overhead of the batch are not counted. To drop a single record that Kafka would reject, set it below the topic limit with some room for them.

The following example assumes a topic limit of 1MB.

<match app.**>
  @type kafka2
  brokers <broker1_host>:<broker1_port>
  default_topic app

  max_send_limit_bytes 1000000

  <format>
    @type json
  </format>

  <buffer topic>
    chunk_limit_size 512k
  </buffer>
</match>

Fluentd may log chunk bytes limit exceeds for an emitted event stream. This warning comes from the Fluentd buffer. It means that one event stream emitted to the output was larger than chunk_limit_size, and Fluentd splits it into several chunks. The number in the warning is the size of that event stream in the buffer, not the growth by formatting. If a single record is larger than chunk_limit_size, that record cannot be stored, and Fluentd raises BufferChunkOverflowError.

kafka_buffered splits a chunk into smaller deliveries by kafka_agg_max_bytes (default: 4096) and kafka_agg_max_messages, so chunk_limit_size does not decide its batch size.

rdkafka2 passes records to librdkafka one by one, and librdkafka builds the record batches within its own batch.size and message.max.bytes. You can set them in rdkafka_options. Keep them below the topic limit. msg_size_too_large is in the default of unrecoverable_error_codes, so Fluentd does not retry a chunk that hits it.

Contributing

1. Fork it
2. Create your feature branch (git checkout -b my-new-feature)
3. Commit your changes (git commit -am 'Added some feature')
4. Push to the branch (git push origin my-new-feature)
5. Create new Pull Request